From a7ae218e89859b5a5b8433f57029590b8877fc18 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 21 Apr 2019 15:55:21 -0400 Subject: Move cuobjdump parser to reentrant Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 15 ++++++++++--- libcuda/cuobjdump.l | 52 ++++++++++++++++++++++----------------------- libcuda/cuobjdump.y | 8 +++++-- 3 files changed, 44 insertions(+), 31 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 27644b3..ef5abce 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1935,7 +1935,10 @@ void setCuobjdumpsassfilename(const char* filename){ } (dynamic_cast(cuobjdumpSectionList.front()))->setSASSfilename(filename); } -extern int cuobjdump_parse(); +typedef void * yyscan_t; +extern int cuobjdump_lex_init(yyscan_t* scanner); +extern int cuobjdump_parse(yyscan_t scanner); +extern int cuobjdump_lex_destroy(yyscan_t scanner); extern FILE *cuobjdump_in; //! Return the executable file of the process containing the PTX/SASS code @@ -2147,7 +2150,10 @@ void extract_code_using_cuobjdump(){ printf("Parsing file %s\n", fname); cuobjdump_in = fopen(fname, "r"); - cuobjdump_parse(); + yyscan_t scanner; + cuobjdump_lex_init(&scanner); + cuobjdump_parse(scanner); + cuobjdump_lex_destroy(scanner); fclose(cuobjdump_in); printf("Done parsing!!!\n"); } else { @@ -2196,7 +2202,10 @@ void extract_code_using_cuobjdump(){ std::cout << "Trying to parse " << libcodfn.str() << std::endl; cuobjdump_in = fopen(libcodfn.str().c_str(), "r"); - cuobjdump_parse(); + yyscan_t scanner; + cuobjdump_lex_init(&scanner); + cuobjdump_parse(scanner); + cuobjdump_lex_destroy(scanner); fclose(cuobjdump_in); std::getline(libsf, line); } diff --git a/libcuda/cuobjdump.l b/libcuda/cuobjdump.l index 2b0dac8..9a468ba 100644 --- a/libcuda/cuobjdump.l +++ b/libcuda/cuobjdump.l @@ -38,9 +38,7 @@ #define YYDEBUG 1 -#define yylval cuobjdump_lval - -void cuobjdump_error(const char*); +void print_error(const char*, const char*, int); %} %option stack @@ -48,6 +46,8 @@ void cuobjdump_error(const char*); %option yylineno %option nounput +%option bison-bridge +%option reentrant %s ptxcode %s sasscode @@ -69,54 +69,54 @@ newlines {newline}+ "ptxasOptions"{notnewline}*{newline} -[1-9]{numeric}* yylval.string_value = strdup(yytext); return DECIMAL; +[1-9]{numeric}* yylval->string_value = strdup(yytext); return DECIMAL; "has debug info"{newline} "Fatbin ptx code:"{newline} { - yy_push_state(ptxcode); - yy_push_state(identifier); - yy_push_state(ptxheader); - yylval.string_value = strdup(yytext); + yy_push_state(ptxcode, yyscanner); + yy_push_state(identifier, yyscanner); + yy_push_state(ptxheader, yyscanner); + yylval->string_value = strdup(yytext); return PTXHEADER; } "Fatbin elf code:"{newline} { - yy_push_state(elfcode); - yy_push_state(identifier); - yy_push_state(elfheader); - yylval.string_value = strdup(yytext); + yy_push_state(elfcode, yyscanner); + yy_push_state(identifier, yyscanner); + yy_push_state(elfheader, yyscanner); + yylval->string_value = strdup(yytext); return ELFHEADER; } /*PTX code tokens*/ -{notnewline}*{newline} yylval.string_value = strdup(yytext); return PTXLINE; +{notnewline}*{newline} yylval->string_value = strdup(yytext); return PTXLINE; /*ELF code tokens*/ {whitespace}*"code for sm_"{numeric}+{newline} { BEGIN(sasscode); - yylval.string_value = strdup(yytext); + yylval->string_value = strdup(yytext); return SASSLINE; } -{notnewline}*{newline} yylval.string_value = strdup(yytext); return ELFLINE; +{notnewline}*{newline} yylval->string_value = strdup(yytext); return ELFLINE; /*SASS code tokens*/ -{notnewline}*{newline} yylval.string_value = strdup(yytext); return SASSLINE; +{notnewline}*{newline} yylval->string_value = strdup(yytext); return SASSLINE; {newline}"compressed"{newline} BEGIN(conidentifier); return H_COMPRESSED; {newline}"identifier = " BEGIN(endidentifier); return H_IDENTIFIER; -{newline}{newline} yy_pop_state(); +{newline}{newline} yy_pop_state(yyscanner); "identifier = " BEGIN(endidentifier); return H_IDENTIFIER; -{newline} yy_pop_state(); +{newline} yy_pop_state(yyscanner); -{notnewline}+ yylval.string_value = strdup(yytext); yy_pop_state(); return FILENAME; +{notnewline}+ yylval->string_value = strdup(yytext); yy_pop_state(yyscanner); return FILENAME; /*Header tokens*/ -[[:alnum:]_]+ yylval.string_value = strdup(yytext); return IDENTIFIER; +[[:alnum:]_]+ yylval->string_value = strdup(yytext); return IDENTIFIER; "================" return H_SEPARATOR; "arch = " return H_ARCH; "code version = " return H_CODEVERSION; @@ -128,7 +128,7 @@ newlines {newline}+ /*Header tokens*/ -[[:alnum:]_]+ yylval.string_value = strdup(yytext); return IDENTIFIER; +[[:alnum:]_]+ yylval->string_value = strdup(yytext); return IDENTIFIER; "================" return H_SEPARATOR; "arch = " return H_ARCH; "code version = " return H_CODEVERSION; @@ -143,7 +143,7 @@ newlines {newline}+ /* Looking for the identifier (filename) then the header is done */ -{notnewline}+ yylval.string_value = strdup(yytext); yy_pop_state(); return IDENTIFIER; +{notnewline}+ yylval->string_value = strdup(yytext); yy_pop_state(yyscanner); return IDENTIFIER; @@ -153,11 +153,11 @@ newlines {newline}+ <> BEGIN(INITIAL);return 0; /*No other rule matched. Throw an error*/ -. cuobjdump_error("Invalid token"); +. print_error("Invalid token", yytext, yylineno); %% -void cuobjdump_error(const char* message) +void print_error(const char* message, const char* text, int lineno) { - printf(" %s near \"%s\"",message, yytext); - printf(" on line %i\n",yylineno); + printf(" %s near \"%s\"",message, text); + printf(" on line %i\n",lineno); } diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y index 31760f7..eea0a1f 100644 --- a/libcuda/cuobjdump.y +++ b/libcuda/cuobjdump.y @@ -29,8 +29,8 @@ %{ #include -int yylex(void); -void yyerror(const char*); +typedef void * yyscan_t; + extern void addCuobjdumpSection(int sectiontype); void setCuobjdumparch(const char* arch); void setCuobjdumpidentifier(const char* identifier); @@ -44,6 +44,10 @@ FILE *elffile; FILE *sassfile; char filename [1024]; %} +%define api.pure full +%parse-param {yyscan_t scanner} +%lex-param {yyscan_t scanner} + %union { char* string_value; } -- cgit v1.3 From fe15e8714972667342e650e8c348e7d6b68e5fbc Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 21 Apr 2019 17:11:07 -0400 Subject: Fix cuobjdump_in Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index ef5abce..1d1b950 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1937,9 +1937,9 @@ void setCuobjdumpsassfilename(const char* filename){ } typedef void * yyscan_t; extern int cuobjdump_lex_init(yyscan_t* scanner); +extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); extern int cuobjdump_parse(yyscan_t scanner); extern int cuobjdump_lex_destroy(yyscan_t scanner); -extern FILE *cuobjdump_in; //! Return the executable file of the process containing the PTX/SASS code //! @@ -2148,11 +2148,13 @@ void extract_code_using_cuobjdump(){ if (parse_output) { printf("Parsing file %s\n", fname); + FILE *cuobjdump_in; cuobjdump_in = fopen(fname, "r"); yyscan_t scanner; cuobjdump_lex_init(&scanner); cuobjdump_parse(scanner); + cuobjdump_set_in(cuobjdump_in, scanner); cuobjdump_lex_destroy(scanner); fclose(cuobjdump_in); printf("Done parsing!!!\n"); @@ -2201,9 +2203,11 @@ void extract_code_using_cuobjdump(){ std::cout << "Done" << std::endl; std::cout << "Trying to parse " << libcodfn.str() << std::endl; + FILE *cuobjdump_in; cuobjdump_in = fopen(libcodfn.str().c_str(), "r"); yyscan_t scanner; cuobjdump_lex_init(&scanner); + cuobjdump_set_in(cuobjdump_in, scanner); cuobjdump_parse(scanner); cuobjdump_lex_destroy(scanner); fclose(cuobjdump_in); -- cgit v1.3 From e6e9181f4c0706cc9d4b5172ba9f932f6a1bd03e Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 21 Apr 2019 20:14:26 -0400 Subject: Fix declaration for yylex and yyerror Signed-off-by: Mengchi Zhang --- libcuda/cuobjdump.y | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'libcuda') diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y index eea0a1f..927d17e 100644 --- a/libcuda/cuobjdump.y +++ b/libcuda/cuobjdump.y @@ -51,6 +51,10 @@ char filename [1024]; %union { char* string_value; } +%{ +int yylex(YYSTYPE * yylval_param, yyscan_t yyscanner); +int yyerror(yyscan_t yyscanner, const char* msg); +%} %token H_SEPARATOR H_ARCH H_CODEVERSION H_PRODUCER H_HOST H_COMPILESIZE H_IDENTIFIER H_UNKNOWN H_COMPRESSED %token CODEVERSION %token STRING -- cgit v1.3 From b6631a07ca4aa0345f569b49116c649feadccc89 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 21 Apr 2019 22:32:26 -0400 Subject: Add yyerror definition Signed-off-by: Mengchi Zhang --- libcuda/cuobjdump.l | 11 ++++++----- libcuda/cuobjdump.y | 2 +- 2 files changed, 7 insertions(+), 6 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuobjdump.l b/libcuda/cuobjdump.l index 9a468ba..9359281 100644 --- a/libcuda/cuobjdump.l +++ b/libcuda/cuobjdump.l @@ -38,7 +38,7 @@ #define YYDEBUG 1 -void print_error(const char*, const char*, int); +void cuobjdump_error(yyscan_t yyscanner, const char* msg); %} %option stack @@ -153,11 +153,12 @@ newlines {newline}+ <> BEGIN(INITIAL);return 0; /*No other rule matched. Throw an error*/ -. print_error("Invalid token", yytext, yylineno); +. cuobjdump_error(yyscanner, "Invalid token"); %% -void print_error(const char* message, const char* text, int lineno) +void cuobjdump_error(yyscan_t yyscanner, const char* msg) { - printf(" %s near \"%s\"",message, text); - printf(" on line %i\n",lineno); + struct yyguts_t * yyg = (struct yyguts_t*)yyscanner; + printf(" %s near \"%s\"",msg, yytext); + printf(" on line %i\n",yylineno); } diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y index 927d17e..66cbace 100644 --- a/libcuda/cuobjdump.y +++ b/libcuda/cuobjdump.y @@ -53,7 +53,7 @@ char filename [1024]; } %{ int yylex(YYSTYPE * yylval_param, yyscan_t yyscanner); -int yyerror(yyscan_t yyscanner, const char* msg); +void yyerror(yyscan_t yyscanner, const char* msg); %} %token H_SEPARATOR H_ARCH H_CODEVERSION H_PRODUCER H_HOST H_COMPILESIZE H_IDENTIFIER H_UNKNOWN H_COMPRESSED %token CODEVERSION -- cgit v1.3 From c4fe0159232673b0c75bacd33b2e96d6a68dfbaf Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 21 Apr 2019 22:53:00 -0400 Subject: Fix order Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 1d1b950..225c402 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -2153,8 +2153,8 @@ void extract_code_using_cuobjdump(){ yyscan_t scanner; cuobjdump_lex_init(&scanner); - cuobjdump_parse(scanner); cuobjdump_set_in(cuobjdump_in, scanner); + cuobjdump_parse(scanner); cuobjdump_lex_destroy(scanner); fclose(cuobjdump_in); printf("Done parsing!!!\n"); -- cgit v1.3 From 10156f6728f10fdc33ca8796e9cc6ade306f06ec Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 16 May 2019 21:08:39 -0400 Subject: Remove unused extern Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 1 - 1 file changed, 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 11dcd5a..fc152c2 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -3261,7 +3261,6 @@ extern int ptx__scan_string(const char*); extern FILE *ptx_in; extern int ptxinfo_parse(); -extern int ptxinfo_debug; extern FILE *ptxinfo_in; /// static functions -- cgit v1.3 From 913865e6ac7d0f7c12faeb8430dc5724fdc4be80 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 27 May 2019 17:27:58 -0400 Subject: Move some cuobjdump parser variables Signed-off-by: Mengchi Zhang --- libcuda/Makefile | 4 ++-- libcuda/cuda_runtime_api.cc | 23 ++++++++++++----------- libcuda/cuobjdump.h | 18 ++++++++++++++++++ libcuda/cuobjdump.l | 9 ++++++--- libcuda/cuobjdump.y | 43 ++++++++++++++++++++----------------------- 5 files changed, 58 insertions(+), 39 deletions(-) create mode 100644 libcuda/cuobjdump.h (limited to 'libcuda') diff --git a/libcuda/Makefile b/libcuda/Makefile index 13932e2..c8ff2e3 100644 --- a/libcuda/Makefile +++ b/libcuda/Makefile @@ -111,10 +111,10 @@ $(OUTPUT_DIR)/%.o: %.cc $(CPP) $(CXXFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ $(OUTPUT_DIR)/%.o: %.c - $(CC) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ + $(CPP) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ $(OUTPUT_DIR)/%.o: $(OUTPUT_DIR)/%.c - $(CC) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ + $(CPP) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ $(OUTPUT_DIR)/cuobjdump_parser.c: cuobjdump.y $(YACC) $(YFLAGS) -p cuobjdump_ -o$@ $< --file-prefix=$(OUTPUT_DIR)/cuobjdump diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index fc152c2..1d4e870 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1972,9 +1972,10 @@ void setCuobjdumpsassfilename(const char* filename){ (dynamic_cast(cuobjdumpSectionList.front()))->setSASSfilename(filename); } typedef void * yyscan_t; +#include "cuobjdump.h" extern int cuobjdump_lex_init(yyscan_t* scanner); extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); -extern int cuobjdump_parse(yyscan_t scanner); +extern int cuobjdump_parse(yyscan_t scanner, cuobjdump_parser* parser); extern int cuobjdump_lex_destroy(yyscan_t scanner); //! Return the executable file of the process containing the PTX/SASS code @@ -2187,11 +2188,11 @@ void extract_code_using_cuobjdump(){ FILE *cuobjdump_in; cuobjdump_in = fopen(fname, "r"); - yyscan_t scanner; - cuobjdump_lex_init(&scanner); - cuobjdump_set_in(cuobjdump_in, scanner); - cuobjdump_parse(scanner); - cuobjdump_lex_destroy(scanner); + cuobjdump_parser parser; + cuobjdump_lex_init(&(parser.scanner)); + cuobjdump_set_in(cuobjdump_in, (parser.scanner)); + cuobjdump_parse(parser.scanner, &parser); + cuobjdump_lex_destroy(parser.scanner); fclose(cuobjdump_in); printf("Done parsing!!!\n"); } else { @@ -2241,11 +2242,11 @@ void extract_code_using_cuobjdump(){ std::cout << "Trying to parse " << libcodfn.str() << std::endl; FILE *cuobjdump_in; cuobjdump_in = fopen(libcodfn.str().c_str(), "r"); - yyscan_t scanner; - cuobjdump_lex_init(&scanner); - cuobjdump_set_in(cuobjdump_in, scanner); - cuobjdump_parse(scanner); - cuobjdump_lex_destroy(scanner); + cuobjdump_parser parser; + cuobjdump_lex_init(&(parser.scanner)); + cuobjdump_set_in(cuobjdump_in, (parser.scanner)); + cuobjdump_parse(parser.scanner, &parser); + cuobjdump_lex_destroy(parser.scanner); fclose(cuobjdump_in); std::getline(libsf, line); } diff --git a/libcuda/cuobjdump.h b/libcuda/cuobjdump.h new file mode 100644 index 0000000..61bf806 --- /dev/null +++ b/libcuda/cuobjdump.h @@ -0,0 +1,18 @@ +#ifndef __cuobjdump_h__ +#define __cuobjdump_h__ +class cuobjdump_parser { + + public: + yyscan_t scanner; + int elfserial; + int ptxserial; + FILE *ptxfile; + FILE *elffile; + FILE *sassfile; + char filename [1024]; + cuobjdump_parser() { + int elfserial = 1; + int ptxserial = 1; + } +}; +#endif /* __cuobjdump_h__ */ diff --git a/libcuda/cuobjdump.l b/libcuda/cuobjdump.l index 9359281..26fbb55 100644 --- a/libcuda/cuobjdump.l +++ b/libcuda/cuobjdump.l @@ -30,6 +30,7 @@ %{ #include #include +#include "cuobjdump.h" #include "cuobjdump_parser.h" #define YY_NEVER_INTERACTIVE 1 @@ -38,7 +39,9 @@ #define YYDEBUG 1 -void cuobjdump_error(yyscan_t yyscanner, const char* msg); +void cuobjdump_error(yyscan_t yyscanner, cuobjdump_parser* parser, const char* msg); +#define YY_DECL int cuobjdump_lex \ + (YYSTYPE * yylval_param , yyscan_t yyscanner, cuobjdump_parser* parser) %} %option stack @@ -153,10 +156,10 @@ newlines {newline}+ <> BEGIN(INITIAL);return 0; /*No other rule matched. Throw an error*/ -. cuobjdump_error(yyscanner, "Invalid token"); +. cuobjdump_error(yyscanner, parser, "Invalid token"); %% -void cuobjdump_error(yyscan_t yyscanner, const char* msg) +void cuobjdump_error(yyscan_t yyscanner, cuobjdump_parser* parser, const char* msg) { struct yyguts_t * yyg = (struct yyguts_t*)yyscanner; printf(" %s near \"%s\"",msg, yytext); diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y index 66cbace..9c0c28d 100644 --- a/libcuda/cuobjdump.y +++ b/libcuda/cuobjdump.y @@ -30,6 +30,7 @@ #include typedef void * yyscan_t; +#include "cuobjdump.h" extern void addCuobjdumpSection(int sectiontype); void setCuobjdumparch(const char* arch); @@ -37,23 +38,19 @@ void setCuobjdumpidentifier(const char* identifier); void setCuobjdumpptxfilename(const char* filename); void setCuobjdumpelffilename(const char* filename); void setCuobjdumpsassfilename(const char* filename); -int elfserial = 1; -int ptxserial = 1; -FILE *ptxfile; -FILE *elffile; -FILE *sassfile; -char filename [1024]; %} %define api.pure full %parse-param {yyscan_t scanner} +%parse-param {cuobjdump_parser* parser} %lex-param {yyscan_t scanner} +%lex-param {cuobjdump_parser* parser} %union { char* string_value; } %{ -int yylex(YYSTYPE * yylval_param, yyscan_t yyscanner); -void yyerror(yyscan_t yyscanner, const char* msg); +int yylex(YYSTYPE * yylval_param, yyscan_t yyscanner, cuobjdump_parser* parser); +void yyerror(yyscan_t yyscanner, cuobjdump_parser* parser, const char* msg); %} %token H_SEPARATOR H_ARCH H_CODEVERSION H_PRODUCER H_HOST H_COMPILESIZE H_IDENTIFIER H_UNKNOWN H_COMPRESSED %token CODEVERSION @@ -79,24 +76,24 @@ emptylines : emptylines NEWLINE section : PTXHEADER { addCuobjdumpSection(0); - snprintf(filename, 1024, "_cuobjdump_%d.ptx", ptxserial++); - ptxfile = fopen(filename, "w"); - setCuobjdumpptxfilename(filename); + snprintf(parser->filename, 1024, "_cuobjdump_%d.ptx", parser->ptxserial++); + parser->ptxfile = fopen(parser->filename, "w"); + setCuobjdumpptxfilename(parser->filename); } headerinfo compressedkeyword identifier ptxcode { - fclose(ptxfile); + fclose(parser->ptxfile); } | ELFHEADER { addCuobjdumpSection(1); - snprintf(filename, 1024, "_cuobjdump_%d.elf", elfserial); - elffile = fopen(filename, "w"); - setCuobjdumpelffilename(filename); + snprintf(parser->filename, 1024, "_cuobjdump_%d.elf", parser->elfserial); + parser->elffile = fopen(parser->filename, "w"); + setCuobjdumpelffilename(parser->filename); } headerinfo compressedkeyword identifier elfcode { - fclose(elffile); - snprintf(filename, 1024, "_cuobjdump_%d.sass", elfserial++); - sassfile = fopen(filename, "w"); - setCuobjdumpsassfilename(filename); + fclose(parser->elffile); + snprintf(parser->filename, 1024, "_cuobjdump_%d.sass", parser->elfserial++); + parser->sassfile = fopen(parser->filename, "w"); + setCuobjdumpsassfilename(parser->filename); } sasscode { - fclose(sassfile); + fclose(parser->sassfile); }; headerinfo : H_SEPARATOR NEWLINE @@ -118,13 +115,13 @@ identifier : H_IDENTIFIER FILENAME emptylines {setCuobjdumpidentifier($2);} compressedkeyword : H_COMPRESSED emptylines | ; -ptxcode : ptxcode PTXLINE {fprintf(ptxfile, "%s", $2);} +ptxcode : ptxcode PTXLINE {fprintf(parser->ptxfile, "%s", $2);} | ; -elfcode : elfcode ELFLINE {fprintf(elffile, "%s", $2);} +elfcode : elfcode ELFLINE {fprintf(parser->elffile, "%s", $2);} | ; -sasscode : sasscode SASSLINE {fprintf(sassfile, "%s", $2);} +sasscode : sasscode SASSLINE {fprintf(parser->sassfile, "%s", $2);} | ; -- cgit v1.3 From 568e3185c58b07ac31fdd59f6c2aa7af7533939e Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 28 May 2019 15:14:52 -0400 Subject: Move some functions to enable C++ Signed-off-by: Mengchi Zhang --- libcuda/Makefile | 4 +- libcuda/cuda_runtime_api.cc | 246 ++++++++++++++++++++++---------------------- 2 files changed, 126 insertions(+), 124 deletions(-) (limited to 'libcuda') diff --git a/libcuda/Makefile b/libcuda/Makefile index 13932e2..c8ff2e3 100644 --- a/libcuda/Makefile +++ b/libcuda/Makefile @@ -111,10 +111,10 @@ $(OUTPUT_DIR)/%.o: %.cc $(CPP) $(CXXFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ $(OUTPUT_DIR)/%.o: %.c - $(CC) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ + $(CPP) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ $(OUTPUT_DIR)/%.o: $(OUTPUT_DIR)/%.c - $(CC) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ + $(CPP) $(CCFLAGS) -I./ -I$(OUTPUT_DIR) -I$(CUDA_INSTALL_PATH)/include -c $< -o $@ $(OUTPUT_DIR)/cuobjdump_parser.c: cuobjdump.y $(YACC) $(YFLAGS) -p cuobjdump_ -o$@ $< --file-prefix=$(OUTPUT_DIR)/cuobjdump diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 168c80f..afe47f8 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -483,6 +483,130 @@ event_tracker_t g_timer_events; int g_active_device = 0; //active gpu that runs the code std::list g_cuda_launch_stack; +typedef void * yyscan_t; +#include "cuobjdump.h" +extern int cuobjdump_lex_init(yyscan_t* scanner); +extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); +extern int cuobjdump_parse(yyscan_t scanner, struct cuobjdump_parser* parser); +extern int cuobjdump_lex_destroy(yyscan_t scanner); + +enum cuobjdumpSectionType { + PTXSECTION=0, + ELFSECTION +}; + + +class cuobjdumpSection { +public: + //Constructor + cuobjdumpSection() { + arch = 0; + identifier = ""; + } + virtual ~cuobjdumpSection() {} + unsigned getArch() {return arch;} + void setArch(unsigned a) {arch = a;} + std::string getIdentifier() {return identifier;} + void setIdentifier(std::string i) {identifier = i;} + virtual void print(){std::cout << "cuobjdump Section: unknown type" << std::endl;} +private: + unsigned arch; + std::string identifier; +}; + +class cuobjdumpELFSection : public cuobjdumpSection +{ +public: + cuobjdumpELFSection() {} + virtual ~cuobjdumpELFSection() { + elffilename = ""; + sassfilename = ""; + } + std::string getELFfilename() {return elffilename;} + void setELFfilename(std::string f) {elffilename = f;} + std::string getSASSfilename() {return sassfilename;} + void setSASSfilename(std::string f) {sassfilename = f;} + virtual void print() { + std::cout << "ELF Section:" << std::endl; + std::cout << "arch: sm_" << getArch() << std::endl; + std::cout << "identifier: " << getIdentifier() << std::endl; + std::cout << "elf filename: " << getELFfilename() << std::endl; + std::cout << "sass filename: " << getSASSfilename() << std::endl; + std::cout << std::endl; + } +private: + std::string elffilename; + std::string sassfilename; +}; + +class cuobjdumpPTXSection : public cuobjdumpSection +{ +public: + cuobjdumpPTXSection(){ + ptxfilename = ""; + } + std::string getPTXfilename() {return ptxfilename;} + void setPTXfilename(std::string f) {ptxfilename = f;} + virtual void print() { + std::cout << "PTX Section:" << std::endl; + std::cout << "arch: sm_" << getArch() << std::endl; + std::cout << "identifier: " << getIdentifier() << std::endl; + std::cout << "ptx filename: " << getPTXfilename() << std::endl; + std::cout << std::endl; + } +private: + std::string ptxfilename; +}; + + +std::list cuobjdumpSectionList; +std::list libSectionList; + +// sectiontype: 0 for ptx, 1 for elf +void addCuobjdumpSection(int sectiontype){ + if (sectiontype) + cuobjdumpSectionList.push_front(new cuobjdumpELFSection()); + else + cuobjdumpSectionList.push_front(new cuobjdumpPTXSection()); + printf("## Adding new section %s\n", sectiontype?"ELF":"PTX"); +} + +void setCuobjdumparch(const char* arch){ + unsigned archnum; + sscanf(arch, "sm_%u", &archnum); + assert (archnum && "cannot have sm_0"); + printf("Adding arch: %s\n", arch); + cuobjdumpSectionList.front()->setArch(archnum); +} + +void setCuobjdumpidentifier(const char* identifier){ + printf("Adding identifier: %s\n", identifier); + cuobjdumpSectionList.front()->setIdentifier(identifier); +} + +void setCuobjdumpptxfilename(const char* filename){ + printf("Adding ptx filename: %s\n", filename); + cuobjdumpSection* x = cuobjdumpSectionList.front(); + if (dynamic_cast(x) == NULL){ + assert (0 && "You shouldn't be trying to add a ptxfilename to an elf section"); + } + (dynamic_cast(x))->setPTXfilename(filename); +} + +void setCuobjdumpelffilename(const char* filename){ + if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ + assert (0 && "You shouldn't be trying to add a elffilename to an ptx section"); + } + (dynamic_cast(cuobjdumpSectionList.front()))->setELFfilename(filename); +} + +void setCuobjdumpsassfilename(const char* filename){ + if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ + assert (0 && "You shouldn't be trying to add a sassfilename to an ptx section"); + } + (dynamic_cast(cuobjdumpSectionList.front()))->setSASSfilename(filename); +} + /******************************************************************************* * * * * @@ -1856,128 +1980,6 @@ __host__ cudaError_t CUDARTAPI cudaGetExportTable(const void **ppExportTable, co //#include "../../cuobjdump_to_ptxplus/cuobjdump_parser.h" -enum cuobjdumpSectionType { - PTXSECTION=0, - ELFSECTION -}; - - -class cuobjdumpSection { -public: - //Constructor - cuobjdumpSection() { - arch = 0; - identifier = ""; - } - virtual ~cuobjdumpSection() {} - unsigned getArch() {return arch;} - void setArch(unsigned a) {arch = a;} - std::string getIdentifier() {return identifier;} - void setIdentifier(std::string i) {identifier = i;} - virtual void print(){std::cout << "cuobjdump Section: unknown type" << std::endl;} -private: - unsigned arch; - std::string identifier; -}; - -class cuobjdumpELFSection : public cuobjdumpSection -{ -public: - cuobjdumpELFSection() {} - virtual ~cuobjdumpELFSection() { - elffilename = ""; - sassfilename = ""; - } - std::string getELFfilename() {return elffilename;} - void setELFfilename(std::string f) {elffilename = f;} - std::string getSASSfilename() {return sassfilename;} - void setSASSfilename(std::string f) {sassfilename = f;} - virtual void print() { - std::cout << "ELF Section:" << std::endl; - std::cout << "arch: sm_" << getArch() << std::endl; - std::cout << "identifier: " << getIdentifier() << std::endl; - std::cout << "elf filename: " << getELFfilename() << std::endl; - std::cout << "sass filename: " << getSASSfilename() << std::endl; - std::cout << std::endl; - } -private: - std::string elffilename; - std::string sassfilename; -}; - -class cuobjdumpPTXSection : public cuobjdumpSection -{ -public: - cuobjdumpPTXSection(){ - ptxfilename = ""; - } - std::string getPTXfilename() {return ptxfilename;} - void setPTXfilename(std::string f) {ptxfilename = f;} - virtual void print() { - std::cout << "PTX Section:" << std::endl; - std::cout << "arch: sm_" << getArch() << std::endl; - std::cout << "identifier: " << getIdentifier() << std::endl; - std::cout << "ptx filename: " << getPTXfilename() << std::endl; - std::cout << std::endl; - } -private: - std::string ptxfilename; -}; - -std::list cuobjdumpSectionList; -std::list libSectionList; - -// sectiontype: 0 for ptx, 1 for elf -void addCuobjdumpSection(int sectiontype){ - if (sectiontype) - cuobjdumpSectionList.push_front(new cuobjdumpELFSection()); - else - cuobjdumpSectionList.push_front(new cuobjdumpPTXSection()); - printf("## Adding new section %s\n", sectiontype?"ELF":"PTX"); -} - -void setCuobjdumparch(const char* arch){ - unsigned archnum; - sscanf(arch, "sm_%u", &archnum); - assert (archnum && "cannot have sm_0"); - printf("Adding arch: %s\n", arch); - cuobjdumpSectionList.front()->setArch(archnum); -} - -void setCuobjdumpidentifier(const char* identifier){ - printf("Adding identifier: %s\n", identifier); - cuobjdumpSectionList.front()->setIdentifier(identifier); -} - -void setCuobjdumpptxfilename(const char* filename){ - printf("Adding ptx filename: %s\n", filename); - cuobjdumpSection* x = cuobjdumpSectionList.front(); - if (dynamic_cast(x) == NULL){ - assert (0 && "You shouldn't be trying to add a ptxfilename to an elf section"); - } - (dynamic_cast(x))->setPTXfilename(filename); -} - -void setCuobjdumpelffilename(const char* filename){ - if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ - assert (0 && "You shouldn't be trying to add a elffilename to an ptx section"); - } - (dynamic_cast(cuobjdumpSectionList.front()))->setELFfilename(filename); -} - -void setCuobjdumpsassfilename(const char* filename){ - if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ - assert (0 && "You shouldn't be trying to add a sassfilename to an ptx section"); - } - (dynamic_cast(cuobjdumpSectionList.front()))->setSASSfilename(filename); -} -typedef void * yyscan_t; -#include "cuobjdump.h" -extern int cuobjdump_lex_init(yyscan_t* scanner); -extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); -extern int cuobjdump_parse(yyscan_t scanner, struct cuobjdump_parser* parser); -extern int cuobjdump_lex_destroy(yyscan_t scanner); - //! Return the executable file of the process containing the PTX/SASS code //! //! This Function returns the executable file ran by the process. This -- cgit v1.3 From 25bdc0dd89932f95ace0fc617649a4e041aaadd9 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 29 May 2019 01:28:29 -0400 Subject: Move SectionList to context Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 109 ++++++++++---------------------------------- libcuda/cuobjdump.h | 67 +++++++++++++++++++++++++++ libcuda/cuobjdump.l | 8 ++-- libcuda/cuobjdump.y | 36 ++++++++------- 4 files changed, 113 insertions(+), 107 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index afe47f8..df7ddc7 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -139,6 +139,8 @@ #include "../src/gpgpusim_entrypoint.h" #include "../src/stream_manager.h" #include "../src/abstract_hardware_model.h" +typedef void * yyscan_t; +#include "cuobjdump.h" #include #include @@ -302,6 +304,7 @@ struct CUctx_st { return i->second; } + std::list cuobjdumpSectionList; private: _cuda_device_id *m_gpu; // selected gpu std::map m_code; // fat binary handle => global symbol table @@ -483,11 +486,9 @@ event_tracker_t g_timer_events; int g_active_device = 0; //active gpu that runs the code std::list g_cuda_launch_stack; -typedef void * yyscan_t; -#include "cuobjdump.h" extern int cuobjdump_lex_init(yyscan_t* scanner); extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); -extern int cuobjdump_parse(yyscan_t scanner, struct cuobjdump_parser* parser); +extern int cuobjdump_parse(yyscan_t scanner, struct cuobjdump_parser* parser, std::list &cuobjdumpSectionList); extern int cuobjdump_lex_destroy(yyscan_t scanner); enum cuobjdumpSectionType { @@ -496,74 +497,10 @@ enum cuobjdumpSectionType { }; -class cuobjdumpSection { -public: - //Constructor - cuobjdumpSection() { - arch = 0; - identifier = ""; - } - virtual ~cuobjdumpSection() {} - unsigned getArch() {return arch;} - void setArch(unsigned a) {arch = a;} - std::string getIdentifier() {return identifier;} - void setIdentifier(std::string i) {identifier = i;} - virtual void print(){std::cout << "cuobjdump Section: unknown type" << std::endl;} -private: - unsigned arch; - std::string identifier; -}; - -class cuobjdumpELFSection : public cuobjdumpSection -{ -public: - cuobjdumpELFSection() {} - virtual ~cuobjdumpELFSection() { - elffilename = ""; - sassfilename = ""; - } - std::string getELFfilename() {return elffilename;} - void setELFfilename(std::string f) {elffilename = f;} - std::string getSASSfilename() {return sassfilename;} - void setSASSfilename(std::string f) {sassfilename = f;} - virtual void print() { - std::cout << "ELF Section:" << std::endl; - std::cout << "arch: sm_" << getArch() << std::endl; - std::cout << "identifier: " << getIdentifier() << std::endl; - std::cout << "elf filename: " << getELFfilename() << std::endl; - std::cout << "sass filename: " << getSASSfilename() << std::endl; - std::cout << std::endl; - } -private: - std::string elffilename; - std::string sassfilename; -}; - -class cuobjdumpPTXSection : public cuobjdumpSection -{ -public: - cuobjdumpPTXSection(){ - ptxfilename = ""; - } - std::string getPTXfilename() {return ptxfilename;} - void setPTXfilename(std::string f) {ptxfilename = f;} - virtual void print() { - std::cout << "PTX Section:" << std::endl; - std::cout << "arch: sm_" << getArch() << std::endl; - std::cout << "identifier: " << getIdentifier() << std::endl; - std::cout << "ptx filename: " << getPTXfilename() << std::endl; - std::cout << std::endl; - } -private: - std::string ptxfilename; -}; - - -std::list cuobjdumpSectionList; std::list libSectionList; // sectiontype: 0 for ptx, 1 for elf -void addCuobjdumpSection(int sectiontype){ +void addCuobjdumpSection(int sectiontype, std::list &cuobjdumpSectionList){ if (sectiontype) cuobjdumpSectionList.push_front(new cuobjdumpELFSection()); else @@ -571,7 +508,7 @@ void addCuobjdumpSection(int sectiontype){ printf("## Adding new section %s\n", sectiontype?"ELF":"PTX"); } -void setCuobjdumparch(const char* arch){ +void setCuobjdumparch(const char* arch, std::list &cuobjdumpSectionList){ unsigned archnum; sscanf(arch, "sm_%u", &archnum); assert (archnum && "cannot have sm_0"); @@ -579,12 +516,12 @@ void setCuobjdumparch(const char* arch){ cuobjdumpSectionList.front()->setArch(archnum); } -void setCuobjdumpidentifier(const char* identifier){ +void setCuobjdumpidentifier(const char* identifier, std::list &cuobjdumpSectionList){ printf("Adding identifier: %s\n", identifier); cuobjdumpSectionList.front()->setIdentifier(identifier); } -void setCuobjdumpptxfilename(const char* filename){ +void setCuobjdumpptxfilename(const char* filename, std::list &cuobjdumpSectionList){ printf("Adding ptx filename: %s\n", filename); cuobjdumpSection* x = cuobjdumpSectionList.front(); if (dynamic_cast(x) == NULL){ @@ -593,14 +530,14 @@ void setCuobjdumpptxfilename(const char* filename){ (dynamic_cast(x))->setPTXfilename(filename); } -void setCuobjdumpelffilename(const char* filename){ +void setCuobjdumpelffilename(const char* filename, std::list &cuobjdumpSectionList){ if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ assert (0 && "You shouldn't be trying to add a elffilename to an ptx section"); } (dynamic_cast(cuobjdumpSectionList.front()))->setELFfilename(filename); } -void setCuobjdumpsassfilename(const char* filename){ +void setCuobjdumpsassfilename(const char* filename, std::list &cuobjdumpSectionList){ if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ assert (0 && "You shouldn't be trying to add a sassfilename to an ptx section"); } @@ -2132,7 +2069,7 @@ static int get_app_cuda_version() { * It is also responsible for extracting the libraries linked to the binary if the option is * enabled * */ -void extract_code_using_cuobjdump(){ +void extract_code_using_cuobjdump(std::list &cuobjdumpSectionList){ CUctx_st *context = GPGPUSim_Context(); unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); @@ -2195,7 +2132,7 @@ void extract_code_using_cuobjdump(){ parser.ptxserial = 1; cuobjdump_lex_init(&(parser.scanner)); cuobjdump_set_in(cuobjdump_in, (parser.scanner)); - cuobjdump_parse(parser.scanner, &parser); + cuobjdump_parse(parser.scanner, &parser, cuobjdumpSectionList); cuobjdump_lex_destroy(parser.scanner); fclose(cuobjdump_in); printf("Done parsing!!!\n"); @@ -2251,7 +2188,7 @@ void extract_code_using_cuobjdump(){ parser.ptxserial = 1; cuobjdump_lex_init(&(parser.scanner)); cuobjdump_set_in(cuobjdump_in, (parser.scanner)); - cuobjdump_parse(parser.scanner, &parser); + cuobjdump_parse(parser.scanner, &parser, cuobjdumpSectionList); cuobjdump_lex_destroy(parser.scanner); fclose(cuobjdump_in); std::getline(libsf, line); @@ -2445,7 +2382,7 @@ cuobjdumpELFSection* findELFSectionInList(std::list sectionli } //! Find an ELF section in all the known lists -cuobjdumpELFSection* findELFSection(const std::string identifier){ +cuobjdumpELFSection* findELFSection(const std::string identifier, std::list cuobjdumpSectionList){ cuobjdumpELFSection* sec = findELFSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; sec = findELFSectionInList(libSectionList, identifier); @@ -2480,7 +2417,7 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list sectionli } //! Find an PTX section in all the known lists -cuobjdumpPTXSection* findPTXSection(const std::string identifier){ +cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list cuobjdumpSectionList){ cuobjdumpPTXSection* sec = findPTXSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; sec = findPTXSectionInList(libSectionList, identifier); @@ -2493,9 +2430,9 @@ cuobjdumpPTXSection* findPTXSection(const std::string identifier){ //! Extract the code using cuobjdump and remove unnecessary sections -void cuobjdumpInit(){ +void cuobjdumpInit(std::list &cuobjdumpSectionList){ CUctx_st *context = GPGPUSim_Context(); - extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.* + extract_code_using_cuobjdump(cuobjdumpSectionList); //extract all the output of cuobjdump to _cuobjdump_*.* const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if (pre_load ==NULL || strlen(pre_load)==0){ cuobjdumpSectionList = pruneSectionList(cuobjdumpSectionList, context); @@ -2513,7 +2450,7 @@ void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename){ } //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it -void cuobjdumpParseBinary(unsigned int handle){ +void cuobjdumpParseBinary(unsigned int handle, std::list &cuobjdumpSectionList){ if(fatbin_registered[handle]) return; fatbin_registered[handle] = true; @@ -2566,7 +2503,7 @@ void cuobjdumpParseBinary(unsigned int handle){ cuobjdumpPTXSection* ptx = NULL; const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if(pre_load==NULL || strlen(pre_load)==0) - ptx = findPTXSection(fname); + ptx = findPTXSection(fname, context->cuobjdumpSectionList); char *ptxcode; const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) { @@ -2576,7 +2513,7 @@ void cuobjdumpParseBinary(unsigned int handle){ ptxcode = readfile(override_ptx_name); } if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier()); + cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier(), context->cuobjdumpSectionList); assert (elfsection!= NULL); char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( ptx->getPTXfilename(), @@ -2664,7 +2601,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) * then for next calls, only returns the appropriate number */ assert(fat_cubin_handle >= 1); - if (fat_cubin_handle==1) cuobjdumpInit(); + if (fat_cubin_handle==1) cuobjdumpInit(context->cuobjdumpSectionList); cuobjdumpRegisterFatBinary(fat_cubin_handle, filename); return (void**)fat_cubin_handle; @@ -2779,7 +2716,7 @@ void CUDARTAPI __cudaRegisterFunction( printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", deviceFun, hostFun, fat_cubin_handle); if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - cuobjdumpParseBinary(fat_cubin_handle); + cuobjdumpParseBinary(fat_cubin_handle, context->cuobjdumpSectionList); context->register_function( fat_cubin_handle, hostFun, deviceFun ); } @@ -2799,7 +2736,7 @@ extern void __cudaRegisterVar( printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); + cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle, GPGPUSim_Context()->cuobjdumpSectionList ); fflush(stdout); if ( constant && !global && !ext ) { gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); diff --git a/libcuda/cuobjdump.h b/libcuda/cuobjdump.h index 66cd736..49af3e2 100644 --- a/libcuda/cuobjdump.h +++ b/libcuda/cuobjdump.h @@ -1,5 +1,9 @@ #ifndef __cuobjdump_h__ #define __cuobjdump_h__ +#include +#include +#include + struct cuobjdump_parser { yyscan_t scanner; int elfserial; @@ -9,4 +13,67 @@ struct cuobjdump_parser { FILE *sassfile; char filename [1024]; }; + +class cuobjdumpSection { +public: + //Constructor + cuobjdumpSection() { + arch = 0; + identifier = ""; + } + virtual ~cuobjdumpSection() {} + unsigned getArch() {return arch;} + void setArch(unsigned a) {arch = a;} + std::string getIdentifier() {return identifier;} + void setIdentifier(std::string i) {identifier = i;} + virtual void print(){std::cout << "cuobjdump Section: unknown type" << std::endl;} +private: + unsigned arch; + std::string identifier; +}; + +class cuobjdumpELFSection : public cuobjdumpSection +{ +public: + cuobjdumpELFSection() {} + virtual ~cuobjdumpELFSection() { + elffilename = ""; + sassfilename = ""; + } + std::string getELFfilename() {return elffilename;} + void setELFfilename(std::string f) {elffilename = f;} + std::string getSASSfilename() {return sassfilename;} + void setSASSfilename(std::string f) {sassfilename = f;} + virtual void print() { + std::cout << "ELF Section:" << std::endl; + std::cout << "arch: sm_" << getArch() << std::endl; + std::cout << "identifier: " << getIdentifier() << std::endl; + std::cout << "elf filename: " << getELFfilename() << std::endl; + std::cout << "sass filename: " << getSASSfilename() << std::endl; + std::cout << std::endl; + } +private: + std::string elffilename; + std::string sassfilename; +}; + +class cuobjdumpPTXSection : public cuobjdumpSection +{ +public: + cuobjdumpPTXSection(){ + ptxfilename = ""; + } + std::string getPTXfilename() {return ptxfilename;} + void setPTXfilename(std::string f) {ptxfilename = f;} + virtual void print() { + std::cout << "PTX Section:" << std::endl; + std::cout << "arch: sm_" << getArch() << std::endl; + std::cout << "identifier: " << getIdentifier() << std::endl; + std::cout << "ptx filename: " << getPTXfilename() << std::endl; + std::cout << std::endl; + } +private: + std::string ptxfilename; +}; + #endif /* __cuobjdump_h__ */ diff --git a/libcuda/cuobjdump.l b/libcuda/cuobjdump.l index eccc1f2..5a19d65 100644 --- a/libcuda/cuobjdump.l +++ b/libcuda/cuobjdump.l @@ -39,9 +39,9 @@ #define YYDEBUG 1 -void cuobjdump_error(yyscan_t yyscanner, struct cuobjdump_parser* parser, const char* msg); +void cuobjdump_error(yyscan_t yyscanner, struct cuobjdump_parser* parser, std::list &cuobjdumpSectionList, const char* msg); #define YY_DECL int cuobjdump_lex \ - (YYSTYPE * yylval_param , yyscan_t yyscanner, struct cuobjdump_parser* parser) + (YYSTYPE * yylval_param , yyscan_t yyscanner, struct cuobjdump_parser* parser, std::list &cuobjdumpSectionList) %} %option stack @@ -156,10 +156,10 @@ newlines {newline}+ <> BEGIN(INITIAL);return 0; /*No other rule matched. Throw an error*/ -. cuobjdump_error(yyscanner, parser, "Invalid token"); +. cuobjdump_error(yyscanner, parser, cuobjdumpSectionList, "Invalid token"); %% -void cuobjdump_error(yyscan_t yyscanner, struct cuobjdump_parser* parser, const char* msg) +void cuobjdump_error(yyscan_t yyscanner, struct cuobjdump_parser* parser, std::list &cuobjdumpSectionList, const char* msg) { struct yyguts_t * yyg = (struct yyguts_t*)yyscanner; printf(" %s near \"%s\"",msg, yytext); diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y index fcc863e..8d1bca6 100644 --- a/libcuda/cuobjdump.y +++ b/libcuda/cuobjdump.y @@ -32,25 +32,27 @@ typedef void * yyscan_t; #include "cuobjdump.h" -extern void addCuobjdumpSection(int sectiontype); -void setCuobjdumparch(const char* arch); -void setCuobjdumpidentifier(const char* identifier); -void setCuobjdumpptxfilename(const char* filename); -void setCuobjdumpelffilename(const char* filename); -void setCuobjdumpsassfilename(const char* filename); +extern void addCuobjdumpSection(int sectiontype, std::list &cuobjdumpSectionList); +void setCuobjdumparch(const char* arch, std::list &cuobjdumpSectionList); +void setCuobjdumpidentifier(const char* identifier, std::list &cuobjdumpSectionList); +void setCuobjdumpptxfilename(const char* filename, std::list &cuobjdumpSectionList); +void setCuobjdumpelffilename(const char* filename, std::list &cuobjdumpSectionList); +void setCuobjdumpsassfilename(const char* filename, std::list &cuobjdumpSectionList); %} %define api.pure full %parse-param {yyscan_t scanner} %parse-param {struct cuobjdump_parser* parser} +%parse-param {std::list &cuobjdumpSectionList} %lex-param {yyscan_t scanner} %lex-param {struct cuobjdump_parser* parser} +%lex-param {std::list &cuobjdumpSectionList} %union { char* string_value; } %{ -int yylex(YYSTYPE * yylval_param, yyscan_t yyscanner, struct cuobjdump_parser* parser); -void yyerror(yyscan_t yyscanner, struct cuobjdump_parser* parser, const char* msg); +int yylex(YYSTYPE * yylval_param, yyscan_t yyscanner, struct cuobjdump_parser* parser, std::list &cuobjdumpSectionList); +void yyerror(yyscan_t yyscanner, struct cuobjdump_parser* parser, std::list &cuobjdumpSectionList, const char* msg); %} %token H_SEPARATOR H_ARCH H_CODEVERSION H_PRODUCER H_HOST H_COMPILESIZE H_IDENTIFIER H_UNKNOWN H_COMPRESSED %token CODEVERSION @@ -75,23 +77,23 @@ emptylines : emptylines NEWLINE | ; section : PTXHEADER { - addCuobjdumpSection(0); + addCuobjdumpSection(0, cuobjdumpSectionList); snprintf(parser->filename, 1024, "_cuobjdump_%d.ptx", parser->ptxserial++); parser->ptxfile = fopen(parser->filename, "w"); - setCuobjdumpptxfilename(parser->filename); + setCuobjdumpptxfilename(parser->filename, cuobjdumpSectionList); } headerinfo compressedkeyword identifier ptxcode { fclose(parser->ptxfile); } | ELFHEADER { - addCuobjdumpSection(1); + addCuobjdumpSection(1, cuobjdumpSectionList); snprintf(parser->filename, 1024, "_cuobjdump_%d.elf", parser->elfserial); parser->elffile = fopen(parser->filename, "w"); - setCuobjdumpelffilename(parser->filename); + setCuobjdumpelffilename(parser->filename, cuobjdumpSectionList); } headerinfo compressedkeyword identifier elfcode { fclose(parser->elffile); snprintf(parser->filename, 1024, "_cuobjdump_%d.sass", parser->elfserial++); parser->sassfile = fopen(parser->filename, "w"); - setCuobjdumpsassfilename(parser->filename); + setCuobjdumpsassfilename(parser->filename, cuobjdumpSectionList); } sasscode { fclose(parser->sassfile); }; @@ -101,16 +103,16 @@ headerinfo : H_SEPARATOR NEWLINE H_CODEVERSION CODEVERSION NEWLINE H_PRODUCER H_UNKNOWN NEWLINE H_HOST IDENTIFIER NEWLINE - H_COMPILESIZE IDENTIFIER {setCuobjdumparch($4);}; + H_COMPILESIZE IDENTIFIER {setCuobjdumparch($4, cuobjdumpSectionList);}; | H_SEPARATOR NEWLINE H_ARCH IDENTIFIER NEWLINE H_CODEVERSION CODEVERSION NEWLINE H_PRODUCER IDENTIFIER NEWLINE H_HOST IDENTIFIER NEWLINE - H_COMPILESIZE IDENTIFIER {setCuobjdumparch($4);}; + H_COMPILESIZE IDENTIFIER {setCuobjdumparch($4, cuobjdumpSectionList);}; -identifier : H_IDENTIFIER FILENAME emptylines {setCuobjdumpidentifier($2);} - | {setCuobjdumpidentifier("default");}; +identifier : H_IDENTIFIER FILENAME emptylines {setCuobjdumpidentifier($2, cuobjdumpSectionList);} + | {setCuobjdumpidentifier("default", cuobjdumpSectionList);}; compressedkeyword : H_COMPRESSED emptylines | ; -- cgit v1.3 From 7864d663823cb5f6af7d9d4eb25c4527a8b4e80f Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 29 May 2019 12:25:48 -0400 Subject: Move libSectionList Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index df7ddc7..225e93b 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -305,6 +305,8 @@ struct CUctx_st { } std::list cuobjdumpSectionList; + std::list libSectionList; + private: _cuda_device_id *m_gpu; // selected gpu std::map m_code; // fat binary handle => global symbol table @@ -497,8 +499,6 @@ enum cuobjdumpSectionType { }; -std::list libSectionList; - // sectiontype: 0 for ptx, 1 for elf void addCuobjdumpSection(int sectiontype, std::list &cuobjdumpSectionList){ if (sectiontype) @@ -2193,7 +2193,7 @@ void extract_code_using_cuobjdump(std::list &cuobjdumpSection fclose(cuobjdump_in); std::getline(libsf, line); } - libSectionList = cuobjdumpSectionList; + context->libSectionList = cuobjdumpSectionList; //Restore the original section list cuobjdumpSectionList = tmpsl; @@ -2382,7 +2382,7 @@ cuobjdumpELFSection* findELFSectionInList(std::list sectionli } //! Find an ELF section in all the known lists -cuobjdumpELFSection* findELFSection(const std::string identifier, std::list cuobjdumpSectionList){ +cuobjdumpELFSection* findELFSection(const std::string identifier, std::list cuobjdumpSectionList, std::list &libSectionList){ cuobjdumpELFSection* sec = findELFSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; sec = findELFSectionInList(libSectionList, identifier); @@ -2417,7 +2417,7 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list sectionli } //! Find an PTX section in all the known lists -cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list cuobjdumpSectionList){ +cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list cuobjdumpSectionList, std::list &libSectionList){ cuobjdumpPTXSection* sec = findPTXSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; sec = findPTXSectionInList(libSectionList, identifier); @@ -2503,7 +2503,7 @@ void cuobjdumpParseBinary(unsigned int handle, std::list &cuo cuobjdumpPTXSection* ptx = NULL; const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if(pre_load==NULL || strlen(pre_load)==0) - ptx = findPTXSection(fname, context->cuobjdumpSectionList); + ptx = findPTXSection(fname, context->cuobjdumpSectionList, context->libSectionList); char *ptxcode; const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) { @@ -2513,7 +2513,7 @@ void cuobjdumpParseBinary(unsigned int handle, std::list &cuo ptxcode = readfile(override_ptx_name); } if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier(), context->cuobjdumpSectionList); + cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier(), context->cuobjdumpSectionList, context->libSectionList); assert (elfsection!= NULL); char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( ptx->getPTXfilename(), -- cgit v1.3 From d9ca3558c774b8b86bb18024bbbf330df53722f7 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 29 May 2019 14:48:14 -0400 Subject: Move version_filename and g_cuda_launch_stack Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 50 ++++++++++++++++++++++++--------------------- 1 file changed, 27 insertions(+), 23 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 225e93b..f3c827c 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -153,8 +153,6 @@ std::map pinned_memory; //support for pinned memories added std::map pinned_memory_size; std::map g_mallocPtr_Size; int no_of_ptx=0; -//maps sm version number to set of filenames -std::map > version_filename; extern void synchronize(); extern void exit_simulation(); @@ -241,6 +239,8 @@ private: struct _cuda_device_id *m_next; }; +class kernel_config; + struct CUctx_st { CUctx_st( _cuda_device_id *gpu ) { @@ -306,6 +306,9 @@ struct CUctx_st { std::list cuobjdumpSectionList; std::list libSectionList; + //maps sm version number to set of filenames + std::map > version_filename; + std::list g_cuda_launch_stack; private: _cuda_device_id *m_gpu; // selected gpu @@ -486,7 +489,6 @@ typedef std::map event_tracker_t; int CUevent_st::m_next_event_uid; event_tracker_t g_timer_events; int g_active_device = 0; //active gpu that runs the code -std::list g_cuda_launch_stack; extern int cuobjdump_lex_init(yyscan_t* scanner); extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); @@ -1505,7 +1507,8 @@ __host__ cudaError_t CUDARTAPI cudaConfigureCall(dim3 gridDim, dim3 blockDim, si announce_call(__my_func__); } struct CUstream_st *s = (struct CUstream_st *)stream; - g_cuda_launch_stack.push_back( kernel_config(gridDim,blockDim,sharedMem,s) ); + CUctx_st *context = GPGPUSim_Context(); + context->g_cuda_launch_stack.push_back( kernel_config(gridDim,blockDim,sharedMem,s) ); return g_last_cudaError = cudaSuccess; } @@ -1514,8 +1517,9 @@ __host__ cudaError_t CUDARTAPI cudaSetupArgument(const void *arg, size_t size, s if(g_debug_execution >= 3){ announce_call(__my_func__); } - gpgpusim_ptx_assert( !g_cuda_launch_stack.empty(), "empty launch stack" ); - kernel_config &config = g_cuda_launch_stack.back(); + CUctx_st *context = GPGPUSim_Context(); + gpgpusim_ptx_assert( !context->g_cuda_launch_stack.empty(), "empty launch stack" ); + kernel_config &config = context->g_cuda_launch_stack.back(); config.set_arg(arg,size,offset); printf("GPGPU-Sim PTX: Setting up arguments for %zu bytes starting at 0x%llx..\n",size, (unsigned long long) arg); @@ -1532,8 +1536,8 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) char *mode = getenv("PTX_SIM_MODE_FUNC"); if( mode ) sscanf(mode,"%u", &g_ptx_sim_mode); - gpgpusim_ptx_assert( !g_cuda_launch_stack.empty(), "empty launch stack" ); - kernel_config config = g_cuda_launch_stack.back(); + gpgpusim_ptx_assert( !context->g_cuda_launch_stack.empty(), "empty launch stack" ); + kernel_config config = context->g_cuda_launch_stack.back(); struct CUstream_st *stream = config.get_stream(); printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); @@ -1574,14 +1578,14 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) g_checkpoint->load_global_mem(global_mem, f1name); printf("Skipping kernel %d as resuming from kernel %d\n",grid->get_uid(),gpu->resume_kernel ); - g_cuda_launch_stack.pop_back(); + context->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; } if(gpu->checkpoint_option==1 && (grid->get_uid()>gpu->checkpoint_kernel)) { printf("Skipping kernel %d as checkpoint from kernel %d\n",grid->get_uid(),gpu->checkpoint_kernel ); - g_cuda_launch_stack.pop_back(); + context->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; } @@ -1589,7 +1593,7 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); stream_operation op(grid,g_ptx_sim_mode,stream); g_stream_manager->push(op); - g_cuda_launch_stack.pop_back(); + context->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; } @@ -1969,7 +1973,7 @@ char* get_app_binary_name(std::string abs_path){ } //extracts all ptx files from binary and dumps into prog_name.unique_no.sm_<>.ptx files -void extract_ptx_files_using_cuobjdump(){ +void extract_ptx_files_using_cuobjdump(CUctx_st *context){ extern bool g_cdp_enabled; char command[1000]; char *pytorch_bin = getenv("PYTORCH_BIN"); @@ -2030,10 +2034,10 @@ void extract_ptx_files_using_cuobjdump(){ } std::string vstr = line.substr(pos1+3,pos2-pos1-3); int version = atoi(vstr.c_str()); - if (version_filename.find(version)==version_filename.end()){ - version_filename[version] = std::set(); + if (context->version_filename.find(version)==context->version_filename.end()){ + context->version_filename[version] = std::set(); } - version_filename[version].insert(line); + context->version_filename[version].insert(line); } } @@ -2091,7 +2095,7 @@ void extract_code_using_cuobjdump(std::list &cuobjdumpSection //dump ptx for all individial ptx files into sepearte files which is later used by ptxas. int result=0; #if (CUDART_VERSION >= 6000) - extract_ptx_files_using_cuobjdump(); + extract_ptx_files_using_cuobjdump(context); return; #endif //TODO: redundant to dump twice. how can it be prevented? @@ -2450,7 +2454,7 @@ void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename){ } //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it -void cuobjdumpParseBinary(unsigned int handle, std::list &cuobjdumpSectionList){ +void cuobjdumpParseBinary(unsigned int handle){ if(fatbin_registered[handle]) return; fatbin_registered[handle] = true; @@ -2467,7 +2471,7 @@ void cuobjdumpParseBinary(unsigned int handle, std::list &cuo #if (CUDART_VERSION >= 6000) //loops through all ptx files from smallest sm version to largest std::map >::iterator itr_m; - for (itr_m = version_filename.begin(); itr_m!=version_filename.end(); itr_m++){ + for (itr_m = context->version_filename.begin(); itr_m!=context->version_filename.end(); itr_m++){ std::set::iterator itr_s; for (itr_s = itr_m->second.begin(); itr_s!=itr_m->second.end(); itr_s++){ std::string ptx_filename = *itr_s; @@ -2479,7 +2483,7 @@ void cuobjdumpParseBinary(unsigned int handle, std::list &cuo context->add_binary(symtab, handle); load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - for (itr_m = version_filename.begin(); itr_m!=version_filename.end(); itr_m++){ + for (itr_m = context->version_filename.begin(); itr_m!=context->version_filename.end(); itr_m++){ std::set::iterator itr_s; for (itr_s = itr_m->second.begin(); itr_s!=itr_m->second.end(); itr_s++){ std::string ptx_filename = *itr_s; @@ -2491,8 +2495,8 @@ void cuobjdumpParseBinary(unsigned int handle, std::list &cuo #endif unsigned max_capability = 0; - for ( std::list::iterator iter = cuobjdumpSectionList.begin(); - iter != cuobjdumpSectionList.end(); + for ( std::list::iterator iter = context->cuobjdumpSectionList.begin(); + iter != context->cuobjdumpSectionList.end(); iter++){ unsigned capability = (*iter)->getArch(); if (capability > max_capability) max_capability = capability; @@ -2716,7 +2720,7 @@ void CUDARTAPI __cudaRegisterFunction( printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", deviceFun, hostFun, fat_cubin_handle); if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - cuobjdumpParseBinary(fat_cubin_handle, context->cuobjdumpSectionList); + cuobjdumpParseBinary(fat_cubin_handle); context->register_function( fat_cubin_handle, hostFun, deviceFun ); } @@ -2736,7 +2740,7 @@ extern void __cudaRegisterVar( printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle, GPGPUSim_Context()->cuobjdumpSectionList ); + cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); fflush(stdout); if ( constant && !global && !ext ) { gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); -- cgit v1.3 From 9e3a9ac5ed0a70ec9b048bd3cb4df781687e85f8 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 29 May 2019 19:23:40 -0400 Subject: Move fatbin etc globals Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 111 +++++++++++++++++++++----------------------- src/cuda-sim/ptx_loader.cc | 2 +- src/cuda-sim/ptx_loader.h | 2 +- 3 files changed, 56 insertions(+), 59 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index f3c827c..fcd5b07 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -149,10 +149,6 @@ typedef void * yyscan_t; #include #endif -std::map pinned_memory; //support for pinned memories added -std::map pinned_memory_size; -std::map g_mallocPtr_Size; -int no_of_ptx=0; extern void synchronize(); extern void exit_simulation(); @@ -241,12 +237,25 @@ private: class kernel_config; +#ifndef OPENGL_SUPPORT +typedef unsigned long GLuint; +#endif + +struct glbmap_entry { + GLuint m_bufferObj; + void *m_devPtr; + size_t m_size; + struct glbmap_entry *m_next; +}; + struct CUctx_st { CUctx_st( _cuda_device_id *gpu ) { m_gpu = gpu; m_binary_info.cmem = 0; m_binary_info.gmem = 0; + no_of_ptx=0; + g_glbmap = NULL; } _cuda_device_id *get_device() { return m_gpu; } @@ -309,6 +318,16 @@ struct CUctx_st { //maps sm version number to set of filenames std::map > version_filename; std::list g_cuda_launch_stack; + std::mapfatbin_registered; + std::map fatbinmap; + std::map g_mallocPtr_Size; + std::map name_symtab; + std::map pinned_memory; //support for pinned memories added + std::map pinned_memory_size; + int no_of_ptx; + typedef struct glbmap_entry glbmap_entry_t; + + glbmap_entry_t* g_glbmap; private: _cuda_device_id *m_gpu; // selected gpu @@ -573,7 +592,7 @@ __host__ cudaError_t CUDARTAPI cudaMalloc(void **devPtr, size_t size) *devPtr = context->get_device()->get_gpgpu()->gpu_malloc(size); if(g_debug_execution >= 3){ printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *devPtr); - g_mallocPtr_Size[(unsigned long long)*devPtr] = size; + context->g_mallocPtr_Size[(unsigned long long)*devPtr] = size; } if ( *devPtr ) { return g_last_cudaError = cudaSuccess; @@ -587,11 +606,11 @@ __host__ cudaError_t CUDARTAPI cudaMallocHost(void **ptr, size_t size) if(g_debug_execution >= 3){ announce_call(__my_func__); } - GPGPUSim_Context(); + CUctx_st* context = GPGPUSim_Context(); *ptr = malloc(size); if ( *ptr ) { //track pinned memory size allocated in the host so that same amount of memory is also allocated in GPU. - pinned_memory_size[*ptr]=size; + context->pinned_memory_size[*ptr]=size; return g_last_cudaError = cudaSuccess; } else { return g_last_cudaError = cudaErrorMemoryAllocation; @@ -2010,11 +2029,11 @@ void extract_ptx_files_using_cuobjdump(CUctx_st *context){ printf("ERROR: command: %s failed \n",command); exit(0); } - no_of_ptx++; + context->no_of_ptx++; } } - if(!no_of_ptx){ + if(!context->no_of_ptx){ printf("WARNING: Number of ptx in the executable file are 0. One of the reasons might be\n"); printf("\t1. CDP is enabled\n"); printf("\t2. When using PyTorch, PYTORCH_BIN is not set correctly\n"); @@ -2444,25 +2463,22 @@ void cuobjdumpInit(std::list &cuobjdumpSectionList){ } } -std::map fatbinmap; -std::mapfatbin_registered; -std::map name_symtab; //! Keep track of the association between filename and cubin handle -void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename){ - fatbinmap[handle] = filename; +void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ + context->fatbinmap[handle] = filename; } //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it void cuobjdumpParseBinary(unsigned int handle){ - if(fatbin_registered[handle]) return; - fatbin_registered[handle] = true; CUctx_st *context = GPGPUSim_Context(); - std::string fname = fatbinmap[handle]; + if(context->fatbin_registered[handle]) return; + context->fatbin_registered[handle] = true; + std::string fname = context->fatbinmap[handle]; - if (name_symtab.find(fname) != name_symtab.end()) { - symbol_table *symtab = name_symtab[fname]; + if (context->name_symtab.find(fname) != context->name_symtab.end()) { + symbol_table *symtab = context->name_symtab[fname]; context->add_binary(symtab, handle); return; } @@ -2479,7 +2495,7 @@ void cuobjdumpParseBinary(unsigned int handle){ symtab = gpgpu_ptx_sim_load_ptx_from_filename( ptx_filename.c_str() ); } } - name_symtab[fname] = symtab; + context->name_symtab[fname] = symtab; context->add_binary(symtab, handle); load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); @@ -2526,18 +2542,18 @@ void cuobjdumpParseBinary(unsigned int handle){ symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxplus_str, handle); printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); context->add_binary(symtab, handle); - gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); + gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability, context->no_of_ptx ); delete[] ptxplus_str; } else { symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxcode, handle); //if CUOBJDUMP_SIM_FILE is not set, ptx is NULL. So comment below. //printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); context->add_binary(symtab, handle); - gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); + gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability, context->no_of_ptx ); } load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - name_symtab[fname] = symtab; + context->name_symtab[fname] = symtab; //TODO: Remove temporarily files as per configurations } @@ -2606,7 +2622,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) */ assert(fat_cubin_handle >= 1); if (fat_cubin_handle==1) cuobjdumpInit(context->cuobjdumpSectionList); - cuobjdumpRegisterFatBinary(fat_cubin_handle, filename); + cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); return (void**)fat_cubin_handle; } @@ -2658,7 +2674,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) } else { symtab=gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); context->add_binary(symtab,fat_cubin_handle); - gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability ); + gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); } source_num++; load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); @@ -2818,10 +2834,6 @@ char __cudaInitModule( } -#ifndef OPENGL_SUPPORT -typedef unsigned long GLuint; -#endif - cudaError_t cudaGLRegisterBufferObject(GLuint bufferObj) { if(g_debug_execution >= 3){ @@ -2831,16 +2843,6 @@ cudaError_t cudaGLRegisterBufferObject(GLuint bufferObj) return g_last_cudaError = cudaSuccess; } -struct glbmap_entry { - GLuint m_bufferObj; - void *m_devPtr; - size_t m_size; - struct glbmap_entry *m_next; -}; -typedef struct glbmap_entry glbmap_entry_t; - -glbmap_entry_t* g_glbmap = NULL; - cudaError_t cudaGLMapBufferObject(void** devPtr, GLuint bufferObj) { if(g_debug_execution >= 3){ @@ -2850,7 +2852,7 @@ cudaError_t cudaGLMapBufferObject(void** devPtr, GLuint bufferObj) GLint buffer_size=0; CUctx_st* ctx = GPGPUSim_Context(); - glbmap_entry_t *p = g_glbmap; + glbmap_entry_t *p = ctx->g_glbmap; while ( p && p->m_bufferObj != bufferObj ) p = p->m_next; if ( p == NULL ) { @@ -2861,8 +2863,8 @@ cudaError_t cudaGLMapBufferObject(void** devPtr, GLuint bufferObj) // create entry and insert to front of list glbmap_entry_t *n = (glbmap_entry_t *) calloc(1,sizeof(glbmap_entry_t)); - n->m_next = g_glbmap; - g_glbmap = n; + n->m_next = ctx->g_glbmap; + ctx->g_glbmap = n; // initialize entry n->m_bufferObj = bufferObj; @@ -2903,7 +2905,8 @@ cudaError_t cudaGLUnmapBufferObject(GLuint bufferObj) announce_call(__my_func__); } #ifdef OPENGL_SUPPORT - glbmap_entry_t *p = g_glbmap; + CUctx_st* ctx = GPGPUSim_Context(); + glbmap_entry_t *p = ctx->g_glbmap; while ( p && p->m_bufferObj != bufferObj ) p = p->m_next; if ( p == NULL ) @@ -2943,7 +2946,8 @@ cudaError_t CUDARTAPI cudaHostAlloc(void **pHost, size_t bytes, unsigned int fl *pHost = malloc(bytes); //need to track the size allocated so that cudaHostGetDevicePointer() can function properly. //TODO: vary this function behavior based on flags value (following nvidia documentation) - pinned_memory_size[*pHost]=bytes; + CUctx_st* context = GPGPUSim_Context(); + context->pinned_memory_size[*pHost]=bytes; if( *pHost ) return g_last_cudaError = cudaSuccess; else @@ -2960,16 +2964,16 @@ cudaError_t CUDARTAPI cudaHostGetDevicePointer(void **pDevice, void *pHost, unsi flags=0; CUctx_st* context = GPGPUSim_Context(); gpgpu_t *gpu = context->get_device()->get_gpgpu(); - std::map::const_iterator i = pinned_memory_size.find(pHost); - assert(i != pinned_memory_size.end()); + std::map::const_iterator i = context->pinned_memory_size.find(pHost); + assert(i != context->pinned_memory_size.end()); size_t size = i->second; *pDevice = gpu->gpu_malloc(size); if(g_debug_execution >= 3){ printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *pDevice); - g_mallocPtr_Size[(unsigned long long)*pDevice] = size; + context->g_mallocPtr_Size[(unsigned long long)*pDevice] = size; } if ( *pDevice ) { - pinned_memory[pHost]=pDevice; + context->pinned_memory[pHost]=pDevice; //Copy contents in cpu to gpu gpu->memcpy_to_gpu((size_t)*pDevice,pHost,size); return g_last_cudaError = cudaSuccess; @@ -3204,13 +3208,6 @@ int CUDARTAPI __cudaSynchronizeThreads(void**, void*) //////// -extern int ptx_parse(); -extern int ptx__scan_string(const char*); -extern FILE *ptx_in; - -extern int ptxinfo_parse(); -extern FILE *ptxinfo_in; - /// static functions static int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr, gpgpu_t *gpu ) @@ -3330,7 +3327,7 @@ kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, fflush(stdout); if(g_debug_execution >= 4){ - entry->ptx_jit_config(g_mallocPtr_Size, result->get_param_memory(), (gpgpu_t *) context->get_device()->get_gpgpu(), gridDim, blockDim); + entry->ptx_jit_config(context->g_mallocPtr_Size, result->get_param_memory(), (gpgpu_t *) context->get_device()->get_gpgpu(), gridDim, blockDim); } return result; @@ -3815,7 +3812,7 @@ cuLinkAddFile(CUlinkState state, CUjitInputType type, const char *path, strcat(file,path); symbol_table *symtab = gpgpu_ptx_sim_load_ptx_from_filename( file ); std::string fname(path); - name_symtab[fname] = symtab; + context->name_symtab[fname] = symtab; context->add_binary(symtab, 1); load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 735ff84..f037c34 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -367,7 +367,7 @@ void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_versio fclose(ptxinfo_in); } -void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version ) +void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version, int no_of_ptx ) { //do ptxas for individual files instead of one big embedded ptx. This prevents the duplicate defs and declarations. char ptx_file[1000]; diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index 36e439e..c4d8292 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -44,7 +44,7 @@ extern int no_of_ptx; //counter to track number of ptx files to be extracted in class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); class symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ); -void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20 ); +void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version ); char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); bool keep_intermediate_files(); -- cgit v1.3 From ed9f0e6b2a99840e9649551825a40a04e236dcd9 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 30 May 2019 18:28:15 -0400 Subject: adding new values to gpu context --- libcuda/cuda_runtime_api.cc | 42 +++++++++++++++++++------------------ src/abstract_hardware_model.cc | 9 ++++---- src/abstract_hardware_model.h | 2 +- src/cuda-sim/cuda-sim.cc | 36 +++++++++++++++---------------- src/cuda-sim/cuda_device_runtime.cc | 5 +++-- src/gpgpusim_entrypoint.cc | 12 +++++++++-- src/gpgpusim_entrypoint.h | 24 ++++++++++++++++----- 7 files changed, 78 insertions(+), 52 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index fcd5b07..4990edf 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -185,7 +185,7 @@ struct cudaArray cudaError_t g_last_cudaError = cudaSuccess; -extern stream_manager *g_stream_manager; +//extern stream_manager *g_stream_manager(); void register_ptx_function( const char *name, function_info *impl ) { @@ -375,7 +375,8 @@ private: struct _cuda_device_id *GPGPUSim_Init() { - static _cuda_device_id *the_device = NULL; + //static _cuda_device_id *the_device = NULL; + _cuda_device_id *the_device = GPGPUsim_ctx_ptr()->the_cude_device; if( !the_device ) { gpgpu_sim *the_gpu = gpgpu_ptx_sim_init_perf(); @@ -428,7 +429,8 @@ struct _cuda_device_id *GPGPUSim_Init() static CUctx_st* GPGPUSim_Context() { - static CUctx_st *the_context = NULL; + //static CUctx_st *the_context = NULL; + CUctx_st *the_context = GPGPUsim_ctx_ptr()->the_context; if( the_context == NULL ) { _cuda_device_id *the_gpu = GPGPUSim_Init(); the_context = new CUctx_st(the_gpu); @@ -699,21 +701,21 @@ __host__ cudaError_t CUDARTAPI cudaMemcpy(void *dst, const void *src, size_t cou if(g_debug_execution >= 3) printf("GPGPU-Sim PTX: cudaMemcpy(): devPtr = %p\n", dst); if( kind == cudaMemcpyHostToDevice ) - g_stream_manager->push( stream_operation(src,(size_t)dst,count,0) ); + g_stream_manager()->push( stream_operation(src,(size_t)dst,count,0) ); else if( kind == cudaMemcpyDeviceToHost ) - g_stream_manager->push( stream_operation((size_t)src,dst,count,0) ); + g_stream_manager()->push( stream_operation((size_t)src,dst,count,0) ); else if( kind == cudaMemcpyDeviceToDevice ) - g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,0) ); + g_stream_manager()->push( stream_operation((size_t)src,(size_t)dst,count,0) ); else if ( kind == cudaMemcpyDefault ) { if ((size_t)src >= GLOBAL_HEAP_START) { if ((size_t)dst >= GLOBAL_HEAP_START) - g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,0) ); // device to device + g_stream_manager()->push( stream_operation((size_t)src,(size_t)dst,count,0) ); // device to device else - g_stream_manager->push( stream_operation((size_t)src,dst,count,0) ); // device to host + g_stream_manager()->push( stream_operation((size_t)src,dst,count,0) ); // device to host } else { if ((size_t)dst >= GLOBAL_HEAP_START) - g_stream_manager->push( stream_operation(src,(size_t)dst,count,0) ); + g_stream_manager()->push( stream_operation(src,(size_t)dst,count,0) ); else { printf("GPGPU-Sim PTX: cudaMemcpy - ERROR : unsupported transfer: host to host\n"); abort(); @@ -855,7 +857,7 @@ __host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const char *symbol, const void assert(kind == cudaMemcpyHostToDevice); printf("GPGPU-Sim PTX: cudaMemcpyToSymbol: symbol = %p\n", symbol); //stream_operation( const char *symbol, const void *src, size_t count, size_t offset ) - g_stream_manager->push( stream_operation(src,symbol,count,offset,0) ); + g_stream_manager()->push( stream_operation(src,symbol,count,offset,0) ); //gpgpu_ptx_sim_memcpy_symbol(symbol,src,count,offset,1,context->get_device()->get_gpgpu()); return g_last_cudaError = cudaSuccess; } @@ -869,7 +871,7 @@ __host__ cudaError_t CUDARTAPI cudaMemcpyFromSymbol(void *dst, const char *symbo //CUctx_st *context = GPGPUSim_Context(); assert(kind == cudaMemcpyDeviceToHost); printf("GPGPU-Sim PTX: cudaMemcpyFromSymbol: symbol = %p\n", symbol); - g_stream_manager->push( stream_operation(symbol,dst,count,offset,0) ); + g_stream_manager()->push( stream_operation(symbol,dst,count,offset,0) ); //gpgpu_ptx_sim_memcpy_symbol(symbol,dst,count,offset,0,context->get_device()->get_gpgpu()); return g_last_cudaError = cudaSuccess; } @@ -898,9 +900,9 @@ __host__ cudaError_t CUDARTAPI cudaMemcpyAsync(void *dst, const void *src, size_ } struct CUstream_st *s = (struct CUstream_st *)stream; switch( kind ) { - case cudaMemcpyHostToDevice: g_stream_manager->push( stream_operation(src,(size_t)dst,count,s) ); break; - case cudaMemcpyDeviceToHost: g_stream_manager->push( stream_operation((size_t)src,dst,count,s) ); break; - case cudaMemcpyDeviceToDevice: g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,s) ); break; + case cudaMemcpyHostToDevice: g_stream_manager()->push( stream_operation(src,(size_t)dst,count,s) ); break; + case cudaMemcpyDeviceToHost: g_stream_manager()->push( stream_operation((size_t)src,dst,count,s) ); break; + case cudaMemcpyDeviceToDevice: g_stream_manager()->push( stream_operation((size_t)src,(size_t)dst,count,s) ); break; default: abort(); } @@ -1611,7 +1613,7 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); stream_operation op(grid,g_ptx_sim_mode,stream); - g_stream_manager->push(op); + g_stream_manager()->push(op); context->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; } @@ -1650,7 +1652,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamCreate(cudaStream_t *stream) printf("GPGPU-Sim PTX: cudaStreamCreate\n"); #if (CUDART_VERSION >= 3000) *stream = new struct CUstream_st(); - g_stream_manager->add_stream(*stream); + g_stream_manager()->add_stream(*stream); #else *stream = 0; printf("GPGPU-Sim PTX: WARNING: Asynchronous kernel execution not supported (%s)\n", __my_func__); @@ -1689,7 +1691,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream) //per-stream synchronization required for application using external libraries without explicit synchronization in the code to //avoid the stream_manager from spinning forever to destroy non-empty streams without making any forward progress. stream->synchronize(); - g_stream_manager->destroy_stream(stream); + g_stream_manager()->destroy_stream(stream); #endif return g_last_cudaError = cudaSuccess; } @@ -1769,7 +1771,7 @@ __host__ cudaError_t CUDARTAPI cudaEventRecord(cudaEvent_t event, cudaStream_t s if( !e ) return g_last_cudaError = cudaErrorUnknown; struct CUstream_st *s = (struct CUstream_st *)stream; stream_operation op(e,s); - g_stream_manager->push(op); + g_stream_manager()->push(op); return g_last_cudaError = cudaSuccess; } @@ -1785,11 +1787,11 @@ __host__ cudaError_t CUDARTAPI cudaStreamWaitEvent(cudaStream_t stream, cudaEven return g_last_cudaError = cudaSuccess; } if (!stream){ - g_stream_manager->pushCudaStreamWaitEventToAllStreams(e, flags); + g_stream_manager()->pushCudaStreamWaitEventToAllStreams(e, flags); } else { struct CUstream_st *s = (struct CUstream_st *)stream; stream_operation op(s,e,flags); - g_stream_manager->push(op); + g_stream_manager()->push(op); } return g_last_cudaError = cudaSuccess; } diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 63b139e..7755477 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -34,6 +34,7 @@ #include "cuda-sim/cuda-sim.h" #include "gpgpu-sim/gpu-sim.h" #include "option_parser.h" +#include "gpgpusim_entrypoint.h" #include #include #include @@ -771,14 +772,14 @@ void kernel_info_t::notify_parent_finished() { extern unsigned long long g_total_param_size; g_total_param_size -= ((m_kernel_entry->get_args_aligned_size() + 255)/256*256); m_parent_kernel->remove_child(this); - g_stream_manager->register_finished_kernel(m_parent_kernel->get_uid()); + g_stream_manager()->register_finished_kernel(m_parent_kernel->get_uid()); } } CUstream_st * kernel_info_t::create_stream_cta(dim3 ctaid) { assert(get_default_stream_cta(ctaid)); CUstream_st * stream = new CUstream_st(); - g_stream_manager->add_stream(stream); + g_stream_manager()->add_stream(stream); assert(m_cta_streams.find(ctaid) != m_cta_streams.end()); assert(m_cta_streams[ctaid].size() >= 1); //must have default stream m_cta_streams[ctaid].push_back(stream); @@ -794,7 +795,7 @@ CUstream_st * kernel_info_t::get_default_stream_cta(dim3 ctaid) { else { m_cta_streams[ctaid] = std::list(); CUstream_st * stream = new CUstream_st(); - g_stream_manager->add_stream(stream); + g_stream_manager()->add_stream(stream); m_cta_streams[ctaid].push_back(stream); return stream; } @@ -826,7 +827,7 @@ void kernel_info_t::destroy_cta_streams() { for(auto s = m_cta_streams.begin(); s != m_cta_streams.end(); s++) { stream_size += s->second.size(); for(auto ss = s->second.begin(); ss != s->second.end(); ss++) - g_stream_manager->destroy_stream(*ss); + g_stream_manager()->destroy_stream(*ss); s->second.clear(); } printf("size %lu\n", stream_size); diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 1735c2f..77d5f58 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -197,7 +197,7 @@ void increment_x_then_y_then_z( dim3 &i, const dim3 &bound); #include "stream_manager.h" class stream_manager; struct CUstream_st; -extern stream_manager * g_stream_manager; +//extern stream_manager * g_stream_manager; //support for pinned memories added extern std::map pinned_memory; extern std::map pinned_memory_size; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index e733b7f..a456978 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -448,8 +448,8 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun m_global_mem->write(dst_start_addr+n,1, src_data+n,NULL,NULL); // Copy into the performance model. - extern gpgpu_sim* g_the_gpu; - g_the_gpu->perf_memcpy_to_gpu(dst_start_addr, count); + //extern gpgpu_sim* g_the_gpu; + g_the_gpu()->perf_memcpy_to_gpu(dst_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -467,8 +467,8 @@ void gpgpu_t::memcpy_from_gpu( void *dst, size_t src_start_addr, size_t count ) m_global_mem->read(src_start_addr+n,1,dst_data+n); // Copy into the performance model. - extern gpgpu_sim* g_the_gpu; - g_the_gpu->perf_memcpy_to_gpu(src_start_addr, count); + //extern gpgpu_sim* g_the_gpu; + g_the_gpu()->perf_memcpy_to_gpu(src_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -1270,8 +1270,8 @@ void function_info::finalize( memory_space *param_mem ) void function_info::param_to_shared( memory_space *shared_mem, symbol_table *symtab ) { // TODO: call this only for PTXPlus with GT200 models - extern gpgpu_sim* g_the_gpu; - if (not g_the_gpu->get_config().convert_to_ptxplus()) return; + //extern gpgpu_sim* g_the_gpu; + if (not g_the_gpu()->get_config().convert_to_ptxplus()) return; // copies parameters into simulated shared memory for( std::map::iterator i=m_ptx_kernel_param_info.begin(); i!=m_ptx_kernel_param_info.end(); i++ ) { @@ -2150,7 +2150,7 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) printf("GPGPU-Sim: Performing Functional Simulation, executing kernel %s...\n",kernel.name().c_str()); //using a shader core object for book keeping, it is not needed but as most function built for performance simulation need it we use it here - extern gpgpu_sim *g_the_gpu; + //extern gpgpu_sim *g_the_gpu; //before we execute, we should do PDOM analysis for functional simulation scenario. function_info *kernel_func_info = kernel.entry(); const struct gpgpu_ptx_sim_info *kernel_info = ptx_sim_kernel_info(kernel_func_info); @@ -2165,7 +2165,7 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) kernel_func_info->set_pdom(); } - unsigned max_cta_tot = max_cta(kernel_info,kernel.threads_per_cta(), g_the_gpu->getShaderCoreConfig()->warp_size, g_the_gpu->getShaderCoreConfig()->n_thread_per_shader, g_the_gpu->getShaderCoreConfig()->gpgpu_shmem_size, g_the_gpu->getShaderCoreConfig()->gpgpu_shader_registers, g_the_gpu->getShaderCoreConfig()->max_cta_per_core); + unsigned max_cta_tot = max_cta(kernel_info,kernel.threads_per_cta(), g_the_gpu()->getShaderCoreConfig()->warp_size, g_the_gpu()->getShaderCoreConfig()->n_thread_per_shader, g_the_gpu()->getShaderCoreConfig()->gpgpu_shmem_size, g_the_gpu()->getShaderCoreConfig()->gpgpu_shader_registers, g_the_gpu()->getShaderCoreConfig()->max_cta_per_core); printf("Max CTA : %d\n",max_cta_tot); @@ -2173,11 +2173,11 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) int inst_count=50; - int cp_op= g_the_gpu->checkpoint_option; - int cp_CTA = g_the_gpu->checkpoint_CTA; - int cp_kernel= g_the_gpu->checkpoint_kernel; - cp_count= g_the_gpu->checkpoint_insn_Y; - cp_cta_resume= g_the_gpu->checkpoint_CTA_t; + int cp_op= g_the_gpu()->checkpoint_option; + int cp_CTA = g_the_gpu()->checkpoint_CTA; + int cp_kernel= g_the_gpu()->checkpoint_kernel; + cp_count= g_the_gpu()->checkpoint_insn_Y; + cp_cta_resume= g_the_gpu()->checkpoint_CTA_t; int cta_launched =0; //we excute the kernel one CTA (Block) at the time, as synchronization functions work block wise @@ -2189,8 +2189,8 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) { functionalCoreSim cta( &kernel, - g_the_gpu, - g_the_gpu->getShaderCoreConfig()->warp_size + g_the_gpu(), + g_the_gpu()->getShaderCoreConfig()->warp_size ); cta.execute(cp_count,temp); @@ -2211,7 +2211,7 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) { char f1name[2048]; snprintf(f1name,2048,"checkpoint_files/global_mem_%d.txt", kernel.get_uid() ); - g_checkpoint->store_global_mem(g_the_gpu->get_global_memory(), f1name , "%08x"); + g_checkpoint->store_global_mem(g_the_gpu()->get_global_memory(), f1name , "%08x"); } @@ -2221,8 +2221,8 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) //openCL kernel simulation calls don't register the kernel so we don't register its exit if(!openCL) { - extern stream_manager *g_stream_manager; - g_stream_manager->register_finished_kernel(kernel.get_uid()); + //extern stream_manager *g_stream_manager; + g_stream_manager()->register_finished_kernel(kernel.get_uid()); } //******PRINTING******* diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index 86e8147..be8369f 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -18,6 +18,7 @@ unsigned long long g_max_total_param_size = 0; #include "cuda-sim.h" #include "ptx_ir.h" #include "../stream_manager.h" +#include "../gpgpusim_entrypoint.h" #include "cuda_device_runtime.h" #define DEV_RUNTIME_REPORT(a) \ @@ -64,7 +65,7 @@ public: std::map g_cuda_device_launch_param_map; std::list g_cuda_device_launch_op; -extern stream_manager *g_stream_manager; +//extern stream_manager *g_stream_manager(); //Handling device runtime api: //void * cudaGetParameterBufferV2(void *func, dim3 gridDimension, dim3 blockDimension, unsigned int sharedMemSize) @@ -322,7 +323,7 @@ void launch_one_device_kernel() { device_launch_operation_t &op = g_cuda_device_launch_op.front(); stream_operation stream_op = stream_operation(op.grid, g_ptx_sim_mode, op.stream); - g_stream_manager->push(stream_op); + g_stream_manager()->push(stream_op); g_cuda_device_launch_op.pop_front(); } } diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index c8770e2..de937b0 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -41,8 +41,6 @@ struct GPGPUsim_ctx* the_gpgpusim = NULL; -static void print_simulation_time(); - struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(){ if(the_gpgpusim == NULL) the_gpgpusim = new GPGPUsim_ctx(); @@ -50,6 +48,16 @@ struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(){ return the_gpgpusim; } +class gpgpu_sim* g_the_gpu() { + return GPGPUsim_ctx_ptr()->g_the_gpu; +} + +class stream_manager* g_stream_manager() { + return GPGPUsim_ctx_ptr()->g_stream_manager; +} + +static void print_simulation_time(); + void *gpgpu_sim_thread_sequential(void*) { // at most one kernel running at a time diff --git a/src/gpgpusim_entrypoint.h b/src/gpgpusim_entrypoint.h index e29159b..2ad0fdf 100644 --- a/src/gpgpusim_entrypoint.h +++ b/src/gpgpusim_entrypoint.h @@ -46,9 +46,15 @@ struct GPGPUsim_ctx { sg_argc = 3; sg_argv = {"", "-config","gpgpusim.config"}; + + g_the_gpu_config=NULL; + g_the_gpu=NULL; + g_stream_manager=NULL; + the_cude_device=NULL; + the_context=NULL; } - struct gpgpu_ptx_sim_arg *grid_params; + //struct gpgpu_ptx_sim_arg *grid_params; sem_t g_sim_signal_start; sem_t g_sim_signal_finish; @@ -60,19 +66,27 @@ struct GPGPUsim_ctx { class gpgpu_sim *g_the_gpu; class stream_manager *g_stream_manager; + struct _cuda_device_id *the_cude_device; + struct CUctx_st* the_context; + + int sg_argc; const char *sg_argv[3]; - pthread_mutex_t g_sim_lock; - bool g_sim_active; - bool g_sim_done; - bool break_limit; + pthread_mutex_t g_sim_lock; + bool g_sim_active; + bool g_sim_done; + bool break_limit; }; class gpgpu_sim *gpgpu_ptx_sim_init_perf(); void start_sim_thread(int api); +class gpgpu_sim* g_the_gpu(); +struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(); +class stream_manager* g_stream_manager(); + int gpgpu_opencl_ptx_sim_main_perf( kernel_info_t *grid ); int gpgpu_opencl_ptx_sim_main_func( kernel_info_t *grid ); -- cgit v1.3 From 5d55ca32e2d81ee2c9187566050c5d399a142373 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 31 May 2019 13:22:29 -0400 Subject: add new params to the global stryct and fixing some bugs --- libcuda/cuda_runtime_api.cc | 6 ++++-- src/gpgpu-sim/gpu-sim.cc | 4 ++-- src/gpgpusim_entrypoint.h | 2 +- 3 files changed, 7 insertions(+), 5 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 4990edf..3d85b62 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -421,7 +421,8 @@ struct _cuda_device_id *GPGPUSim_Init() prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); #endif the_gpu->set_prop(prop); - the_device = new _cuda_device_id(the_gpu); + GPGPUsim_ctx_ptr()->the_cude_device = new _cuda_device_id(the_gpu); + the_device = GPGPUsim_ctx_ptr()->the_cude_device; } start_sim_thread(1); return the_device; @@ -433,7 +434,8 @@ static CUctx_st* GPGPUSim_Context() CUctx_st *the_context = GPGPUsim_ctx_ptr()->the_context; if( the_context == NULL ) { _cuda_device_id *the_gpu = GPGPUSim_Init(); - the_context = new CUctx_st(the_gpu); + GPGPUsim_ctx_ptr()->the_context = new CUctx_st(the_gpu); + the_context = GPGPUsim_ctx_ptr()->the_context; } return the_context; } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 6f19640..a557d6f 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1131,7 +1131,7 @@ void gpgpu_sim::gpu_print_stat() time_t curr_time; time(&curr_time); - unsigned long long elapsed_time = MAX( curr_time - g_simulation_starttime, 1 ); + unsigned long long elapsed_time = MAX( curr_time - GPGPUsim_ctx_ptr()->g_simulation_starttime, 1 ); printf( "gpu_total_sim_rate=%u\n", (unsigned)( ( gpu_tot_sim_insn + gpu_sim_insn ) / elapsed_time ) ); //shader_print_l1_miss_stat( stdout ); @@ -1701,7 +1701,7 @@ void gpgpu_sim::cycle() time_t days, hrs, minutes, sec; time_t curr_time; time(&curr_time); - unsigned long long elapsed_time = MAX(curr_time - g_simulation_starttime, 1); + unsigned long long elapsed_time = MAX(curr_time - GPGPUsim_ctx_ptr()->g_simulation_starttime, 1); if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq && DTRACE(LIVENESS) ) { days = elapsed_time/(3600*24); hrs = elapsed_time/3600 - 24*days; diff --git a/src/gpgpusim_entrypoint.h b/src/gpgpusim_entrypoint.h index 2ad0fdf..406dd00 100644 --- a/src/gpgpusim_entrypoint.h +++ b/src/gpgpusim_entrypoint.h @@ -33,7 +33,7 @@ #include #include -extern time_t g_simulation_starttime; +//extern time_t g_simulation_starttime; struct GPGPUsim_ctx { -- cgit v1.3 From 8ad9dbc77947b4abd51ebb55ef4bbe80be01caaa Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 4 Jun 2019 14:26:42 -0400 Subject: Add gpgpu_context as the top level class Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 182 ++++++++++++++++++++++++++++++++------------ libcuda/gpgpu_context.h | 18 +++++ 2 files changed, 150 insertions(+), 50 deletions(-) create mode 100644 libcuda/gpgpu_context.h (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 3d85b62..c39571c 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -131,6 +131,7 @@ #if (CUDART_VERSION < 8000) #include "__cudaFatFormat.h" #endif +#include "gpgpu_context.h" #include "../src/gpgpu-sim/gpu-sim.h" #include "../src/cuda-sim/ptx_loader.h" #include "../src/cuda-sim/cuda-sim.h" @@ -313,7 +314,6 @@ struct CUctx_st { return i->second; } - std::list cuobjdumpSectionList; std::list libSectionList; //maps sm version number to set of filenames std::map > version_filename; @@ -440,6 +440,15 @@ static CUctx_st* GPGPUSim_Context() return the_context; } +static gpgpu_context* GPGPU_Context() +{ + static gpgpu_context *gpgpu_ctx = NULL; + if( gpgpu_ctx == NULL ) { + gpgpu_ctx = new gpgpu_context(); + } + return gpgpu_ctx; +} + void ptxinfo_addinfo() { if(!get_ptxinfo_kname()){ @@ -2096,7 +2105,7 @@ static int get_app_cuda_version() { * It is also responsible for extracting the libraries linked to the binary if the option is * enabled * */ -void extract_code_using_cuobjdump(std::list &cuobjdumpSectionList){ +void gpgpu_context::extract_code_using_cuobjdump(){ CUctx_st *context = GPGPUSim_Context(); unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); @@ -2266,7 +2275,7 @@ void printSectionList(std::list sl) { } //! Remove unecessary sm versions from the section list -std::list pruneSectionList(std::list cuobjdumpSectionList, CUctx_st *context) { +std::list gpgpu_context::pruneSectionList(CUctx_st *context) { unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); //For ptxplus, force the max capability to 19 if it's higher or unspecified(0) @@ -2319,7 +2328,7 @@ std::list pruneSectionList(std::list cuobj } //! Merge all PTX sections that have a specific identifier into one file -std::list mergeMatchingSections(std::list cuobjdumpSectionList, std::string identifier){ +std::list gpgpu_context::mergeMatchingSections(std::string identifier){ const char *ptxcode = ""; std::list::iterator old_iter; cuobjdumpPTXSection* old_ptxsection = NULL; @@ -2362,7 +2371,7 @@ std::list mergeMatchingSections(std::list } //! Merge any PTX sections with matching identifiers -std::list mergeSections(std::list cuobjdumpSectionList){ +std::list gpgpu_context::mergeSections(){ std::vector identifier; cuobjdumpPTXSection* ptxsection; @@ -2384,7 +2393,7 @@ std::list mergeSections(std::list cuobjdum for ( std::vector::iterator iter = identifier.begin(); iter != identifier.end(); iter++) { - cuobjdumpSectionList = mergeMatchingSections(cuobjdumpSectionList, *iter); + cuobjdumpSectionList = mergeMatchingSections(*iter); } return cuobjdumpSectionList; @@ -2409,7 +2418,7 @@ cuobjdumpELFSection* findELFSectionInList(std::list sectionli } //! Find an ELF section in all the known lists -cuobjdumpELFSection* findELFSection(const std::string identifier, std::list cuobjdumpSectionList, std::list &libSectionList){ +cuobjdumpELFSection* gpgpu_context::findELFSection(const std::string identifier, std::list &libSectionList){ cuobjdumpELFSection* sec = findELFSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; sec = findELFSectionInList(libSectionList, identifier); @@ -2420,7 +2429,7 @@ cuobjdumpELFSection* findELFSection(const std::string identifier, std::list sectionlist, const std::string identifier){ +cuobjdumpPTXSection* findPTXSectionInList(std::list §ionlist, const std::string identifier){ std::list::iterator iter; for ( iter = sectionlist.begin(); iter != sectionlist.end(); @@ -2444,7 +2453,7 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list sectionli } //! Find an PTX section in all the known lists -cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list cuobjdumpSectionList, std::list &libSectionList){ +cuobjdumpPTXSection* gpgpu_context::findPTXSection(const std::string identifier, std::list &libSectionList){ cuobjdumpPTXSection* sec = findPTXSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; sec = findPTXSectionInList(libSectionList, identifier); @@ -2457,13 +2466,13 @@ cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list &cuobjdumpSectionList){ +void gpgpu_context::cuobjdumpInit(){ CUctx_st *context = GPGPUSim_Context(); - extract_code_using_cuobjdump(cuobjdumpSectionList); //extract all the output of cuobjdump to _cuobjdump_*.* + extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.* const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if (pre_load ==NULL || strlen(pre_load)==0){ - cuobjdumpSectionList = pruneSectionList(cuobjdumpSectionList, context); - cuobjdumpSectionList = mergeSections(cuobjdumpSectionList); + cuobjdumpSectionList = pruneSectionList(context); + cuobjdumpSectionList = mergeSections(); } } @@ -2474,7 +2483,7 @@ void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx } //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it -void cuobjdumpParseBinary(unsigned int handle){ +void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ CUctx_st *context = GPGPUSim_Context(); if(context->fatbin_registered[handle]) return; @@ -2515,8 +2524,8 @@ void cuobjdumpParseBinary(unsigned int handle){ #endif unsigned max_capability = 0; - for ( std::list::iterator iter = context->cuobjdumpSectionList.begin(); - iter != context->cuobjdumpSectionList.end(); + for ( std::list::iterator iter = cuobjdumpSectionList.begin(); + iter != cuobjdumpSectionList.end(); iter++){ unsigned capability = (*iter)->getArch(); if (capability > max_capability) max_capability = capability; @@ -2527,7 +2536,7 @@ void cuobjdumpParseBinary(unsigned int handle){ cuobjdumpPTXSection* ptx = NULL; const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if(pre_load==NULL || strlen(pre_load)==0) - ptx = findPTXSection(fname, context->cuobjdumpSectionList, context->libSectionList); + ptx = findPTXSection(fname, context->libSectionList); char *ptxcode; const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) { @@ -2537,7 +2546,7 @@ void cuobjdumpParseBinary(unsigned int handle){ ptxcode = readfile(override_ptx_name); } if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier(), context->cuobjdumpSectionList, context->libSectionList); + cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier(), context->libSectionList); assert (elfsection!= NULL); char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( ptx->getPTXfilename(), @@ -2561,9 +2570,16 @@ void cuobjdumpParseBinary(unsigned int handle){ //TODO: Remove temporarily files as per configurations } +} -void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) +void** cudaRegisterFatBinary( void *fatCubin, gpgpu_context* gpgpu_ctx = NULL) { + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } if(g_debug_execution >= 3){ announce_call(__my_func__); } @@ -2625,7 +2641,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) * then for next calls, only returns the appropriate number */ assert(fat_cubin_handle >= 1); - if (fat_cubin_handle==1) cuobjdumpInit(context->cuobjdumpSectionList); + if (fat_cubin_handle==1) ctx->cuobjdumpInit(); cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); return (void**)fat_cubin_handle; @@ -2696,31 +2712,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) #endif } -void __cudaUnregisterFatBinary(void **fatCubinHandle) -{ - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - ; -} - -cudaError_t cudaDeviceReset ( void ) { - // Should reset the simulated GPU - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - return g_last_cudaError = cudaSuccess; -} -cudaError_t CUDARTAPI cudaDeviceSynchronize(void){ - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - //Blocks until the device has completed all preceding requested tasks - synchronize(); - return g_last_cudaError = cudaSuccess; -} - -void CUDARTAPI __cudaRegisterFunction( +void cudaRegisterFunction( void **fatCubinHandle, const char *hostFun, char *deviceFun, @@ -2729,9 +2721,16 @@ void CUDARTAPI __cudaRegisterFunction( uint3 *tid, uint3 *bid, dim3 *bDim, - dim3 *gDim + dim3 *gDim, + gpgpu_context *gpgpu_ctx = NULL ) { + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } if(g_debug_execution >= 3){ announce_call(__my_func__); } @@ -2740,11 +2739,11 @@ void CUDARTAPI __cudaRegisterFunction( printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", deviceFun, hostFun, fat_cubin_handle); if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - cuobjdumpParseBinary(fat_cubin_handle); + ctx->cuobjdumpParseBinary(fat_cubin_handle); context->register_function( fat_cubin_handle, hostFun, deviceFun ); } -extern void __cudaRegisterVar( +void cudaRegisterVar( void **fatCubinHandle, char *hostVar, //pointer to...something char *deviceAddress, //name of variable @@ -2752,15 +2751,22 @@ extern void __cudaRegisterVar( int ext, int size, int constant, - int global ) + int global, + gpgpu_context *gpgpu_ctx = NULL) { + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } if(g_debug_execution >= 3){ announce_call(__my_func__); } printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); + ctx->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); fflush(stdout); if ( constant && !global && !ext ) { gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); @@ -2769,6 +2775,82 @@ extern void __cudaRegisterVar( } else cuda_not_implemented(__my_func__,__LINE__); } +extern "C" { + +void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) +{ + return cudaRegisterFatBinary(fatCubin); +} + +void CUDARTAPI __cudaRegisterFunction( + void **fatCubinHandle, + const char *hostFun, + char *deviceFun, + const char *deviceName, + int thread_limit, + uint3 *tid, + uint3 *bid, + dim3 *bDim, + dim3 *gDim +) { + cudaRegisterFunction( + fatCubinHandle, + hostFun, + deviceFun, + deviceName, + thread_limit, + tid, + bid, + bDim, + gDim + ); + +} + +extern void __cudaRegisterVar( + void **fatCubinHandle, + char *hostVar, //pointer to...something + char *deviceAddress, //name of variable + const char *deviceName, //name of variable (same as above) + int ext, + int size, + int constant, + int global ) +{ + cudaRegisterVar( + fatCubinHandle, + hostVar, + deviceAddress, + deviceName, + ext, + size, + constant, + global ); +} +void __cudaUnregisterFatBinary(void **fatCubinHandle) +{ + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + ; +} + +cudaError_t cudaDeviceReset ( void ) { + // Should reset the simulated GPU + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + return g_last_cudaError = cudaSuccess; +} +cudaError_t CUDARTAPI cudaDeviceSynchronize(void){ + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + //Blocks until the device has completed all preceding requested tasks + synchronize(); + return g_last_cudaError = cudaSuccess; +} + void __cudaRegisterShared( void **fatCubinHandle, diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h new file mode 100644 index 0000000..d6e564b --- /dev/null +++ b/libcuda/gpgpu_context.h @@ -0,0 +1,18 @@ +#include + +class cuobjdumpSection; +class cuobjdumpELFSection; +class cuobjdumpPTXSection; + +class gpgpu_context { + public: + std::list cuobjdumpSectionList; + void cuobjdumpInit(); + void cuobjdumpParseBinary(unsigned int handle); + void extract_code_using_cuobjdump(); + std::list pruneSectionList(CUctx_st *context); + std::list mergeMatchingSections(std::string identifier); + std::list mergeSections(); + cuobjdumpELFSection* findELFSection(const std::string identifier, std::list &libSectionList); + cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list &libSectionList); +}; -- cgit v1.3 From d553124832aca461dda4dd7d503748d22b44bbe2 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 5 Jun 2019 23:28:12 -0400 Subject: Add cuda_api_object and move some var Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 11 +++++++++++ libcuda/cuda_runtime_api.cc | 30 ++++++++++++++---------------- libcuda/gpgpu_context.h | 19 +++++++++++++++++-- 3 files changed, 42 insertions(+), 18 deletions(-) create mode 100644 libcuda/cuda_api_object.h (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h new file mode 100644 index 0000000..86ffa98 --- /dev/null +++ b/libcuda/cuda_api_object.h @@ -0,0 +1,11 @@ +#ifndef __cuda_api_object_h__ +#define __cuda_api_object_h__ +class cuobjdumpSection; + +class cuda_runtime_api { + public: + // global list + std::list libSectionList; + // member function list +}; +#endif /* __cuda_api_object_h__ */ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index c39571c..4221634 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -132,6 +132,7 @@ #include "__cudaFatFormat.h" #endif #include "gpgpu_context.h" +#include "cuda_api_object.h" #include "../src/gpgpu-sim/gpu-sim.h" #include "../src/cuda-sim/ptx_loader.h" #include "../src/cuda-sim/cuda-sim.h" @@ -314,9 +315,6 @@ struct CUctx_st { return i->second; } - std::list libSectionList; - //maps sm version number to set of filenames - std::map > version_filename; std::list g_cuda_launch_stack; std::mapfatbin_registered; std::map fatbinmap; @@ -2005,7 +2003,7 @@ char* get_app_binary_name(std::string abs_path){ } //extracts all ptx files from binary and dumps into prog_name.unique_no.sm_<>.ptx files -void extract_ptx_files_using_cuobjdump(CUctx_st *context){ +void gpgpu_context::extract_ptx_files_using_cuobjdump(CUctx_st *context){ extern bool g_cdp_enabled; char command[1000]; char *pytorch_bin = getenv("PYTORCH_BIN"); @@ -2066,10 +2064,10 @@ void extract_ptx_files_using_cuobjdump(CUctx_st *context){ } std::string vstr = line.substr(pos1+3,pos2-pos1-3); int version = atoi(vstr.c_str()); - if (context->version_filename.find(version)==context->version_filename.end()){ - context->version_filename[version] = std::set(); + if (version_filename.find(version)==version_filename.end()){ + version_filename[version] = std::set(); } - context->version_filename[version].insert(line); + version_filename[version].insert(line); } } @@ -2229,7 +2227,7 @@ void gpgpu_context::extract_code_using_cuobjdump(){ fclose(cuobjdump_in); std::getline(libsf, line); } - context->libSectionList = cuobjdumpSectionList; + api->libSectionList = cuobjdumpSectionList; //Restore the original section list cuobjdumpSectionList = tmpsl; @@ -2418,10 +2416,10 @@ cuobjdumpELFSection* findELFSectionInList(std::list sectionli } //! Find an ELF section in all the known lists -cuobjdumpELFSection* gpgpu_context::findELFSection(const std::string identifier, std::list &libSectionList){ +cuobjdumpELFSection* gpgpu_context::findELFSection(const std::string identifier){ cuobjdumpELFSection* sec = findELFSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; - sec = findELFSectionInList(libSectionList, identifier); + sec = findELFSectionInList(api->libSectionList, identifier); if (sec!=NULL)return sec; std::cout << "Could not find " << identifier << std::endl; assert(0 && "Could not find the required ELF section"); @@ -2453,10 +2451,10 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list §ionl } //! Find an PTX section in all the known lists -cuobjdumpPTXSection* gpgpu_context::findPTXSection(const std::string identifier, std::list &libSectionList){ +cuobjdumpPTXSection* gpgpu_context::findPTXSection(const std::string identifier){ cuobjdumpPTXSection* sec = findPTXSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; - sec = findPTXSectionInList(libSectionList, identifier); + sec = findPTXSectionInList(api->libSectionList, identifier); if (sec!=NULL)return sec; std::cout << "Could not find " << identifier << std::endl; assert(0 && "Could not find the required PTX section"); @@ -2500,7 +2498,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ #if (CUDART_VERSION >= 6000) //loops through all ptx files from smallest sm version to largest std::map >::iterator itr_m; - for (itr_m = context->version_filename.begin(); itr_m!=context->version_filename.end(); itr_m++){ + for (itr_m = version_filename.begin(); itr_m!=version_filename.end(); itr_m++){ std::set::iterator itr_s; for (itr_s = itr_m->second.begin(); itr_s!=itr_m->second.end(); itr_s++){ std::string ptx_filename = *itr_s; @@ -2512,7 +2510,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ context->add_binary(symtab, handle); load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - for (itr_m = context->version_filename.begin(); itr_m!=context->version_filename.end(); itr_m++){ + for (itr_m = version_filename.begin(); itr_m!=version_filename.end(); itr_m++){ std::set::iterator itr_s; for (itr_s = itr_m->second.begin(); itr_s!=itr_m->second.end(); itr_s++){ std::string ptx_filename = *itr_s; @@ -2536,7 +2534,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ cuobjdumpPTXSection* ptx = NULL; const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if(pre_load==NULL || strlen(pre_load)==0) - ptx = findPTXSection(fname, context->libSectionList); + ptx = findPTXSection(fname); char *ptxcode; const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) { @@ -2546,7 +2544,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ ptxcode = readfile(override_ptx_name); } if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier(), context->libSectionList); + cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier()); assert (elfsection!= NULL); char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( ptx->getPTXfilename(), diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index d6e564b..576ec67 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -1,4 +1,9 @@ +#ifndef __gpgpu_context_h__ +#define __gpgpu_context_h__ #include +#include +#include +#include "cuda_api_object.h" class cuobjdumpSection; class cuobjdumpELFSection; @@ -6,13 +11,23 @@ class cuobjdumpPTXSection; class gpgpu_context { public: + gpgpu_context() { + api = new cuda_runtime_api(); + } + // global list std::list cuobjdumpSectionList; + //maps sm version number to set of filenames + std::map > version_filename; + cuda_runtime_api* api; + // member function list void cuobjdumpInit(); void cuobjdumpParseBinary(unsigned int handle); void extract_code_using_cuobjdump(); std::list pruneSectionList(CUctx_st *context); std::list mergeMatchingSections(std::string identifier); std::list mergeSections(); - cuobjdumpELFSection* findELFSection(const std::string identifier, std::list &libSectionList); - cuobjdumpPTXSection* findPTXSection(const std::string identifier, std::list &libSectionList); + cuobjdumpELFSection* findELFSection(const std::string identifier); + cuobjdumpPTXSection* findPTXSection(const std::string identifier); + void extract_ptx_files_using_cuobjdump(CUctx_st *context); }; +#endif /* __gpgpu_context_h__ */ -- cgit v1.3 From bd5bbc6b1d56436dbcc0cfd84e96c2d514ab4ccc Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Jun 2019 01:07:21 -0400 Subject: Move g_cuda_launch_stack Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 3 + libcuda/cuda_runtime_api.cc | 910 +++++++++++++++++++++++--------------------- 2 files changed, 475 insertions(+), 438 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index 86ffa98..2001f91 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -1,11 +1,14 @@ #ifndef __cuda_api_object_h__ #define __cuda_api_object_h__ class cuobjdumpSection; +class kernel_config; + class cuda_runtime_api { public: // global list std::list libSectionList; + std::list g_cuda_launch_stack; // member function list }; #endif /* __cuda_api_object_h__ */ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 4221634..ea96570 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -237,8 +237,6 @@ private: struct _cuda_device_id *m_next; }; -class kernel_config; - #ifndef OPENGL_SUPPORT typedef unsigned long GLuint; #endif @@ -315,7 +313,6 @@ struct CUctx_st { return i->second; } - std::list g_cuda_launch_stack; std::mapfatbin_registered; std::map fatbinmap; std::map g_mallocPtr_Size; @@ -531,51 +528,448 @@ enum cuobjdumpSectionType { }; -// sectiontype: 0 for ptx, 1 for elf -void addCuobjdumpSection(int sectiontype, std::list &cuobjdumpSectionList){ - if (sectiontype) - cuobjdumpSectionList.push_front(new cuobjdumpELFSection()); - else - cuobjdumpSectionList.push_front(new cuobjdumpPTXSection()); - printf("## Adding new section %s\n", sectiontype?"ELF":"PTX"); +// sectiontype: 0 for ptx, 1 for elf +void addCuobjdumpSection(int sectiontype, std::list &cuobjdumpSectionList){ + if (sectiontype) + cuobjdumpSectionList.push_front(new cuobjdumpELFSection()); + else + cuobjdumpSectionList.push_front(new cuobjdumpPTXSection()); + printf("## Adding new section %s\n", sectiontype?"ELF":"PTX"); +} + +void setCuobjdumparch(const char* arch, std::list &cuobjdumpSectionList){ + unsigned archnum; + sscanf(arch, "sm_%u", &archnum); + assert (archnum && "cannot have sm_0"); + printf("Adding arch: %s\n", arch); + cuobjdumpSectionList.front()->setArch(archnum); +} + +void setCuobjdumpidentifier(const char* identifier, std::list &cuobjdumpSectionList){ + printf("Adding identifier: %s\n", identifier); + cuobjdumpSectionList.front()->setIdentifier(identifier); +} + +void setCuobjdumpptxfilename(const char* filename, std::list &cuobjdumpSectionList){ + printf("Adding ptx filename: %s\n", filename); + cuobjdumpSection* x = cuobjdumpSectionList.front(); + if (dynamic_cast(x) == NULL){ + assert (0 && "You shouldn't be trying to add a ptxfilename to an elf section"); + } + (dynamic_cast(x))->setPTXfilename(filename); +} + +void setCuobjdumpelffilename(const char* filename, std::list &cuobjdumpSectionList){ + if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ + assert (0 && "You shouldn't be trying to add a elffilename to an ptx section"); + } + (dynamic_cast(cuobjdumpSectionList.front()))->setELFfilename(filename); +} + +void setCuobjdumpsassfilename(const char* filename, std::list &cuobjdumpSectionList){ + if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ + assert (0 && "You shouldn't be trying to add a sassfilename to an ptx section"); + } + (dynamic_cast(cuobjdumpSectionList.front()))->setSASSfilename(filename); +} + +//! Return the executable file of the process containing the PTX/SASS code +//! +//! This Function returns the executable file ran by the process. This +//! executable is supposed to contain the PTX/SASS code. It provides workaround +//! for processes running on valgrind by dereferencing /proc//exe within the +//! GPGPU-Sim process before calling cuobjdump to extract PTX/SASS. This is +//! needed because valgrind uses x86 emulation to detect memory leak. Other +//! processes (e.g. cuobjdump) reading /proc//exe will see the emulator +//! executable instead of the application binary. +//! +std::string get_app_binary(){ + char self_exe_path[1025]; +#ifdef __APPLE__ + uint32_t size = sizeof(self_exe_path); + if( _NSGetExecutablePath(self_exe_path,&size) != 0 ) { + printf("GPGPU-Sim ** ERROR: _NSGetExecutablePath input buffer too small\n"); + exit(1); + } +#else + std::stringstream exec_link; + exec_link << "/proc/self/exe"; + + ssize_t path_length = readlink(exec_link.str().c_str(), self_exe_path, 1024); + assert(path_length != -1); + self_exe_path[path_length] = '\0'; +#endif + + printf("self exe links to: %s\n", self_exe_path); + return self_exe_path; +} + +//above func gives abs path whereas this give just the name of application. +char* get_app_binary_name(std::string abs_path){ + char *self_exe_path; +#ifdef __APPLE__ + //TODO: get apple device and check the result. + printf("WARNING: not tested for Apple-mac devices \n"); + abort(); +#else + char* buf = strdup(abs_path.c_str()); + char *token = strtok(buf, "/"); + while(token !=NULL){ + self_exe_path = token; + token = strtok(NULL,"/"); + } +#endif + self_exe_path = strtok(self_exe_path, "."); + printf("self exe links to: %s\n", self_exe_path); + return self_exe_path; +} + +static int get_app_cuda_version() { + int app_cuda_version = 0; + char fname[1024]; + snprintf(fname,1024,"_app_cuda_version_XXXXXX"); + int fd=mkstemp(fname); + close(fd); + std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; + system(app_cuda_version_command.c_str()); + FILE * cmd = fopen(fname, "r"); + char buf[256]; + while (fgets(buf, sizeof(buf), cmd) != 0) { + std::cout << buf; + app_cuda_version = atoi(buf); + } + fclose(cmd); + if ( app_cuda_version == 0 ) { + printf( "Error - Cannot detect the app's CUDA version.\n" ); + exit(1); + } + return app_cuda_version; +} + +//! Keep track of the association between filename and cubin handle +void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ + context->fatbinmap[handle] = filename; +} + +/******************************************************************************* + * Add internal cuda runtime API call to accept gpgpu_context * + *******************************************************************************/ + +void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } +#if (CUDART_VERSION < 2010) + printf("GPGPU-Sim PTX: ERROR ** this version of GPGPU-Sim requires CUDA 2.1 or higher\n"); + exit(1); +#endif + CUctx_st *context = GPGPUSim_Context(); + static unsigned next_fat_bin_handle = 1; + if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) { + // The following workaround has only been verified on 64-bit systems. + if (sizeof(void*) == 4) + printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); + + // This code will get the CUDA version the app was compiled with. + // We need this to determine how to handle the parsing of the binary. + // Making this a runtime variable based on the app, enables GPGPU-Sim compiled + // with a newer version of CUDA to run apps compiled with older versions of + // CUDA. This is especially useful for PTXPLUS execution. + //Skip cuda version check for pytorch application + std::string app_binary_path = get_app_binary(); + int pos = app_binary_path.find("python"); + if (pos==std::string::npos){ + // Not pytorch app : checking cuda version + int app_cuda_version = get_app_cuda_version(); + assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator." ); + } + + //int app_cuda_version = get_app_cuda_version(); + //assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator." ); + const char* filename; +#if CUDART_VERSION < 6000 + // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) + typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); + __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; + + // Extract the source code file name that generate the given FatBin. + // - Obtains the pointer to the actual fatbin structure from the FatBin handle (fatCubin). + // - An integer inside the fatbin structure contains the relative offset to the source code file name. + // - This offset differs among different CUDA and GCC versions. + char * pfatbin = (char*) fatDeviceText->d; + int offset = *((int*)(pfatbin+48)); + filename = (pfatbin+16+offset); +#else + filename = "default"; +#endif + + // The extracted file name is associated with a fat_cubin_handle passed + // into cudaLaunch(). Inside cudaLaunch(), the associated file name is + // used to find the PTX/SASS section from cuobjdump, which contains the + // PTX/SASS code for the launched kernel function. + // This allows us to work around the fact that cuobjdump only outputs the + // file name associated with each section. + unsigned long long fat_cubin_handle = next_fat_bin_handle; + next_fat_bin_handle++; + printf("GPGPU-Sim PTX: __cudaRegisterFatBinary, fat_cubin_handle = %llu, filename=%s\n", fat_cubin_handle, filename); + /*! + * This function extracts all data from all files in first call + * then for next calls, only returns the appropriate number + */ + assert(fat_cubin_handle >= 1); + if (fat_cubin_handle==1) ctx->cuobjdumpInit(); + cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); + + return (void**)fat_cubin_handle; + } +#if (CUDART_VERSION < 8000) + else { + static unsigned source_num=1; + unsigned long long fat_cubin_handle = next_fat_bin_handle++; + __cudaFatCudaBinary *info = (__cudaFatCudaBinary *)fatCubin; + assert( info->version >= 3 ); + unsigned num_ptx_versions=0; + unsigned max_capability=0; + unsigned selected_capability=0; + bool found=false; + unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); + if (!info->ptx){ + printf("ERROR: Cannot find ptx code in cubin file\n" + "\tIf you are using CUDA 4.0 or higher, please enable -gpgpu_ptx_use_cuobjdump or downgrade to CUDA 3.1\n"); + exit(1); + } + while( info->ptx[num_ptx_versions].gpuProfileName != NULL ) { + unsigned capability=0; + sscanf(info->ptx[num_ptx_versions].gpuProfileName,"compute_%u",&capability); + printf("GPGPU-Sim PTX: __cudaRegisterFatBinary found PTX versions for '%s', ", info->ident); + printf("capability = %s\n", info->ptx[num_ptx_versions].gpuProfileName ); + if( forced_max_capability ) { + if( capability > max_capability && capability <= forced_max_capability ) { + found = true; + max_capability=capability; + selected_capability = num_ptx_versions; + } + } else { + if( capability > max_capability ) { + found = true; + max_capability=capability; + selected_capability = num_ptx_versions; + } + } + num_ptx_versions++; + } + if( found ) { + printf("GPGPU-Sim PTX: Loading PTX for %s, capability = %s\n", + info->ident, info->ptx[selected_capability].gpuProfileName ); + symbol_table *symtab; + const char *ptx = info->ptx[selected_capability].ptx; + if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { + printf("GPGPU-Sim PTX: ERROR ** PTXPlus is only supported through cuobjdump\n" + "\tEither enable cuobjdump or disable PTXPlus in your configuration file\n"); + exit(1); + } else { + symtab=gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); + context->add_binary(symtab,fat_cubin_handle); + gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); + } + source_num++; + load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + } else { + printf("GPGPU-Sim PTX: warning -- did not find an appropriate PTX in cubin\n"); + } + return (void**)fat_cubin_handle; + } +#else + else { + printf("ERROR ** __cudaRegisterFatBinary() needs to be updated\n"); + abort(); + } +#endif +} + +void cudaRegisterFunctionInternal( + void **fatCubinHandle, + const char *hostFun, + char *deviceFun, + const char *deviceName, + int thread_limit, + uint3 *tid, + uint3 *bid, + dim3 *bDim, + dim3 *gDim, + gpgpu_context *gpgpu_ctx = NULL +) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + CUctx_st *context = GPGPUSim_Context(); + unsigned fat_cubin_handle = (unsigned)(unsigned long long)fatCubinHandle; + printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", + deviceFun, hostFun, fat_cubin_handle); + if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) + ctx->cuobjdumpParseBinary(fat_cubin_handle); + context->register_function( fat_cubin_handle, hostFun, deviceFun ); +} + +void cudaRegisterVarInternal( + void **fatCubinHandle, + char *hostVar, //pointer to...something + char *deviceAddress, //name of variable + const char *deviceName, //name of variable (same as above) + int ext, + int size, + int constant, + int global, + gpgpu_context *gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); + printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); + if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) + ctx->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); + fflush(stdout); + if ( constant && !global && !ext ) { + gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); + } else if ( !constant && !global && !ext ) { + gpgpu_ptx_sim_register_global_variable(hostVar,deviceName,size); + } else cuda_not_implemented(__my_func__,__LINE__); +} + +cudaError_t cudaConfigureCallInternal(dim3 gridDim, dim3 blockDim, size_t sharedMem, cudaStream_t stream, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + struct CUstream_st *s = (struct CUstream_st *)stream; + ctx->api->g_cuda_launch_stack.push_back( kernel_config(gridDim,blockDim,sharedMem,s) ); + return g_last_cudaError = cudaSuccess; +} + +cudaError_t cudaSetupArgumentInternal(const void *arg, size_t size, size_t offset, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + gpgpusim_ptx_assert( !ctx->api->g_cuda_launch_stack.empty(), "empty launch stack" ); + kernel_config &config = ctx->api->g_cuda_launch_stack.back(); + config.set_arg(arg,size,offset); + printf("GPGPU-Sim PTX: Setting up arguments for %zu bytes starting at 0x%llx..\n",size, (unsigned long long) arg); + + return g_last_cudaError = cudaSuccess; } -void setCuobjdumparch(const char* arch, std::list &cuobjdumpSectionList){ - unsigned archnum; - sscanf(arch, "sm_%u", &archnum); - assert (archnum && "cannot have sm_0"); - printf("Adding arch: %s\n", arch); - cuobjdumpSectionList.front()->setArch(archnum); -} +cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = NULL ) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + CUctx_st* context = GPGPUSim_Context(); + char *mode = getenv("PTX_SIM_MODE_FUNC"); + if( mode ) + sscanf(mode,"%u", &g_ptx_sim_mode); + gpgpusim_ptx_assert( !ctx->api->g_cuda_launch_stack.empty(), "empty launch stack" ); + kernel_config config = ctx->api->g_cuda_launch_stack.back(); + struct CUstream_st *stream = config.get_stream(); + printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, + g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); + kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); + //do dynamic PDOM analysis for performance simulation scenario + std::string kname = grid->name(); + function_info *kernel_func_info = grid->entry(); + if (kernel_func_info->is_pdom_set()) { + printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", kname.c_str() ); + } else { + printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", kname.c_str() ); + kernel_func_info->do_pdom(); + kernel_func_info->set_pdom(); + } + dim3 gridDim = config.grid_dim(); + dim3 blockDim = config.block_dim(); + + gpgpu_t *gpu = context->get_device()->get_gpgpu(); + checkpoint *g_checkpoint; + g_checkpoint = new checkpoint(); + class memory_space *global_mem; + global_mem = gpu->get_global_memory(); -void setCuobjdumpidentifier(const char* identifier, std::list &cuobjdumpSectionList){ - printf("Adding identifier: %s\n", identifier); - cuobjdumpSectionList.front()->setIdentifier(identifier); -} + if(gpu->resume_option ==1 && (grid->get_uid()==gpu->resume_kernel)) + { + + char f1name[2048]; + snprintf(f1name,2048,"checkpoint_files/global_mem_%d.txt", grid->get_uid()); -void setCuobjdumpptxfilename(const char* filename, std::list &cuobjdumpSectionList){ - printf("Adding ptx filename: %s\n", filename); - cuobjdumpSection* x = cuobjdumpSectionList.front(); - if (dynamic_cast(x) == NULL){ - assert (0 && "You shouldn't be trying to add a ptxfilename to an elf section"); + g_checkpoint->load_global_mem(global_mem, f1name); + for (int i=0;iresume_CTA;i++) + grid->increment_cta_id(); } - (dynamic_cast(x))->setPTXfilename(filename); -} + if(gpu->resume_option==1 && (grid->get_uid()resume_kernel)) + { + char f1name[2048]; + snprintf(f1name,2048,"checkpoint_files/global_mem_%d.txt", grid->get_uid()); -void setCuobjdumpelffilename(const char* filename, std::list &cuobjdumpSectionList){ - if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ - assert (0 && "You shouldn't be trying to add a elffilename to an ptx section"); + g_checkpoint->load_global_mem(global_mem, f1name); + printf("Skipping kernel %d as resuming from kernel %d\n",grid->get_uid(),gpu->resume_kernel ); + ctx->api->g_cuda_launch_stack.pop_back(); + return g_last_cudaError = cudaSuccess; + } - (dynamic_cast(cuobjdumpSectionList.front()))->setELFfilename(filename); -} - -void setCuobjdumpsassfilename(const char* filename, std::list &cuobjdumpSectionList){ - if (dynamic_cast(cuobjdumpSectionList.front()) == NULL){ - assert (0 && "You shouldn't be trying to add a sassfilename to an ptx section"); + if(gpu->checkpoint_option==1 && (grid->get_uid()>gpu->checkpoint_kernel)) + { + printf("Skipping kernel %d as checkpoint from kernel %d\n",grid->get_uid(),gpu->checkpoint_kernel ); + ctx->api->g_cuda_launch_stack.pop_back(); + return g_last_cudaError = cudaSuccess; + } - (dynamic_cast(cuobjdumpSectionList.front()))->setSASSfilename(filename); + printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", + kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); + stream_operation op(grid,g_ptx_sim_mode,stream); + g_stream_manager()->push(op); + ctx->api->g_cuda_launch_stack.pop_back(); + return g_last_cudaError = cudaSuccess; } + /******************************************************************************* * * * * @@ -1531,100 +1925,15 @@ __host__ const char* CUDARTAPI cudaGetErrorString(cudaError_t error) return strdup(buf); } -__host__ cudaError_t CUDARTAPI cudaConfigureCall(dim3 gridDim, dim3 blockDim, size_t sharedMem, cudaStream_t stream) -{ - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - struct CUstream_st *s = (struct CUstream_st *)stream; - CUctx_st *context = GPGPUSim_Context(); - context->g_cuda_launch_stack.push_back( kernel_config(gridDim,blockDim,sharedMem,s) ); - return g_last_cudaError = cudaSuccess; -} - __host__ cudaError_t CUDARTAPI cudaSetupArgument(const void *arg, size_t size, size_t offset) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st *context = GPGPUSim_Context(); - gpgpusim_ptx_assert( !context->g_cuda_launch_stack.empty(), "empty launch stack" ); - kernel_config &config = context->g_cuda_launch_stack.back(); - config.set_arg(arg,size,offset); - printf("GPGPU-Sim PTX: Setting up arguments for %zu bytes starting at 0x%llx..\n",size, (unsigned long long) arg); - - return g_last_cudaError = cudaSuccess; + return cudaSetupArgumentInternal(arg, size, offset); } __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st* context = GPGPUSim_Context(); - char *mode = getenv("PTX_SIM_MODE_FUNC"); - if( mode ) - sscanf(mode,"%u", &g_ptx_sim_mode); - gpgpusim_ptx_assert( !context->g_cuda_launch_stack.empty(), "empty launch stack" ); - kernel_config config = context->g_cuda_launch_stack.back(); - struct CUstream_st *stream = config.get_stream(); - printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, - g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); - kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); - //do dynamic PDOM analysis for performance simulation scenario - std::string kname = grid->name(); - function_info *kernel_func_info = grid->entry(); - if (kernel_func_info->is_pdom_set()) { - printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", kname.c_str() ); - } else { - printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", kname.c_str() ); - kernel_func_info->do_pdom(); - kernel_func_info->set_pdom(); - } - dim3 gridDim = config.grid_dim(); - dim3 blockDim = config.block_dim(); - - gpgpu_t *gpu = context->get_device()->get_gpgpu(); - checkpoint *g_checkpoint; - g_checkpoint = new checkpoint(); - class memory_space *global_mem; - global_mem = gpu->get_global_memory(); - - if(gpu->resume_option ==1 && (grid->get_uid()==gpu->resume_kernel)) - { - - char f1name[2048]; - snprintf(f1name,2048,"checkpoint_files/global_mem_%d.txt", grid->get_uid()); - - g_checkpoint->load_global_mem(global_mem, f1name); - for (int i=0;iresume_CTA;i++) - grid->increment_cta_id(); - } - if(gpu->resume_option==1 && (grid->get_uid()resume_kernel)) - { - char f1name[2048]; - snprintf(f1name,2048,"checkpoint_files/global_mem_%d.txt", grid->get_uid()); - - g_checkpoint->load_global_mem(global_mem, f1name); - printf("Skipping kernel %d as resuming from kernel %d\n",grid->get_uid(),gpu->resume_kernel ); - context->g_cuda_launch_stack.pop_back(); - return g_last_cudaError = cudaSuccess; - - } - if(gpu->checkpoint_option==1 && (grid->get_uid()>gpu->checkpoint_kernel)) - { - printf("Skipping kernel %d as checkpoint from kernel %d\n",grid->get_uid(),gpu->checkpoint_kernel ); - context->g_cuda_launch_stack.pop_back(); - return g_last_cudaError = cudaSuccess; - - } - printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", - kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); - stream_operation op(grid,g_ptx_sim_mode,stream); - g_stream_manager()->push(op); - context->g_cuda_launch_stack.pop_back(); - return g_last_cudaError = cudaSuccess; + return cudaLaunchInternal( hostFun ); } __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 gridDim, dim3 blockDim, const void** args, size_t sharedMem, cudaStream_t stream ) @@ -1636,13 +1945,13 @@ __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 grid CUctx_st *context = GPGPUSim_Context(); function_info *entry = context->get_kernel(hostFun); - cudaConfigureCall(gridDim, blockDim, sharedMem, stream); + cudaConfigureCallInternal(gridDim, blockDim, sharedMem, stream); for(unsigned i = 0; i < entry->num_args(); i++){ std::pair p = entry->get_param_config(i); - cudaSetupArgument(args[i], p.first, p.second); + cudaSetupArgumentInternal(args[i], p.first, p.second); } - cudaLaunch(hostFun); + cudaLaunchInternal(hostFun); return g_last_cudaError = cudaSuccess; } @@ -1925,83 +2234,32 @@ typedef int (*ExportedFunction)(); static ExportedFunction exportTable[3] = {&dummy0, &dummy0, &dummy0}; -__host__ cudaError_t CUDARTAPI cudaGetExportTable(const void **ppExportTable, const cudaUUID_t *pExportTableId) -{ - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - printf("cudaGetExportTable: UUID = "); - for (int s = 0; s < 16; s++) { - printf("%#2x ", (unsigned char) (pExportTableId->bytes[s])); - } - *ppExportTable = &exportTable; - - printf("\n"); - return g_last_cudaError = cudaSuccess; -} - -#endif - - -/******************************************************************************* - * * - * * - * * - *******************************************************************************/ - -//#include "../../cuobjdump_to_ptxplus/cuobjdump_parser.h" - -//! Return the executable file of the process containing the PTX/SASS code -//! -//! This Function returns the executable file ran by the process. This -//! executable is supposed to contain the PTX/SASS code. It provides workaround -//! for processes running on valgrind by dereferencing /proc//exe within the -//! GPGPU-Sim process before calling cuobjdump to extract PTX/SASS. This is -//! needed because valgrind uses x86 emulation to detect memory leak. Other -//! processes (e.g. cuobjdump) reading /proc//exe will see the emulator -//! executable instead of the application binary. -//! -std::string get_app_binary(){ - char self_exe_path[1025]; -#ifdef __APPLE__ - uint32_t size = sizeof(self_exe_path); - if( _NSGetExecutablePath(self_exe_path,&size) != 0 ) { - printf("GPGPU-Sim ** ERROR: _NSGetExecutablePath input buffer too small\n"); - exit(1); - } -#else - std::stringstream exec_link; - exec_link << "/proc/self/exe"; - - ssize_t path_length = readlink(exec_link.str().c_str(), self_exe_path, 1024); - assert(path_length != -1); - self_exe_path[path_length] = '\0'; -#endif - - printf("self exe links to: %s\n", self_exe_path); - return self_exe_path; -} - -//above func gives abs path whereas this give just the name of application. -char* get_app_binary_name(std::string abs_path){ - char *self_exe_path; -#ifdef __APPLE__ - //TODO: get apple device and check the result. - printf("WARNING: not tested for Apple-mac devices \n"); - abort(); -#else - char* buf = strdup(abs_path.c_str()); - char *token = strtok(buf, "/"); - while(token !=NULL){ - self_exe_path = token; - token = strtok(NULL,"/"); - } -#endif - self_exe_path = strtok(self_exe_path, "."); - printf("self exe links to: %s\n", self_exe_path); - return self_exe_path; +__host__ cudaError_t CUDARTAPI cudaGetExportTable(const void **ppExportTable, const cudaUUID_t *pExportTableId) +{ + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + printf("cudaGetExportTable: UUID = "); + for (int s = 0; s < 16; s++) { + printf("%#2x ", (unsigned char) (pExportTableId->bytes[s])); + } + *ppExportTable = &exportTable; + + printf("\n"); + return g_last_cudaError = cudaSuccess; } +#endif + + +/******************************************************************************* + * * + * * + * * + *******************************************************************************/ + +//#include "../../cuobjdump_to_ptxplus/cuobjdump_parser.h" + //extracts all ptx files from binary and dumps into prog_name.unique_no.sm_<>.ptx files void gpgpu_context::extract_ptx_files_using_cuobjdump(CUctx_st *context){ extern bool g_cdp_enabled; @@ -2072,28 +2330,6 @@ void gpgpu_context::extract_ptx_files_using_cuobjdump(CUctx_st *context){ } -static int get_app_cuda_version() { - int app_cuda_version = 0; - char fname[1024]; - snprintf(fname,1024,"_app_cuda_version_XXXXXX"); - int fd=mkstemp(fname); - close(fd); - std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; - system(app_cuda_version_command.c_str()); - FILE * cmd = fopen(fname, "r"); - char buf[256]; - while (fgets(buf, sizeof(buf), cmd) != 0) { - std::cout << buf; - app_cuda_version = atoi(buf); - } - fclose(cmd); - if ( app_cuda_version == 0 ) { - printf( "Error - Cannot detect the app's CUDA version.\n" ); - exit(1); - } - return app_cuda_version; -} - //! Call cuobjdump to extract everything (-elf -sass -ptx) /*! @@ -2475,11 +2711,6 @@ void gpgpu_context::cuobjdumpInit(){ } -//! Keep track of the association between filename and cubin handle -void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ - context->fatbinmap[handle] = filename; -} - //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ @@ -2570,214 +2801,11 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ } } -void** cudaRegisterFatBinary( void *fatCubin, gpgpu_context* gpgpu_ctx = NULL) -{ - gpgpu_context *ctx; - if (gpgpu_ctx){ - ctx = gpgpu_ctx; - } else { - ctx = GPGPU_Context(); - } - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } -#if (CUDART_VERSION < 2010) - printf("GPGPU-Sim PTX: ERROR ** this version of GPGPU-Sim requires CUDA 2.1 or higher\n"); - exit(1); -#endif - CUctx_st *context = GPGPUSim_Context(); - static unsigned next_fat_bin_handle = 1; - if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) { - // The following workaround has only been verified on 64-bit systems. - if (sizeof(void*) == 4) - printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); - - // This code will get the CUDA version the app was compiled with. - // We need this to determine how to handle the parsing of the binary. - // Making this a runtime variable based on the app, enables GPGPU-Sim compiled - // with a newer version of CUDA to run apps compiled with older versions of - // CUDA. This is especially useful for PTXPLUS execution. - //Skip cuda version check for pytorch application - std::string app_binary_path = get_app_binary(); - int pos = app_binary_path.find("python"); - if (pos==std::string::npos){ - // Not pytorch app : checking cuda version - int app_cuda_version = get_app_cuda_version(); - assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator." ); - } - - //int app_cuda_version = get_app_cuda_version(); - //assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator." ); - const char* filename; -#if CUDART_VERSION < 6000 - // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) - typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); - __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; - - // Extract the source code file name that generate the given FatBin. - // - Obtains the pointer to the actual fatbin structure from the FatBin handle (fatCubin). - // - An integer inside the fatbin structure contains the relative offset to the source code file name. - // - This offset differs among different CUDA and GCC versions. - char * pfatbin = (char*) fatDeviceText->d; - int offset = *((int*)(pfatbin+48)); - filename = (pfatbin+16+offset); -#else - filename = "default"; -#endif - - // The extracted file name is associated with a fat_cubin_handle passed - // into cudaLaunch(). Inside cudaLaunch(), the associated file name is - // used to find the PTX/SASS section from cuobjdump, which contains the - // PTX/SASS code for the launched kernel function. - // This allows us to work around the fact that cuobjdump only outputs the - // file name associated with each section. - unsigned long long fat_cubin_handle = next_fat_bin_handle; - next_fat_bin_handle++; - printf("GPGPU-Sim PTX: __cudaRegisterFatBinary, fat_cubin_handle = %llu, filename=%s\n", fat_cubin_handle, filename); - /*! - * This function extracts all data from all files in first call - * then for next calls, only returns the appropriate number - */ - assert(fat_cubin_handle >= 1); - if (fat_cubin_handle==1) ctx->cuobjdumpInit(); - cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); - - return (void**)fat_cubin_handle; - } -#if (CUDART_VERSION < 8000) - else { - static unsigned source_num=1; - unsigned long long fat_cubin_handle = next_fat_bin_handle++; - __cudaFatCudaBinary *info = (__cudaFatCudaBinary *)fatCubin; - assert( info->version >= 3 ); - unsigned num_ptx_versions=0; - unsigned max_capability=0; - unsigned selected_capability=0; - bool found=false; - unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); - if (!info->ptx){ - printf("ERROR: Cannot find ptx code in cubin file\n" - "\tIf you are using CUDA 4.0 or higher, please enable -gpgpu_ptx_use_cuobjdump or downgrade to CUDA 3.1\n"); - exit(1); - } - while( info->ptx[num_ptx_versions].gpuProfileName != NULL ) { - unsigned capability=0; - sscanf(info->ptx[num_ptx_versions].gpuProfileName,"compute_%u",&capability); - printf("GPGPU-Sim PTX: __cudaRegisterFatBinary found PTX versions for '%s', ", info->ident); - printf("capability = %s\n", info->ptx[num_ptx_versions].gpuProfileName ); - if( forced_max_capability ) { - if( capability > max_capability && capability <= forced_max_capability ) { - found = true; - max_capability=capability; - selected_capability = num_ptx_versions; - } - } else { - if( capability > max_capability ) { - found = true; - max_capability=capability; - selected_capability = num_ptx_versions; - } - } - num_ptx_versions++; - } - if( found ) { - printf("GPGPU-Sim PTX: Loading PTX for %s, capability = %s\n", - info->ident, info->ptx[selected_capability].gpuProfileName ); - symbol_table *symtab; - const char *ptx = info->ptx[selected_capability].ptx; - if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - printf("GPGPU-Sim PTX: ERROR ** PTXPlus is only supported through cuobjdump\n" - "\tEither enable cuobjdump or disable PTXPlus in your configuration file\n"); - exit(1); - } else { - symtab=gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); - context->add_binary(symtab,fat_cubin_handle); - gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); - } - source_num++; - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - } else { - printf("GPGPU-Sim PTX: warning -- did not find an appropriate PTX in cubin\n"); - } - return (void**)fat_cubin_handle; - } -#else - else { - printf("ERROR ** __cudaRegisterFatBinary() needs to be updated\n"); - abort(); - } -#endif -} - -void cudaRegisterFunction( - void **fatCubinHandle, - const char *hostFun, - char *deviceFun, - const char *deviceName, - int thread_limit, - uint3 *tid, - uint3 *bid, - dim3 *bDim, - dim3 *gDim, - gpgpu_context *gpgpu_ctx = NULL -) -{ - gpgpu_context *ctx; - if (gpgpu_ctx){ - ctx = gpgpu_ctx; - } else { - ctx = GPGPU_Context(); - } - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st *context = GPGPUSim_Context(); - unsigned fat_cubin_handle = (unsigned)(unsigned long long)fatCubinHandle; - printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", - deviceFun, hostFun, fat_cubin_handle); - if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - ctx->cuobjdumpParseBinary(fat_cubin_handle); - context->register_function( fat_cubin_handle, hostFun, deviceFun ); -} - -void cudaRegisterVar( - void **fatCubinHandle, - char *hostVar, //pointer to...something - char *deviceAddress, //name of variable - const char *deviceName, //name of variable (same as above) - int ext, - int size, - int constant, - int global, - gpgpu_context *gpgpu_ctx = NULL) -{ - gpgpu_context *ctx; - if (gpgpu_ctx){ - ctx = gpgpu_ctx; - } else { - ctx = GPGPU_Context(); - } - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); - printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); - if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - ctx->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); - fflush(stdout); - if ( constant && !global && !ext ) { - gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); - } else if ( !constant && !global && !ext ) { - gpgpu_ptx_sim_register_global_variable(hostVar,deviceName,size); - } else cuda_not_implemented(__my_func__,__LINE__); -} - extern "C" { void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) { - return cudaRegisterFatBinary(fatCubin); + return cudaRegisterFatBinaryInternal(fatCubin); } void CUDARTAPI __cudaRegisterFunction( @@ -2791,7 +2819,7 @@ void CUDARTAPI __cudaRegisterFunction( dim3 *bDim, dim3 *gDim ) { - cudaRegisterFunction( + cudaRegisterFunctionInternal( fatCubinHandle, hostFun, deviceFun, @@ -2815,7 +2843,7 @@ extern void __cudaRegisterVar( int constant, int global ) { - cudaRegisterVar( + cudaRegisterVarInternal( fatCubinHandle, hostVar, deviceAddress, @@ -2825,6 +2853,12 @@ extern void __cudaRegisterVar( constant, global ); } + +__host__ cudaError_t CUDARTAPI cudaConfigureCall(dim3 gridDim, dim3 blockDim, size_t sharedMem, cudaStream_t stream) +{ + return cudaConfigureCallInternal(gridDim, blockDim, sharedMem, stream); +} + void __cudaUnregisterFatBinary(void **fatCubinHandle) { if(g_debug_execution >= 3){ @@ -4868,12 +4902,12 @@ CUresult CUDAAPI cuLaunchKernel(CUfunction f, const char *hostFun = (const char*) f; CUctx_st *context = GPGPUSim_Context(); function_info *entry = context->get_kernel(hostFun); - cudaConfigureCall(dim3(gridDimX, gridDimY, gridDimZ), dim3(blockDimX, blockDimY, blockDimZ), sharedMemBytes, (cudaStream_t) hStream); + cudaConfigureCallInternal(dim3(gridDimX, gridDimY, gridDimZ), dim3(blockDimX, blockDimY, blockDimZ), sharedMemBytes, (cudaStream_t) hStream); for(unsigned i = 0; i < entry->num_args(); i++){ std::pair p = entry->get_param_config(i); cudaSetupArgument(kernelParams[i], p.first, p.second); } - cudaLaunch(hostFun); + cudaLaunchInternal(hostFun); return CUDA_SUCCESS; } #endif /* CUDART_VERSION >= 4000 */ -- cgit v1.3 From 570ce0a3d6e049ec3ee42329fbf25019fd4eb2e5 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Jun 2019 01:57:40 -0400 Subject: Move 3 more var Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 92 +++++++++++++++++++++++++-------------------- libcuda/gpgpu_context.h | 6 +++ 2 files changed, 58 insertions(+), 40 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index ea96570..855f11d 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -313,10 +313,7 @@ struct CUctx_st { return i->second; } - std::mapfatbin_registered; - std::map fatbinmap; std::map g_mallocPtr_Size; - std::map name_symtab; std::map pinned_memory; //support for pinned memories added std::map pinned_memory_size; int no_of_ptx; @@ -647,8 +644,8 @@ static int get_app_cuda_version() { } //! Keep track of the association between filename and cubin handle -void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ - context->fatbinmap[handle] = filename; +void gpgpu_context::cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ + fatbinmap[handle] = filename; } /******************************************************************************* @@ -725,7 +722,7 @@ void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = */ assert(fat_cubin_handle >= 1); if (fat_cubin_handle==1) ctx->cuobjdumpInit(); - cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); + ctx->cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); return (void**)fat_cubin_handle; } @@ -969,6 +966,46 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = return g_last_cudaError = cudaSuccess; } +#if CUDART_VERSION >= 6050 +CUresult +cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, + unsigned int numOptions, CUjit_option *options, void **optionValues, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + static bool addedFile = false; + if (addedFile){ + printf("GPGPU-Sim PTX: ERROR: cuLinkAddFile does not support multiple files\n"); + abort(); + } + + //blocking + assert(type==CU_JIT_INPUT_PTX); + CUctx_st *context = GPGPUSim_Context(); + char *file = getenv("PTX_JIT_PATH"); + if(file==NULL){ + printf("GPGPU-Sim PTX: ERROR: PTX_JIT_PATH has not been set\n"); + abort(); + } + strcat(file,"/"); + strcat(file,path); + symbol_table *symtab = gpgpu_ptx_sim_load_ptx_from_filename( file ); + std::string fname(path); + ctx->name_symtab[fname] = symtab; + context->add_binary(symtab, 1); + load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + addedFile = true; + return CUDA_SUCCESS; +} +#endif /******************************************************************************* * * @@ -2715,12 +2752,12 @@ void gpgpu_context::cuobjdumpInit(){ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ CUctx_st *context = GPGPUSim_Context(); - if(context->fatbin_registered[handle]) return; - context->fatbin_registered[handle] = true; - std::string fname = context->fatbinmap[handle]; + if(fatbin_registered[handle]) return; + fatbin_registered[handle] = true; + std::string fname = fatbinmap[handle]; - if (context->name_symtab.find(fname) != context->name_symtab.end()) { - symbol_table *symtab = context->name_symtab[fname]; + if (name_symtab.find(fname) != name_symtab.end()) { + symbol_table *symtab = name_symtab[fname]; context->add_binary(symtab, handle); return; } @@ -2737,7 +2774,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ symtab = gpgpu_ptx_sim_load_ptx_from_filename( ptx_filename.c_str() ); } } - context->name_symtab[fname] = symtab; + name_symtab[fname] = symtab; context->add_binary(symtab, handle); load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); @@ -2795,7 +2832,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ } load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - context->name_symtab[fname] = symtab; + name_symtab[fname] = symtab; //TODO: Remove temporarily files as per configurations } @@ -3909,33 +3946,8 @@ CUresult CUDAAPI cuLinkAddFile(CUlinkState state, CUjitInputType type, const char *path, unsigned int numOptions, CUjit_option *options, void **optionValues) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - static bool addedFile = false; - if (addedFile){ - printf("GPGPU-Sim PTX: ERROR: cuLinkAddFile does not support multiple files\n"); - abort(); - } - - //blocking - assert(type==CU_JIT_INPUT_PTX); - CUctx_st *context = GPGPUSim_Context(); - char *file = getenv("PTX_JIT_PATH"); - if(file==NULL){ - printf("GPGPU-Sim PTX: ERROR: PTX_JIT_PATH has not been set\n"); - abort(); - } - strcat(file,"/"); - strcat(file,path); - symbol_table *symtab = gpgpu_ptx_sim_load_ptx_from_filename( file ); - std::string fname(path); - context->name_symtab[fname] = symtab; - context->add_binary(symtab, 1); - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - addedFile = true; - return CUDA_SUCCESS; + return cuLinkAddFileInternal(state, type, path, + numOptions, options, optionValues) } #endif diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 576ec67..0543ff8 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -3,11 +3,13 @@ #include #include #include +#include #include "cuda_api_object.h" class cuobjdumpSection; class cuobjdumpELFSection; class cuobjdumpPTXSection; +class symbol_table; class gpgpu_context { public: @@ -16,6 +18,9 @@ class gpgpu_context { } // global list std::list cuobjdumpSectionList; + std::mapfatbin_registered; + std::map fatbinmap; + std::map name_symtab; //maps sm version number to set of filenames std::map > version_filename; cuda_runtime_api* api; @@ -29,5 +34,6 @@ class gpgpu_context { cuobjdumpELFSection* findELFSection(const std::string identifier); cuobjdumpPTXSection* findPTXSection(const std::string identifier); void extract_ptx_files_using_cuobjdump(CUctx_st *context); + void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context); }; #endif /* __gpgpu_context_h__ */ -- cgit v1.3 From ca08abd4340e990d0c135f93672184a8e2116ecd Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Jun 2019 02:28:12 -0400 Subject: Miss semicolon Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 855f11d..66562aa 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -3947,7 +3947,7 @@ cuLinkAddFile(CUlinkState state, CUjitInputType type, const char *path, unsigned int numOptions, CUjit_option *options, void **optionValues) { return cuLinkAddFileInternal(state, type, path, - numOptions, options, optionValues) + numOptions, options, optionValues); } #endif -- cgit v1.3 From bc8d1e10507f043c916ee051dc9a687adc6d9b4b Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Jun 2019 02:47:07 -0400 Subject: Move more vars Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 117 ++++++++++++++++++++++++++------------------ libcuda/gpgpu_context.h | 11 +++++ 2 files changed, 81 insertions(+), 47 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 66562aa..02e2b2e 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -158,11 +158,11 @@ extern void exit_simulation(); static int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr, gpgpu_t *gpu ); static int load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu ); -static kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, - gpgpu_ptx_sim_arg_list_t args, - struct dim3 gridDim, - struct dim3 blockDim, - struct CUctx_st* context ); +//static kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, +// gpgpu_ptx_sim_arg_list_t args, +// struct dim3 gridDim, +// struct dim3 blockDim, +// struct CUctx_st* context ); /*DEVICE_BUILTIN*/ struct cudaArray @@ -313,7 +313,6 @@ struct CUctx_st { return i->second; } - std::map g_mallocPtr_Size; std::map pinned_memory; //support for pinned memories added std::map pinned_memory_size; int no_of_ptx; @@ -910,7 +909,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = struct CUstream_st *stream = config.get_stream(); printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); - kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); + kernel_info_t *grid = ctx->gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); //do dynamic PDOM analysis for performance simulation scenario std::string kname = grid->name(); function_info *kernel_func_info = grid->entry(); @@ -966,6 +965,66 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = return g_last_cudaError = cudaSuccess; } +cudaError_t cudaMallocInternal(void **devPtr, size_t size, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + CUctx_st* context = GPGPUSim_Context(); + *devPtr = context->get_device()->get_gpgpu()->gpu_malloc(size); + if(g_debug_execution >= 3){ + printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *devPtr); + ctx->g_mallocPtr_Size[(unsigned long long)*devPtr] = size; + } + if ( *devPtr ) { + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorMemoryAllocation; + } +} + +cudaError_t cudaHostGetDevicePointerInternal(void **pDevice, void *pHost, unsigned int flags, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + //only cpu memory allocation happens in cudaHostAlloc. Linking with device pointer to pinned memory happens here. + //TODO: once kernel is executed, the contents in global pointer of GPU must be copied back to CPU host pointer! + flags=0; + CUctx_st* context = GPGPUSim_Context(); + gpgpu_t *gpu = context->get_device()->get_gpgpu(); + std::map::const_iterator i = context->pinned_memory_size.find(pHost); + assert(i != context->pinned_memory_size.end()); + size_t size = i->second; + *pDevice = gpu->gpu_malloc(size); + if(g_debug_execution >= 3){ + printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *pDevice); + ctx->g_mallocPtr_Size[(unsigned long long)*pDevice] = size; + } + if ( *pDevice ) { + context->pinned_memory[pHost]=pDevice; + //Copy contents in cpu to gpu + gpu->memcpy_to_gpu((size_t)*pDevice,pHost,size); + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorMemoryAllocation; + } +} #if CUDART_VERSION >= 6050 CUresult cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, @@ -1027,20 +1086,7 @@ cudaError_t cudaPeekAtLastError(void) __host__ cudaError_t CUDARTAPI cudaMalloc(void **devPtr, size_t size) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st* context = GPGPUSim_Context(); - *devPtr = context->get_device()->get_gpgpu()->gpu_malloc(size); - if(g_debug_execution >= 3){ - printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *devPtr); - context->g_mallocPtr_Size[(unsigned long long)*devPtr] = size; - } - if ( *devPtr ) { - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorMemoryAllocation; - } + return cudaMallocInternal(devPtr, size); } __host__ cudaError_t CUDARTAPI cudaMallocHost(void **ptr, size_t size) @@ -3111,30 +3157,7 @@ cudaError_t CUDARTAPI cudaHostAlloc(void **pHost, size_t bytes, unsigned int fl cudaError_t CUDARTAPI cudaHostGetDevicePointer(void **pDevice, void *pHost, unsigned int flags) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - //only cpu memory allocation happens in cudaHostAlloc. Linking with device pointer to pinned memory happens here. - //TODO: once kernel is executed, the contents in global pointer of GPU must be copied back to CPU host pointer! - flags=0; - CUctx_st* context = GPGPUSim_Context(); - gpgpu_t *gpu = context->get_device()->get_gpgpu(); - std::map::const_iterator i = context->pinned_memory_size.find(pHost); - assert(i != context->pinned_memory_size.end()); - size_t size = i->second; - *pDevice = gpu->gpu_malloc(size); - if(g_debug_execution >= 3){ - printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *pDevice); - context->g_mallocPtr_Size[(unsigned long long)*pDevice] = size; - } - if ( *pDevice ) { - context->pinned_memory[pHost]=pDevice; - //Copy contents in cpu to gpu - gpu->memcpy_to_gpu((size_t)*pDevice,pHost,size); - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorMemoryAllocation; - } + return cudaHostGetDevicePointerInternal(pDevice, pHost, flags); } __host__ cudaError_t CUDARTAPI cudaPointerGetAttributes( @@ -3450,7 +3473,7 @@ static int load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu return nc_bytes; } -kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, +kernel_info_t * gpgpu_context::gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, gpgpu_ptx_sim_arg_list_t args, struct dim3 gridDim, struct dim3 blockDim, @@ -3482,7 +3505,7 @@ kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, fflush(stdout); if(g_debug_execution >= 4){ - entry->ptx_jit_config(context->g_mallocPtr_Size, result->get_param_memory(), (gpgpu_t *) context->get_device()->get_gpgpu(), gridDim, blockDim); + entry->ptx_jit_config(g_mallocPtr_Size, result->get_param_memory(), (gpgpu_t *) context->get_device()->get_gpgpu(), gridDim, blockDim); } return result; diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 0543ff8..f29e2e0 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -10,6 +10,10 @@ class cuobjdumpSection; class cuobjdumpELFSection; class cuobjdumpPTXSection; class symbol_table; +class gpgpu_ptx_sim_arg; +class kernel_info_t; + +typedef std::list gpgpu_ptx_sim_arg_list_t; class gpgpu_context { public: @@ -21,6 +25,7 @@ class gpgpu_context { std::mapfatbin_registered; std::map fatbinmap; std::map name_symtab; + std::map g_mallocPtr_Size; //maps sm version number to set of filenames std::map > version_filename; cuda_runtime_api* api; @@ -35,5 +40,11 @@ class gpgpu_context { cuobjdumpPTXSection* findPTXSection(const std::string identifier); void extract_ptx_files_using_cuobjdump(CUctx_st *context); void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context); + kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, + gpgpu_ptx_sim_arg_list_t args, + struct dim3 gridDim, + struct dim3 blockDim, + struct CUctx_st* context ); + }; #endif /* __gpgpu_context_h__ */ -- cgit v1.3 From 7e286ec0ffc963d307551caced7f4a52241dced4 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Jun 2019 12:53:34 -0400 Subject: Move pinned_memory etc Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 80 +++++++++++++++++++++++++++++---------------- libcuda/gpgpu_context.h | 3 ++ 2 files changed, 54 insertions(+), 29 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 02e2b2e..41f27bf 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -313,8 +313,6 @@ struct CUctx_st { return i->second; } - std::map pinned_memory; //support for pinned memories added - std::map pinned_memory_size; int no_of_ptx; typedef struct glbmap_entry glbmap_entry_t; @@ -989,6 +987,27 @@ cudaError_t cudaMallocInternal(void **devPtr, size_t size, gpgpu_context* gpgpu_ } } +cudaError_t cudaMallocHostInternal(void **ptr, size_t size, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + *ptr = malloc(size); + if ( *ptr ) { + //track pinned memory size allocated in the host so that same amount of memory is also allocated in GPU. + ctx->pinned_memory_size[*ptr]=size; + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorMemoryAllocation; + } +} + cudaError_t cudaHostGetDevicePointerInternal(void **pDevice, void *pHost, unsigned int flags, gpgpu_context* gpgpu_ctx = NULL) { gpgpu_context *ctx; @@ -1008,8 +1027,8 @@ cudaError_t cudaHostGetDevicePointerInternal(void **pDevice, void *pHost, unsign flags=0; CUctx_st* context = GPGPUSim_Context(); gpgpu_t *gpu = context->get_device()->get_gpgpu(); - std::map::const_iterator i = context->pinned_memory_size.find(pHost); - assert(i != context->pinned_memory_size.end()); + std::map::const_iterator i = ctx->pinned_memory_size.find(pHost); + assert(i != ctx->pinned_memory_size.end()); size_t size = i->second; *pDevice = gpu->gpu_malloc(size); if(g_debug_execution >= 3){ @@ -1017,7 +1036,7 @@ cudaError_t cudaHostGetDevicePointerInternal(void **pDevice, void *pHost, unsign ctx->g_mallocPtr_Size[(unsigned long long)*pDevice] = size; } if ( *pDevice ) { - context->pinned_memory[pHost]=pDevice; + ctx->pinned_memory[pHost]=pDevice; //Copy contents in cpu to gpu gpu->memcpy_to_gpu((size_t)*pDevice,pHost,size); return g_last_cudaError = cudaSuccess; @@ -1066,6 +1085,31 @@ cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, } #endif +#if (CUDART_VERSION >= 2010) + +cudaError_t cudaHostAllocInternal(void **pHost, size_t bytes, unsigned int flags, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + *pHost = malloc(bytes); + //need to track the size allocated so that cudaHostGetDevicePointer() can function properly. + //TODO: vary this function behavior based on flags value (following nvidia documentation) + ctx->pinned_memory_size[*pHost]=bytes; + if( *pHost ) + return g_last_cudaError = cudaSuccess; + else + return g_last_cudaError = cudaErrorMemoryAllocation; +} + +#endif + /******************************************************************************* * * * * @@ -1091,18 +1135,7 @@ __host__ cudaError_t CUDARTAPI cudaMalloc(void **devPtr, size_t size) __host__ cudaError_t CUDARTAPI cudaMallocHost(void **ptr, size_t size) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st* context = GPGPUSim_Context(); - *ptr = malloc(size); - if ( *ptr ) { - //track pinned memory size allocated in the host so that same amount of memory is also allocated in GPU. - context->pinned_memory_size[*ptr]=size; - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorMemoryAllocation; - } + return cudaMallocHostInternal(ptr, size); } __host__ cudaError_t CUDARTAPI cudaMallocPitch(void **devPtr, size_t *pitch, size_t width, size_t height) { @@ -3141,18 +3174,7 @@ cudaError_t cudaGLUnregisterBufferObject(GLuint bufferObj) cudaError_t CUDARTAPI cudaHostAlloc(void **pHost, size_t bytes, unsigned int flags) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - *pHost = malloc(bytes); - //need to track the size allocated so that cudaHostGetDevicePointer() can function properly. - //TODO: vary this function behavior based on flags value (following nvidia documentation) - CUctx_st* context = GPGPUSim_Context(); - context->pinned_memory_size[*pHost]=bytes; - if( *pHost ) - return g_last_cudaError = cudaSuccess; - else - return g_last_cudaError = cudaErrorMemoryAllocation; + return cudaHostAllocInternal(pHost, bytes, flags); } cudaError_t CUDARTAPI cudaHostGetDevicePointer(void **pDevice, void *pHost, unsigned int flags) diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index f29e2e0..6878d5c 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -28,6 +28,9 @@ class gpgpu_context { std::map g_mallocPtr_Size; //maps sm version number to set of filenames std::map > version_filename; + std::map pinned_memory; //support for pinned memories added + std::map pinned_memory_size; + // objects pointers for each file cuda_runtime_api* api; // member function list void cuobjdumpInit(); -- cgit v1.3 From ba8374c72558e4b89a0bddc973bcc87a10e2ab5f Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Jun 2019 13:20:29 -0400 Subject: Move g_glbmap Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 134 ++++++++++++++++++++++---------------------- libcuda/gpgpu_context.h | 15 +++++ 2 files changed, 82 insertions(+), 67 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 41f27bf..18a9abb 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -237,17 +237,6 @@ private: struct _cuda_device_id *m_next; }; -#ifndef OPENGL_SUPPORT -typedef unsigned long GLuint; -#endif - -struct glbmap_entry { - GLuint m_bufferObj; - void *m_devPtr; - size_t m_size; - struct glbmap_entry *m_next; -}; - struct CUctx_st { CUctx_st( _cuda_device_id *gpu ) { @@ -255,7 +244,6 @@ struct CUctx_st { m_binary_info.cmem = 0; m_binary_info.gmem = 0; no_of_ptx=0; - g_glbmap = NULL; } _cuda_device_id *get_device() { return m_gpu; } @@ -314,9 +302,6 @@ struct CUctx_st { } int no_of_ptx; - typedef struct glbmap_entry glbmap_entry_t; - - glbmap_entry_t* g_glbmap; private: _cuda_device_id *m_gpu; // selected gpu @@ -1044,6 +1029,72 @@ cudaError_t cudaHostGetDevicePointerInternal(void **pDevice, void *pHost, unsign return g_last_cudaError = cudaErrorMemoryAllocation; } } + +cudaError_t cudaGLMapBufferObjectInternal(void** devPtr, GLuint bufferObj, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } +#ifdef OPENGL_SUPPORT + GLint buffer_size=0; + CUctx_st* context = GPGPUSim_Context(); + + glbmap_entry_t *p = ctx->g_glbmap; + while ( p && p->m_bufferObj != bufferObj ) + p = p->m_next; + if ( p == NULL ) { + glBindBuffer(GL_ARRAY_BUFFER,bufferObj); + glGetBufferParameteriv(GL_ARRAY_BUFFER,GL_BUFFER_SIZE,&buffer_size); + assert( buffer_size != 0 ); + *devPtr = context->get_device()->get_gpgpu()->gpu_malloc(buffer_size); + + // create entry and insert to front of list + glbmap_entry_t *n = (glbmap_entry_t *) calloc(1,sizeof(glbmap_entry_t)); + n->m_next = ctx->g_glbmap; + ctx->g_glbmap = n; + + // initialize entry + n->m_bufferObj = bufferObj; + n->m_devPtr = *devPtr; + n->m_size = buffer_size; + + p = n; + } else { + buffer_size = p->m_size; + *devPtr = p->m_devPtr; + } + + if ( *devPtr ) { + char *data = (char *) calloc(p->m_size,1); + glGetBufferSubData(GL_ARRAY_BUFFER,0,buffer_size,data); + memcpy_to_gpu( (size_t) *devPtr, data, buffer_size ); + free(data); + printf("GPGPU-Sim PTX: cudaGLMapBufferObject %zu bytes starting at 0x%llx..\n", (size_t)buffer_size, + (unsigned long long) *devPtr); + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorMemoryAllocation; + } + + return g_last_cudaError = cudaSuccess; +#else + fflush(stdout); + fflush(stderr); + printf("GPGPU-Sim PTX: GPGPU-Sim support for OpenGL integration disabled -- exiting\n"); + fflush(stdout); + exit(50); +#endif +} + #if CUDART_VERSION >= 6050 CUresult cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, @@ -3079,58 +3130,7 @@ cudaError_t cudaGLRegisterBufferObject(GLuint bufferObj) cudaError_t cudaGLMapBufferObject(void** devPtr, GLuint bufferObj) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } -#ifdef OPENGL_SUPPORT - GLint buffer_size=0; - CUctx_st* ctx = GPGPUSim_Context(); - - glbmap_entry_t *p = ctx->g_glbmap; - while ( p && p->m_bufferObj != bufferObj ) - p = p->m_next; - if ( p == NULL ) { - glBindBuffer(GL_ARRAY_BUFFER,bufferObj); - glGetBufferParameteriv(GL_ARRAY_BUFFER,GL_BUFFER_SIZE,&buffer_size); - assert( buffer_size != 0 ); - *devPtr = ctx->get_device()->get_gpgpu()->gpu_malloc(buffer_size); - - // create entry and insert to front of list - glbmap_entry_t *n = (glbmap_entry_t *) calloc(1,sizeof(glbmap_entry_t)); - n->m_next = ctx->g_glbmap; - ctx->g_glbmap = n; - - // initialize entry - n->m_bufferObj = bufferObj; - n->m_devPtr = *devPtr; - n->m_size = buffer_size; - - p = n; - } else { - buffer_size = p->m_size; - *devPtr = p->m_devPtr; - } - - if ( *devPtr ) { - char *data = (char *) calloc(p->m_size,1); - glGetBufferSubData(GL_ARRAY_BUFFER,0,buffer_size,data); - memcpy_to_gpu( (size_t) *devPtr, data, buffer_size ); - free(data); - printf("GPGPU-Sim PTX: cudaGLMapBufferObject %zu bytes starting at 0x%llx..\n", (size_t)buffer_size, - (unsigned long long) *devPtr); - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorMemoryAllocation; - } - - return g_last_cudaError = cudaSuccess; -#else - fflush(stdout); - fflush(stderr); - printf("GPGPU-Sim PTX: GPGPU-Sim support for OpenGL integration disabled -- exiting\n"); - fflush(stdout); - exit(50); -#endif + return cudaGLMapBufferObjectInternal(devPtr, bufferObj); } cudaError_t cudaGLUnmapBufferObject(GLuint bufferObj) diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 6878d5c..7569ea6 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -15,10 +15,24 @@ class kernel_info_t; typedef std::list gpgpu_ptx_sim_arg_list_t; +#ifndef OPENGL_SUPPORT +typedef unsigned long GLuint; +#endif + +struct glbmap_entry { + GLuint m_bufferObj; + void *m_devPtr; + size_t m_size; + struct glbmap_entry *m_next; +}; + +typedef struct glbmap_entry glbmap_entry_t; + class gpgpu_context { public: gpgpu_context() { api = new cuda_runtime_api(); + g_glbmap = NULL; } // global list std::list cuobjdumpSectionList; @@ -30,6 +44,7 @@ class gpgpu_context { std::map > version_filename; std::map pinned_memory; //support for pinned memories added std::map pinned_memory_size; + glbmap_entry_t* g_glbmap; // objects pointers for each file cuda_runtime_api* api; // member function list -- cgit v1.3 From 87409183125b863be0c3e0470b2417c3c92a748b Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 9 Jun 2019 02:47:39 -0400 Subject: Move some vars back to cuda_api_object Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 54 +++++++++++++++++++++++++++++++++++++++ libcuda/cuda_runtime_api.cc | 62 ++++++++++++++++++++++----------------------- libcuda/gpgpu_context.h | 53 -------------------------------------- 3 files changed, 85 insertions(+), 84 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index 2001f91..41337c6 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -1,14 +1,68 @@ #ifndef __cuda_api_object_h__ #define __cuda_api_object_h__ + +#include +#include +#include +#include + class cuobjdumpSection; +class cuobjdumpELFSection; +class cuobjdumpPTXSection; +class symbol_table; +class gpgpu_ptx_sim_arg; class kernel_config; +class kernel_info_t; + +typedef std::list gpgpu_ptx_sim_arg_list_t; + +#ifndef OPENGL_SUPPORT +typedef unsigned long GLuint; +#endif + +struct glbmap_entry { + GLuint m_bufferObj; + void *m_devPtr; + size_t m_size; + struct glbmap_entry *m_next; +}; + +typedef struct glbmap_entry glbmap_entry_t; class cuda_runtime_api { public: + cuda_runtime_api() { + g_glbmap = NULL; + } // global list + std::list cuobjdumpSectionList; std::list libSectionList; std::list g_cuda_launch_stack; + std::mapfatbin_registered; + std::map fatbinmap; + std::map name_symtab; + std::map g_mallocPtr_Size; + //maps sm version number to set of filenames + std::map > version_filename; + std::map pinned_memory; //support for pinned memories added + std::map pinned_memory_size; + glbmap_entry_t* g_glbmap; // member function list + void cuobjdumpInit(); + void extract_code_using_cuobjdump(); + void extract_ptx_files_using_cuobjdump(CUctx_st *context); + void cuobjdumpParseBinary(unsigned int handle); + std::list pruneSectionList(CUctx_st *context); + std::list mergeMatchingSections(std::string identifier); + std::list mergeSections(); + cuobjdumpELFSection* findELFSection(const std::string identifier); + cuobjdumpPTXSection* findPTXSection(const std::string identifier); + void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context); + kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, + gpgpu_ptx_sim_arg_list_t args, + struct dim3 gridDim, + struct dim3 blockDim, + struct CUctx_st* context ); }; #endif /* __cuda_api_object_h__ */ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 18a9abb..72ab002 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -626,7 +626,7 @@ static int get_app_cuda_version() { } //! Keep track of the association between filename and cubin handle -void gpgpu_context::cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ +void cuda_runtime_api::cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context){ fatbinmap[handle] = filename; } @@ -703,8 +703,8 @@ void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = * then for next calls, only returns the appropriate number */ assert(fat_cubin_handle >= 1); - if (fat_cubin_handle==1) ctx->cuobjdumpInit(); - ctx->cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); + if (fat_cubin_handle==1) ctx->api->cuobjdumpInit(); + ctx->api->cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); return (void**)fat_cubin_handle; } @@ -801,7 +801,7 @@ void cudaRegisterFunctionInternal( printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", deviceFun, hostFun, fat_cubin_handle); if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - ctx->cuobjdumpParseBinary(fat_cubin_handle); + ctx->api->cuobjdumpParseBinary(fat_cubin_handle); context->register_function( fat_cubin_handle, hostFun, deviceFun ); } @@ -828,7 +828,7 @@ void cudaRegisterVarInternal( printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - ctx->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); + ctx->api->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); fflush(stdout); if ( constant && !global && !ext ) { gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); @@ -892,7 +892,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = struct CUstream_st *stream = config.get_stream(); printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); - kernel_info_t *grid = ctx->gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); + kernel_info_t *grid = ctx->api->gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); //do dynamic PDOM analysis for performance simulation scenario std::string kname = grid->name(); function_info *kernel_func_info = grid->entry(); @@ -963,7 +963,7 @@ cudaError_t cudaMallocInternal(void **devPtr, size_t size, gpgpu_context* gpgpu_ *devPtr = context->get_device()->get_gpgpu()->gpu_malloc(size); if(g_debug_execution >= 3){ printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *devPtr); - ctx->g_mallocPtr_Size[(unsigned long long)*devPtr] = size; + ctx->api->g_mallocPtr_Size[(unsigned long long)*devPtr] = size; } if ( *devPtr ) { return g_last_cudaError = cudaSuccess; @@ -986,7 +986,7 @@ cudaError_t cudaMallocHostInternal(void **ptr, size_t size, gpgpu_context* gpgpu *ptr = malloc(size); if ( *ptr ) { //track pinned memory size allocated in the host so that same amount of memory is also allocated in GPU. - ctx->pinned_memory_size[*ptr]=size; + ctx->api->pinned_memory_size[*ptr]=size; return g_last_cudaError = cudaSuccess; } else { return g_last_cudaError = cudaErrorMemoryAllocation; @@ -1012,16 +1012,16 @@ cudaError_t cudaHostGetDevicePointerInternal(void **pDevice, void *pHost, unsign flags=0; CUctx_st* context = GPGPUSim_Context(); gpgpu_t *gpu = context->get_device()->get_gpgpu(); - std::map::const_iterator i = ctx->pinned_memory_size.find(pHost); - assert(i != ctx->pinned_memory_size.end()); + std::map::const_iterator i = ctx->api->pinned_memory_size.find(pHost); + assert(i != ctx->api->pinned_memory_size.end()); size_t size = i->second; *pDevice = gpu->gpu_malloc(size); if(g_debug_execution >= 3){ printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *pDevice); - ctx->g_mallocPtr_Size[(unsigned long long)*pDevice] = size; + ctx->api->g_mallocPtr_Size[(unsigned long long)*pDevice] = size; } if ( *pDevice ) { - ctx->pinned_memory[pHost]=pDevice; + ctx->api->pinned_memory[pHost]=pDevice; //Copy contents in cpu to gpu gpu->memcpy_to_gpu((size_t)*pDevice,pHost,size); return g_last_cudaError = cudaSuccess; @@ -1048,7 +1048,7 @@ cudaError_t cudaGLMapBufferObjectInternal(void** devPtr, GLuint bufferObj, gpgpu GLint buffer_size=0; CUctx_st* context = GPGPUSim_Context(); - glbmap_entry_t *p = ctx->g_glbmap; + glbmap_entry_t *p = ctx->api->g_glbmap; while ( p && p->m_bufferObj != bufferObj ) p = p->m_next; if ( p == NULL ) { @@ -1059,8 +1059,8 @@ cudaError_t cudaGLMapBufferObjectInternal(void** devPtr, GLuint bufferObj, gpgpu // create entry and insert to front of list glbmap_entry_t *n = (glbmap_entry_t *) calloc(1,sizeof(glbmap_entry_t)); - n->m_next = ctx->g_glbmap; - ctx->g_glbmap = n; + n->m_next = ctx->api->g_glbmap; + ctx->api->g_glbmap = n; // initialize entry n->m_bufferObj = bufferObj; @@ -1127,7 +1127,7 @@ cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, strcat(file,path); symbol_table *symtab = gpgpu_ptx_sim_load_ptx_from_filename( file ); std::string fname(path); - ctx->name_symtab[fname] = symtab; + ctx->api->name_symtab[fname] = symtab; context->add_binary(symtab, 1); load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); @@ -1152,7 +1152,7 @@ cudaError_t cudaHostAllocInternal(void **pHost, size_t bytes, unsigned int flag *pHost = malloc(bytes); //need to track the size allocated so that cudaHostGetDevicePointer() can function properly. //TODO: vary this function behavior based on flags value (following nvidia documentation) - ctx->pinned_memory_size[*pHost]=bytes; + ctx->api->pinned_memory_size[*pHost]=bytes; if( *pHost ) return g_last_cudaError = cudaSuccess; else @@ -2428,7 +2428,7 @@ __host__ cudaError_t CUDARTAPI cudaGetExportTable(const void **ppExportTable, co //#include "../../cuobjdump_to_ptxplus/cuobjdump_parser.h" //extracts all ptx files from binary and dumps into prog_name.unique_no.sm_<>.ptx files -void gpgpu_context::extract_ptx_files_using_cuobjdump(CUctx_st *context){ +void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context){ extern bool g_cdp_enabled; char command[1000]; char *pytorch_bin = getenv("PYTORCH_BIN"); @@ -2506,7 +2506,7 @@ void gpgpu_context::extract_ptx_files_using_cuobjdump(CUctx_st *context){ * It is also responsible for extracting the libraries linked to the binary if the option is * enabled * */ -void gpgpu_context::extract_code_using_cuobjdump(){ +void cuda_runtime_api::extract_code_using_cuobjdump(){ CUctx_st *context = GPGPUSim_Context(); unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); @@ -2630,7 +2630,7 @@ void gpgpu_context::extract_code_using_cuobjdump(){ fclose(cuobjdump_in); std::getline(libsf, line); } - api->libSectionList = cuobjdumpSectionList; + libSectionList = cuobjdumpSectionList; //Restore the original section list cuobjdumpSectionList = tmpsl; @@ -2676,7 +2676,7 @@ void printSectionList(std::list sl) { } //! Remove unecessary sm versions from the section list -std::list gpgpu_context::pruneSectionList(CUctx_st *context) { +std::list cuda_runtime_api::pruneSectionList(CUctx_st *context) { unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); //For ptxplus, force the max capability to 19 if it's higher or unspecified(0) @@ -2729,7 +2729,7 @@ std::list gpgpu_context::pruneSectionList(CUctx_st *context) } //! Merge all PTX sections that have a specific identifier into one file -std::list gpgpu_context::mergeMatchingSections(std::string identifier){ +std::list cuda_runtime_api::mergeMatchingSections(std::string identifier){ const char *ptxcode = ""; std::list::iterator old_iter; cuobjdumpPTXSection* old_ptxsection = NULL; @@ -2772,7 +2772,7 @@ std::list gpgpu_context::mergeMatchingSections(std::string id } //! Merge any PTX sections with matching identifiers -std::list gpgpu_context::mergeSections(){ +std::list cuda_runtime_api::mergeSections(){ std::vector identifier; cuobjdumpPTXSection* ptxsection; @@ -2819,10 +2819,10 @@ cuobjdumpELFSection* findELFSectionInList(std::list sectionli } //! Find an ELF section in all the known lists -cuobjdumpELFSection* gpgpu_context::findELFSection(const std::string identifier){ +cuobjdumpELFSection* cuda_runtime_api::findELFSection(const std::string identifier){ cuobjdumpELFSection* sec = findELFSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; - sec = findELFSectionInList(api->libSectionList, identifier); + sec = findELFSectionInList(libSectionList, identifier); if (sec!=NULL)return sec; std::cout << "Could not find " << identifier << std::endl; assert(0 && "Could not find the required ELF section"); @@ -2854,10 +2854,10 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list §ionl } //! Find an PTX section in all the known lists -cuobjdumpPTXSection* gpgpu_context::findPTXSection(const std::string identifier){ +cuobjdumpPTXSection* cuda_runtime_api::findPTXSection(const std::string identifier){ cuobjdumpPTXSection* sec = findPTXSectionInList(cuobjdumpSectionList, identifier); if (sec!=NULL)return sec; - sec = findPTXSectionInList(api->libSectionList, identifier); + sec = findPTXSectionInList(libSectionList, identifier); if (sec!=NULL)return sec; std::cout << "Could not find " << identifier << std::endl; assert(0 && "Could not find the required PTX section"); @@ -2867,7 +2867,7 @@ cuobjdumpPTXSection* gpgpu_context::findPTXSection(const std::string identifier) //! Extract the code using cuobjdump and remove unnecessary sections -void gpgpu_context::cuobjdumpInit(){ +void cuda_runtime_api::cuobjdumpInit(){ CUctx_st *context = GPGPUSim_Context(); extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.* const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); @@ -2879,7 +2879,7 @@ void gpgpu_context::cuobjdumpInit(){ //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it -void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ +void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ CUctx_st *context = GPGPUSim_Context(); if(fatbin_registered[handle]) return; @@ -3140,7 +3140,7 @@ cudaError_t cudaGLUnmapBufferObject(GLuint bufferObj) } #ifdef OPENGL_SUPPORT CUctx_st* ctx = GPGPUSim_Context(); - glbmap_entry_t *p = ctx->g_glbmap; + glbmap_entry_t *p = ctx->api->g_glbmap; while ( p && p->m_bufferObj != bufferObj ) p = p->m_next; if ( p == NULL ) @@ -3495,7 +3495,7 @@ static int load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu return nc_bytes; } -kernel_info_t * gpgpu_context::gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, +kernel_info_t * cuda_runtime_api::gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, gpgpu_ptx_sim_arg_list_t args, struct dim3 gridDim, struct dim3 blockDim, diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 7569ea6..4622f00 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -1,68 +1,15 @@ #ifndef __gpgpu_context_h__ #define __gpgpu_context_h__ -#include -#include -#include -#include #include "cuda_api_object.h" -class cuobjdumpSection; -class cuobjdumpELFSection; -class cuobjdumpPTXSection; -class symbol_table; -class gpgpu_ptx_sim_arg; -class kernel_info_t; - -typedef std::list gpgpu_ptx_sim_arg_list_t; - -#ifndef OPENGL_SUPPORT -typedef unsigned long GLuint; -#endif - -struct glbmap_entry { - GLuint m_bufferObj; - void *m_devPtr; - size_t m_size; - struct glbmap_entry *m_next; -}; - -typedef struct glbmap_entry glbmap_entry_t; - class gpgpu_context { public: gpgpu_context() { api = new cuda_runtime_api(); - g_glbmap = NULL; } // global list - std::list cuobjdumpSectionList; - std::mapfatbin_registered; - std::map fatbinmap; - std::map name_symtab; - std::map g_mallocPtr_Size; - //maps sm version number to set of filenames - std::map > version_filename; - std::map pinned_memory; //support for pinned memories added - std::map pinned_memory_size; - glbmap_entry_t* g_glbmap; // objects pointers for each file cuda_runtime_api* api; // member function list - void cuobjdumpInit(); - void cuobjdumpParseBinary(unsigned int handle); - void extract_code_using_cuobjdump(); - std::list pruneSectionList(CUctx_st *context); - std::list mergeMatchingSections(std::string identifier); - std::list mergeSections(); - cuobjdumpELFSection* findELFSection(const std::string identifier); - cuobjdumpPTXSection* findPTXSection(const std::string identifier); - void extract_ptx_files_using_cuobjdump(CUctx_st *context); - void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context); - kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, - gpgpu_ptx_sim_arg_list_t args, - struct dim3 gridDim, - struct dim3 blockDim, - struct CUctx_st* context ); - }; #endif /* __gpgpu_context_h__ */ -- cgit v1.3 From c7f515f6f5325c65f32dd64e1ad479660c751e99 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 9 Jun 2019 22:20:15 -0400 Subject: A bunch of boilerbplate to get 10.1 to compile. Still does not yet run. The way CUDA calls kerenels (even on old code) has changed. --- Makefile | 2 ++ libcuda/cuda_api.h | 2 ++ libcuda/cuda_runtime_api.cc | 25 +++++++++++++++++++++++++ linux-so-version.txt | 4 ++++ setup_environment | 2 +- 5 files changed, 34 insertions(+), 1 deletion(-) (limited to 'libcuda') diff --git a/Makefile b/Makefile index 3db8ce8..a69130c 100644 --- a/Makefile +++ b/Makefile @@ -164,6 +164,8 @@ $(SIM_LIB_DIR)/libcudart.so: makedirs $(LIBS) cudalib if [ ! -f $(SIM_LIB_DIR)/libcudart.so.9.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.9.0; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.9.1 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.9.1; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.9.2 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.9.2; fi + if [ ! -f $(SIM_LIB_DIR)/libcudart.so.10.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.10.0; fi + if [ ! -f $(SIM_LIB_DIR)/libcudart.so.10.1 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.10.1; fi $(SIM_LIB_DIR)/libcudart.dylib: makedirs $(LIBS) cudalib g++ -dynamiclib -Wl,-headerpad_max_install_names,-undefined,dynamic_lookup,-compatibility_version,1.1,-current_version,1.1\ diff --git a/libcuda/cuda_api.h b/libcuda/cuda_api.h index 3808e8a..7ee26dc 100644 --- a/libcuda/cuda_api.h +++ b/libcuda/cuda_api.h @@ -234,9 +234,11 @@ typedef struct CUgraphicsResource_st *CUgraphicsResource; /**< CUDA graphics int typedef unsigned long long CUtexObject; /**< An opaque value that represents a CUDA texture object */ typedef unsigned long long CUsurfObject; /**< An opaque value that represents a CUDA surface object */ +#if __CUDA_API_VERSION < 1010 typedef struct CUuuid_st { /**< CUDA definition of UUID */ char bytes[16]; } CUuuid; +#endif #if __CUDA_API_VERSION >= 4010 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 18a9abb..718db49 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -630,6 +630,8 @@ void gpgpu_context::cuobjdumpRegisterFatBinary(unsigned int handle, const char* fatbinmap[handle] = filename; } + + /******************************************************************************* * Add internal cuda runtime API call to accept gpgpu_context * *******************************************************************************/ @@ -2975,6 +2977,29 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) return cudaRegisterFatBinaryInternal(fatCubin); } +void CUDARTAPI __cudaRegisterFatBinaryEnd( void **fatCubinHandle ) +{ + +} + +unsigned CUDARTAPI __cudaPushCallConfiguration(dim3 gridDim, + dim3 blockDim, + size_t sharedMem = 0, + struct CUstream_st *stream = 0) +{ + +} + +cudaError_t CUDARTAPI __cudaPopCallConfiguration( + dim3 *gridDim, + dim3 *blockDim, + size_t *sharedMem, + void *stream +) +{ + return g_last_cudaError = cudaSuccess; +} + void CUDARTAPI __cudaRegisterFunction( void **fatCubinHandle, const char *hostFun, diff --git a/linux-so-version.txt b/linux-so-version.txt index a7c2d3c..45c40dd 100644 --- a/linux-so-version.txt +++ b/linux-so-version.txt @@ -4,5 +4,9 @@ libcudart.so.9.1{ }; libcudart.so.9.2{ }; +libcudart.so.10.0{ +}; +libcudart.so.10.1{ +}; libcuda.so.1{ }; diff --git a/setup_environment b/setup_environment index b420584..17891ce 100644 --- a/setup_environment +++ b/setup_environment @@ -51,7 +51,7 @@ CC_VERSION=`gcc --version | head -1 | awk '{for(i=1;i<=NF;i++){ if(match($i,/^[0 CUDA_VERSION_STRING=`$CUDA_INSTALL_PATH/bin/nvcc --version | awk '/release/ {print $5;}' | sed 's/,//'`; export CUDA_VERSION_NUMBER=`echo $CUDA_VERSION_STRING | sed 's/\./ /' | awk '{printf("%02u%02u", 10*int($1), 10*$2);}'` -if [ $CUDA_VERSION_NUMBER -gt 9100 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then +if [ $CUDA_VERSION_NUMBER -gt 10100 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then echo "ERROR ** GPGPU-Sim version $GPGPUSIM_VERSION_STRING not tested with CUDA version $CUDA_VERSION_STRING (please see README)"; return fi -- cgit v1.3 From 8a6d828b9adfe8bf690f671e9fc5902315b030b7 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 9 Jun 2019 22:35:34 -0400 Subject: updating the APU --- libcuda/cuda_api.h | 5470 +++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 4371 insertions(+), 1099 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api.h b/libcuda/cuda_api.h index 7ee26dc..27983b4 100644 --- a/libcuda/cuda_api.h +++ b/libcuda/cuda_api.h @@ -1,5 +1,5 @@ /* - * Copyright 1993-2014 NVIDIA Corporation. All rights reserved. + * Copyright 1993-2018 NVIDIA Corporation. All rights reserved. * * NOTICE TO LICENSEE: * @@ -63,6 +63,16 @@ typedef uint64_t cuuint64_t; /** * CUDA API versioning support */ +#if defined(__CUDA_API_VERSION_INTERNAL) || defined(__DOXYGEN_ONLY__) || defined(CUDA_ENABLE_DEPRECATED) +#define __CUDA_DEPRECATED +#elif defined(_MSC_VER) +#define __CUDA_DEPRECATED __declspec(deprecated) +#elif defined(__GNUC__) +#define __CUDA_DEPRECATED __attribute__((deprecated)) +#else +#define __CUDA_DEPRECATED +#endif + #if defined(CUDA_FORCE_API_VERSION) #if (CUDA_FORCE_API_VERSION == 3010) #define __CUDA_API_VERSION 3010 @@ -70,7 +80,7 @@ typedef uint64_t cuuint64_t; #error "Unsupported value of CUDA_FORCE_API_VERSION" #endif #else - #define __CUDA_API_VERSION 8000 + #define __CUDA_API_VERSION 10010 #endif /* CUDA_FORCE_API_VERSION */ #if defined(__CUDA_API_VERSION_INTERNAL) || defined(CUDA_API_PER_THREAD_DEFAULT_STREAM) @@ -135,10 +145,20 @@ typedef uint64_t cuuint64_t; #if defined(__CUDA_API_VERSION_INTERNAL) || __CUDA_API_VERSION >= 4010 #define cuTexRefSetAddress2D cuTexRefSetAddress2D_v3 #endif /* __CUDA_API_VERSION_INTERNAL || __CUDA_API_VERSION >= 4010 */ +#if defined(__CUDA_API_VERSION_INTERNAL) || __CUDA_API_VERSION >= 6050 + #define cuLinkCreate cuLinkCreate_v2 + #define cuLinkAddData cuLinkAddData_v2 + #define cuLinkAddFile cuLinkAddFile_v2 +#endif /* __CUDA_API_VERSION_INTERNAL || __CUDA_API_VERSION >= 6050 */ #if defined(__CUDA_API_VERSION_INTERNAL) || __CUDA_API_VERSION >= 6050 #define cuMemHostRegister cuMemHostRegister_v2 #define cuGraphicsResourceSetMapFlags cuGraphicsResourceSetMapFlags_v2 #endif /* __CUDA_API_VERSION_INTERNAL || __CUDA_API_VERSION >= 6050 */ +#if defined(__CUDA_API_VERSION_INTERNAL) || __CUDA_API_VERSION >= 10010 + #define cuStreamBeginCapture __CUDA_API_PTSZ(cuStreamBeginCapture_v2) +#elif defined(__CUDA_API_PER_THREAD_DEFAULT_STREAM) + #define cuStreamBeginCapture __CUDA_API_PTSZ(cuStreamBeginCapture) +#endif /* __CUDA_API_VERSION_INTERNAL || __CUDA_API_VERSION >= 10010 */ #if !defined(__CUDA_API_VERSION_INTERNAL) #if defined(__CUDA_API_VERSION) && __CUDA_API_VERSION >= 3020 && __CUDA_API_VERSION < 4010 @@ -164,19 +184,33 @@ typedef uint64_t cuuint64_t; #define cuStreamGetPriority __CUDA_API_PTSZ(cuStreamGetPriority) #define cuStreamGetFlags __CUDA_API_PTSZ(cuStreamGetFlags) + #define cuStreamGetCtx __CUDA_API_PTSZ(cuStreamGetCtx) #define cuStreamWaitEvent __CUDA_API_PTSZ(cuStreamWaitEvent) + #define cuStreamEndCapture __CUDA_API_PTSZ(cuStreamEndCapture) + #define cuStreamIsCapturing __CUDA_API_PTSZ(cuStreamIsCapturing) + #define cuStreamGetCaptureInfo __CUDA_API_PTSZ(cuStreamGetCaptureInfo) #define cuStreamAddCallback __CUDA_API_PTSZ(cuStreamAddCallback) #define cuStreamAttachMemAsync __CUDA_API_PTSZ(cuStreamAttachMemAsync) #define cuStreamQuery __CUDA_API_PTSZ(cuStreamQuery) #define cuStreamSynchronize __CUDA_API_PTSZ(cuStreamSynchronize) #define cuEventRecord __CUDA_API_PTSZ(cuEventRecord) #define cuLaunchKernel __CUDA_API_PTSZ(cuLaunchKernel) + #define cuLaunchHostFunc __CUDA_API_PTSZ(cuLaunchHostFunc) #define cuGraphicsMapResources __CUDA_API_PTSZ(cuGraphicsMapResources) #define cuGraphicsUnmapResources __CUDA_API_PTSZ(cuGraphicsUnmapResources) #define cuStreamWriteValue32 __CUDA_API_PTSZ(cuStreamWriteValue32) #define cuStreamWaitValue32 __CUDA_API_PTSZ(cuStreamWaitValue32) + #define cuStreamWriteValue64 __CUDA_API_PTSZ(cuStreamWriteValue64) + #define cuStreamWaitValue64 __CUDA_API_PTSZ(cuStreamWaitValue64) #define cuStreamBatchMemOp __CUDA_API_PTSZ(cuStreamBatchMemOp) + + #define cuLaunchCooperativeKernel __CUDA_API_PTSZ(cuLaunchCooperativeKernel) + + #define cuSignalExternalSemaphoresAsync __CUDA_API_PTSZ(cuSignalExternalSemaphoresAsync) + #define cuWaitExternalSemaphoresAsync __CUDA_API_PTSZ(cuWaitExternalSemaphoresAsync) + + #define cuGraphLaunch __CUDA_API_PTSZ(cuGraphLaunch) #endif /** @@ -200,7 +234,7 @@ typedef uint64_t cuuint64_t; /** * CUDA API version number */ -#define CUDA_VERSION 8000 +#define CUDA_VERSION 10010 #ifdef __cplusplus extern "C" { @@ -209,7 +243,7 @@ extern "C" { /** * CUDA device pointer * CUdeviceptr is defined as an unsigned integer type whose size matches the size of a pointer on the target platform. - */ + */ #if __CUDA_API_VERSION >= 3020 #if defined(_WIN64) || defined(__LP64__) @@ -233,18 +267,23 @@ typedef struct CUstream_st *CUstream; /**< CUDA stream */ typedef struct CUgraphicsResource_st *CUgraphicsResource; /**< CUDA graphics interop resource */ typedef unsigned long long CUtexObject; /**< An opaque value that represents a CUDA texture object */ typedef unsigned long long CUsurfObject; /**< An opaque value that represents a CUDA surface object */ - -#if __CUDA_API_VERSION < 1010 +typedef struct CUextMemory_st *CUexternalMemory; /**< CUDA external memory */ +typedef struct CUextSemaphore_st *CUexternalSemaphore; /**< CUDA external semaphore */ +typedef struct CUgraph_st *CUgraph; /**< CUDA graph */ +typedef struct CUgraphNode_st *CUgraphNode; /**< CUDA graph node */ +typedef struct CUgraphExec_st *CUgraphExec; /**< CUDA executable graph */ + +#ifndef CU_UUID_HAS_BEEN_DEFINED +#define CU_UUID_HAS_BEEN_DEFINED typedef struct CUuuid_st { /**< CUDA definition of UUID */ char bytes[16]; } CUuuid; #endif - #if __CUDA_API_VERSION >= 4010 /** - * CUDA IPC handle size + * CUDA IPC handle size */ #define CU_IPC_HANDLE_SIZE 64 @@ -291,7 +330,7 @@ typedef enum CUctx_flags_enum { CU_CTX_BLOCKING_SYNC = 0x04, /**< Set blocking synchronization as default scheduling * \deprecated This flag was deprecated as of CUDA 4.0 * and was replaced with ::CU_CTX_SCHED_BLOCKING_SYNC. */ - CU_CTX_SCHED_MASK = 0x07, + CU_CTX_SCHED_MASK = 0x07, CU_CTX_MAP_HOST = 0x08, /**< Support mapped pinned allocations */ CU_CTX_LMEM_RESIZE_TO_MAX = 0x10, /**< Keep local memory allocation after launch */ CU_CTX_FLAGS_MASK = 0x1f @@ -337,20 +376,26 @@ typedef enum CUevent_flags_enum { #if __CUDA_API_VERSION >= 8000 /** - * Flags for ::cuStreamWaitValue32 + * Flags for ::cuStreamWaitValue32 and ::cuStreamWaitValue64 */ typedef enum CUstreamWaitValue_flags_enum { - CU_STREAM_WAIT_VALUE_GEQ = 0x0, /**< Wait until (int32_t)(*addr - value) >= 0. Note this is a - cyclic comparison which ignores wraparound. (Default behavior.) */ + CU_STREAM_WAIT_VALUE_GEQ = 0x0, /**< Wait until (int32_t)(*addr - value) >= 0 (or int64_t for 64 bit + values). Note this is a cyclic comparison which ignores wraparound. + (Default behavior.) */ CU_STREAM_WAIT_VALUE_EQ = 0x1, /**< Wait until *addr == value. */ CU_STREAM_WAIT_VALUE_AND = 0x2, /**< Wait until (*addr & value) != 0. */ + CU_STREAM_WAIT_VALUE_NOR = 0x3, /**< Wait until ~(*addr | value) != 0. Support for this operation can be + queried with ::cuDeviceGetAttribute() and + ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR.*/ CU_STREAM_WAIT_VALUE_FLUSH = 1<<30 /**< Follow the wait operation with a flush of outstanding remote writes. This means that, if a remote write operation is guaranteed to have reached the device before the wait can be satisfied, that write is guaranteed to be visible to downstream device work. The device is permitted to reorder remote writes internally. For example, this flag would be required if two remote writes arrive in a defined order, the wait is satisfied by the - second write, and downstream work needs to observe the first write. */ + second write, and downstream work needs to observe the first write. + Support for this operation is restricted to selected platforms and can be + queried with ::CU_DEVICE_ATTRIBUTE_CAN_USE_WAIT_VALUE_FLUSH.*/ } CUstreamWaitValue_flags; /** @@ -372,6 +417,8 @@ typedef enum CUstreamWriteValue_flags_enum { typedef enum CUstreamBatchMemOpType_enum { CU_STREAM_MEM_OP_WAIT_VALUE_32 = 1, /**< Represents a ::cuStreamWaitValue32 operation */ CU_STREAM_MEM_OP_WRITE_VALUE_32 = 2, /**< Represents a ::cuStreamWriteValue32 operation */ + CU_STREAM_MEM_OP_WAIT_VALUE_64 = 4, /**< Represents a ::cuStreamWaitValue64 operation */ + CU_STREAM_MEM_OP_WRITE_VALUE_64 = 5, /**< Represents a ::cuStreamWriteValue64 operation */ CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES = 3 /**< This has the same effect as ::CU_STREAM_WAIT_VALUE_FLUSH, but as a standalone operation. */ } CUstreamBatchMemOpType; @@ -386,7 +433,7 @@ typedef union CUstreamBatchMemOpParams_union { CUdeviceptr address; union { cuuint32_t value; - cuuint64_t pad; + cuuint64_t value64; }; unsigned int flags; CUdeviceptr alias; /**< For driver internal use. Initial value is unimportant. */ @@ -396,7 +443,7 @@ typedef union CUstreamBatchMemOpParams_union { CUdeviceptr address; union { cuuint32_t value; - cuuint64_t pad; + cuuint64_t value64; }; unsigned int flags; CUdeviceptr alias; /**< For driver internal use. Initial value is unimportant. */ @@ -498,7 +545,7 @@ typedef enum CUdevice_attribute_enum { CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE = 38, /**< Size of L2 cache in bytes */ CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR = 39, /**< Maximum resident threads per multiprocessor */ CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT = 40, /**< Number of asynchronous engines */ - CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING = 41, /**< Device shares a unified address space with the host */ + CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING = 41, /**< Device shares a unified address space with the host */ CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH = 42, /**< Maximum 1D layered texture width */ CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS = 43, /**< Maximum layers in a 1D layered texture */ CU_DEVICE_ATTRIBUTE_CAN_TEX2D_GATHER = 44, /**< Deprecated, do not use. */ @@ -532,7 +579,7 @@ typedef enum CUdevice_attribute_enum { CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH = 72, /**< Maximum 2D linear texture pitch in bytes */ CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH = 73, /**< Maximum mipmapped 2D texture width */ CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT = 74,/**< Maximum mipmapped 2D texture height */ - CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75, /**< Major compute capability version number */ + CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75, /**< Major compute capability version number */ CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76, /**< Minor compute capability version number */ CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH = 77, /**< Maximum mipmapped 1D texture width */ CU_DEVICE_ATTRIBUTE_STREAM_PRIORITIES_SUPPORTED = 78, /**< Device supports stream priorities */ @@ -541,7 +588,7 @@ typedef enum CUdevice_attribute_enum { CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR = 81, /**< Maximum shared memory available per multiprocessor in bytes */ CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR = 82, /**< Maximum number of 32-bit registers available per multiprocessor */ CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY = 83, /**< Device can allocate managed memory on this system */ - CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD = 84, /**< Device is on a multi-GPU board */ + CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD = 84, /**< Device is on a multi-GPU board */ CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD_GROUP_ID = 85, /**< Unique id for a group of devices on the same multi-GPU board */ CU_DEVICE_ATTRIBUTE_HOST_NATIVE_ATOMIC_SUPPORTED = 86, /**< Link between the device and the host supports native atomic operations (this is a placeholder attribute, and is not supported on any current hardware)*/ CU_DEVICE_ATTRIBUTE_SINGLE_TO_DOUBLE_PRECISION_PERF_RATIO = 87, /**< Ratio of single precision performance (in floating-point operations per second) to double precision performance */ @@ -549,6 +596,16 @@ typedef enum CUdevice_attribute_enum { CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS = 89, /**< Device can coherently access managed memory concurrently with the CPU */ CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED = 90, /**< Device supports compute preemption. */ CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM = 91, /**< Device can access host registered memory at the same virtual address as the CPU */ + CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS = 92, /**< ::cuStreamBatchMemOp and related APIs are supported. */ + CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS = 93, /**< 64-bit operations are supported in ::cuStreamBatchMemOp and related APIs. */ + CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR = 94, /**< ::CU_STREAM_WAIT_VALUE_NOR is supported. */ + CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH = 95, /**< Device supports launching cooperative kernels via ::cuLaunchCooperativeKernel */ + CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH = 96, /**< Device can participate in cooperative kernels launched via ::cuLaunchCooperativeKernelMultiDevice */ + CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN = 97, /**< Maximum optin shared memory per block */ + CU_DEVICE_ATTRIBUTE_CAN_FLUSH_REMOTE_WRITES = 98, /**< Both the ::CU_STREAM_WAIT_VALUE_FLUSH flag and the ::CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES MemOp are supported on the device. See \ref CUDA_MEMOP for additional details. */ + CU_DEVICE_ATTRIBUTE_HOST_REGISTER_SUPPORTED = 99, /**< Device supports host memory registration via ::cudaHostRegister. */ + CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES = 100, /**< Device accesses pageable memory via the host's page tables. */ + CU_DEVICE_ATTRIBUTE_DIRECT_MANAGED_MEM_ACCESS_FROM_HOST = 101, /**< The host can directly access managed memory on the device without migration. */ CU_DEVICE_ATTRIBUTE_MAX } CUdevice_attribute; @@ -579,7 +636,8 @@ typedef enum CUpointer_attribute_enum { CU_POINTER_ATTRIBUTE_P2P_TOKENS = 5, /**< A pair of tokens for use with the nv-p2p.h Linux kernel interface */ CU_POINTER_ATTRIBUTE_SYNC_MEMOPS = 6, /**< Synchronize every synchronous memory operation initiated on this region */ CU_POINTER_ATTRIBUTE_BUFFER_ID = 7, /**< A process-wide unique ID for an allocated memory region*/ - CU_POINTER_ATTRIBUTE_IS_MANAGED = 8 /**< Indicates if the pointer points to managed memory */ + CU_POINTER_ATTRIBUTE_IS_MANAGED = 8, /**< Indicates if the pointer points to managed memory */ + CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL = 9 /**< A device ordinal of a device on which a pointer was allocated or registered */ } CUpointer_attribute; /** @@ -635,11 +693,28 @@ typedef enum CUfunction_attribute_enum { CU_FUNC_ATTRIBUTE_BINARY_VERSION = 6, /** - * The attribute to indicate whether the function has been compiled with + * The attribute to indicate whether the function has been compiled with * user specified option "-Xptxas --dlcm=ca" set . */ CU_FUNC_ATTRIBUTE_CACHE_MODE_CA = 7, + /** + * The maximum size in bytes of dynamically-allocated shared memory that can be used by + * this function. If the user-specified dynamic shared memory size is larger than this + * value, the launch will fail. + * See ::cuFuncSetAttribute + */ + CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES = 8, + + /** + * On devices where the L1 cache and shared memory use the same hardware resources, + * this sets the shared memory carveout preference, in percent of the total shared memory. + * Refer to ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR. + * This is only a hint, and the driver can choose a different ratio if required to execute the function. + * See ::cuFuncSetAttribute + */ + CU_FUNC_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT = 9, + CU_FUNC_ATTRIBUTE_MAX } CUfunction_attribute; @@ -662,6 +737,15 @@ typedef enum CUsharedconfig_enum { CU_SHARED_MEM_CONFIG_EIGHT_BYTE_BANK_SIZE = 0x02 /**< set shared memory bank width to eight bytes */ } CUsharedconfig; +/** + * Shared memory carveout configurations. These may be passed to ::cuFuncSetAttribute + */ +typedef enum CUshared_carveout_enum { + CU_SHAREDMEM_CARVEOUT_DEFAULT = -1, /**< No preference for shared memory or L1 (default) */ + CU_SHAREDMEM_CARVEOUT_MAX_SHARED = 100, /**< Prefer maximum available shared memory, minimum L1 cache */ + CU_SHAREDMEM_CARVEOUT_MAX_L1 = 0 /**< Prefer maximum available L1 cache, minimum shared memory */ +} CUshared_carveout; + /** * Memory types */ @@ -840,6 +924,37 @@ typedef enum CUjit_option_enum CU_JIT_NEW_SM3X_OPT, CU_JIT_FAST_COMPILE, + /** + * Array of device symbol names that will be relocated to the corresponing + * host addresses stored in ::CU_JIT_GLOBAL_SYMBOL_ADDRESSES.\n + * Must contain ::CU_JIT_GLOBAL_SYMBOL_COUNT entries.\n + * When loding a device module, driver will relocate all encountered + * unresolved symbols to the host addresses.\n + * It is only allowed to register symbols that correspond to unresolved + * global variables.\n + * It is illegal to register the same device symbol at multiple addresses.\n + * Option type: const char **\n + * Applies to: dynamic linker only + */ + CU_JIT_GLOBAL_SYMBOL_NAMES, + + /** + * Array of host addresses that will be used to relocate corresponding + * device symbols stored in ::CU_JIT_GLOBAL_SYMBOL_NAMES.\n + * Must contain ::CU_JIT_GLOBAL_SYMBOL_COUNT entries.\n + * Option type: void **\n + * Applies to: dynamic linker only + */ + CU_JIT_GLOBAL_SYMBOL_ADDRESSES, + + /** + * Number of entries in ::CU_JIT_GLOBAL_SYMBOL_NAMES and + * ::CU_JIT_GLOBAL_SYMBOL_ADDRESSES arrays.\n + * Option type: unsigned int\n + * Applies to: dynamic linker only + */ + CU_JIT_GLOBAL_SYMBOL_COUNT, + CU_JIT_NUM_OPTIONS } CUjit_option; @@ -849,10 +964,6 @@ typedef enum CUjit_option_enum */ typedef enum CUjit_target_enum { - CU_TARGET_COMPUTE_10 = 10, /**< Compute device class 1.0 */ - CU_TARGET_COMPUTE_11 = 11, /**< Compute device class 1.1 */ - CU_TARGET_COMPUTE_12 = 12, /**< Compute device class 1.2 */ - CU_TARGET_COMPUTE_13 = 13, /**< Compute device class 1.3 */ CU_TARGET_COMPUTE_20 = 20, /**< Compute device class 2.0 */ CU_TARGET_COMPUTE_21 = 21, /**< Compute device class 2.1 */ CU_TARGET_COMPUTE_30 = 30, /**< Compute device class 3.0 */ @@ -862,9 +973,12 @@ typedef enum CUjit_target_enum CU_TARGET_COMPUTE_50 = 50, /**< Compute device class 5.0 */ CU_TARGET_COMPUTE_52 = 52, /**< Compute device class 5.2 */ CU_TARGET_COMPUTE_53 = 53, /**< Compute device class 5.3 */ - CU_TARGET_COMPUTE_60 = 60, /**< Compute device class 6.0. This must be removed for CUDA 7.0 toolkit. See bug 1518217. */ - CU_TARGET_COMPUTE_61 = 61, /**< Compute device class 6.1. This must be removed for CUDA 7.0 toolkit.*/ - CU_TARGET_COMPUTE_62 = 62 /**< Compute device class 6.2. This must be removed for CUDA 7.0 toolkit.*/ + CU_TARGET_COMPUTE_60 = 60, /**< Compute device class 6.0.*/ + CU_TARGET_COMPUTE_61 = 61, /**< Compute device class 6.1.*/ + CU_TARGET_COMPUTE_62 = 62, /**< Compute device class 6.2.*/ + CU_TARGET_COMPUTE_70 = 70, /**< Compute device class 7.0.*/ + CU_TARGET_COMPUTE_72 = 72, /**< Compute device class 7.2.*/ + CU_TARGET_COMPUTE_75 = 75 /**< Compute device class 7.5.*/ } CUjit_target; /** @@ -879,7 +993,7 @@ typedef enum CUjit_fallback_enum } CUjit_fallback; /** - * Caching modes for dlcm + * Caching modes for dlcm */ typedef enum CUjit_cacheMode_enum { @@ -971,6 +1085,7 @@ typedef enum CUlimit_enum { CU_LIMIT_MALLOC_HEAP_SIZE = 0x02, /**< GPU malloc heap size */ CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH = 0x03, /**< GPU device runtime launch synchronize depth */ CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT = 0x04, /**< GPU device runtime pending launch count */ + CU_LIMIT_MAX_L2_FETCH_GRANULARITY = 0x05, /**< A value between 0 and 128 that indicates the maximum fetch granularity of L2 (in Bytes). This is a hint */ CU_LIMIT_MAX } CUlimit; @@ -984,13 +1099,102 @@ typedef enum CUresourcetype_enum { CU_RESOURCE_TYPE_PITCH2D = 0x03 /**< Pitch 2D resource */ } CUresourcetype; +#ifdef _WIN32 +#define CUDA_CB __stdcall +#else +#define CUDA_CB +#endif + +#if __CUDA_API_VERSION >= 10000 + +/** + * CUDA host function + * \param userData Argument value passed to the function + */ +typedef void (CUDA_CB *CUhostFn)(void *userData); + +/** + * GPU kernel node parameters + */ +typedef struct CUDA_KERNEL_NODE_PARAMS_st { + CUfunction func; /**< Kernel to launch */ + unsigned int gridDimX; /**< Width of grid in blocks */ + unsigned int gridDimY; /**< Height of grid in blocks */ + unsigned int gridDimZ; /**< Depth of grid in blocks */ + unsigned int blockDimX; /**< X dimension of each thread block */ + unsigned int blockDimY; /**< Y dimension of each thread block */ + unsigned int blockDimZ; /**< Z dimension of each thread block */ + unsigned int sharedMemBytes; /**< Dynamic shared-memory size per thread block in bytes */ + void **kernelParams; /**< Array of pointers to kernel parameters */ + void **extra; /**< Extra options */ +} CUDA_KERNEL_NODE_PARAMS; + +/** + * Memset node parameters + */ +typedef struct CUDA_MEMSET_NODE_PARAMS_st { + CUdeviceptr dst; /**< Destination device pointer */ + size_t pitch; /**< Pitch of destination device pointer. Unused if height is 1 */ + unsigned int value; /**< Value to be set */ + unsigned int elementSize; /**< Size of each element in bytes. Must be 1, 2, or 4. */ + size_t width; /**< Width in bytes, of the row */ + size_t height; /**< Number of rows */ +} CUDA_MEMSET_NODE_PARAMS; + +/** + * Host node parameters + */ +typedef struct CUDA_HOST_NODE_PARAMS_st { + CUhostFn fn; /**< The function to call when the node executes */ + void* userData; /**< Argument to pass to the function */ +} CUDA_HOST_NODE_PARAMS; + +/** + * Graph node types + */ +typedef enum CUgraphNodeType_enum { + CU_GRAPH_NODE_TYPE_KERNEL = 0, /**< GPU kernel node */ + CU_GRAPH_NODE_TYPE_MEMCPY = 1, /**< Memcpy node */ + CU_GRAPH_NODE_TYPE_MEMSET = 2, /**< Memset node */ + CU_GRAPH_NODE_TYPE_HOST = 3, /**< Host (executable) node */ + CU_GRAPH_NODE_TYPE_GRAPH = 4, /**< Node which executes an embedded graph */ + CU_GRAPH_NODE_TYPE_EMPTY = 5, /**< Empty (no-op) node */ + CU_GRAPH_NODE_TYPE_COUNT +} CUgraphNodeType; + +/** + * Possible stream capture statuses returned by ::cuStreamIsCapturing + */ +typedef enum CUstreamCaptureStatus_enum { + CU_STREAM_CAPTURE_STATUS_NONE = 0, /**< Stream is not capturing */ + CU_STREAM_CAPTURE_STATUS_ACTIVE = 1, /**< Stream is actively capturing */ + CU_STREAM_CAPTURE_STATUS_INVALIDATED = 2 /**< Stream is part of a capture sequence that + has been invalidated, but not terminated */ +} CUstreamCaptureStatus; + +#endif /* __CUDA_API_VERSION >= 10000 */ + +#if __CUDA_API_VERSION >= 10010 + +/** + * Possible modes for stream capture thread interactions. For more details see + * ::cuStreamBeginCapture and ::cuThreadExchangeStreamCaptureMode + */ +typedef enum CUstreamCaptureMode_enum { + CU_STREAM_CAPTURE_MODE_GLOBAL = 0, + CU_STREAM_CAPTURE_MODE_THREAD_LOCAL = 1, + CU_STREAM_CAPTURE_MODE_RELAXED = 2 +} CUstreamCaptureMode; + +#endif /* __CUDA_API_VERSION >= 10010 */ + /** * Error codes */ typedef enum cudaError_enum { /** * The API call returned with no errors. In the case of query calls, this - * can also mean that the operation being queried is complete (see + * also means that the operation being queried is complete (see * ::cuEventQuery() and ::cuStreamQuery()). */ CUDA_SUCCESS = 0, @@ -1150,7 +1354,7 @@ typedef enum cudaError_enum { /** * This indicates that the ::CUcontext passed to the API call can - * only be bound to a single CPU thread at a time but is already + * only be bound to a single CPU thread at a time but is already * bound to a CPU thread. */ CUDA_ERROR_CONTEXT_ALREADY_IN_USE = 216, @@ -1177,6 +1381,11 @@ typedef enum cudaError_enum { */ CUDA_ERROR_NVLINK_UNCORRECTABLE = 220, + /** + * This indicates that the PTX JIT compiler library was not found. + */ + CUDA_ERROR_JIT_COMPILER_NOT_FOUND = 221, + /** * This indicates that the device kernel source is invalid. */ @@ -1208,6 +1417,12 @@ typedef enum cudaError_enum { */ CUDA_ERROR_INVALID_HANDLE = 400, + /** + * This indicates that a resource required by the API call is not in a + * valid state to perform the requested operation. + */ + CUDA_ERROR_ILLEGAL_STATE = 401, + /** * This indicates that a named symbol was not found. Examples of symbols * are global/constant variable names, texture names, and surface names. @@ -1257,7 +1472,7 @@ typedef enum cudaError_enum { * mode. */ CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING = 703, - + /** * This error indicates that a call to ::cuCtxEnablePeerAccess() is * trying to re-enable peer access to a context which has already @@ -1266,9 +1481,9 @@ typedef enum cudaError_enum { CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED = 704, /** - * This error indicates that ::cuCtxDisablePeerAccess() is - * trying to disable peer access which has not been enabled yet - * via ::cuCtxEnablePeerAccess(). + * This error indicates that ::cuCtxDisablePeerAccess() is + * trying to disable peer access which has not been enabled yet + * via ::cuCtxEnablePeerAccess(). */ CUDA_ERROR_PEER_ACCESS_NOT_ENABLED = 705, @@ -1287,15 +1502,15 @@ typedef enum cudaError_enum { /** * A device-side assert triggered during kernel execution. The context - * cannot be used anymore, and must be destroyed. All existing device - * memory allocations from this context are invalid and must be + * cannot be used anymore, and must be destroyed. All existing device + * memory allocations from this context are invalid and must be * reconstructed if the program is to continue using CUDA. */ CUDA_ERROR_ASSERT = 710, /** * This error indicates that the hardware resources required to enable - * peer access have been exhausted for one or more of the devices + * peer access have been exhausted for one or more of the devices * passed to ::cuCtxEnablePeerAccess(). */ CUDA_ERROR_TOO_MANY_PEERS = 711, @@ -1360,13 +1575,22 @@ typedef enum cudaError_enum { /** * An exception occurred on the device while executing a kernel. Common * causes include dereferencing an invalid device pointer and accessing - * out of bounds shared memory. + * out of bounds shared memory. Less common cases can be system specific - more + * information about these cases can be found in the system specific user guide. * This leaves the process in an inconsistent state and any further CUDA work * will return the same error. To continue using CUDA, the process must be terminated * and relaunched. */ CUDA_ERROR_LAUNCH_FAILED = 719, + /** + * This error indicates that the number of blocks launched per grid for a kernel that was + * launched via either ::cuLaunchCooperativeKernel or ::cuLaunchCooperativeKernelMultiDevice + * exceeds the maximum number of blocks as allowed by ::cuOccupancyMaxActiveBlocksPerMultiprocessor + * or ::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags times the number of multiprocessors + * as specified by the device attribute ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT. + */ + CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE = 720, /** * This error indicates that the attempted operation is not permitted. @@ -1379,6 +1603,86 @@ typedef enum cudaError_enum { */ CUDA_ERROR_NOT_SUPPORTED = 801, + /** + * This error indicates that the system is not yet ready to start any CUDA + * work. To continue using CUDA, verify the system configuration is in a + * valid state and all required driver daemons are actively running. + * More information about this error can be found in the system specific + * user guide. + */ + CUDA_ERROR_SYSTEM_NOT_READY = 802, + + /** + * This error indicates that there is a mismatch between the versions of + * the display driver and the CUDA driver. Refer to the compatibility documentation + * for supported versions. + */ + CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803, + + /** + * This error indicates that the system was upgraded to run with forward compatibility + * but the visible hardware detected by CUDA does not support this configuration. + * Refer to the compatibility documentation for the supported hardware matrix or ensure + * that only supported hardware is visible during initialization via the CUDA_VISIBLE_DEVICES + * environment variable. + */ + CUDA_ERROR_COMPAT_NOT_SUPPORTED_ON_DEVICE = 804, + + /** + * This error indicates that the operation is not permitted when + * the stream is capturing. + */ + CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED = 900, + + /** + * This error indicates that the current capture sequence on the stream + * has been invalidated due to a previous error. + */ + CUDA_ERROR_STREAM_CAPTURE_INVALIDATED = 901, + + /** + * This error indicates that the operation would have resulted in a merge + * of two independent capture sequences. + */ + CUDA_ERROR_STREAM_CAPTURE_MERGE = 902, + + /** + * This error indicates that the capture was not initiated in this stream. + */ + CUDA_ERROR_STREAM_CAPTURE_UNMATCHED = 903, + + /** + * This error indicates that the capture sequence contains a fork that was + * not joined to the primary stream. + */ + CUDA_ERROR_STREAM_CAPTURE_UNJOINED = 904, + + /** + * This error indicates that a dependency would have been created which + * crosses the capture sequence boundary. Only implicit in-stream ordering + * dependencies are allowed to cross the boundary. + */ + CUDA_ERROR_STREAM_CAPTURE_ISOLATION = 905, + + /** + * This error indicates a disallowed implicit dependency on a current capture + * sequence from cudaStreamLegacy. + */ + CUDA_ERROR_STREAM_CAPTURE_IMPLICIT = 906, + + /** + * This error indicates that the operation is not permitted on an event which + * was last recorded in a capturing stream. + */ + CUDA_ERROR_CAPTURED_EVENT = 907, + + /** + * A stream capture sequence not initiated with the ::CU_STREAM_CAPTURE_MODE_RELAXED + * argument to ::cuStreamBeginCapture was passed to ::cuStreamEndCapture in a + * different thread. + */ + CUDA_ERROR_STREAM_CAPTURE_WRONG_THREAD = 908, + /** * This indicates that an unknown internal error has occurred. */ @@ -1389,17 +1693,13 @@ typedef enum cudaError_enum { * P2P Attributes */ typedef enum CUdevice_P2PAttribute_enum { - CU_DEVICE_P2P_ATTRIBUTE_PERFORMANCE_RANK = 0x01, /**< A relative value indicating the performance of the link between two devices */ - CU_DEVICE_P2P_ATTRIBUTE_ACCESS_SUPPORTED = 0x02, /**< P2P Access is enable */ - CU_DEVICE_P2P_ATTRIBUTE_NATIVE_ATOMIC_SUPPORTED = 0x03 /**< Atomic operation over the link supported */ + CU_DEVICE_P2P_ATTRIBUTE_PERFORMANCE_RANK = 0x01, /**< A relative value indicating the performance of the link between two devices */ + CU_DEVICE_P2P_ATTRIBUTE_ACCESS_SUPPORTED = 0x02, /**< P2P Access is enable */ + CU_DEVICE_P2P_ATTRIBUTE_NATIVE_ATOMIC_SUPPORTED = 0x03, /**< Atomic operation over the link supported */ + CU_DEVICE_P2P_ATTRIBUTE_ACCESS_ACCESS_SUPPORTED = 0x04, /**< \deprecated use CU_DEVICE_P2P_ATTRIBUTE_CUDA_ARRAY_ACCESS_SUPPORTED instead */ + CU_DEVICE_P2P_ATTRIBUTE_CUDA_ARRAY_ACCESS_SUPPORTED = 0x04 /**< Accessing CUDA arrays over the link supported */ } CUdevice_P2PAttribute; -#ifdef _WIN32 -#define CUDA_CB __stdcall -#else -#define CUDA_CB -#endif - /** * CUDA stream callback * \param hStream The stream the callback was added to, as passed to ::cuStreamAddCallback. May be NULL. @@ -1635,7 +1935,7 @@ typedef struct CUDA_TEXTURE_DESC_st { CUfilter_mode mipmapFilterMode; /**< Mipmap filter mode */ float mipmapLevelBias; /**< Mipmap level bias */ float minMipmapLevelClamp; /**< Mipmap minimum level clamp */ - float maxMipmapLevelClamp; /**< Mipmap maximum level clamp */ + float maxMipmapLevelClamp; /**< Mipmap maximum level clamp */ float borderColor[4]; /**< Border Color */ int reserved[12]; } CUDA_TEXTURE_DESC; @@ -1708,9 +2008,281 @@ typedef struct CUDA_POINTER_ATTRIBUTE_P2P_TOKENS_st { #endif /* __CUDA_API_VERSION >= 5000 */ +#if __CUDA_API_VERSION >= 9000 + +/** + * Kernel launch parameters + */ +typedef struct CUDA_LAUNCH_PARAMS_st { + CUfunction function; /**< Kernel to launch */ + unsigned int gridDimX; /**< Width of grid in blocks */ + unsigned int gridDimY; /**< Height of grid in blocks */ + unsigned int gridDimZ; /**< Depth of grid in blocks */ + unsigned int blockDimX; /**< X dimension of each thread block */ + unsigned int blockDimY; /**< Y dimension of each thread block */ + unsigned int blockDimZ; /**< Z dimension of each thread block */ + unsigned int sharedMemBytes; /**< Dynamic shared-memory size per thread block in bytes */ + CUstream hStream; /**< Stream identifier */ + void **kernelParams; /**< Array of pointers to kernel parameters */ +} CUDA_LAUNCH_PARAMS; + +#endif /* __CUDA_API_VERSION >= 9000 */ + +#if __CUDA_API_VERSION >= 10000 + +/** + * External memory handle types + */ +typedef enum CUexternalMemoryHandleType_enum { + /** + * Handle is an opaque file descriptor + */ + CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD = 1, + /** + * Handle is an opaque shared NT handle + */ + CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 = 2, + /** + * Handle is an opaque, globally shared handle + */ + CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3, + /** + * Handle is a D3D12 heap object + */ + CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP = 4, + /** + * Handle is a D3D12 committed resource + */ + CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE = 5 +} CUexternalMemoryHandleType; + +/** + * Indicates that the external memory object is a dedicated resource + */ +#define CUDA_EXTERNAL_MEMORY_DEDICATED 0x1 + +/** + * External memory handle descriptor + */ +typedef struct CUDA_EXTERNAL_MEMORY_HANDLE_DESC_st { + /** + * Type of the handle + */ + CUexternalMemoryHandleType type; + union { + /** + * File descriptor referencing the memory object. Valid + * when type is + * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD + */ + int fd; + /** + * Win32 handle referencing the semaphore object. Valid when + * type is one of the following: + * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 + * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT + * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP + * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE + * Exactly one of 'handle' and 'name' must be non-NULL. If + * type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT + * then 'name' must be NULL. + */ + struct { + /** + * Valid NT handle. Must be NULL if 'name' is non-NULL + */ + void *handle; + /** + * Name of a valid memory object. + * Must be NULL if 'handle' is non-NULL. + */ + const void *name; + } win32; + } handle; + /** + * Size of the memory allocation + */ + unsigned long long size; + /** + * Flags must either be zero or ::CUDA_EXTERNAL_MEMORY_DEDICATED + */ + unsigned int flags; + unsigned int reserved[16]; +} CUDA_EXTERNAL_MEMORY_HANDLE_DESC; + +/** + * External memory buffer descriptor + */ +typedef struct CUDA_EXTERNAL_MEMORY_BUFFER_DESC_st { + /** + * Offset into the memory object where the buffer's base is + */ + unsigned long long offset; + /** + * Size of the buffer + */ + unsigned long long size; + /** + * Flags reserved for future use. Must be zero. + */ + unsigned int flags; + unsigned int reserved[16]; +} CUDA_EXTERNAL_MEMORY_BUFFER_DESC; + +/** + * External memory mipmap descriptor + */ +typedef struct CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_st { + /** + * Offset into the memory object where the base level of the + * mipmap chain is. + */ + unsigned long long offset; + /** + * Format, dimension and type of base level of the mipmap chain + */ + CUDA_ARRAY3D_DESCRIPTOR arrayDesc; + /** + * Total number of levels in the mipmap chain + */ + unsigned int numLevels; + unsigned int reserved[16]; +} CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC; + +/** + * External semaphore handle types + */ +typedef enum CUexternalSemaphoreHandleType_enum { + /** + * Handle is an opaque file descriptor + */ + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD = 1, + /** + * Handle is an opaque shared NT handle + */ + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32 = 2, + /** + * Handle is an opaque, globally shared handle + */ + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3, + /** + * Handle is a shared NT handle referencing a D3D12 fence object + */ + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE = 4 +} CUexternalSemaphoreHandleType; + +/** + * External semaphore handle descriptor + */ +typedef struct CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_st { + /** + * Type of the handle + */ + CUexternalSemaphoreHandleType type; + union { + /** + * File descriptor referencing the semaphore object. Valid + * when type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD + */ + int fd; + /** + * Win32 handle referencing the semaphore object. Valid when + * type is one of the following: + * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32 + * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT + * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE + * Exactly one of 'handle' and 'name' must be non-NULL. If + * type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT + * then 'name' must be NULL. + */ + struct { + /** + * Valid NT handle. Must be NULL if 'name' is non-NULL + */ + void *handle; + /** + * Name of a valid synchronization primitive. + * Must be NULL if 'handle' is non-NULL. + */ + const void *name; + } win32; + } handle; + /** + * Flags reserved for the future. Must be zero. + */ + unsigned int flags; + unsigned int reserved[16]; +} CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC; + +/** + * External semaphore signal parameters + */ +typedef struct CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_st { + struct { + /** + * Parameters for fence objects + */ + struct { + /** + * Value of fence to be signaled + */ + unsigned long long value; + } fence; + unsigned int reserved[16]; + } params; + /** + * Flags reserved for the future. Must be zero. + */ + unsigned int flags; + unsigned int reserved[16]; +} CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS; + +/** + * External semaphore wait parameters + */ +typedef struct CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS_st { + struct { + /** + * Parameters for fence objects + */ + struct { + /** + * Value of fence to be waited on + */ + unsigned long long value; + } fence; + unsigned int reserved[16]; + } params; + /** + * Flags reserved for the future. Must be zero. + */ + unsigned int flags; + unsigned int reserved[16]; +} CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS; + + +#endif /* __CUDA_API_VERSION >= 10000 */ + +/** + * If set, each kernel launched as part of ::cuLaunchCooperativeKernelMultiDevice only + * waits for prior work in the stream corresponding to that GPU to complete before the + * kernel begins execution. + */ +#define CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC 0x01 + +/** + * If set, any subsequent work pushed in a stream that participated in a call to + * ::cuLaunchCooperativeKernelMultiDevice will only wait for the kernel launched on + * the GPU corresponding to that stream to complete before it begins execution. + */ +#define CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC 0x02 + /** * If set, the CUDA array is a collection of layers, where each layer is either a 1D - * or a 2D array and the Depth member of CUDA_ARRAY3D_DESCRIPTOR specifies the number + * or a 2D array and the Depth member of CUDA_ARRAY3D_DESCRIPTOR specifies the number * of layers, not the depth of a 3D array. */ #define CUDA_ARRAY3D_LAYERED 0x01 @@ -1743,9 +2315,15 @@ typedef struct CUDA_POINTER_ATTRIBUTE_P2P_TOKENS_st { /** * This flag if set indicates that the CUDA * array is a DEPTH_TEXTURE. -*/ + */ #define CUDA_ARRAY3D_DEPTH_TEXTURE 0x10 +/** + * This flag indicates that the CUDA array may be bound as a color target + * in an external graphics API + */ +#define CUDA_ARRAY3D_COLOR_ATTACHMENT 0x20 + /** * Override the texref format with a format inferred from the array. * Flag for ::cuTexRefSetArray() @@ -1849,7 +2427,9 @@ typedef struct CUDA_POINTER_ATTRIBUTE_P2P_TOKENS_st { * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::CUresult + * \sa + * ::CUresult, + * ::cudaGetErrorString */ CUresult CUDAAPI cuGetErrorString(CUresult error, const char **pStr); @@ -1868,7 +2448,9 @@ CUresult CUDAAPI cuGetErrorString(CUresult error, const char **pStr); * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::CUresult + * \sa + * ::CUresult, + * ::cudaGetErrorName */ CUresult CUDAAPI cuGetErrorName(CUresult error, const char **pStr); @@ -1899,7 +2481,9 @@ CUresult CUDAAPI cuGetErrorName(CUresult error, const char **pStr); * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_INVALID_DEVICE + * ::CUDA_ERROR_INVALID_DEVICE, + * ::CUDA_ERROR_SYSTEM_DRIVER_MISMATCH, + * ::CUDA_ERROR_COMPAT_NOT_SUPPORTED_ON_DEVICE * \notefnerr */ CUresult CUDAAPI cuInit(unsigned int Flags); @@ -1919,11 +2503,15 @@ CUresult CUDAAPI cuInit(unsigned int Flags); */ /** - * \brief Returns the CUDA driver version + * \brief Returns the latest CUDA version supported by driver + * + * Returns in \p *driverVersion the version of CUDA supported by + * the driver. The version is returned as + * (1000 × major + 10 × minor). For example, CUDA 9.2 + * would be represented by 9020. * - * Returns in \p *driverVersion the version number of the installed CUDA - * driver. This function automatically returns ::CUDA_ERROR_INVALID_VALUE if - * the \p driverVersion argument is NULL. + * This function automatically returns ::CUDA_ERROR_INVALID_VALUE if + * \p driverVersion is NULL. * * \param driverVersion - Returns the CUDA driver version * @@ -1931,6 +2519,10 @@ CUresult CUDAAPI cuInit(unsigned int Flags); * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_VALUE * \notefnerr + * + * \sa + * ::cudaDriverGetVersion, + * ::cudaRuntimeGetVersion */ CUresult CUDAAPI cuDriverGetVersion(int *driverVersion); @@ -1970,6 +2562,8 @@ CUresult CUDAAPI cuDriverGetVersion(int *driverVersion); * ::cuDeviceGetAttribute, * ::cuDeviceGetCount, * ::cuDeviceGetName, + * ::cuDeviceGetUuid, + * ::cuDeviceGetLuid, * ::cuDeviceTotalMem */ CUresult CUDAAPI cuDeviceGet(CUdevice *device, int ordinal); @@ -1978,7 +2572,7 @@ CUresult CUDAAPI cuDeviceGet(CUdevice *device, int ordinal); * \brief Returns the number of compute-capable devices * * Returns in \p *count the number of devices with compute capability greater - * than or equal to 1.0 that are available for execution. If there is no such + * than or equal to 2.0 that are available for execution. If there is no such * device, ::cuDeviceGetCount() returns 0. * * \param count - Returned number of compute-capable devices @@ -1994,8 +2588,11 @@ CUresult CUDAAPI cuDeviceGet(CUdevice *device, int ordinal); * \sa * ::cuDeviceGetAttribute, * ::cuDeviceGetName, + * ::cuDeviceGetUuid, + * ::cuDeviceGetLuid, * ::cuDeviceGet, - * ::cuDeviceTotalMem + * ::cuDeviceTotalMem, + * ::cudaGetDeviceCount */ CUresult CUDAAPI cuDeviceGetCount(int *count); @@ -2021,27 +2618,29 @@ CUresult CUDAAPI cuDeviceGetCount(int *count); * * \sa * ::cuDeviceGetAttribute, + * ::cuDeviceGetUuid, + * ::cuDeviceGetLuid, * ::cuDeviceGetCount, * ::cuDeviceGet, - * ::cuDeviceTotalMem + * ::cuDeviceTotalMem, + * ::cudaGetDeviceProperties */ CUresult CUDAAPI cuDeviceGetName(char *name, int len, CUdevice dev); -#if __CUDA_API_VERSION >= 3020 +#if __CUDA_API_VERSION >= 9020 /** - * \brief Returns the total amount of memory on the device + * \brief Return an UUID for the device * - * Returns in \p *bytes the total amount of memory available on the device - * \p dev in bytes. + * Returns 16-octets identifing the device \p dev in the structure + * pointed by the \p uuid. * - * \param bytes - Returned memory available on device in bytes - * \param dev - Device handle + * \param uuid - Returned UUID + * \param dev - Device to get identifier string for * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE, * ::CUDA_ERROR_INVALID_DEVICE * \notefnerr @@ -2050,41 +2649,104 @@ CUresult CUDAAPI cuDeviceGetName(char *name, int len, CUdevice dev); * ::cuDeviceGetAttribute, * ::cuDeviceGetCount, * ::cuDeviceGetName, + * ::cuDeviceGetLuid, * ::cuDeviceGet, + * ::cuDeviceTotalMem, + * ::cudaGetDeviceProperties */ -CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev); -#endif /* __CUDA_API_VERSION >= 3020 */ +CUresult CUDAAPI cuDeviceGetUuid(CUuuid *uuid, CUdevice dev); +#endif +#if defined(_WIN32) && __CUDA_API_VERSION >= 10000 /** - * \brief Returns information about the device + * \brief Return an LUID and device node mask for the device * - * Returns in \p *pi the integer value of the attribute \p attrib on device - * \p dev. The supported attributes are: - * - ::CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK: Maximum number of threads per - * block; - * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X: Maximum x-dimension of a block; - * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y: Maximum y-dimension of a block; - * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z: Maximum z-dimension of a block; - * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X: Maximum x-dimension of a grid; - * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y: Maximum y-dimension of a grid; - * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z: Maximum z-dimension of a grid; - * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK: Maximum amount of - * shared memory available to a thread block in bytes; - * - ::CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY: Memory available on device for - * __constant__ variables in a CUDA C kernel in bytes; - * - ::CU_DEVICE_ATTRIBUTE_WARP_SIZE: Warp size in threads; - * - ::CU_DEVICE_ATTRIBUTE_MAX_PITCH: Maximum pitch in bytes allowed by the - * memory copy functions that involve memory regions allocated through - * ::cuMemAllocPitch(); - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH: Maximum 1D - * texture width; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH: Maximum width - * for a 1D texture bound to linear memory; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH: Maximum - * mipmapped 1D texture width; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_WIDTH: Maximum 2D + * Return identifying information (\p luid and \p deviceNodeMask) to allow + * matching device with graphics APIs. + * + * \param luid - Returned LUID + * \param deviceNodeMask - Returned device node mask + * \param dev - Device to get identifier string for + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_DEVICE + * \notefnerr + * + * \sa + * ::cuDeviceGetAttribute, + * ::cuDeviceGetCount, + * ::cuDeviceGetName, + * ::cuDeviceGet, + * ::cuDeviceTotalMem, + * ::cudaGetDeviceProperties + */ +CUresult CUDAAPI cuDeviceGetLuid(char *luid, unsigned int *deviceNodeMask, CUdevice dev); +#endif + +#if __CUDA_API_VERSION >= 3020 +/** + * \brief Returns the total amount of memory on the device + * + * Returns in \p *bytes the total amount of memory available on the device + * \p dev in bytes. + * + * \param bytes - Returned memory available on device in bytes + * \param dev - Device handle + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_DEVICE + * \notefnerr + * + * \sa + * ::cuDeviceGetAttribute, + * ::cuDeviceGetCount, + * ::cuDeviceGetName, + * ::cuDeviceGetUuid, + * ::cuDeviceGet, + * ::cudaMemGetInfo + */ +CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev); +#endif /* __CUDA_API_VERSION >= 3020 */ + +/** + * \brief Returns information about the device + * + * Returns in \p *pi the integer value of the attribute \p attrib on device + * \p dev. The supported attributes are: + * - ::CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK: Maximum number of threads per + * block; + * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X: Maximum x-dimension of a block; + * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y: Maximum y-dimension of a block; + * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z: Maximum z-dimension of a block; + * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X: Maximum x-dimension of a grid; + * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y: Maximum y-dimension of a grid; + * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z: Maximum z-dimension of a grid; + * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK: Maximum amount of + * shared memory available to a thread block in bytes; + * - ::CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY: Memory available on device for + * __constant__ variables in a CUDA C kernel in bytes; + * - ::CU_DEVICE_ATTRIBUTE_WARP_SIZE: Warp size in threads; + * - ::CU_DEVICE_ATTRIBUTE_MAX_PITCH: Maximum pitch in bytes allowed by the + * memory copy functions that involve memory regions allocated through + * ::cuMemAllocPitch(); + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH: Maximum 1D + * texture width; + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH: Maximum width + * for a 1D texture bound to linear memory; + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH: Maximum + * mipmapped 1D texture width; + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_WIDTH: Maximum 2D * texture width; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_HEIGHT: Maximum 2D + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_HEIGHT: Maximum 2D * texture height; * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH: Maximum width * for a 2D texture bound to linear memory; @@ -2092,40 +2754,40 @@ CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev); * for a 2D texture bound to linear memory; * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH: Maximum pitch * in bytes for a 2D texture bound to linear memory; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH: Maximum + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH: Maximum * mipmapped 2D texture width; * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT: Maximum * mipmapped 2D texture height; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH: Maximum 3D + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH: Maximum 3D * texture width; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT: Maximum 3D + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT: Maximum 3D * texture height; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH: Maximum 3D + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH: Maximum 3D * texture depth; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE: * Alternate maximum 3D texture width, 0 if no alternate * maximum 3D texture size is supported; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE: * Alternate maximum 3D texture height, 0 if no alternate * maximum 3D texture size is supported; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE: * Alternate maximum 3D texture depth, 0 if no alternate * maximum 3D texture size is supported; * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_WIDTH: * Maximum cubemap texture width or height; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH: * Maximum 1D layered texture width; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS: * Maximum layers in a 1D layered texture; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH: * Maximum 2D layered texture width; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT: * Maximum 2D layered texture height; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS: * Maximum layers in a 2D layered texture; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH: * Maximum cubemap layered texture width or height; - * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS: + * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS: * Maximum layers in a cubemap layered texture; * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_WIDTH: * Maximum 1D surface width; @@ -2191,19 +2853,20 @@ CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev); * - ::CU_DEVICE_ATTRIBUTE_PCI_BUS_ID: PCI bus identifier of the device; * - ::CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID: PCI device (also known as slot) identifier * of the device; + * - ::CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID: PCI domain identifier of the device * - ::CU_DEVICE_ATTRIBUTE_TCC_DRIVER: 1 if the device is using a TCC driver. TCC * is only available on Tesla hardware running Windows Vista or later; * - ::CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE: Peak memory clock frequency in kilohertz; * - ::CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH: Global memory bus width in bits; * - ::CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE: Size of L2 cache in bytes. 0 if the device doesn't have L2 cache; * - ::CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR: Maximum resident threads per multiprocessor; - * - ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING: 1 if the device shares a unified address space with + * - ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING: 1 if the device shares a unified address space with * the host, or 0 if not; * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR: Major compute capability version number; * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR: Minor compute capability version number; - * - ::CU_DEVICE_ATTRIBUTE_GLOBAL_L1_CACHE_SUPPORTED: 1 if device supports caching globals + * - ::CU_DEVICE_ATTRIBUTE_GLOBAL_L1_CACHE_SUPPORTED: 1 if device supports caching globals * in L1 cache, 0 if caching globals in L1 cache is not supported by the device; - * - ::CU_DEVICE_ATTRIBUTE_LOCAL_L1_CACHE_SUPPORTED: 1 if device supports caching locals + * - ::CU_DEVICE_ATTRIBUTE_LOCAL_L1_CACHE_SUPPORTED: 1 if device supports caching locals * in L1 cache, 0 if caching locals in L1 cache is not supported by the device; * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR: Maximum amount of * shared memory available to a multiprocessor in bytes; this amount is shared @@ -2227,6 +2890,12 @@ CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev); * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED: Device supports Compute Preemption. * - ::CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM: Device can access host registered * memory at the same virtual address as the CPU. + * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN: The maximum per block shared memory size + * suported on this device. This is the maximum value that can be opted into when using the cuFuncSetAttribute() call. + * For more details see ::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES + * - ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES: Device accesses pageable memory via the host's + * page tables. + * - ::CU_DEVICE_ATTRIBUTE_DIRECT_MANAGED_MEM_ACCESS_FROM_HOST: The host can directly access managed memory on the device without migration. * * \param pi - Returned device attribute value * \param attrib - Device attribute to query @@ -2244,8 +2913,11 @@ CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev); * \sa * ::cuDeviceGetCount, * ::cuDeviceGetName, + * ::cuDeviceGetUuid, * ::cuDeviceGet, - * ::cuDeviceTotalMem + * ::cuDeviceTotalMem, + * ::cudaDeviceGetAttribute, + * ::cudaGetDeviceProperties */ CUresult CUDAAPI cuDeviceGetAttribute(int *pi, CUdevice_attribute attrib, CUdevice dev); @@ -2321,10 +2993,11 @@ CUresult CUDAAPI cuDeviceGetAttribute(int *pi, CUdevice_attribute attrib, CUdevi * ::cuDeviceGetAttribute, * ::cuDeviceGetCount, * ::cuDeviceGetName, + * ::cuDeviceGetUuid, * ::cuDeviceGet, * ::cuDeviceTotalMem */ -CUresult CUDAAPI cuDeviceGetProperties(CUdevprop *prop, CUdevice dev); +__CUDA_DEPRECATED CUresult CUDAAPI cuDeviceGetProperties(CUdevprop *prop, CUdevice dev); /** * \brief Returns the compute capability of the device @@ -2332,7 +3005,7 @@ CUresult CUDAAPI cuDeviceGetProperties(CUdevprop *prop, CUdevice dev); * \deprecated * * This function was deprecated as of CUDA 5.0 and its functionality superceded - * by ::cuDeviceGetAttribute(). + * by ::cuDeviceGetAttribute(). * * Returns in \p *major and \p *minor the major and minor revision numbers that * define the compute capability of the device \p dev. @@ -2354,10 +3027,11 @@ CUresult CUDAAPI cuDeviceGetProperties(CUdevprop *prop, CUdevice dev); * ::cuDeviceGetAttribute, * ::cuDeviceGetCount, * ::cuDeviceGetName, + * ::cuDeviceGetUuid, * ::cuDeviceGet, * ::cuDeviceTotalMem */ -CUresult CUDAAPI cuDeviceComputeCapability(int *major, int *minor, CUdevice dev); +__CUDA_DEPRECATED CUresult CUDAAPI cuDeviceComputeCapability(int *major, int *minor, CUdevice dev); /** @} */ /* END CUDA_DEVICE_DEPRECATED */ @@ -2370,8 +3044,8 @@ CUresult CUDAAPI cuDeviceComputeCapability(int *major, int *minor, CUdevice dev) * This section describes the primary context management functions of the low-level * CUDA driver application programming interface. * - * The primary context unique per device and it's shared with CUDA runtime API. - * Those functions allows seemless integration with other libraries using CUDA. + * The primary context is unique per device and shared with the CUDA runtime API. + * These functions allow integration with other libraries using CUDA. * * @{ */ @@ -2387,9 +3061,9 @@ CUresult CUDAAPI cuDeviceComputeCapability(int *major, int *minor, CUdevice dev) * Unlike ::cuCtxCreate() the newly created context is not pushed onto the stack. * * Context creation will fail with ::CUDA_ERROR_UNKNOWN if the compute mode of - * the device is ::CU_COMPUTEMODE_PROHIBITED. The function ::cuDeviceGetAttribute() - * can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to determine the compute mode - * of the device. + * the device is ::CU_COMPUTEMODE_PROHIBITED. The function ::cuDeviceGetAttribute() + * can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to determine the compute mode + * of the device. * The nvidia-smi tool can be used to set the compute mode for * devices. Documentation for nvidia-smi can be obtained by passing a * -h option to it. @@ -2497,8 +3171,9 @@ CUresult CUDAAPI cuDevicePrimaryCtxRelease(CUdevice dev); * \e C > \e P, then CUDA will yield to other OS threads when waiting for * the GPU (::CU_CTX_SCHED_YIELD), otherwise CUDA will not yield while * waiting for results and actively spin on the processor (::CU_CTX_SCHED_SPIN). - * However, on low power devices like Tegra, it always defaults to - * ::CU_CTX_SCHED_BLOCKING_SYNC. + * Additionally, on Tegra devices, ::CU_CTX_SCHED_AUTO uses a heuristic based on + * the power profile of the platform and may choose ::CU_CTX_SCHED_BLOCKING_SYNC + * for low-powered devices. * * - ::CU_CTX_LMEM_RESIZE_TO_MAX: Instruct CUDA to not reduce local memory * after resizing local memory for a kernel. This can prevent thrashing by @@ -2520,7 +3195,8 @@ CUresult CUDAAPI cuDevicePrimaryCtxRelease(CUdevice dev); * \sa ::cuDevicePrimaryCtxRetain, * ::cuDevicePrimaryCtxGetState, * ::cuCtxCreate, - * ::cuCtxGetFlags + * ::cuCtxGetFlags, + * ::cudaSetDeviceFlags */ CUresult CUDAAPI cuDevicePrimaryCtxSetFlags(CUdevice dev, unsigned int flags); @@ -2543,8 +3219,10 @@ CUresult CUDAAPI cuDevicePrimaryCtxSetFlags(CUdevice dev, unsigned int flags); * ::CUDA_ERROR_INVALID_VALUE, * \notefnerr * - * \sa ::cuDevicePrimaryCtxSetFlags, - * ::cuCtxGetFlags + * \sa + * ::cuDevicePrimaryCtxSetFlags, + * ::cuCtxGetFlags, + * ::cudaGetDeviceFlags */ CUresult CUDAAPI cuDevicePrimaryCtxGetState(CUdevice dev, unsigned int *flags, int *active); @@ -2581,8 +3259,8 @@ CUresult CUDAAPI cuDevicePrimaryCtxGetState(CUdevice dev, unsigned int *flags, i * ::cuCtxPushCurrent, * ::cuCtxSetCacheConfig, * ::cuCtxSetLimit, - * ::cuCtxSynchronize - * + * ::cuCtxSynchronize, + * ::cudaDeviceReset */ CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev); @@ -2600,6 +3278,9 @@ CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev); * This section describes the context management functions of the low-level * CUDA driver application programming interface. * + * Please note that some functions are described in + * \ref CUDA_PRIMARY_CTX "Primary Context Management" section. + * * @{ */ @@ -2607,11 +3288,13 @@ CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev); /** * \brief Create a CUDA context * + * \note In most cases it is recommended to use ::cuDevicePrimaryCtxRetain. + * * Creates a new CUDA context and associates it with the calling thread. The * \p flags parameter is described below. The context is created with a usage - * count of 1 and the caller of ::cuCtxCreate() must call ::cuCtxDestroy() or - * when done using the context. If a context is already current to the thread, - * it is supplanted by the newly created context and may be restored by a subsequent + * count of 1 and the caller of ::cuCtxCreate() must call ::cuCtxDestroy() + * when done using the context. If a context is already current to the thread, + * it is supplanted by the newly created context and may be restored by a subsequent * call to ::cuCtxPopCurrent(). * * The three LSBs of the \p flags parameter can be used to control how the OS @@ -2628,23 +3311,24 @@ CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev); * results from the GPU. This can increase latency when waiting for the GPU, * but can increase the performance of CPU threads performing work in parallel * with the GPU. - * + * * - ::CU_CTX_SCHED_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a * synchronization primitive when waiting for the GPU to finish work. * * - ::CU_CTX_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a * synchronization primitive when waiting for the GPU to finish work.
* Deprecated: This flag was deprecated as of CUDA 4.0 and was - * replaced with ::CU_CTX_SCHED_BLOCKING_SYNC. + * replaced with ::CU_CTX_SCHED_BLOCKING_SYNC. * * - ::CU_CTX_SCHED_AUTO: The default value if the \p flags parameter is zero, * uses a heuristic based on the number of active CUDA contexts in the * process \e C and the number of logical processors in the system \e P. If - * \e C > \e P, then CUDA will yield to other OS threads when waiting for - * the GPU (::CU_CTX_SCHED_YIELD), otherwise CUDA will not yield while - * waiting for results and actively spin on the processor (::CU_CTX_SCHED_SPIN). - * However, on low power devices like Tegra, it always defaults to - * ::CU_CTX_SCHED_BLOCKING_SYNC. + * \e C > \e P, then CUDA will yield to other OS threads when waiting for + * the GPU (::CU_CTX_SCHED_YIELD), otherwise CUDA will not yield while + * waiting for results and actively spin on the processor (::CU_CTX_SCHED_SPIN). + * Additionally, on Tegra devices, ::CU_CTX_SCHED_AUTO uses a heuristic based on + * the power profile of the platform and may choose ::CU_CTX_SCHED_BLOCKING_SYNC + * for low-powered devices. * * - ::CU_CTX_MAP_HOST: Instruct CUDA to support mapped pinned allocations. * This flag must be set in order to allocate pinned host memory that is @@ -2656,10 +3340,10 @@ CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev); * memory usage at the cost of potentially increased memory usage. * * Context creation will fail with ::CUDA_ERROR_UNKNOWN if the compute mode of - * the device is ::CU_COMPUTEMODE_PROHIBITED. The function ::cuDeviceGetAttribute() - * can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to determine the - * compute mode of the device. The nvidia-smi tool can be used to set - * the compute mode for * devices. + * the device is ::CU_COMPUTEMODE_PROHIBITED. The function ::cuDeviceGetAttribute() + * can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to determine the + * compute mode of the device. The nvidia-smi tool can be used to set + * the compute mode for * devices. * Documentation for nvidia-smi can be obtained by passing a * -h option to it. * @@ -2702,7 +3386,7 @@ CUresult CUDAAPI cuCtxCreate(CUcontext *pctx, unsigned int flags, CUdevice dev); * It is the responsibility of the calling function to ensure that no API * call issues using \p ctx while ::cuCtxDestroy() is executing. * - * If \p ctx is current to the calling thread then \p ctx will also be + * If \p ctx is current to the calling thread then \p ctx will also be * popped from the current thread's context stack (as though ::cuCtxPopCurrent() * were called). If \p ctx is current to other threads, then \p ctx will * remain current to those threads, and attempting to access \p ctx from @@ -2771,8 +3455,8 @@ CUresult CUDAAPI cuCtxPushCurrent(CUcontext ctx); /** * \brief Pops the current CUDA context from the current CPU thread. * - * Pops the current CUDA context from the CPU thread and passes back the - * old context handle in \p *pctx. That context may then be made current + * Pops the current CUDA context from the CPU thread and passes back the + * old context handle in \p *pctx. That context may then be made current * to a different CPU thread by calling ::cuCtxPushCurrent(). * * If a context was current to the CPU thread before ::cuCtxCreate() or @@ -2810,7 +3494,7 @@ CUresult CUDAAPI cuCtxPopCurrent(CUcontext *pctx); * calling CPU thread is unbound and ::CUDA_SUCCESS is returned. * * If there exists a CUDA context stack on the calling CPU thread, this - * will replace the top of that stack with \p ctx. + * will replace the top of that stack with \p ctx. * If \p ctx is NULL then this will be equivalent to popping the top * of the calling CPU thread's CUDA context stack (or a no-op if the * calling CPU thread's CUDA context stack is empty). @@ -2824,7 +3508,11 @@ CUresult CUDAAPI cuCtxPopCurrent(CUcontext *pctx); * ::CUDA_ERROR_INVALID_CONTEXT * \notefnerr * - * \sa ::cuCtxGetCurrent, ::cuCtxCreate, ::cuCtxDestroy + * \sa + * ::cuCtxGetCurrent, + * ::cuCtxCreate, + * ::cuCtxDestroy, + * ::cudaSetDevice */ CUresult CUDAAPI cuCtxSetCurrent(CUcontext ctx); @@ -2843,7 +3531,11 @@ CUresult CUDAAPI cuCtxSetCurrent(CUcontext ctx); * ::CUDA_ERROR_NOT_INITIALIZED, * \notefnerr * - * \sa ::cuCtxSetCurrent, ::cuCtxCreate, ::cuCtxDestroy + * \sa + * ::cuCtxSetCurrent, + * ::cuCtxCreate, + * ::cuCtxDestroy, + * ::cudaGetDevice */ CUresult CUDAAPI cuCtxGetCurrent(CUcontext *pctx); #endif /* __CUDA_API_VERSION >= 4000 */ @@ -2873,7 +3565,8 @@ CUresult CUDAAPI cuCtxGetCurrent(CUcontext *pctx); * ::cuCtxPushCurrent, * ::cuCtxSetCacheConfig, * ::cuCtxSetLimit, - * ::cuCtxSynchronize + * ::cuCtxSynchronize, + * ::cudaGetDevice */ CUresult CUDAAPI cuCtxGetDevice(CUdevice *device); @@ -2901,7 +3594,8 @@ CUresult CUDAAPI cuCtxGetDevice(CUdevice *device); * ::cuCtxGetDevice * ::cuCtxGetLimit, * ::cuCtxGetSharedMemConfig, - * ::cuCtxGetStreamPriorityRange + * ::cuCtxGetStreamPriorityRange, + * ::cudaGetDeviceFlags */ CUresult CUDAAPI cuCtxGetFlags(unsigned int *flags); #endif /* __CUDA_API_VERSION >= 7000 */ @@ -2911,7 +3605,7 @@ CUresult CUDAAPI cuCtxGetFlags(unsigned int *flags); * * Blocks until the device has completed all preceding requested tasks. * ::cuCtxSynchronize() returns an error if one of the preceding tasks failed. - * If the context was created with the ::CU_CTX_SCHED_BLOCKING_SYNC flag, the + * If the context was created with the ::CU_CTX_SCHED_BLOCKING_SYNC flag, the * CPU thread will block until the GPU context has finished its work. * * \return @@ -2931,7 +3625,8 @@ CUresult CUDAAPI cuCtxGetFlags(unsigned int *flags); * ::cuCtxPopCurrent, * ::cuCtxPushCurrent, * ::cuCtxSetCacheConfig, - * ::cuCtxSetLimit + * ::cuCtxSetLimit, + * ::cudaDeviceSynchronize */ CUresult CUDAAPI cuCtxSynchronize(void); @@ -2949,50 +3644,41 @@ CUresult CUDAAPI cuCtxSynchronize(void); * discussed here. * * - ::CU_LIMIT_STACK_SIZE controls the stack size in bytes of each GPU thread. - * This limit is only applicable to devices of compute capability 2.0 and - * higher. Attempting to set this limit on devices of compute capability - * less than 2.0 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT - * being returned. + * Note that the CUDA driver will set the \p limit to the maximum of \p value + * and what the kernel function requires. * * - ::CU_LIMIT_PRINTF_FIFO_SIZE controls the size in bytes of the FIFO used * by the ::printf() device system call. Setting ::CU_LIMIT_PRINTF_FIFO_SIZE * must be performed before launching any kernel that uses the ::printf() * device system call, otherwise ::CUDA_ERROR_INVALID_VALUE will be returned. - * This limit is only applicable to devices of compute capability 2.0 and - * higher. Attempting to set this limit on devices of compute capability - * less than 2.0 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT - * being returned. * * - ::CU_LIMIT_MALLOC_HEAP_SIZE controls the size in bytes of the heap used * by the ::malloc() and ::free() device system calls. Setting * ::CU_LIMIT_MALLOC_HEAP_SIZE must be performed before launching any kernel * that uses the ::malloc() or ::free() device system calls, otherwise - * ::CUDA_ERROR_INVALID_VALUE will be returned. This limit is only applicable - * to devices of compute capability 2.0 and higher. Attempting to set this - * limit on devices of compute capability less than 2.0 will result in the - * error ::CUDA_ERROR_UNSUPPORTED_LIMIT being returned. + * ::CUDA_ERROR_INVALID_VALUE will be returned. * * - ::CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH controls the maximum nesting depth of * a grid at which a thread can safely call ::cudaDeviceSynchronize(). Setting - * this limit must be performed before any launch of a kernel that uses the + * this limit must be performed before any launch of a kernel that uses the * device runtime and calls ::cudaDeviceSynchronize() above the default sync - * depth, two levels of grids. Calls to ::cudaDeviceSynchronize() will fail - * with error code ::cudaErrorSyncDepthExceeded if the limitation is + * depth, two levels of grids. Calls to ::cudaDeviceSynchronize() will fail + * with error code ::cudaErrorSyncDepthExceeded if the limitation is * violated. This limit can be set smaller than the default or up the maximum * launch depth of 24. When setting this limit, keep in mind that additional * levels of sync depth require the driver to reserve large amounts of device - * memory which can no longer be used for user allocations. If these - * reservations of device memory fail, ::cuCtxSetLimit will return + * memory which can no longer be used for user allocations. If these + * reservations of device memory fail, ::cuCtxSetLimit will return * ::CUDA_ERROR_OUT_OF_MEMORY, and the limit can be reset to a lower value. * This limit is only applicable to devices of compute capability 3.5 and * higher. Attempting to set this limit on devices of compute capability less - * than 3.5 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT being + * than 3.5 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT being * returned. * * - ::CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT controls the maximum number of * outstanding device runtime launches that can be made from the current * context. A grid is outstanding from the point of launch up until the grid - * is known to have been completed. Device runtime launches which violate + * is known to have been completed. Device runtime launches which violate * this limitation fail and return ::cudaErrorLaunchPendingCountExceeded when * ::cudaGetLastError() is called after launch. If more pending launches than * the default (2048 launches) are needed for a module using the device @@ -3006,6 +3692,10 @@ CUresult CUDAAPI cuCtxSynchronize(void); * than 3.5 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT being * returned. * + * - ::CU_LIMIT_MAX_L2_FETCH_GRANULARITY controls the L2 cache fetch granularity. + * Values can range from 0B to 128B. This is purely a performance hint and + * it can be ignored or clamped depending on the platform. + * * \param limit - Limit to set * \param value - Size of limit * @@ -3013,7 +3703,8 @@ CUresult CUDAAPI cuCtxSynchronize(void); * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_VALUE, * ::CUDA_ERROR_UNSUPPORTED_LIMIT, - * ::CUDA_ERROR_OUT_OF_MEMORY + * ::CUDA_ERROR_OUT_OF_MEMORY, + * ::CUDA_ERROR_INVALID_CONTEXT * \notefnerr * * \sa ::cuCtxCreate, @@ -3026,7 +3717,8 @@ CUresult CUDAAPI cuCtxSynchronize(void); * ::cuCtxPopCurrent, * ::cuCtxPushCurrent, * ::cuCtxSetCacheConfig, - * ::cuCtxSynchronize + * ::cuCtxSynchronize, + * ::cudaDeviceSetLimit */ CUresult CUDAAPI cuCtxSetLimit(CUlimit limit, size_t value); @@ -3045,6 +3737,7 @@ CUresult CUDAAPI cuCtxSetLimit(CUlimit limit, size_t value); * child grid launches to complete. * - ::CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT: maximum number of outstanding * device runtime launches that can be made from this context. + * - ::CU_LIMIT_MAX_L2_FETCH_GRANULARITY: L2 cache fetch granularity. * * \param limit - Limit to query * \param pvalue - Returned size of limit @@ -3065,7 +3758,8 @@ CUresult CUDAAPI cuCtxSetLimit(CUlimit limit, size_t value); * ::cuCtxPushCurrent, * ::cuCtxSetCacheConfig, * ::cuCtxSetLimit, - * ::cuCtxSynchronize + * ::cuCtxSynchronize, + * ::cudaDeviceGetLimit */ CUresult CUDAAPI cuCtxGetLimit(size_t *pvalue, CUlimit limit); @@ -3108,7 +3802,8 @@ CUresult CUDAAPI cuCtxGetLimit(size_t *pvalue, CUlimit limit); * ::cuCtxSetCacheConfig, * ::cuCtxSetLimit, * ::cuCtxSynchronize, - * ::cuFuncSetCacheConfig + * ::cuFuncSetCacheConfig, + * ::cudaDeviceGetCacheConfig */ CUresult CUDAAPI cuCtxGetCacheConfig(CUfunc_cache *pconfig); @@ -3158,7 +3853,8 @@ CUresult CUDAAPI cuCtxGetCacheConfig(CUfunc_cache *pconfig); * ::cuCtxPushCurrent, * ::cuCtxSetLimit, * ::cuCtxSynchronize, - * ::cuFuncSetCacheConfig + * ::cuFuncSetCacheConfig, + * ::cudaDeviceSetCacheConfig */ CUresult CUDAAPI cuCtxSetCacheConfig(CUfunc_cache config); @@ -3167,20 +3863,20 @@ CUresult CUDAAPI cuCtxSetCacheConfig(CUfunc_cache config); * \brief Returns the current shared memory configuration for the current context. * * This function will return in \p pConfig the current size of shared memory banks - * in the current context. On devices with configurable shared memory banks, - * ::cuCtxSetSharedMemConfig can be used to change this setting, so that all - * subsequent kernel launches will by default use the new bank size. When - * ::cuCtxGetSharedMemConfig is called on devices without configurable shared + * in the current context. On devices with configurable shared memory banks, + * ::cuCtxSetSharedMemConfig can be used to change this setting, so that all + * subsequent kernel launches will by default use the new bank size. When + * ::cuCtxGetSharedMemConfig is called on devices without configurable shared * memory, it will return the fixed bank size of the hardware. * * The returned bank configurations can be either: - * - ::CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE: shared memory bank width is + * - ::CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE: shared memory bank width is * four bytes. * - ::CU_SHARED_MEM_CONFIG_EIGHT_BYTE_BANK_SIZE: shared memory bank width will * eight bytes. * * \param pConfig - returned shared memory configuration - * \return + * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, @@ -3201,6 +3897,7 @@ CUresult CUDAAPI cuCtxSetCacheConfig(CUfunc_cache config); * ::cuCtxSynchronize, * ::cuCtxGetSharedMemConfig, * ::cuFuncSetCacheConfig, + * ::cudaDeviceGetSharedMemConfig */ CUresult CUDAAPI cuCtxGetSharedMemConfig(CUsharedconfig *pConfig); @@ -3208,16 +3905,16 @@ CUresult CUDAAPI cuCtxGetSharedMemConfig(CUsharedconfig *pConfig); * \brief Sets the shared memory configuration for the current context. * * On devices with configurable shared memory banks, this function will set - * the context's shared memory bank size which is used for subsequent kernel - * launches. + * the context's shared memory bank size which is used for subsequent kernel + * launches. * * Changed the shared memory configuration between launches may insert a device * side synchronization point between those launches. * * Changing the shared memory bank size will not increase shared memory usage - * or affect occupancy of kernels, but may have major effects on performance. + * or affect occupancy of kernels, but may have major effects on performance. * Larger bank sizes will allow for greater potential bandwidth to shared memory, - * but will change what kinds of accesses to shared memory will result in bank + * but will change what kinds of accesses to shared memory will result in bank * conflicts. * * This function will do nothing on devices with fixed shared memory bank size. @@ -3253,6 +3950,7 @@ CUresult CUDAAPI cuCtxGetSharedMemConfig(CUsharedconfig *pConfig); * ::cuCtxSynchronize, * ::cuCtxGetSharedMemConfig, * ::cuFuncSetCacheConfig, + * ::cudaDeviceSetSharedMemConfig */ CUresult CUDAAPI cuCtxSetSharedMemConfig(CUsharedconfig config); #endif @@ -3278,6 +3976,7 @@ CUresult CUDAAPI cuCtxSetSharedMemConfig(CUsharedconfig config); * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE, * ::CUDA_ERROR_UNKNOWN * \notefnerr * @@ -3329,7 +4028,8 @@ CUresult CUDAAPI cuCtxGetApiVersion(CUcontext ctx, unsigned int *version); * ::cuCtxGetDevice, * ::cuCtxGetFlags, * ::cuCtxSetLimit, - * ::cuCtxSynchronize + * ::cuCtxSynchronize, + * ::cudaDeviceGetStreamPriorityRange */ CUresult CUDAAPI cuCtxGetStreamPriorityRange(int *leastPriority, int *greatestPriority); @@ -3386,7 +4086,7 @@ CUresult CUDAAPI cuCtxGetStreamPriorityRange(int *leastPriority, int *greatestPr * ::cuCtxSetLimit, * ::cuCtxSynchronize */ -CUresult CUDAAPI cuCtxAttach(CUcontext *pctx, unsigned int flags); +__CUDA_DEPRECATED CUresult CUDAAPI cuCtxAttach(CUcontext *pctx, unsigned int flags); /** * \brief Decrement a context's usage-count @@ -3422,7 +4122,7 @@ CUresult CUDAAPI cuCtxAttach(CUcontext *pctx, unsigned int flags); * ::cuCtxSetLimit, * ::cuCtxSynchronize */ -CUresult CUDAAPI cuCtxDetach(CUcontext ctx); +__CUDA_DEPRECATED CUresult CUDAAPI cuCtxDetach(CUcontext ctx); /** @} */ /* END CUDA_CTX_DEPRECATED */ @@ -3465,7 +4165,8 @@ CUresult CUDAAPI cuCtxDetach(CUcontext ctx); * ::CUDA_ERROR_FILE_NOT_FOUND, * ::CUDA_ERROR_NO_BINARY_FOR_GPU, * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED, + * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND * \notefnerr * * \sa ::cuModuleGetFunction, @@ -3501,7 +4202,8 @@ CUresult CUDAAPI cuModuleLoad(CUmodule *module, const char *fname); * ::CUDA_ERROR_OUT_OF_MEMORY, * ::CUDA_ERROR_NO_BINARY_FOR_GPU, * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED, + * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND * \notefnerr * * \sa ::cuModuleGetFunction, @@ -3525,7 +4227,7 @@ CUresult CUDAAPI cuModuleLoadData(CUmodule *module, const void *image); * as Windows \c FindResource() to obtain the pointer. Options are passed as * an array via \p options and any corresponding parameters are passed in * \p optionValues. The number of total options is supplied via \p numOptions. - * Any outputs will be returned via \p optionValues. + * Any outputs will be returned via \p optionValues. * * \param module - Returned module * \param image - Module data to load @@ -3543,7 +4245,8 @@ CUresult CUDAAPI cuModuleLoadData(CUmodule *module, const void *image); * ::CUDA_ERROR_OUT_OF_MEMORY, * ::CUDA_ERROR_NO_BINARY_FOR_GPU, * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED, + * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND * \notefnerr * * \sa ::cuModuleGetFunction, @@ -3584,7 +4287,8 @@ CUresult CUDAAPI cuModuleLoadDataEx(CUmodule *module, const void *image, unsigne * ::CUDA_ERROR_OUT_OF_MEMORY, * ::CUDA_ERROR_NO_BINARY_FOR_GPU, * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED, + * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND * \notefnerr * * \sa ::cuModuleGetFunction, @@ -3682,7 +4386,9 @@ CUresult CUDAAPI cuModuleGetFunction(CUfunction *hfunc, CUmodule hmod, const cha * ::cuModuleLoadData, * ::cuModuleLoadDataEx, * ::cuModuleLoadFatBinary, - * ::cuModuleUnload + * ::cuModuleUnload, + * ::cudaGetSymbolAddress, + * ::cudaGetSymbolSize */ CUresult CUDAAPI cuModuleGetGlobal(CUdeviceptr *dptr, size_t *bytes, CUmodule hmod, const char *name); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -3716,7 +4422,8 @@ CUresult CUDAAPI cuModuleGetGlobal(CUdeviceptr *dptr, size_t *bytes, CUmodule hm * ::cuModuleLoadData, * ::cuModuleLoadDataEx, * ::cuModuleLoadFatBinary, - * ::cuModuleUnload + * ::cuModuleUnload, + * ::cudaGetTextureReference */ CUresult CUDAAPI cuModuleGetTexRef(CUtexref *pTexRef, CUmodule hmod, const char *name); @@ -3747,7 +4454,8 @@ CUresult CUDAAPI cuModuleGetTexRef(CUtexref *pTexRef, CUmodule hmod, const char * ::cuModuleLoadData, * ::cuModuleLoadDataEx, * ::cuModuleLoadFatBinary, - * ::cuModuleUnload + * ::cuModuleUnload, + * ::cudaGetSurfaceReference */ CUresult CUDAAPI cuModuleGetSurfRef(CUsurfref *pSurfRef, CUmodule hmod, const char *name); @@ -3782,7 +4490,8 @@ CUresult CUDAAPI cuModuleGetSurfRef(CUsurfref *pSurfRef, CUmodule hmod, const ch * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_OUT_OF_MEMORY + * ::CUDA_ERROR_OUT_OF_MEMORY, + * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND * \notefnerr * * \sa ::cuLinkAddData, @@ -3954,7 +4663,8 @@ cuLinkDestroy(CUlinkState state); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemGetInfo */ CUresult CUDAAPI cuMemGetInfo(size_t *free, size_t *total); @@ -3987,7 +4697,8 @@ CUresult CUDAAPI cuMemGetInfo(size_t *free, size_t *total); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMalloc */ CUresult CUDAAPI cuMemAlloc(CUdeviceptr *dptr, size_t bytesize); @@ -4048,7 +4759,8 @@ CUresult CUDAAPI cuMemAlloc(CUdeviceptr *dptr, size_t bytesize); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMallocPitch */ CUresult CUDAAPI cuMemAllocPitch(CUdeviceptr *dptr, size_t *pPitch, size_t WidthInBytes, size_t Height, unsigned int ElementSizeBytes); @@ -4077,7 +4789,8 @@ CUresult CUDAAPI cuMemAllocPitch(CUdeviceptr *dptr, size_t *pPitch, size_t Width * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaFree */ CUresult CUDAAPI cuMemFree(CUdeviceptr dptr); @@ -4098,6 +4811,7 @@ CUresult CUDAAPI cuMemFree(CUdeviceptr dptr); * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_NOT_FOUND, * ::CUDA_ERROR_INVALID_VALUE * \notefnerr * @@ -4131,7 +4845,7 @@ CUresult CUDAAPI cuMemGetAddressRange(CUdeviceptr *pbase, size_t *psize, CUdevic * Note all host memory allocated using ::cuMemHostAlloc() will automatically * be immediately accessible to all contexts on all devices which support unified * addressing (as may be queried using ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING). - * The device pointer that may be used to access this host memory from those + * The device pointer that may be used to access this host memory from those * contexts is always equal to the returned host pointer \p *pp. * See \ref CUDA_UNIFIED for additional details. * @@ -4156,7 +4870,8 @@ CUresult CUDAAPI cuMemGetAddressRange(CUdeviceptr *pbase, size_t *psize, CUdevic * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMallocHost */ CUresult CUDAAPI cuMemAllocHost(void **pp, size_t bytesize); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -4186,7 +4901,8 @@ CUresult CUDAAPI cuMemAllocHost(void **pp, size_t bytesize); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaFreeHost */ CUresult CUDAAPI cuMemFreeHost(void *p); @@ -4213,8 +4929,7 @@ CUresult CUDAAPI cuMemFreeHost(void *p); * * - ::CU_MEMHOSTALLOC_DEVICEMAP: Maps the allocation into the CUDA address * space. The device pointer to the memory may be obtained by calling - * ::cuMemHostGetDevicePointer(). This feature is available only on GPUs - * with compute capability greater than or equal to 1.1. + * ::cuMemHostGetDevicePointer(). * * - ::CU_MEMHOSTALLOC_WRITECOMBINED: Allocates the memory as write-combined * (WC). WC memory can be transferred across the PCI Express bus more @@ -4239,8 +4954,8 @@ CUresult CUDAAPI cuMemFreeHost(void *p); * Note all host memory allocated using ::cuMemHostAlloc() will automatically * be immediately accessible to all contexts on all devices which support unified * addressing (as may be queried using ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING). - * Unless the flag ::CU_MEMHOSTALLOC_WRITECOMBINED is specified, the device pointer - * that may be used to access this host memory from those contexts is always equal + * Unless the flag ::CU_MEMHOSTALLOC_WRITECOMBINED is specified, the device pointer + * that may be used to access this host memory from those contexts is always equal * to the returned host pointer \p *pp. If the flag ::CU_MEMHOSTALLOC_WRITECOMBINED * is specified, then the function ::cuMemHostGetDevicePointer() must be used * to query the device pointer, even if the context supports unified addressing. @@ -4268,7 +4983,8 @@ CUresult CUDAAPI cuMemFreeHost(void *p); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaHostAlloc */ CUresult CUDAAPI cuMemHostAlloc(void **pp, size_t bytesize, unsigned int Flags); @@ -4321,7 +5037,8 @@ CUresult CUDAAPI cuMemHostAlloc(void **pp, size_t bytesize, unsigned int Flags); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaHostGetDevicePointer */ CUresult CUDAAPI cuMemHostGetDevicePointer(CUdeviceptr *pdptr, void *p, unsigned int Flags); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -4346,7 +5063,10 @@ CUresult CUDAAPI cuMemHostGetDevicePointer(CUdeviceptr *pdptr, void *p, unsigned * ::CUDA_ERROR_INVALID_VALUE * \notefnerr * - * \sa ::cuMemAllocHost, ::cuMemHostAlloc + * \sa + * ::cuMemAllocHost, + * ::cuMemHostAlloc, + * ::cudaHostGetFlags */ CUresult CUDAAPI cuMemHostGetFlags(unsigned int *pFlags, void *p); @@ -4431,6 +5151,7 @@ CUresult CUDAAPI cuMemHostGetFlags(unsigned int *pFlags, void *p); * managed memory supporting devices on which contexts were previously created, even if * those contexts have been destroyed. These environment variables are described * in the CUDA programming guide under the "CUDA environment variables" section. + * - On ARM, managed memory is not available on discrete gpu with Drive PX-2. * * \param dptr - Returned device pointer * \param bytesize - Requested allocation size in bytes @@ -4456,7 +5177,8 @@ CUresult CUDAAPI cuMemHostGetFlags(unsigned int *pFlags, void *p); * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, - * ::cuDeviceGetAttribute, ::cuStreamAttachMemAsync + * ::cuDeviceGetAttribute, ::cuStreamAttachMemAsync, + * ::cudaMallocManaged */ CUresult CUDAAPI cuMemAllocManaged(CUdeviceptr *dptr, size_t bytesize, unsigned int flags); @@ -4471,7 +5193,7 @@ CUresult CUDAAPI cuMemAllocManaged(CUdeviceptr *dptr, size_t bytesize, unsigned * * \param dev - Returned device handle * - * \param pciBusId - String in one of the following forms: + * \param pciBusId - String in one of the following forms: * [domain]:[bus]:[device].[function] * [domain]:[bus]:[device] * [bus]:[device].[function] @@ -4485,7 +5207,11 @@ CUresult CUDAAPI cuMemAllocManaged(CUdeviceptr *dptr, size_t bytesize, unsigned * ::CUDA_ERROR_INVALID_DEVICE * \notefnerr * - * \sa ::cuDeviceGet, ::cuDeviceGetAttribute, ::cuDeviceGetPCIBusId + * \sa + * ::cuDeviceGet, + * ::cuDeviceGetAttribute, + * ::cuDeviceGetPCIBusId, + * ::cudaDeviceGetByPCIBusId */ CUresult CUDAAPI cuDeviceGetByPCIBusId(CUdevice *dev, const char *pciBusId); @@ -4513,65 +5239,73 @@ CUresult CUDAAPI cuDeviceGetByPCIBusId(CUdevice *dev, const char *pciBusId); * ::CUDA_ERROR_INVALID_DEVICE * \notefnerr * - * \sa ::cuDeviceGet, ::cuDeviceGetAttribute, ::cuDeviceGetByPCIBusId + * \sa + * ::cuDeviceGet, + * ::cuDeviceGetAttribute, + * ::cuDeviceGetByPCIBusId, + * ::cudaDeviceGetPCIBusId */ CUresult CUDAAPI cuDeviceGetPCIBusId(char *pciBusId, int len, CUdevice dev); /** * \brief Gets an interprocess handle for a previously allocated event * - * Takes as input a previously allocated event. This event must have been - * created with the ::CU_EVENT_INTERPROCESS and ::CU_EVENT_DISABLE_TIMING + * Takes as input a previously allocated event. This event must have been + * created with the ::CU_EVENT_INTERPROCESS and ::CU_EVENT_DISABLE_TIMING * flags set. This opaque handle may be copied into other processes and * opened with ::cuIpcOpenEventHandle to allow efficient hardware * synchronization between GPU work in different processes. * - * After the event has been opened in the importing process, - * ::cuEventRecord, ::cuEventSynchronize, ::cuStreamWaitEvent and - * ::cuEventQuery may be used in either process. Performing operations - * on the imported event after the exported event has been freed + * After the event has been opened in the importing process, + * ::cuEventRecord, ::cuEventSynchronize, ::cuStreamWaitEvent and + * ::cuEventQuery may be used in either process. Performing operations + * on the imported event after the exported event has been freed * with ::cuEventDestroy will result in undefined behavior. * - * IPC functionality is restricted to devices with support for unified - * addressing on Linux operating systems. + * IPC functionality is restricted to devices with support for unified + * addressing on Linux and Windows operating systems. + * IPC functionality on Windows is restricted to GPUs in TCC mode * * \param pHandle - Pointer to a user allocated CUipcEventHandle * in which to return the opaque event handle - * \param event - Event allocated with ::CU_EVENT_INTERPROCESS and + * \param event - Event allocated with ::CU_EVENT_INTERPROCESS and * ::CU_EVENT_DISABLE_TIMING flags. * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_HANDLE, * ::CUDA_ERROR_OUT_OF_MEMORY, - * ::CUDA_ERROR_MAP_FAILED + * ::CUDA_ERROR_MAP_FAILED, + * ::CUDA_ERROR_INVALID_VALUE * - * \sa - * ::cuEventCreate, - * ::cuEventDestroy, + * \sa + * ::cuEventCreate, + * ::cuEventDestroy, * ::cuEventSynchronize, * ::cuEventQuery, * ::cuStreamWaitEvent, * ::cuIpcOpenEventHandle, * ::cuIpcGetMemHandle, * ::cuIpcOpenMemHandle, - * ::cuIpcCloseMemHandle + * ::cuIpcCloseMemHandle, + * ::cudaIpcGetEventHandle */ CUresult CUDAAPI cuIpcGetEventHandle(CUipcEventHandle *pHandle, CUevent event); /** * \brief Opens an interprocess event handle for use in the current process * - * Opens an interprocess event handle exported from another process with - * ::cuIpcGetEventHandle. This function returns a ::CUevent that behaves like - * a locally created event with the ::CU_EVENT_DISABLE_TIMING flag specified. + * Opens an interprocess event handle exported from another process with + * ::cuIpcGetEventHandle. This function returns a ::CUevent that behaves like + * a locally created event with the ::CU_EVENT_DISABLE_TIMING flag specified. * This event must be freed with ::cuEventDestroy. * - * Performing operations on the imported event after the exported event has + * Performing operations on the imported event after the exported event has * been freed with ::cuEventDestroy will result in undefined behavior. * - * IPC functionality is restricted to devices with support for unified - * addressing on Linux operating systems. + * IPC functionality is restricted to devices with support for unified + * addressing on Linux and Windows operating systems. + * IPC functionality on Windows is restricted to GPUs in TCC mode * * \param phEvent - Returns the imported event * \param handle - Interprocess handle to open @@ -4581,18 +5315,20 @@ CUresult CUDAAPI cuIpcGetEventHandle(CUipcEventHandle *pHandle, CUevent event); * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_MAP_FAILED, * ::CUDA_ERROR_PEER_ACCESS_UNSUPPORTED, - * ::CUDA_ERROR_INVALID_HANDLE + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_VALUE * * \sa - * ::cuEventCreate, - * ::cuEventDestroy, + * ::cuEventCreate, + * ::cuEventDestroy, * ::cuEventSynchronize, * ::cuEventQuery, * ::cuStreamWaitEvent, * ::cuIpcGetEventHandle, * ::cuIpcGetMemHandle, * ::cuIpcOpenMemHandle, - * ::cuIpcCloseMemHandle + * ::cuIpcCloseMemHandle, + * ::cudaIpcOpenEventHandle */ CUresult CUDAAPI cuIpcOpenEventHandle(CUevent *phEvent, CUipcEventHandle handle); @@ -4600,36 +5336,39 @@ CUresult CUDAAPI cuIpcOpenEventHandle(CUevent *phEvent, CUipcEventHandle handle) * \brief Gets an interprocess memory handle for an existing device memory * allocation * - * Takes a pointer to the base of an existing device memory allocation created - * with ::cuMemAlloc and exports it for use in another process. This is a + * Takes a pointer to the base of an existing device memory allocation created + * with ::cuMemAlloc and exports it for use in another process. This is a * lightweight operation and may be called multiple times on an allocation - * without adverse effects. + * without adverse effects. * * If a region of memory is freed with ::cuMemFree and a subsequent call * to ::cuMemAlloc returns memory with the same device address, * ::cuIpcGetMemHandle will return a unique handle for the - * new memory. + * new memory. * - * IPC functionality is restricted to devices with support for unified - * addressing on Linux operating systems. + * IPC functionality is restricted to devices with support for unified + * addressing on Linux and Windows operating systems. + * IPC functionality on Windows is restricted to GPUs in TCC mode * * \param pHandle - Pointer to user allocated ::CUipcMemHandle to return * the handle in. - * \param dptr - Base pointer to previously allocated device memory + * \param dptr - Base pointer to previously allocated device memory * * \returns * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_HANDLE, * ::CUDA_ERROR_OUT_OF_MEMORY, * ::CUDA_ERROR_MAP_FAILED, - * + * ::CUDA_ERROR_INVALID_VALUE + * * \sa * ::cuMemAlloc, * ::cuMemFree, * ::cuIpcGetEventHandle, * ::cuIpcOpenEventHandle, * ::cuIpcOpenMemHandle, - * ::cuIpcCloseMemHandle + * ::cuIpcCloseMemHandle, + * ::cudaIpcGetMemHandle */ CUresult CUDAAPI cuIpcGetMemHandle(CUipcMemHandle *pHandle, CUdeviceptr dptr); @@ -4638,14 +5377,17 @@ CUresult CUDAAPI cuIpcGetMemHandle(CUipcMemHandle *pHandle, CUdeviceptr dptr); * and returns a device pointer usable in the local process. * * Maps memory exported from another process with ::cuIpcGetMemHandle into - * the current device address space. For contexts on different devices + * the current device address space. For contexts on different devices * ::cuIpcOpenMemHandle can attempt to enable peer access between the - * devices as if the user called ::cuCtxEnablePeerAccess. This behavior is - * controlled by the ::CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS flag. + * devices as if the user called ::cuCtxEnablePeerAccess. This behavior is + * controlled by the ::CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS flag. * ::cuDeviceCanAccessPeer can determine if a mapping is possible. * + * ::cuIpcOpenMemHandle can open handles to devices that may not be visible + * in the process calling the API. + * * Contexts that may open ::CUipcMemHandles are restricted in the following way. - * ::CUipcMemHandles from each ::CUdevice in a given process may only be opened + * ::CUipcMemHandles from each ::CUdevice in a given process may only be opened * by one ::CUcontext per ::CUdevice per other process. * * Memory returned from ::cuIpcOpenMemHandle must be freed with @@ -4655,9 +5397,10 @@ CUresult CUDAAPI cuIpcGetMemHandle(CUipcMemHandle *pHandle, CUdeviceptr dptr); * ::cuIpcCloseMemHandle in the importing context will result in undefined * behavior. * - * IPC functionality is restricted to devices with support for unified - * addressing on Linux operating systems. - * + * IPC functionality is restricted to devices with support for unified + * addressing on Linux and Windows operating systems. + * IPC functionality on Windows is restricted to GPUs in TCC mode + * * \param pdptr - Returned device pointer * \param handle - ::CUipcMemHandle to open * \param Flags - Flags for this operation. Must be specified as ::CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS @@ -4667,9 +5410,10 @@ CUresult CUDAAPI cuIpcGetMemHandle(CUipcMemHandle *pHandle, CUdeviceptr dptr); * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_MAP_FAILED, * ::CUDA_ERROR_INVALID_HANDLE, - * ::CUDA_ERROR_TOO_MANY_PEERS + * ::CUDA_ERROR_TOO_MANY_PEERS, + * ::CUDA_ERROR_INVALID_VALUE * - * \note No guarantees are made about the address returned in \p *pdptr. + * \note No guarantees are made about the address returned in \p *pdptr. * In particular, multiple processes may not receive the same address for the same \p handle. * * \sa @@ -4681,12 +5425,13 @@ CUresult CUDAAPI cuIpcGetMemHandle(CUipcMemHandle *pHandle, CUdeviceptr dptr); * ::cuIpcCloseMemHandle, * ::cuCtxEnablePeerAccess, * ::cuDeviceCanAccessPeer, + * ::cudaIpcOpenMemHandle */ CUresult CUDAAPI cuIpcOpenMemHandle(CUdeviceptr *pdptr, CUipcMemHandle handle, unsigned int Flags); /** * \brief Close memory mapped with ::cuIpcOpenMemHandle - * + * * Unmaps memory returnd by ::cuIpcOpenMemHandle. The original allocation * in the exporting process as well as imported mappings in other processes * will be unaffected. @@ -4694,17 +5439,18 @@ CUresult CUDAAPI cuIpcOpenMemHandle(CUdeviceptr *pdptr, CUipcMemHandle handle, u * Any resources used to enable peer access will be freed if this is the * last mapping using them. * - * IPC functionality is restricted to devices with support for unified - * addressing on Linux operating systems. + * IPC functionality is restricted to devices with support for unified + * addressing on Linux and Windows operating systems. + * IPC functionality on Windows is restricted to GPUs in TCC mode * * \param dptr - Device pointer returned by ::cuIpcOpenMemHandle - * + * * \returns * ::CUDA_SUCCESS, * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_MAP_FAILED, * ::CUDA_ERROR_INVALID_HANDLE, - * + * ::CUDA_ERROR_INVALID_VALUE * \sa * ::cuMemAlloc, * ::cuMemFree, @@ -4712,6 +5458,7 @@ CUresult CUDAAPI cuIpcOpenMemHandle(CUdeviceptr *pdptr, CUipcMemHandle handle, u * ::cuIpcOpenEventHandle, * ::cuIpcGetMemHandle, * ::cuIpcOpenMemHandle, + * ::cudaIpcCloseMemHandle */ CUresult CUDAAPI cuIpcCloseMemHandle(CUdeviceptr dptr); @@ -4724,8 +5471,8 @@ CUresult CUDAAPI cuIpcCloseMemHandle(CUdeviceptr dptr); * Page-locks the memory range specified by \p p and \p bytesize and maps it * for the device(s) as specified by \p Flags. This memory range also is added * to the same tracking mechanism as ::cuMemHostAlloc to automatically accelerate - * calls to functions such as ::cuMemcpyHtoD(). Since the memory can be accessed - * directly by the device, it can be read or written with much higher bandwidth + * calls to functions such as ::cuMemcpyHtoD(). Since the memory can be accessed + * directly by the device, it can be read or written with much higher bandwidth * than pageable memory that has not been registered. Page-locking excessive * amounts of memory may degrade system performance, since it reduces the amount * of memory available to the system for paging. As a result, this function is @@ -4743,8 +5490,7 @@ CUresult CUDAAPI cuIpcCloseMemHandle(CUdeviceptr dptr); * * - ::CU_MEMHOSTREGISTER_DEVICEMAP: Maps the allocation into the CUDA address * space. The device pointer to the memory may be obtained by calling - * ::cuMemHostGetDevicePointer(). This feature is available only on GPUs - * with compute capability greater than or equal to 1.1. + * ::cuMemHostGetDevicePointer(). * * - ::CU_MEMHOSTREGISTER_IOMEMORY: The pointer is treated as pointing to some * I/O memory space, e.g. the PCI Express resource of a 3rd party device. @@ -4775,7 +5521,7 @@ CUresult CUDAAPI cuIpcCloseMemHandle(CUdeviceptr dptr); * on devices that have a non-zero value for the device attribute. Note however that * such devices should access the memory using only of the two pointers and not both. * - * The memory page-locked by this function must be unregistered with + * The memory page-locked by this function must be unregistered with * ::cuMemHostUnregister(). * * \param p - Host pointer to memory to page-lock @@ -4794,7 +5540,11 @@ CUresult CUDAAPI cuIpcCloseMemHandle(CUdeviceptr dptr); * ::CUDA_ERROR_NOT_SUPPORTED * \notefnerr * - * \sa ::cuMemHostUnregister, ::cuMemHostGetFlags, ::cuMemHostGetDevicePointer + * \sa + * ::cuMemHostUnregister, + * ::cuMemHostGetFlags, + * ::cuMemHostGetDevicePointer, + * ::cudaHostRegister */ CUresult CUDAAPI cuMemHostRegister(void *p, size_t bytesize, unsigned int Flags); @@ -4818,17 +5568,19 @@ CUresult CUDAAPI cuMemHostRegister(void *p, size_t bytesize, unsigned int Flags) * ::CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED, * \notefnerr * - * \sa ::cuMemHostRegister + * \sa + * ::cuMemHostRegister, + * ::cudaHostUnregister */ CUresult CUDAAPI cuMemHostUnregister(void *p); /** * \brief Copies memory * - * Copies data between two pointers. - * \p dst and \p src are base pointers of the destination and source, respectively. + * Copies data between two pointers. + * \p dst and \p src are base pointers of the destination and source, respectively. * \p ByteCount specifies the number of bytes to copy. - * Note that this function infers the type of the transfer (host to host, host to + * Note that this function infers the type of the transfer (host to host, host to * device, device to device, or device to host) from the pointer values. This * function is only allowed in contexts which support unified addressing. * @@ -4854,7 +5606,10 @@ CUresult CUDAAPI cuMemHostUnregister(void *p); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy, + * ::cudaMemcpyToSymbol, + * ::cudaMemcpyFromSymbol */ CUresult CUDAAPI cuMemcpy(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount); @@ -4862,9 +5617,9 @@ CUresult CUDAAPI cuMemcpy(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount); * \brief Copies device memory between two contexts * * Copies from device memory in one context to device memory in another - * context. \p dstDevice is the base device pointer of the destination memory - * and \p dstContext is the destination context. \p srcDevice is the base - * device pointer of the source memory and \p srcContext is the source pointer. + * context. \p dstDevice is the base device pointer of the destination memory + * and \p dstContext is the destination context. \p srcDevice is the base + * device pointer of the source memory and \p srcContext is the source pointer. * \p ByteCount specifies the number of bytes to copy. * * \param dstDevice - Destination device pointer @@ -4883,7 +5638,8 @@ CUresult CUDAAPI cuMemcpy(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount); * \note_sync * * \sa ::cuMemcpyDtoD, ::cuMemcpy3DPeer, ::cuMemcpyDtoDAsync, ::cuMemcpyPeerAsync, - * ::cuMemcpy3DPeerAsync + * ::cuMemcpy3DPeerAsync, + * ::cudaMemcpyPeer */ CUresult CUDAAPI cuMemcpyPeer(CUdeviceptr dstDevice, CUcontext dstContext, CUdeviceptr srcDevice, CUcontext srcContext, size_t ByteCount); @@ -4919,7 +5675,9 @@ CUresult CUDAAPI cuMemcpyPeer(CUdeviceptr dstDevice, CUcontext dstContext, CUdev * ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy, + * ::cudaMemcpyToSymbol */ CUresult CUDAAPI cuMemcpyHtoD(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount); @@ -4952,7 +5710,9 @@ CUresult CUDAAPI cuMemcpyHtoD(CUdeviceptr dstDevice, const void *srcHost, size_t * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy, + * ::cudaMemcpyFromSymbol */ CUresult CUDAAPI cuMemcpyDtoH(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount); @@ -4985,7 +5745,10 @@ CUresult CUDAAPI cuMemcpyDtoH(void *dstHost, CUdeviceptr srcDevice, size_t ByteC * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy, + * ::cudaMemcpyToSymbol, + * ::cudaMemcpyFromSymbol */ CUresult CUDAAPI cuMemcpyDtoD(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount); @@ -5020,7 +5783,8 @@ CUresult CUDAAPI cuMemcpyDtoD(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpyToArray */ CUresult CUDAAPI cuMemcpyDtoA(CUarray dstArray, size_t dstOffset, CUdeviceptr srcDevice, size_t ByteCount); @@ -5057,7 +5821,8 @@ CUresult CUDAAPI cuMemcpyDtoA(CUarray dstArray, size_t dstOffset, CUdeviceptr sr * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpyFromArray */ CUresult CUDAAPI cuMemcpyAtoD(CUdeviceptr dstDevice, CUarray srcArray, size_t srcOffset, size_t ByteCount); @@ -5092,7 +5857,8 @@ CUresult CUDAAPI cuMemcpyAtoD(CUdeviceptr dstDevice, CUarray srcArray, size_t sr * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpyToArray */ CUresult CUDAAPI cuMemcpyHtoA(CUarray dstArray, size_t dstOffset, const void *srcHost, size_t ByteCount); @@ -5127,7 +5893,8 @@ CUresult CUDAAPI cuMemcpyHtoA(CUarray dstArray, size_t dstOffset, const void *sr * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpyFromArray */ CUresult CUDAAPI cuMemcpyAtoH(void *dstHost, CUarray srcArray, size_t srcOffset, size_t ByteCount); @@ -5166,7 +5933,8 @@ CUresult CUDAAPI cuMemcpyAtoH(void *dstHost, CUarray srcArray, size_t srcOffset, * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpyArrayToArray */ CUresult CUDAAPI cuMemcpyAtoA(CUarray dstArray, size_t dstOffset, CUarray srcArray, size_t srcOffset, size_t ByteCount); @@ -5211,9 +5979,9 @@ CUresult CUDAAPI cuMemcpyAtoA(CUarray dstArray, size_t dstOffset, CUarray srcArr * * \par * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::srcArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::srcArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -5238,9 +6006,9 @@ CUresult CUDAAPI cuMemcpyAtoA(CUarray dstArray, size_t dstOffset, CUarray srcArr * * \par * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::dstArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::dstArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -5327,7 +6095,10 @@ CUresult CUDAAPI cuMemcpyAtoA(CUarray dstArray, size_t dstOffset, CUarray srcArr * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy2D, + * ::cudaMemcpy2DToArray, + * ::cudaMemcpy2DFromArray */ CUresult CUDAAPI cuMemcpy2D(const CUDA_MEMCPY2D *pCopy); @@ -5370,9 +6141,9 @@ CUresult CUDAAPI cuMemcpy2D(const CUDA_MEMCPY2D *pCopy); * * \par * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::srcArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::srcArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -5392,9 +6163,9 @@ CUresult CUDAAPI cuMemcpy2D(const CUDA_MEMCPY2D *pCopy); * * \par * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::dstArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::dstArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -5486,7 +6257,10 @@ CUresult CUDAAPI cuMemcpy2D(const CUDA_MEMCPY2D *pCopy); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy2D, + * ::cudaMemcpy2DToArray, + * ::cudaMemcpy2DFromArray */ CUresult CUDAAPI cuMemcpy2DUnaligned(const CUDA_MEMCPY2D *pCopy); @@ -5537,9 +6311,9 @@ CUresult CUDAAPI cuMemcpy2DUnaligned(const CUDA_MEMCPY2D *pCopy); * * \par * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::srcArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::srcArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -5561,9 +6335,9 @@ CUresult CUDAAPI cuMemcpy2DUnaligned(const CUDA_MEMCPY2D *pCopy); * * \par * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::dstArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::dstArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -5654,7 +6428,8 @@ CUresult CUDAAPI cuMemcpy2DUnaligned(const CUDA_MEMCPY2D *pCopy); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMemcpy3D */ CUresult CUDAAPI cuMemcpy3D(const CUDA_MEMCPY3D *pCopy); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -5679,17 +6454,18 @@ CUresult CUDAAPI cuMemcpy3D(const CUDA_MEMCPY3D *pCopy); * \note_sync * * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpyDtoDAsync, ::cuMemcpyPeerAsync, - * ::cuMemcpy3DPeerAsync + * ::cuMemcpy3DPeerAsync, + * ::cudaMemcpy3DPeer */ CUresult CUDAAPI cuMemcpy3DPeer(const CUDA_MEMCPY3D_PEER *pCopy); /** * \brief Copies memory asynchronously * - * Copies data between two pointers. - * \p dst and \p src are base pointers of the destination and source, respectively. + * Copies data between two pointers. + * \p dst and \p src are base pointers of the destination and source, respectively. * \p ByteCount specifies the number of bytes to copy. - * Note that this function infers the type of the transfer (host to host, host to + * Note that this function infers the type of the transfer (host to host, host to * device, device to device, or device to host) from the pointer values. This * function is only allowed in contexts which support unified addressing. * @@ -5703,7 +6479,8 @@ CUresult CUDAAPI cuMemcpy3DPeer(const CUDA_MEMCPY3D_PEER *pCopy); * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -5719,7 +6496,10 @@ CUresult CUDAAPI cuMemcpy3DPeer(const CUDA_MEMCPY3D_PEER *pCopy); * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpyAsync, + * ::cudaMemcpyToSymbolAsync, + * ::cudaMemcpyFromSymbolAsync */ CUresult CUDAAPI cuMemcpyAsync(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount, CUstream hStream); @@ -5727,9 +6507,9 @@ CUresult CUDAAPI cuMemcpyAsync(CUdeviceptr dst, CUdeviceptr src, size_t ByteCoun * \brief Copies device memory between two contexts asynchronously. * * Copies from device memory in one context to device memory in another - * context. \p dstDevice is the base device pointer of the destination memory - * and \p dstContext is the destination context. \p srcDevice is the base - * device pointer of the source memory and \p srcContext is the source pointer. + * context. \p dstDevice is the base device pointer of the destination memory + * and \p dstContext is the destination context. \p srcDevice is the base + * device pointer of the source memory and \p srcContext is the source pointer. * \p ByteCount specifies the number of bytes to copy. * * \param dstDevice - Destination device pointer @@ -5744,13 +6524,15 @@ CUresult CUDAAPI cuMemcpyAsync(CUdeviceptr dst, CUdeviceptr src, size_t ByteCoun * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream * - * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpy3DPeer, ::cuMemcpyDtoDAsync, - * ::cuMemcpy3DPeerAsync + * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpy3DPeer, ::cuMemcpyDtoDAsync, + * ::cuMemcpy3DPeerAsync, + * ::cudaMemcpyPeerAsync */ CUresult CUDAAPI cuMemcpyPeerAsync(CUdeviceptr dstDevice, CUcontext dstContext, CUdeviceptr srcDevice, CUcontext srcContext, size_t ByteCount, CUstream hStream); #endif /* __CUDA_API_VERSION >= 4000 */ @@ -5773,7 +6555,8 @@ CUresult CUDAAPI cuMemcpyPeerAsync(CUdeviceptr dstDevice, CUcontext dstContext, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -5789,7 +6572,9 @@ CUresult CUDAAPI cuMemcpyPeerAsync(CUdeviceptr dstDevice, CUcontext dstContext, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpyAsync, + * ::cudaMemcpyToSymbolAsync */ CUresult CUDAAPI cuMemcpyHtoDAsync(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream hStream); @@ -5810,7 +6595,8 @@ CUresult CUDAAPI cuMemcpyHtoDAsync(CUdeviceptr dstDevice, const void *srcHost, s * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -5826,7 +6612,9 @@ CUresult CUDAAPI cuMemcpyHtoDAsync(CUdeviceptr dstDevice, const void *srcHost, s * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpyAsync, + * ::cudaMemcpyFromSymbolAsync */ CUresult CUDAAPI cuMemcpyDtoHAsync(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream); @@ -5847,7 +6635,8 @@ CUresult CUDAAPI cuMemcpyDtoHAsync(void *dstHost, CUdeviceptr srcDevice, size_t * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -5863,7 +6652,10 @@ CUresult CUDAAPI cuMemcpyDtoHAsync(void *dstHost, CUdeviceptr srcDevice, size_t * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpyAsync, + * ::cudaMemcpyToSymbolAsync, + * ::cudaMemcpyFromSymbolAsync */ CUresult CUDAAPI cuMemcpyDtoDAsync(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream); @@ -5886,7 +6678,8 @@ CUresult CUDAAPI cuMemcpyDtoDAsync(CUdeviceptr dstDevice, CUdeviceptr srcDevice, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -5902,7 +6695,8 @@ CUresult CUDAAPI cuMemcpyDtoDAsync(CUdeviceptr dstDevice, CUdeviceptr srcDevice, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpyToArrayAsync */ CUresult CUDAAPI cuMemcpyHtoAAsync(CUarray dstArray, size_t dstOffset, const void *srcHost, size_t ByteCount, CUstream hStream); @@ -5925,7 +6719,8 @@ CUresult CUDAAPI cuMemcpyHtoAAsync(CUarray dstArray, size_t dstOffset, const voi * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -5941,7 +6736,8 @@ CUresult CUDAAPI cuMemcpyHtoAAsync(CUarray dstArray, size_t dstOffset, const voi * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpyFromArrayAsync */ CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, size_t srcOffset, size_t ByteCount, CUstream hStream); @@ -5989,9 +6785,9 @@ CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, size_t srcOf * * \par * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::srcArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::srcArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -6006,9 +6802,9 @@ CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, size_t srcOf * * \par * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::dstArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::dstArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -6090,7 +6886,8 @@ CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, size_t srcOf * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -6106,7 +6903,10 @@ CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, size_t srcOf * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpy2DAsync, + * ::cudaMemcpy2DToArrayAsync, + * ::cudaMemcpy2DFromArrayAsync */ CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D *pCopy, CUstream hStream); @@ -6157,9 +6957,9 @@ CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D *pCopy, CUstream hStream); * * \par * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::srcArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::srcArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -6181,9 +6981,9 @@ CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D *pCopy, CUstream hStream); * * \par * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch - * specify the (unified virtual address space) base address of the source data - * and the bytes per row to apply. ::dstArray is ignored. - * This value may be used only if unified addressing is supported in the calling + * specify the (unified virtual address space) base address of the source data + * and the bytes per row to apply. ::dstArray is ignored. + * This value may be used only if unified addressing is supported in the calling * context. * * \par @@ -6262,7 +7062,8 @@ CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D *pCopy, CUstream hStream); * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * \note_async * \note_null_stream @@ -6278,7 +7079,8 @@ CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D *pCopy, CUstream hStream); * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemcpy3DAsync */ CUresult CUDAAPI cuMemcpy3DAsync(const CUDA_MEMCPY3D *pCopy, CUstream hStream); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -6305,7 +7107,8 @@ CUresult CUDAAPI cuMemcpy3DAsync(const CUDA_MEMCPY3D *pCopy, CUstream hStream); * \note_null_stream * * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpyDtoDAsync, ::cuMemcpyPeerAsync, - * ::cuMemcpy3DPeerAsync + * ::cuMemcpy3DPeerAsync, + * ::cudaMemcpy3DPeerAsync */ CUresult CUDAAPI cuMemcpy3DPeerAsync(const CUDA_MEMCPY3D_PEER *pCopy, CUstream hStream); #endif /* __CUDA_API_VERSION >= 4000 */ @@ -6341,7 +7144,8 @@ CUresult CUDAAPI cuMemcpy3DPeerAsync(const CUDA_MEMCPY3D_PEER *pCopy, CUstream h * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset */ CUresult CUDAAPI cuMemsetD8(CUdeviceptr dstDevice, unsigned char uc, size_t N); @@ -6375,7 +7179,8 @@ CUresult CUDAAPI cuMemsetD8(CUdeviceptr dstDevice, unsigned char uc, size_t N); * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset */ CUresult CUDAAPI cuMemsetD16(CUdeviceptr dstDevice, unsigned short us, size_t N); @@ -6409,7 +7214,8 @@ CUresult CUDAAPI cuMemsetD16(CUdeviceptr dstDevice, unsigned short us, size_t N) * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32Async + * ::cuMemsetD32Async, + * ::cudaMemset */ CUresult CUDAAPI cuMemsetD32(CUdeviceptr dstDevice, unsigned int ui, size_t N); @@ -6448,7 +7254,8 @@ CUresult CUDAAPI cuMemsetD32(CUdeviceptr dstDevice, unsigned int ui, size_t N); * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset2D */ CUresult CUDAAPI cuMemsetD2D8(CUdeviceptr dstDevice, size_t dstPitch, unsigned char uc, size_t Width, size_t Height); @@ -6488,7 +7295,8 @@ CUresult CUDAAPI cuMemsetD2D8(CUdeviceptr dstDevice, size_t dstPitch, unsigned c * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset2D */ CUresult CUDAAPI cuMemsetD2D16(CUdeviceptr dstDevice, size_t dstPitch, unsigned short us, size_t Width, size_t Height); @@ -6528,7 +7336,8 @@ CUresult CUDAAPI cuMemsetD2D16(CUdeviceptr dstDevice, size_t dstPitch, unsigned * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset2D */ CUresult CUDAAPI cuMemsetD2D32(CUdeviceptr dstDevice, size_t dstPitch, unsigned int ui, size_t Width, size_t Height); @@ -6564,7 +7373,8 @@ CUresult CUDAAPI cuMemsetD2D32(CUdeviceptr dstDevice, size_t dstPitch, unsigned * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemsetAsync */ CUresult CUDAAPI cuMemsetD8Async(CUdeviceptr dstDevice, unsigned char uc, size_t N, CUstream hStream); @@ -6600,7 +7410,8 @@ CUresult CUDAAPI cuMemsetD8Async(CUdeviceptr dstDevice, unsigned char uc, size_t * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemsetAsync */ CUresult CUDAAPI cuMemsetD16Async(CUdeviceptr dstDevice, unsigned short us, size_t N, CUstream hStream); @@ -6635,7 +7446,8 @@ CUresult CUDAAPI cuMemsetD16Async(CUdeviceptr dstDevice, unsigned short us, size * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, - * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, ::cuMemsetD32 + * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, ::cuMemsetD32, + * ::cudaMemsetAsync */ CUresult CUDAAPI cuMemsetD32Async(CUdeviceptr dstDevice, unsigned int ui, size_t N, CUstream hStream); @@ -6676,7 +7488,8 @@ CUresult CUDAAPI cuMemsetD32Async(CUdeviceptr dstDevice, unsigned int ui, size_t * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset2DAsync */ CUresult CUDAAPI cuMemsetD2D8Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned char uc, size_t Width, size_t Height, CUstream hStream); @@ -6685,7 +7498,7 @@ CUresult CUDAAPI cuMemsetD2D8Async(CUdeviceptr dstDevice, size_t dstPitch, unsig * * Sets the 2D memory range of \p Width 16-bit values to the specified value * \p us. \p Height specifies the number of rows to set, and \p dstPitch - * specifies the number of bytes between each row. The \p dstDevice pointer + * specifies the number of bytes between each row. The \p dstDevice pointer * and \p dstPitch offset must be two byte aligned. This function performs * fastest when the pitch is one that has been passed back by * ::cuMemAllocPitch(). @@ -6718,7 +7531,8 @@ CUresult CUDAAPI cuMemsetD2D8Async(CUdeviceptr dstDevice, size_t dstPitch, unsig * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D32, ::cuMemsetD2D32Async, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset2DAsync */ CUresult CUDAAPI cuMemsetD2D16Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned short us, size_t Width, size_t Height, CUstream hStream); @@ -6760,7 +7574,8 @@ CUresult CUDAAPI cuMemsetD2D16Async(CUdeviceptr dstDevice, size_t dstPitch, unsi * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async, * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, - * ::cuMemsetD32, ::cuMemsetD32Async + * ::cuMemsetD32, ::cuMemsetD32Async, + * ::cudaMemset2DAsync */ CUresult CUDAAPI cuMemsetD2D32Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned int ui, size_t Width, size_t Height, CUstream hStream); @@ -6863,7 +7678,8 @@ CUresult CUDAAPI cuMemsetD2D32Async(CUdeviceptr dstDevice, size_t dstPitch, unsi * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMallocArray */ CUresult CUDAAPI cuArrayCreate(CUarray *pHandle, const CUDA_ARRAY_DESCRIPTOR *pAllocateArray); @@ -6896,7 +7712,8 @@ CUresult CUDAAPI cuArrayCreate(CUarray *pHandle, const CUDA_ARRAY_DESCRIPTOR *pA * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaArrayGetInfo */ CUresult CUDAAPI cuArrayGetDescriptor(CUDA_ARRAY_DESCRIPTOR *pArrayDescriptor, CUarray hArray); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -6915,7 +7732,8 @@ CUresult CUDAAPI cuArrayGetDescriptor(CUDA_ARRAY_DESCRIPTOR *pArrayDescriptor, C * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_HANDLE, - * ::CUDA_ERROR_ARRAY_IS_MAPPED + * ::CUDA_ERROR_ARRAY_IS_MAPPED, + * ::CUDA_ERROR_CONTEXT_IS_DESTROYED * \notefnerr * * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate, @@ -6927,7 +7745,8 @@ CUresult CUDAAPI cuArrayGetDescriptor(CUDA_ARRAY_DESCRIPTOR *pArrayDescriptor, C * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaFreeArray */ CUresult CUDAAPI cuArrayDestroy(CUarray hArray); @@ -6956,22 +7775,22 @@ CUresult CUDAAPI cuArrayDestroy(CUarray hArray); * - A 1D array is allocated if \p Height and \p Depth extents are both zero. * - A 2D array is allocated if only \p Depth extent is zero. * - A 3D array is allocated if all three extents are non-zero. - * - A 1D layered CUDA array is allocated if only \p Height is zero and the - * ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 1D array. The number + * - A 1D layered CUDA array is allocated if only \p Height is zero and the + * ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 1D array. The number * of layers is determined by the depth extent. - * - A 2D layered CUDA array is allocated if all three extents are non-zero and - * the ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 2D array. The number + * - A 2D layered CUDA array is allocated if all three extents are non-zero and + * the ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 2D array. The number * of layers is determined by the depth extent. * - A cubemap CUDA array is allocated if all three extents are non-zero and the - * ::CUDA_ARRAY3D_CUBEMAP flag is set. \p Width must be equal to \p Height, and - * \p Depth must be six. A cubemap is a special type of 2D layered CUDA array, - * where the six layers represent the six faces of a cube. The order of the six + * ::CUDA_ARRAY3D_CUBEMAP flag is set. \p Width must be equal to \p Height, and + * \p Depth must be six. A cubemap is a special type of 2D layered CUDA array, + * where the six layers represent the six faces of a cube. The order of the six * layers in memory is the same as that listed in ::CUarray_cubemap_face. - * - A cubemap layered CUDA array is allocated if all three extents are non-zero, - * and both, ::CUDA_ARRAY3D_CUBEMAP and ::CUDA_ARRAY3D_LAYERED flags are set. - * \p Width must be equal to \p Height, and \p Depth must be a multiple of six. - * A cubemap layered CUDA array is a special type of 2D layered CUDA array that - * consists of a collection of cubemaps. The first six layers represent the first + * - A cubemap layered CUDA array is allocated if all three extents are non-zero, + * and both, ::CUDA_ARRAY3D_CUBEMAP and ::CUDA_ARRAY3D_LAYERED flags are set. + * \p Width must be equal to \p Height, and \p Depth must be a multiple of six. + * A cubemap layered CUDA array is a special type of 2D layered CUDA array that + * consists of a collection of cubemaps. The first six layers represent the first * cubemap, the next six layers form the second cubemap, and so on. * * - ::Format specifies the format of the elements; ::CUarray_format is @@ -6992,11 +7811,11 @@ CUresult CUDAAPI cuArrayDestroy(CUarray hArray); * - \p NumChannels specifies the number of packed components per CUDA array * element; it may be 1, 2, or 4; * - * - ::Flags may be set to - * - ::CUDA_ARRAY3D_LAYERED to enable creation of layered CUDA arrays. If this flag is set, + * - ::Flags may be set to + * - ::CUDA_ARRAY3D_LAYERED to enable creation of layered CUDA arrays. If this flag is set, * \p Depth specifies the number of layers, not the depth of a 3D array. - * - ::CUDA_ARRAY3D_SURFACE_LDST to enable surface references to be bound to the CUDA array. - * If this flag is not set, ::cuSurfRefSetArray will fail when attempting to bind the CUDA array + * - ::CUDA_ARRAY3D_SURFACE_LDST to enable surface references to be bound to the CUDA array. + * If this flag is not set, ::cuSurfRefSetArray will fail when attempting to bind the CUDA array * to a surface reference. * - ::CUDA_ARRAY3D_CUBEMAP to enable creation of cubemaps. If this flag is set, \p Width must be * equal to \p Height, and \p Depth must be six. If the ::CUDA_ARRAY3D_LAYERED flag is also set, @@ -7004,20 +7823,20 @@ CUresult CUDAAPI cuArrayDestroy(CUarray hArray); * - ::CUDA_ARRAY3D_TEXTURE_GATHER to indicate that the CUDA array will be used for texture gather. * Texture gather can only be performed on 2D CUDA arrays. * - * \p Width, \p Height and \p Depth must meet certain size requirements as listed in the following table. - * All values are specified in elements. Note that for brevity's sake, the full name of the device attribute - * is not specified. For ex., TEXTURE1D_WIDTH refers to the device attribute + * \p Width, \p Height and \p Depth must meet certain size requirements as listed in the following table. + * All values are specified in elements. Note that for brevity's sake, the full name of the device attribute + * is not specified. For ex., TEXTURE1D_WIDTH refers to the device attribute * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH. * - * Note that 2D CUDA arrays have different size requirements if the ::CUDA_ARRAY3D_TEXTURE_GATHER flag - * is set. \p Width and \p Height must not be greater than ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_WIDTH + * Note that 2D CUDA arrays have different size requirements if the ::CUDA_ARRAY3D_TEXTURE_GATHER flag + * is set. \p Width and \p Height must not be greater than ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_WIDTH * and ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_HEIGHT respectively, in that case. * * * - * - * * * @@ -7027,28 +7846,28 @@ CUresult CUDAAPI cuArrayDestroy(CUarray hArray); * * * - * * - * - * * - * - * * * - * * - * - * *
CUDA array typeValid extents that must always be met
{(width range in elements), (height range), + * 		Valid extents that must always be met
{(width range in elements), (height range), * (depth range)}		Valid extents with CUDA_ARRAY3D_SURFACE_LDST set
+ * 		Valid extents with CUDA_ARRAY3D_SURFACE_LDST set
* {(width range in elements), (height range), (depth range)}
1D{ (1,TEXTURE1D_WIDTH), 0, 0 }{ (1,SURFACE2D_WIDTH), (1,SURFACE2D_HEIGHT), 0 }
3D{ (1,TEXTURE3D_WIDTH), (1,TEXTURE3D_HEIGHT), (1,TEXTURE3D_DEPTH) } - *
OR
{ (1,TEXTURE3D_WIDTH_ALTERNATE), (1,TEXTURE3D_HEIGHT_ALTERNATE), + *
OR
{ (1,TEXTURE3D_WIDTH_ALTERNATE), (1,TEXTURE3D_HEIGHT_ALTERNATE), * (1,TEXTURE3D_DEPTH_ALTERNATE) }		{ (1,SURFACE3D_WIDTH), (1,SURFACE3D_HEIGHT), + * { (1,SURFACE3D_WIDTH), (1,SURFACE3D_HEIGHT), * (1,SURFACE3D_DEPTH) }
1D Layered{ (1,TEXTURE1D_LAYERED_WIDTH), 0, + * { (1,TEXTURE1D_LAYERED_WIDTH), 0, * (1,TEXTURE1D_LAYERED_LAYERS) }{ (1,SURFACE1D_LAYERED_WIDTH), 0, + * { (1,SURFACE1D_LAYERED_WIDTH), 0, * (1,SURFACE1D_LAYERED_LAYERS) }
2D Layered{ (1,TEXTURE2D_LAYERED_WIDTH), (1,TEXTURE2D_LAYERED_HEIGHT), + * { (1,TEXTURE2D_LAYERED_WIDTH), (1,TEXTURE2D_LAYERED_HEIGHT), * (1,TEXTURE2D_LAYERED_LAYERS) }{ (1,SURFACE2D_LAYERED_WIDTH), (1,SURFACE2D_LAYERED_HEIGHT), + * { (1,SURFACE2D_LAYERED_WIDTH), (1,SURFACE2D_LAYERED_HEIGHT), * (1,SURFACE2D_LAYERED_LAYERS) }
Cubemap{ (1,TEXTURECUBEMAP_WIDTH), (1,TEXTURECUBEMAP_WIDTH), 6 }{ (1,SURFACECUBEMAP_WIDTH), + * { (1,SURFACECUBEMAP_WIDTH), * (1,SURFACECUBEMAP_WIDTH), 6 }
Cubemap Layered{ (1,TEXTURECUBEMAP_LAYERED_WIDTH), (1,TEXTURECUBEMAP_LAYERED_WIDTH), + * { (1,TEXTURECUBEMAP_LAYERED_WIDTH), (1,TEXTURECUBEMAP_LAYERED_WIDTH), * (1,TEXTURECUBEMAP_LAYERED_LAYERS) }{ (1,SURFACECUBEMAP_LAYERED_WIDTH), (1,SURFACECUBEMAP_LAYERED_WIDTH), + * { (1,SURFACECUBEMAP_LAYERED_WIDTH), (1,SURFACECUBEMAP_LAYERED_WIDTH), * (1,SURFACECUBEMAP_LAYERED_LAYERS) }
* @@ -7107,7 +7926,8 @@ CUresult CUDAAPI cuArrayDestroy(CUarray hArray); * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaMalloc3DArray */ CUresult CUDAAPI cuArray3DCreate(CUarray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR *pAllocateArray); @@ -7131,7 +7951,8 @@ CUresult CUDAAPI cuArray3DCreate(CUarray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_INVALID_HANDLE + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_CONTEXT_IS_DESTROYED * \notefnerr * * \sa ::cuArray3DCreate, ::cuArrayCreate, @@ -7143,7 +7964,8 @@ CUresult CUDAAPI cuArray3DCreate(CUarray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost, * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc, * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16, - * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32 + * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32, + * ::cudaArrayGetInfo */ CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR *pArrayDescriptor, CUarray hArray); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -7177,22 +7999,22 @@ CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR *pArrayDescripto * - A 1D mipmapped array is allocated if \p Height and \p Depth extents are both zero. * - A 2D mipmapped array is allocated if only \p Depth extent is zero. * - A 3D mipmapped array is allocated if all three extents are non-zero. - * - A 1D layered CUDA mipmapped array is allocated if only \p Height is zero and the - * ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 1D array. The number + * - A 1D layered CUDA mipmapped array is allocated if only \p Height is zero and the + * ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 1D array. The number * of layers is determined by the depth extent. - * - A 2D layered CUDA mipmapped array is allocated if all three extents are non-zero and - * the ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 2D array. The number + * - A 2D layered CUDA mipmapped array is allocated if all three extents are non-zero and + * the ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 2D array. The number * of layers is determined by the depth extent. * - A cubemap CUDA mipmapped array is allocated if all three extents are non-zero and the - * ::CUDA_ARRAY3D_CUBEMAP flag is set. \p Width must be equal to \p Height, and - * \p Depth must be six. A cubemap is a special type of 2D layered CUDA array, - * where the six layers represent the six faces of a cube. The order of the six + * ::CUDA_ARRAY3D_CUBEMAP flag is set. \p Width must be equal to \p Height, and + * \p Depth must be six. A cubemap is a special type of 2D layered CUDA array, + * where the six layers represent the six faces of a cube. The order of the six * layers in memory is the same as that listed in ::CUarray_cubemap_face. - * - A cubemap layered CUDA mipmapped array is allocated if all three extents are non-zero, - * and both, ::CUDA_ARRAY3D_CUBEMAP and ::CUDA_ARRAY3D_LAYERED flags are set. - * \p Width must be equal to \p Height, and \p Depth must be a multiple of six. - * A cubemap layered CUDA array is a special type of 2D layered CUDA array that - * consists of a collection of cubemaps. The first six layers represent the first + * - A cubemap layered CUDA mipmapped array is allocated if all three extents are non-zero, + * and both, ::CUDA_ARRAY3D_CUBEMAP and ::CUDA_ARRAY3D_LAYERED flags are set. + * \p Width must be equal to \p Height, and \p Depth must be a multiple of six. + * A cubemap layered CUDA array is a special type of 2D layered CUDA array that + * consists of a collection of cubemaps. The first six layers represent the first * cubemap, the next six layers form the second cubemap, and so on. * * - ::Format specifies the format of the elements; ::CUarray_format is @@ -7213,11 +8035,11 @@ CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR *pArrayDescripto * - \p NumChannels specifies the number of packed components per CUDA array * element; it may be 1, 2, or 4; * - * - ::Flags may be set to - * - ::CUDA_ARRAY3D_LAYERED to enable creation of layered CUDA mipmapped arrays. If this flag is set, + * - ::Flags may be set to + * - ::CUDA_ARRAY3D_LAYERED to enable creation of layered CUDA mipmapped arrays. If this flag is set, * \p Depth specifies the number of layers, not the depth of a 3D array. * - ::CUDA_ARRAY3D_SURFACE_LDST to enable surface references to be bound to individual mipmap levels of - * the CUDA mipmapped array. If this flag is not set, ::cuSurfRefSetArray will fail when attempting to + * the CUDA mipmapped array. If this flag is not set, ::cuSurfRefSetArray will fail when attempting to * bind a mipmap level of the CUDA mipmapped array to a surface reference. * - ::CUDA_ARRAY3D_CUBEMAP to enable creation of mipmapped cubemaps. If this flag is set, \p Width must be * equal to \p Height, and \p Depth must be six. If the ::CUDA_ARRAY3D_LAYERED flag is also set, @@ -7225,34 +8047,48 @@ CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR *pArrayDescripto * - ::CUDA_ARRAY3D_TEXTURE_GATHER to indicate that the CUDA mipmapped array will be used for texture gather. * Texture gather can only be performed on 2D CUDA mipmapped arrays. * - * \p Width, \p Height and \p Depth must meet certain size requirements as listed in the following table. - * All values are specified in elements. Note that for brevity's sake, the full name of the device attribute - * is not specified. For ex., TEXTURE1D_MIPMAPPED_WIDTH refers to the device attribute + * \p Width, \p Height and \p Depth must meet certain size requirements as listed in the following table. + * All values are specified in elements. Note that for brevity's sake, the full name of the device attribute + * is not specified. For ex., TEXTURE1D_MIPMAPPED_WIDTH refers to the device attribute * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH. * * * - * + * + * * - * + * + * * - * + * + * * * + *
OR
{ (1,TEXTURE3D_WIDTH_ALTERNATE), (1,TEXTURE3D_HEIGHT_ALTERNATE), + * (1,TEXTURE3D_DEPTH_ALTERNATE) } + * * - * + * + * * - * + * + * * - * + * + * * - * + * + * *
CUDA array typeValid extents that must always be met
{(width range in elements), (height range), - * (depth range)}
Valid extents that must always be met
{(width range in elements), (height range), + * (depth range)}		Valid extents with CUDA_ARRAY3D_SURFACE_LDST set
+ * {(width range in elements), (height range), (depth range)}
1D{ (1,TEXTURE1D_MIPMAPPED_WIDTH), 0, 0 }
{ (1,TEXTURE1D_MIPMAPPED_WIDTH), 0, 0 }{ (1,SURFACE1D_WIDTH), 0, 0 }
2D{ (1,TEXTURE2D_MIPMAPPED_WIDTH), (1,TEXTURE2D_MIPMAPPED_HEIGHT), 0 }
{ (1,TEXTURE2D_MIPMAPPED_WIDTH), (1,TEXTURE2D_MIPMAPPED_HEIGHT), 0 }{ (1,SURFACE2D_WIDTH), (1,SURFACE2D_HEIGHT), 0 }
3D{ (1,TEXTURE3D_WIDTH), (1,TEXTURE3D_HEIGHT), (1,TEXTURE3D_DEPTH) } - *
OR
{ (1,TEXTURE3D_WIDTH_ALTERNATE), (1,TEXTURE3D_HEIGHT_ALTERNATE), - * (1,TEXTURE3D_DEPTH_ALTERNATE) }
{ (1,SURFACE3D_WIDTH), (1,SURFACE3D_HEIGHT), + * (1,SURFACE3D_DEPTH) }
1D Layered{ (1,TEXTURE1D_LAYERED_WIDTH), 0, - * (1,TEXTURE1D_LAYERED_LAYERS) }
{ (1,TEXTURE1D_LAYERED_WIDTH), 0, + * (1,TEXTURE1D_LAYERED_LAYERS) }{ (1,SURFACE1D_LAYERED_WIDTH), 0, + * (1,SURFACE1D_LAYERED_LAYERS) }
2D Layered{ (1,TEXTURE2D_LAYERED_WIDTH), (1,TEXTURE2D_LAYERED_HEIGHT), - * (1,TEXTURE2D_LAYERED_LAYERS) }
{ (1,TEXTURE2D_LAYERED_WIDTH), (1,TEXTURE2D_LAYERED_HEIGHT), + * (1,TEXTURE2D_LAYERED_LAYERS) }{ (1,SURFACE2D_LAYERED_WIDTH), (1,SURFACE2D_LAYERED_HEIGHT), + * (1,SURFACE2D_LAYERED_LAYERS) }
Cubemap{ (1,TEXTURECUBEMAP_WIDTH), (1,TEXTURECUBEMAP_WIDTH), 6 }
{ (1,TEXTURECUBEMAP_WIDTH), (1,TEXTURECUBEMAP_WIDTH), 6 }{ (1,SURFACECUBEMAP_WIDTH), + * (1,SURFACECUBEMAP_WIDTH), 6 }
Cubemap Layered{ (1,TEXTURECUBEMAP_LAYERED_WIDTH), (1,TEXTURECUBEMAP_LAYERED_WIDTH), - * (1,TEXTURECUBEMAP_LAYERED_LAYERS) }
{ (1,TEXTURECUBEMAP_LAYERED_WIDTH), (1,TEXTURECUBEMAP_LAYERED_WIDTH), + * (1,TEXTURECUBEMAP_LAYERED_LAYERS) }{ (1,SURFACECUBEMAP_LAYERED_WIDTH), (1,SURFACECUBEMAP_LAYERED_WIDTH), + * (1,SURFACECUBEMAP_LAYERED_LAYERS) }
* * @@ -7270,7 +8106,11 @@ CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR *pArrayDescripto * ::CUDA_ERROR_UNKNOWN * \notefnerr * - * \sa ::cuMipmappedArrayDestroy, ::cuMipmappedArrayGetLevel, ::cuArrayCreate, + * \sa + * ::cuMipmappedArrayDestroy, + * ::cuMipmappedArrayGetLevel, + * ::cuArrayCreate, + * ::cudaMallocMipmappedArray */ CUresult CUDAAPI cuMipmappedArrayCreate(CUmipmappedArray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR *pMipmappedArrayDesc, unsigned int numMipmapLevels); @@ -7296,7 +8136,11 @@ CUresult CUDAAPI cuMipmappedArrayCreate(CUmipmappedArray *pHandle, const CUDA_AR * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * - * \sa ::cuMipmappedArrayCreate, ::cuMipmappedArrayDestroy, ::cuArrayCreate, + * \sa + * ::cuMipmappedArrayCreate, + * ::cuMipmappedArrayDestroy, + * ::cuArrayCreate, + * ::cudaGetMipmappedArrayLevel */ CUresult CUDAAPI cuMipmappedArrayGetLevel(CUarray *pLevelArray, CUmipmappedArray hMipmappedArray, unsigned int level); @@ -7313,10 +8157,15 @@ CUresult CUDAAPI cuMipmappedArrayGetLevel(CUarray *pLevelArray, CUmipmappedArray * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_HANDLE, - * ::CUDA_ERROR_ARRAY_IS_MAPPED + * ::CUDA_ERROR_ARRAY_IS_MAPPED, + * ::CUDA_ERROR_CONTEXT_IS_DESTROYED * \notefnerr * - * \sa ::cuMipmappedArrayCreate, ::cuMipmappedArrayGetLevel, ::cuArrayCreate, + * \sa + * ::cuMipmappedArrayCreate, + * ::cuMipmappedArrayGetLevel, + * ::cuArrayCreate, + * ::cudaFreeMipmappedArray */ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); @@ -7330,43 +8179,42 @@ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); * ___MANBRIEF___ unified addressing functions of the low-level CUDA driver * API (___CURRENT_FILE___) ___ENDMANBRIEF___ * - * This section describes the unified addressing functions of the + * This section describes the unified addressing functions of the * low-level CUDA driver application programming interface. * * @{ * * \section CUDA_UNIFIED_overview Overview * - * CUDA devices can share a unified address space with the host. + * CUDA devices can share a unified address space with the host. * For these devices there is no distinction between a device - * pointer and a host pointer -- the same pointer value may be - * used to access memory from the host program and from a kernel + * pointer and a host pointer -- the same pointer value may be + * used to access memory from the host program and from a kernel * running on the device (with exceptions enumerated below). * * \section CUDA_UNIFIED_support Supported Platforms - * - * Whether or not a device supports unified addressing may be - * queried by calling ::cuDeviceGetAttribute() with the device + * + * Whether or not a device supports unified addressing may be + * queried by calling ::cuDeviceGetAttribute() with the device * attribute ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING. * - * Unified addressing is automatically enabled in 64-bit processes - * on devices with compute capability greater than or equal to 2.0. + * Unified addressing is automatically enabled in 64-bit processes * * \section CUDA_UNIFIED_lookup Looking Up Information from Pointer Values * - * It is possible to look up information about the memory which backs a + * It is possible to look up information about the memory which backs a * pointer value. For instance, one may want to know if a pointer points - * to host or device memory. As another example, in the case of device - * memory, one may want to know on which CUDA device the memory - * resides. These properties may be queried using the function + * to host or device memory. As another example, in the case of device + * memory, one may want to know on which CUDA device the memory + * resides. These properties may be queried using the function * ::cuPointerGetAttribute() * * Since pointers are unique, it is not necessary to specify information - * about the pointers specified to the various copy functions in the + * about the pointers specified to the various copy functions in the * CUDA API. The function ::cuMemcpy() may be used to perform a copy * between two pointers, ignoring whether they point to host or device * memory (making ::cuMemcpyHtoD(), ::cuMemcpyDtoD(), and ::cuMemcpyDtoH() - * unnecessary for devices supporting unified addressing). For + * unnecessary for devices supporting unified addressing). For * multidimensional copies, the memory type ::CU_MEMORYTYPE_UNIFIED may be * used to specify that the CUDA driver should infer the location of the * pointer from its value. @@ -7375,45 +8223,45 @@ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); * * All host memory allocated in all contexts using ::cuMemAllocHost() and * ::cuMemHostAlloc() is always directly accessible from all contexts on - * all devices that support unified addressing. This is the case regardless + * all devices that support unified addressing. This is the case regardless * of whether or not the flags ::CU_MEMHOSTALLOC_PORTABLE and * ::CU_MEMHOSTALLOC_DEVICEMAP are specified. * - * The pointer value through which allocated host memory may be accessed - * in kernels on all devices that support unified addressing is the same + * The pointer value through which allocated host memory may be accessed + * in kernels on all devices that support unified addressing is the same * as the pointer value through which that memory is accessed on the host, - * so it is not necessary to call ::cuMemHostGetDevicePointer() to get the device + * so it is not necessary to call ::cuMemHostGetDevicePointer() to get the device * pointer for these allocations. - * + * * Note that this is not the case for memory allocated using the flag * ::CU_MEMHOSTALLOC_WRITECOMBINED, as discussed below. * * \section CUDA_UNIFIED_autopeerregister Automatic Registration of Peer Memory * - * Upon enabling direct access from a context that supports unified addressing - * to another peer context that supports unified addressing using - * ::cuCtxEnablePeerAccess() all memory allocated in the peer context using - * ::cuMemAlloc() and ::cuMemAllocPitch() will immediately be accessible + * Upon enabling direct access from a context that supports unified addressing + * to another peer context that supports unified addressing using + * ::cuCtxEnablePeerAccess() all memory allocated in the peer context using + * ::cuMemAlloc() and ::cuMemAllocPitch() will immediately be accessible * by the current context. The device pointer value through * which any peer memory may be accessed in the current context * is the same pointer value through which that memory may be * accessed in the peer context. * * \section CUDA_UNIFIED_exceptions Exceptions, Disjoint Addressing - * + * * Not all memory may be accessed on devices through the same pointer * value through which they are accessed on the host. These exceptions * are host memory registered using ::cuMemHostRegister() and host memory - * allocated using the flag ::CU_MEMHOSTALLOC_WRITECOMBINED. For these + * allocated using the flag ::CU_MEMHOSTALLOC_WRITECOMBINED. For these * exceptions, there exists a distinct host and device address for the * memory. The device address is guaranteed to not overlap any valid host - * pointer range and is guaranteed to have the same value across all - * contexts that support unified addressing. - * - * This device address may be queried using ::cuMemHostGetDevicePointer() - * when a context using unified addressing is current. Either the host - * or the unified device pointer value may be used to refer to this memory - * through ::cuMemcpy() and similar functions using the + * pointer range and is guaranteed to have the same value across all + * contexts that support unified addressing. + * + * This device address may be queried using ::cuMemHostGetDevicePointer() + * when a context using unified addressing is current. Either the host + * or the unified device pointer value may be used to refer to this memory + * through ::cuMemcpy() and similar functions using the * ::CU_MEMORYTYPE_UNIFIED memory type. * */ @@ -7421,69 +8269,69 @@ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); #if __CUDA_API_VERSION >= 4000 /** * \brief Returns information about a pointer - * + * * The supported attributes are: - * - * - ::CU_POINTER_ATTRIBUTE_CONTEXT: - * - * Returns in \p *data the ::CUcontext in which \p ptr was allocated or - * registered. - * The type of \p data must be ::CUcontext *. - * + * + * - ::CU_POINTER_ATTRIBUTE_CONTEXT: + * + * Returns in \p *data the ::CUcontext in which \p ptr was allocated or + * registered. + * The type of \p data must be ::CUcontext *. + * * If \p ptr was not allocated by, mapped by, or registered with - * a ::CUcontext which uses unified virtual addressing then + * a ::CUcontext which uses unified virtual addressing then * ::CUDA_ERROR_INVALID_VALUE is returned. - * - * - ::CU_POINTER_ATTRIBUTE_MEMORY_TYPE: - * - * Returns in \p *data the physical memory type of the memory that + * + * - ::CU_POINTER_ATTRIBUTE_MEMORY_TYPE: + * + * Returns in \p *data the physical memory type of the memory that * \p ptr addresses as a ::CUmemorytype enumerated value. * The type of \p data must be unsigned int. - * - * If \p ptr addresses device memory then \p *data is set to - * ::CU_MEMORYTYPE_DEVICE. The particular ::CUdevice on which the - * memory resides is the ::CUdevice of the ::CUcontext returned by the + * + * If \p ptr addresses device memory then \p *data is set to + * ::CU_MEMORYTYPE_DEVICE. The particular ::CUdevice on which the + * memory resides is the ::CUdevice of the ::CUcontext returned by the * ::CU_POINTER_ATTRIBUTE_CONTEXT attribute of \p ptr. - * - * If \p ptr addresses host memory then \p *data is set to + * + * If \p ptr addresses host memory then \p *data is set to * ::CU_MEMORYTYPE_HOST. - * + * * If \p ptr was not allocated by, mapped by, or registered with - * a ::CUcontext which uses unified virtual addressing then + * a ::CUcontext which uses unified virtual addressing then * ::CUDA_ERROR_INVALID_VALUE is returned. * - * If the current ::CUcontext does not support unified virtual + * If the current ::CUcontext does not support unified virtual * addressing then ::CUDA_ERROR_INVALID_CONTEXT is returned. - * + * * - ::CU_POINTER_ATTRIBUTE_DEVICE_POINTER: - * + * * Returns in \p *data the device pointer value through which - * \p ptr may be accessed by kernels running in the current + * \p ptr may be accessed by kernels running in the current * ::CUcontext. * The type of \p data must be CUdeviceptr *. - * + * * If there exists no device pointer value through which * kernels running in the current ::CUcontext may access * \p ptr then ::CUDA_ERROR_INVALID_VALUE is returned. - * - * If there is no current ::CUcontext then + * + * If there is no current ::CUcontext then * ::CUDA_ERROR_INVALID_CONTEXT is returned. - * - * Except in the exceptional disjoint addressing cases discussed - * below, the value returned in \p *data will equal the input + * + * Except in the exceptional disjoint addressing cases discussed + * below, the value returned in \p *data will equal the input * value \p ptr. - * + * * - ::CU_POINTER_ATTRIBUTE_HOST_POINTER: - * - * Returns in \p *data the host pointer value through which + * + * Returns in \p *data the host pointer value through which * \p ptr may be accessed by by the host program. * The type of \p data must be void **. * If there exists no host pointer value through which - * the host program may directly access \p ptr then + * the host program may directly access \p ptr then * ::CUDA_ERROR_INVALID_VALUE is returned. - * - * Except in the exceptional disjoint addressing cases discussed - * below, the value returned in \p *data will equal the input + * + * Except in the exceptional disjoint addressing cases discussed + * below, the value returned in \p *data will equal the input * value \p ptr. * * - ::CU_POINTER_ATTRIBUTE_P2P_TOKENS: @@ -7499,7 +8347,7 @@ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); * Querying this attribute has a side effect of setting the attribute * ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS for the region of memory that * \p ptr points to. - * + * * - ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS: * * A boolean attribute which when set, ensures that synchronous memory operations @@ -7524,22 +8372,27 @@ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); * Returns in \p *data a boolean that indicates whether the pointer points to * managed memory or not. * + * - ::CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL: + * + * Returns in \p *data an integer representing a device ordinal of a device against + * which the memory was allocated or registered. + * * \par * * Note that for most allocations in the unified virtual address space - * the host and device pointer for accessing the allocation will be the + * the host and device pointer for accessing the allocation will be the * same. The exceptions to this are - * - user memory registered using ::cuMemHostRegister - * - host memory allocated using ::cuMemHostAlloc with the + * - user memory registered using ::cuMemHostRegister + * - host memory allocated using ::cuMemHostAlloc with the * ::CU_MEMHOSTALLOC_WRITECOMBINED flag - * For these types of allocation there will exist separate, disjoint host - * and device addresses for accessing the allocation. In particular - * - The host address will correspond to an invalid unmapped device address - * (which will result in an exception if accessed from the device) - * - The device address will correspond to an invalid unmapped host address + * For these types of allocation there will exist separate, disjoint host + * and device addresses for accessing the allocation. In particular + * - The host address will correspond to an invalid unmapped device address + * (which will result in an exception if accessed from the device) + * - The device address will correspond to an invalid unmapped host address * (which will result in an exception if accessed from the host). - * For these types of allocations, querying ::CU_POINTER_ATTRIBUTE_HOST_POINTER - * and ::CU_POINTER_ATTRIBUTE_DEVICE_POINTER may be used to retrieve the host + * For these types of allocations, querying ::CU_POINTER_ATTRIBUTE_HOST_POINTER + * and ::CU_POINTER_ATTRIBUTE_DEVICE_POINTER may be used to retrieve the host * and device addresses from either address. * * \param data - Returned pointer attribute value @@ -7555,14 +8408,16 @@ CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray); * ::CUDA_ERROR_INVALID_DEVICE * \notefnerr * - * \sa cuPointerSetAttribute, + * \sa + * ::cuPointerSetAttribute, * ::cuMemAlloc, * ::cuMemFree, * ::cuMemAllocHost, * ::cuMemFreeHost, * ::cuMemHostAlloc, * ::cuMemHostRegister, - * ::cuMemHostUnregister + * ::cuMemHostUnregister, + * ::cudaPointerGetAttributes */ CUresult CUDAAPI cuPointerGetAttribute(void *data, CUpointer_attribute attribute, CUdeviceptr ptr); #endif /* __CUDA_API_VERSION >= 4000 */ @@ -7571,8 +8426,8 @@ CUresult CUDAAPI cuPointerGetAttribute(void *data, CUpointer_attribute attribute /** * \brief Prefetches memory to the specified destination device * - * Prefetches memory to the specified destination device. \p devPtr is the - * base device pointer of the memory to be prefetched and \p dstDevice is the + * Prefetches memory to the specified destination device. \p devPtr is the + * base device pointer of the memory to be prefetched and \p dstDevice is the * destination device. \p count specifies the number of bytes to copy. \p hStream * is the stream in which the operation is enqueued. The memory range must refer * to managed memory allocated via ::cuMemAllocManaged or declared via __managed__ variables. @@ -7631,7 +8486,8 @@ CUresult CUDAAPI cuPointerGetAttribute(void *data, CUpointer_attribute attribute * \note_null_stream * * \sa ::cuMemcpy, ::cuMemcpyPeer, ::cuMemcpyAsync, - * ::cuMemcpy3DPeerAsync, ::cuMemAdvise + * ::cuMemcpy3DPeerAsync, ::cuMemAdvise, + * ::cudaMemPrefetchAsync */ CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream); @@ -7642,7 +8498,10 @@ CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice d * starting at \p devPtr with a size of \p count bytes. The start address and end address of the memory * range will be rounded down and rounded up respectively to be aligned to CPU page size before the * advice is applied. The memory range must refer to managed memory allocated via ::cuMemAllocManaged - * or declared via __managed__ variables. + * or declared via __managed__ variables. The memory range could also refer to system-allocated pageable + * memory provided it represents a valid, host-accessible region of memory and all additional constraints + * imposed by \p advice as outlined below are also satisfied. Specifying an invalid system-allocated pageable + * memory range results in an error being returned. * * The \p advice parameter can take the following values: * - ::CU_MEM_ADVISE_SET_READ_MOSTLY: This implies that the data is mostly going to be read @@ -7656,11 +8515,18 @@ CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice d * Also, if a context is created on a device that does not have the device attribute * ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS set, then read-duplication will not occur until * all such contexts are destroyed. + * If the memory region refers to valid system-allocated pageable memory, then the accessing device must + * have a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS for a read-only + * copy to be created on that device. Note however that if the accessing device also has a non-zero value for the + * device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, then setting this advice + * will not create a read-only copy when that device accesses this memory region. + * * - ::CU_MEM_ADVISE_UNSET_READ_MOSTLY: Undoes the effect of ::CU_MEM_ADVISE_SET_READ_MOSTLY and also prevents the * Unified Memory driver from attempting heuristic read-duplication on the memory range. Any read-duplicated * copies of the data will be collapsed into a single copy. The location for the collapsed * copy will be the preferred location if the page has a preferred location and one of the read-duplicated * copies was resident at that location. Otherwise, the location chosen is arbitrary. + * * - ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION: This advice sets the preferred location for the * data to be the memory belonging to \p device. Passing in CU_DEVICE_CPU for \p device sets the * preferred location as host memory. If \p device is a GPU, then it must have a non-zero value for the @@ -7677,9 +8543,17 @@ CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice d * memory, the page may eventually be pinned to host memory by the Unified Memory driver. But * if the preferred location is set as device memory, then the page will continue to thrash indefinitely. * If ::CU_MEM_ADVISE_SET_READ_MOSTLY is also set on this memory region or any subset of it, then the - * policies associated with that advice will override the policies of this advice. + * policies associated with that advice will override the policies of this advice, unless read accesses from + * \p device will not result in a read-only copy being created on that device as outlined in description for + * the advice ::CU_MEM_ADVISE_SET_READ_MOSTLY. + * If the memory region refers to valid system-allocated pageable memory, then \p device must have a non-zero + * value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. Additionally, if \p device has + * a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, + * then this call has no effect. Note however that this behavior may change in the future. + * * - ::CU_MEM_ADVISE_UNSET_PREFERRED_LOCATION: Undoes the effect of ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION * and changes the preferred location to none. + * * - ::CU_MEM_ADVISE_SET_ACCESSED_BY: This advice implies that the data will be accessed by \p device. * Passing in ::CU_DEVICE_CPU for \p device will set the advice for the CPU. If \p device is a GPU, then * the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS must be non-zero. @@ -7700,8 +8574,17 @@ CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice d * policies associated with that advice will override the policies of this advice. Additionally, if the * preferred location of this memory region or any subset of it is also \p device, then the policies * associated with ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION will override the policies of this advice. + * If the memory region refers to valid system-allocated pageable memory, then \p device must have a non-zero + * value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. Additionally, if \p device has + * a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, + * then this call has no effect. + * * - ::CU_MEM_ADVISE_UNSET_ACCESSED_BY: Undoes the effect of ::CU_MEM_ADVISE_SET_ACCESSED_BY. Any mappings to * the data from \p device may be removed at any time causing accesses to result in non-fatal page faults. + * If the memory region refers to valid system-allocated pageable memory, then \p device must have a non-zero + * value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. Additionally, if \p device has + * a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, + * then this call has no effect. * * \param devPtr - Pointer to memory to set the advice for * \param count - Size in bytes of the memory range @@ -7717,13 +8600,14 @@ CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice d * \note_null_stream * * \sa ::cuMemcpy, ::cuMemcpyPeer, ::cuMemcpyAsync, - * ::cuMemcpy3DPeerAsync, ::cuMemPrefetchAsync + * ::cuMemcpy3DPeerAsync, ::cuMemPrefetchAsync, + * ::cudaMemAdvise */ CUresult CUDAAPI cuMemAdvise(CUdeviceptr devPtr, size_t count, CUmem_advise advice, CUdevice device); /** * \brief Query an attribute of a given memory range - * + * * Query an attribute about the memory range starting at \p devPtr with a size of \p count bytes. The * memory range must refer to managed memory allocated via ::cuMemAllocManaged or declared via * __managed__ variables. @@ -7738,7 +8622,7 @@ CUresult CUDAAPI cuMemAdvise(CUdeviceptr devPtr, size_t count, CUmem_advise advi * if all pages in the memory range have the CPU as their preferred location, or it will be CU_DEVICE_INVALID * if either all the pages don't have the same preferred location or some of the pages don't have a * preferred location at all. Note that the actual location of the pages in the memory range at the time of - * the query may be different from the preferred location. + * the query may be different from the preferred location. * - ::CU_MEM_RANGE_ATTRIBUTE_ACCESSED_BY: If this attribute is specified, \p data will be interpreted * as an array of 32-bit integers, and \p dataSize must be a non-zero multiple of 4. The result returned * will be a list of device ids that had ::CU_MEM_ADVISE_SET_ACCESSED_BY set for that entire memory range. @@ -7774,7 +8658,8 @@ CUresult CUDAAPI cuMemAdvise(CUdeviceptr devPtr, size_t count, CUmem_advise advi * \note_null_stream * * \sa ::cuMemRangeGetAttributes, ::cuMemPrefetchAsync, - * ::cuMemAdvise + * ::cuMemAdvise, + * ::cudaMemRangeGetAttribute */ CUresult CUDAAPI cuMemRangeGetAttribute(void *data, size_t dataSize, CUmem_range_attribute attribute, CUdeviceptr devPtr, size_t count); @@ -7813,7 +8698,8 @@ CUresult CUDAAPI cuMemRangeGetAttribute(void *data, size_t dataSize, CUmem_range * \notefnerr * * \sa ::cuMemRangeGetAttribute, ::cuMemAdvise - * ::cuMemPrefetchAsync + * ::cuMemPrefetchAsync, + * ::cudaMemRangeGetAttributes */ CUresult CUDAAPI cuMemRangeGetAttributes(void **data, size_t *dataSizes, CUmem_range_attribute *attributes, size_t numAttributes, CUdeviceptr devPtr, size_t count); #endif /* __CUDA_API_VERSION >= 8000 */ @@ -7875,6 +8761,7 @@ CUresult CUDAAPI cuPointerSetAttribute(const void *value, CUpointer_attribute at * - ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS * - ::CU_POINTER_ATTRIBUTE_BUFFER_ID * - ::CU_POINTER_ATTRIBUTE_IS_MANAGED + * - ::CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL * * \param numAttributes - Number of attributes to query * \param attributes - An array of attributes to query @@ -7898,8 +8785,10 @@ CUresult CUDAAPI cuPointerSetAttribute(const void *value, CUpointer_attribute at * ::CUDA_ERROR_INVALID_DEVICE * \notefnerr * - * \sa ::cuPointerGetAttribute, - * ::cuPointerSetAttribute + * \sa + * ::cuPointerGetAttribute, + * ::cuPointerSetAttribute, + * ::cudaPointerGetAttributes */ CUresult CUDAAPI cuPointerGetAttributes(unsigned int numAttributes, CUpointer_attribute *attributes, void **data, CUdeviceptr ptr); #endif /* __CUDA_API_VERSION >= 7000 */ @@ -7924,7 +8813,7 @@ CUresult CUDAAPI cuPointerGetAttributes(unsigned int numAttributes, CUpointer_at * Creates a stream and returns a handle in \p phStream. The \p Flags argument * determines behaviors of the stream. Valid values for \p Flags are: * - ::CU_STREAM_DEFAULT: Default stream creation flag. - * - ::CU_STREAM_NON_BLOCKING: Specifies that work running in the created + * - ::CU_STREAM_NON_BLOCKING: Specifies that work running in the created * stream may run concurrently with work in stream 0 (the NULL stream), and that * the created stream should perform no implicit synchronization with stream 0. * @@ -7947,7 +8836,9 @@ CUresult CUDAAPI cuPointerGetAttributes(unsigned int numAttributes, CUpointer_at * ::cuStreamWaitEvent, * ::cuStreamQuery, * ::cuStreamSynchronize, - * ::cuStreamAddCallback + * ::cuStreamAddCallback, + * ::cudaStreamCreate, + * ::cudaStreamCreateWithFlags */ CUresult CUDAAPI cuStreamCreate(CUstream *phStream, unsigned int Flags); @@ -7980,7 +8871,7 @@ CUresult CUDAAPI cuStreamCreate(CUstream *phStream, unsigned int Flags); * ::CUDA_ERROR_OUT_OF_MEMORY * \notefnerr * - * \note Stream priorities are supported only on Quadro and Tesla GPUs + * \note Stream priorities are supported only on GPUs * with compute capability 3.5 or higher. * * \note In the current implementation, only compute kernels launched in @@ -7995,7 +8886,8 @@ CUresult CUDAAPI cuStreamCreate(CUstream *phStream, unsigned int Flags); * ::cuStreamWaitEvent, * ::cuStreamQuery, * ::cuStreamSynchronize, - * ::cuStreamAddCallback + * ::cuStreamAddCallback, + * ::cudaStreamCreateWithPriority */ CUresult CUDAAPI cuStreamCreateWithPriority(CUstream *phStream, unsigned int flags, int priority); @@ -8025,7 +8917,8 @@ CUresult CUDAAPI cuStreamCreateWithPriority(CUstream *phStream, unsigned int fla * ::cuStreamCreate, * ::cuStreamCreateWithPriority, * ::cuCtxGetStreamPriorityRange, - * ::cuStreamGetFlags + * ::cuStreamGetFlags, + * ::cudaStreamGetPriority */ CUresult CUDAAPI cuStreamGetPriority(CUstream hStream, int *priority); @@ -8052,28 +8945,66 @@ CUresult CUDAAPI cuStreamGetPriority(CUstream hStream, int *priority); * * \sa ::cuStreamDestroy, * ::cuStreamCreate, - * ::cuStreamGetPriority + * ::cuStreamGetPriority, + * ::cudaStreamGetFlags */ CUresult CUDAAPI cuStreamGetFlags(CUstream hStream, unsigned int *flags); +#if __CUDA_API_VERSION >= 9020 /** - * \brief Make a compute stream wait on an event + * \brief Query the context associated with a stream + * + * Returns the CUDA context that the stream is associated with. + * + * The stream handle \p hStream can refer to any of the following: + *
    + *
  • a stream created via any of the CUDA driver APIs such as ::cuStreamCreate + * and ::cuStreamCreateWithPriority, or their runtime API equivalents such as + * ::cudaStreamCreate, ::cudaStreamCreateWithFlags and ::cudaStreamCreateWithPriority. + * The returned context is the context that was active in the calling thread when the + * stream was created. Passing an invalid handle will result in undefined behavior.
    • + *
  • any of the special streams such as the NULL stream, ::CU_STREAM_LEGACY and + * ::CU_STREAM_PER_THREAD. The runtime API equivalents of these are also accepted, + * which are NULL, ::cudaStreamLegacy and ::cudaStreamPerThread respectively. + * Specifying any of the special handles will return the context current to the + * calling thread. If no context is current to the calling thread, + * ::CUDA_ERROR_INVALID_CONTEXT is returned.
    • + *
* - * Makes all future work submitted to \p hStream wait until \p hEvent - * reports completion before beginning execution. This synchronization - * will be performed efficiently on the device. The event \p hEvent may - * be from a different context than \p hStream, in which case this function - * will perform cross-device synchronization. + * \param hStream - Handle to the stream to be queried + * \param pctx - Returned context associated with the stream + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * \notefnerr * - * The stream \p hStream will wait only for the completion of the most recent - * host call to ::cuEventRecord() on \p hEvent. Once this call has returned, - * any functions (including ::cuEventRecord() and ::cuEventDestroy()) may be - * called on \p hEvent again, and subsequent calls will not have any - * effect on \p hStream. + * \sa ::cuStreamDestroy, + * ::cuStreamCreateWithPriority, + * ::cuStreamGetPriority, + * ::cuStreamGetFlags, + * ::cuStreamWaitEvent, + * ::cuStreamQuery, + * ::cuStreamSynchronize, + * ::cuStreamAddCallback, + * ::cudaStreamCreate, + * ::cudaStreamCreateWithFlags + */ +CUresult CUDAAPI cuStreamGetCtx(CUstream hStream, CUcontext *pctx); + +#endif /* __CUDA_API_VERSION >= 9020 */ + +/** + * \brief Make a compute stream wait on an event * - * If ::cuEventRecord() has not been called on \p hEvent, this call acts as if - * the record has already completed, and so is a functional no-op. + * Makes all future work submitted to \p hStream wait for all work captured in + * \p hEvent. See ::cuEventRecord() for details on what is captured by an event. + * The synchronization will be performed efficiently on the device when applicable. + * \p hEvent may be from a different context or device than \p hStream. * * \param hStream - Stream to wait * \param hEvent - Event to wait on (may not be NULL) @@ -8093,15 +9024,22 @@ CUresult CUDAAPI cuStreamGetFlags(CUstream hStream, unsigned int *flags); * ::cuStreamQuery, * ::cuStreamSynchronize, * ::cuStreamAddCallback, - * ::cuStreamDestroy + * ::cuStreamDestroy, + * ::cudaStreamWaitEvent */ CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int Flags); /** * \brief Add a callback to a compute stream * + * \note This function is slated for eventual deprecation and removal. If + * you do not require the callback to execute in case of a device error, + * consider using ::cuLaunchHostFunc. Additionally, this function is not + * supported with ::cuStreamBeginCapture and ::cuStreamEndCapture, unlike + * ::cuLaunchHostFunc. + * * Adds a callback to be called on the host after all currently enqueued - * items in the stream have completed. For each + * items in the stream have completed. For each * cuStreamAddCallback call, the callback will be executed exactly once. * The callback will block later work in the stream until it is finished. * @@ -8115,11 +9053,6 @@ CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned in * that are not mandated to run earlier. Callbacks without a mandated order * (in independent streams) execute in undefined order and may be serialized. * - * This API requires compute capability 1.1 or greater. See - * ::cuDeviceGetAttribute or ::cuDeviceGetProperties to query compute - * capability. Attempting to use this API with earlier compute versions will - * return ::CUDA_ERROR_NOT_SUPPORTED. - * * For the purposes of Unified Memory, callback execution makes a number of * guarantees: *
    @@ -8131,10 +9064,11 @@ CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned in * the callback. It thus synchronizes streams which have been "joined" * prior to the callback. *
  • Adding device work to any stream does not have the effect of making - * the stream active until all preceding callbacks have executed. Thus, for + * the stream active until all preceding host functions and stream callbacks + * have executed. Thus, for * example, a callback might use global attached memory even if work has - * been added to another stream, if it has been properly ordered with an - * event.
    • + * been added to another stream, if the work has been ordered behind the + * callback with an event. *
  • Completion of a callback does not cause a stream to become * active except as described above. The callback stream will remain idle * if no device work follows the callback, and will remain idle across @@ -8164,10 +9098,211 @@ CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned in * ::cuStreamWaitEvent, * ::cuStreamDestroy, * ::cuMemAllocManaged, - * ::cuStreamAttachMemAsync + * ::cuStreamAttachMemAsync, + * ::cuStreamLaunchHostFunc, + * ::cudaStreamAddCallback */ CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback, void *userData, unsigned int flags); +#if __CUDA_API_VERSION >= 10000 + +/** + * \brief Begins graph capture on a stream + * + * Begin graph capture on \p hStream. When a stream is in capture mode, all operations + * pushed into the stream will not be executed, but will instead be captured into + * a graph, which will be returned via ::cuStreamEndCapture. Capture may not be initiated + * if \p stream is CU_STREAM_LEGACY. Capture must be ended on the same stream in which + * it was initiated, and it may only be initiated if the stream is not already in capture + * mode. The capture mode may be queried via ::cuStreamIsCapturing. A unique id + * representing the capture sequence may be queried via ::cuStreamGetCaptureInfo. + * + * If \p mode is not ::CU_STREAM_CAPTURE_MODE_RELAXED, ::cuStreamEndCapture must be + * called on this stream from the same thread. + * + * \param hStream - Stream in which to initiate capture + * \param mode - Controls the interaction of this capture sequence with other API + * calls that are potentially unsafe. For more details see + * ::cuThreadExchangeStreamCaptureMode. + * + * \note Kernels captured using this API must not use texture and surface references. + * Reading or writing through any texture or surface reference is undefined + * behavior. This restriction does not apply to texture and surface objects. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa + * ::cuStreamCreate, + * ::cuStreamIsCapturing, + * ::cuStreamEndCapture, + * ::cuThreadExchangeStreamCaptureMode + */ +CUresult CUDAAPI cuStreamBeginCapture(CUstream hStream, CUstreamCaptureMode mode); + +#endif /* __CUDA_API_VERSION >= 10000 */ +#if __CUDA_API_VERSION >= 10010 + +/** + * \brief Swaps the stream capture interaction mode for a thread + * + * Sets the calling thread's stream capture interaction mode to the value contained + * in \p *mode, and overwrites \p *mode with the previous mode for the thread. To + * facilitate deterministic behavior across function or module boundaries, callers + * are encouraged to use this API in a push-pop fashion: \code + CUstreamCaptureMode mode = desiredMode; + cuThreadExchangeStreamCaptureMode(&mode); + ... + cuThreadExchangeStreamCaptureMode(&mode); // restore previous mode + * \endcode + * + * During stream capture (see ::cuStreamBeginCapture), some actions, such as a call + * to ::cudaMalloc, may be unsafe. In the case of ::cudaMalloc, the operation is + * not enqueued asynchronously to a stream, and is not observed by stream capture. + * Therefore, if the sequence of operations captured via ::cuStreamBeginCapture + * depended on the allocation being replayed whenever the graph is launched, the + * captured graph would be invalid. + * + * Therefore, stream capture places restrictions on API calls that can be made within + * or concurrently to a ::cuStreamBeginCapture-::cuStreamEndCapture sequence. This + * behavior can be controlled via this API and flags to ::cuStreamBeginCapture. + * + * A thread's mode is one of the following: + * - \p CU_STREAM_CAPTURE_MODE_GLOBAL: This is the default mode. If the local thread has + * an ongoing capture sequence that was not initiated with + * \p CU_STREAM_CAPTURE_MODE_RELAXED at \p cuStreamBeginCapture, or if any other thread + * has a concurrent capture sequence initiated with \p CU_STREAM_CAPTURE_MODE_GLOBAL, + * this thread is prohibited from potentially unsafe API calls. + * - \p CU_STREAM_CAPTURE_MODE_THREAD_LOCAL: If the local thread has an ongoing capture + * sequence not initiated with \p CU_STREAM_CAPTURE_MODE_RELAXED, it is prohibited + * from potentially unsafe API calls. Concurrent capture sequences in other threads + * are ignored. + * - \p CU_STREAM_CAPTURE_MODE_RELAXED: The local thread is not prohibited from potentially + * unsafe API calls. Note that the thread is still prohibited from API calls which + * necessarily conflict with stream capture, for example, attempting ::cuEventQuery + * on an event that was last recorded inside a capture sequence. + * + * \param mode - Pointer to mode value to swap with the current mode + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa + * ::cuStreamBeginCapture + */ +CUresult CUDAAPI cuThreadExchangeStreamCaptureMode(CUstreamCaptureMode *mode); + +#endif /* __CUDA_API_VERSION >= 10010 */ +#if __CUDA_API_VERSION >= 10000 + +/** + * \brief Ends capture on a stream, returning the captured graph + * + * End capture on \p hStream, returning the captured graph via \p phGraph. + * Capture must have been initiated on \p hStream via a call to ::cuStreamBeginCapture. + * If capture was invalidated, due to a violation of the rules of stream capture, then + * a NULL graph will be returned. + * + * If the \p mode argument to ::cuStreamBeginCapture was not + * ::CU_STREAM_CAPTURE_MODE_RELAXED, this call must be from the same thread as + * ::cuStreamBeginCapture. + * + * \param hStream - Stream to query + * \param phGraph - The captured graph + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_STREAM_CAPTURE_WRONG_THREAD + * \notefnerr + * + * \sa + * ::cuStreamCreate, + * ::cuStreamBeginCapture, + * ::cuStreamIsCapturing + */ +CUresult CUDAAPI cuStreamEndCapture(CUstream hStream, CUgraph *phGraph); + +/** + * \brief Returns a stream's capture status + * + * Return the capture status of \p hStream via \p captureStatus. After a successful + * call, \p *captureStatus will contain one of the following: + * - ::CU_STREAM_CAPTURE_STATUS_NONE: The stream is not capturing. + * - ::CU_STREAM_CAPTURE_STATUS_ACTIVE: The stream is capturing. + * - ::CU_STREAM_CAPTURE_STATUS_INVALIDATED: The stream was capturing but an error + * has invalidated the capture sequence. The capture sequence must be terminated + * with ::cuStreamEndCapture on the stream where it was initiated in order to + * continue using \p hStream. + * + * Note that, if this is called on ::CU_STREAM_LEGACY (the "null stream") while + * a blocking stream in the same context is capturing, it will return + * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT and \p *captureStatus is unspecified + * after the call. The blocking stream capture is not invalidated. + * + * When a blocking stream is capturing, the legacy stream is in an + * unusable state until the blocking stream capture is terminated. The legacy + * stream is not supported for stream capture, but attempted use would have an + * implicit dependency on the capturing stream(s). + * + * \param hStream - Stream to query + * \param captureStatus - Returns the stream's capture status + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT + * \notefnerr + * + * \sa + * ::cuStreamCreate, + * ::cuStreamBeginCapture, + * ::cuStreamEndCapture + */ +CUresult CUDAAPI cuStreamIsCapturing(CUstream hStream, CUstreamCaptureStatus *captureStatus); + +#endif /* __CUDA_API_VERSION >= 10000 */ + +#if __CUDA_API_VERSION >= 10010 + +/** + * \brief Query capture status of a stream + * + * Query the capture status of a stream and and get an id for + * the capture sequence, which is unique over the lifetime of the process. + * + * If called on ::CU_STREAM_LEGACY (the "null stream") while a stream not created + * with ::CU_STREAM_NON_BLOCKING is capturing, returns ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT. + * + * A valid id is returned only if both of the following are true: + * - the call returns CUDA_SUCCESS + * - captureStatus is set to ::CU_STREAM_CAPTURE_STATUS_ACTIVE + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT + * \notefnerr + * + * \sa + * ::cuStreamBeginCapture, + * ::cuStreamIsCapturing + */ + CUresult CUDAAPI cuStreamGetCaptureInfo(CUstream hStream, CUstreamCaptureStatus *captureStatus, cuuint64_t *id); + +#endif /* __CUDA_API_VERSION >= 10010 */ + #if __CUDA_API_VERSION >= 6000 /** @@ -8179,12 +9314,20 @@ CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback * only take effect when, previous work in stream has completed. Any * previous association is automatically replaced. * - * \p dptr must point to an address within managed memory space declared - * using the __managed__ keyword or allocated with ::cuMemAllocManaged. + * \p dptr must point to one of the following types of memories: + * - managed memory declared using the __managed__ keyword or allocated with + * ::cuMemAllocManaged. + * - a valid host-accessible region of system-allocated pageable memory. This + * type of memory may only be specified if the device associated with the + * stream reports a non-zero value for the device attribute + * ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. * - * \p length must be zero, to indicate that the entire allocation's - * stream association is being changed. Currently, it's not possible - * to change stream association for a portion of an allocation. + * For managed allocations, \p length must be either zero or the entire + * allocation's size. Both indicate that the entire allocation's stream + * association is being changed. Currently, it is not possible to change stream + * association for a portion of a managed allocation. + * + * For pageable host allocations, \p length must be non-zero. * * The stream association is specified using \p flags which must be * one of ::CUmemAttach_flags. @@ -8209,7 +9352,7 @@ CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback * Accessing memory on the device from streams that are not associated with * it will produce undefined results. No error checking is performed by the * Unified Memory system to ensure that kernels launched into other streams - * do not access this region. + * do not access this region. * * It is a program's responsibility to order calls to ::cuStreamAttachMemAsync * via events, synchronization or other means to ensure legal access to memory @@ -8224,8 +9367,10 @@ CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback * happen until all work in the stream has completed. * * \param hStream - Stream in which to enqueue the attach operation - * \param dptr - Pointer to memory (must be a pointer to managed memory) - * \param length - Length of memory (must be zero) + * \param dptr - Pointer to memory (must be a pointer to managed memory or + * to a valid host-accessible region of system-allocated + * pageable memory) + * \param length - Length of memory * \param flags - Must be one of ::CUmemAttach_flags * * \return @@ -8243,7 +9388,8 @@ CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback * ::cuStreamSynchronize, * ::cuStreamWaitEvent, * ::cuStreamDestroy, - * ::cuMemAllocManaged + * ::cuMemAllocManaged, + * ::cudaStreamAttachMemAsync */ CUresult CUDAAPI cuStreamAttachMemAsync(CUstream hStream, CUdeviceptr dptr, size_t length, unsigned int flags); @@ -8274,7 +9420,8 @@ CUresult CUDAAPI cuStreamAttachMemAsync(CUstream hStream, CUdeviceptr dptr, size * ::cuStreamWaitEvent, * ::cuStreamDestroy, * ::cuStreamSynchronize, - * ::cuStreamAddCallback + * ::cuStreamAddCallback, + * ::cudaStreamQuery */ CUresult CUDAAPI cuStreamQuery(CUstream hStream); @@ -8282,7 +9429,7 @@ CUresult CUDAAPI cuStreamQuery(CUstream hStream); * \brief Wait until a stream's tasks are completed * * Waits until the device has completed all operations in the stream specified - * by \p hStream. If the context was created with the + * by \p hStream. If the context was created with the * ::CU_CTX_SCHED_BLOCKING_SYNC flag, the CPU thread will block until the * stream is finished with all of its tasks. * @@ -8294,6 +9441,7 @@ CUresult CUDAAPI cuStreamQuery(CUstream hStream); * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_HANDLE + * \note_null_stream * \notefnerr * @@ -8301,7 +9449,8 @@ CUresult CUDAAPI cuStreamQuery(CUstream hStream); * ::cuStreamDestroy, * ::cuStreamWaitEvent, * ::cuStreamQuery, - * ::cuStreamAddCallback + * ::cuStreamAddCallback, + * ::cudaStreamSynchronize */ CUresult CUDAAPI cuStreamSynchronize(CUstream hStream); @@ -8309,11 +9458,11 @@ CUresult CUDAAPI cuStreamSynchronize(CUstream hStream); /** * \brief Destroys a stream * - * Destroys the stream specified by \p hStream. + * Destroys the stream specified by \p hStream. * * In case the device is still doing work in the stream \p hStream - * when ::cuStreamDestroy() is called, the function will return immediately - * and the resources associated with \p hStream will be released automatically + * when ::cuStreamDestroy() is called, the function will return immediately + * and the resources associated with \p hStream will be released automatically * once the device has completed all work in \p hStream. * * \param hStream - Stream to destroy @@ -8323,14 +9472,16 @@ CUresult CUDAAPI cuStreamSynchronize(CUstream hStream); * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * * \sa ::cuStreamCreate, * ::cuStreamWaitEvent, * ::cuStreamQuery, * ::cuStreamSynchronize, - * ::cuStreamAddCallback + * ::cuStreamAddCallback, + * ::cudaStreamDestroy */ CUresult CUDAAPI cuStreamDestroy(CUstream hStream); #endif /* __CUDA_API_VERSION >= 4000 */ @@ -8353,8 +9504,8 @@ CUresult CUDAAPI cuStreamDestroy(CUstream hStream); /** * \brief Creates an event * - * Creates an event *phEvent with the flags specified via \p Flags. Valid flags - * include: + * Creates an event *phEvent for the current context with the flags specified via + * \p Flags. Valid flags include: * - ::CU_EVENT_DEFAULT: Default event creation flag. * - ::CU_EVENT_BLOCKING_SYNC: Specifies that the created event should use blocking * synchronization. A CPU thread that uses ::cuEventSynchronize() to wait on @@ -8385,23 +9536,29 @@ CUresult CUDAAPI cuStreamDestroy(CUstream hStream); * ::cuEventQuery, * ::cuEventSynchronize, * ::cuEventDestroy, - * ::cuEventElapsedTime + * ::cuEventElapsedTime, + * ::cudaEventCreate, + * ::cudaEventCreateWithFlags */ CUresult CUDAAPI cuEventCreate(CUevent *phEvent, unsigned int Flags); /** * \brief Records an event * - * Records an event. See note on NULL stream behavior. Since operation is - * asynchronous, ::cuEventQuery or ::cuEventSynchronize() must be used - * to determine when the event has actually been recorded. - * - * If ::cuEventRecord() has previously been called on \p hEvent, then this - * call will overwrite any existing state in \p hEvent. Any subsequent calls - * which examine the status of \p hEvent will only examine the completion of - * this most recent call to ::cuEventRecord(). - * - * It is necessary that \p hEvent and \p hStream be created on the same context. + * Captures in \p hEvent the contents of \p hStream at the time of this call. + * \p hEvent and \p hStream must be from the same context. + * Calls such as ::cuEventQuery() or ::cuStreamWaitEvent() will then + * examine or wait for completion of the work that was captured. Uses of + * \p hStream after this call do not modify \p hEvent. See note on default + * stream behavior for what is captured in the default case. + * + * ::cuEventRecord() can be called multiple times on the same event and + * will overwrite the previously captured state. Other APIs such as + * ::cuStreamWaitEvent() use the most recently captured state at the time + * of the API call, and are not affected by later calls to + * ::cuEventRecord(). Before the first call to ::cuEventRecord(), an + * event represents an empty set of work, so for example ::cuEventQuery() + * would return ::CUDA_SUCCESS. * * \param hEvent - Event to record * \param hStream - Stream to record event for @@ -8421,21 +9578,19 @@ CUresult CUDAAPI cuEventCreate(CUevent *phEvent, unsigned int Flags); * ::cuEventSynchronize, * ::cuStreamWaitEvent, * ::cuEventDestroy, - * ::cuEventElapsedTime + * ::cuEventElapsedTime, + * ::cudaEventRecord */ CUresult CUDAAPI cuEventRecord(CUevent hEvent, CUstream hStream); /** * \brief Queries an event's status * - * Query the status of all device work preceding the most recent - * call to ::cuEventRecord() (in the appropriate compute streams, - * as specified by the arguments to ::cuEventRecord()). + * Queries the status of all work currently captured by \p hEvent. See + * ::cuEventRecord() for details on what is captured by an event. * - * If this work has successfully been completed by the device, or if - * ::cuEventRecord() has not been called on \p hEvent, then ::CUDA_SUCCESS is - * returned. If this work has not yet been completed by the device then - * ::CUDA_ERROR_NOT_READY is returned. + * Returns ::CUDA_SUCCESS if all captured work has been completed, or + * ::CUDA_ERROR_NOT_READY if any captured work is incomplete. * * For the purposes of Unified Memory, a return value of ::CUDA_SUCCESS * is equivalent to having called ::cuEventSynchronize(). @@ -8455,19 +9610,16 @@ CUresult CUDAAPI cuEventRecord(CUevent hEvent, CUstream hStream); * ::cuEventRecord, * ::cuEventSynchronize, * ::cuEventDestroy, - * ::cuEventElapsedTime + * ::cuEventElapsedTime, + * ::cudaEventQuery */ CUresult CUDAAPI cuEventQuery(CUevent hEvent); /** * \brief Waits for an event to complete * - * Wait until the completion of all device work preceding the most recent - * call to ::cuEventRecord() (in the appropriate compute streams, as specified - * by the arguments to ::cuEventRecord()). - * - * If ::cuEventRecord() has not been called on \p hEvent, ::CUDA_SUCCESS is - * returned immediately. + * Waits until the completion of all work currently captured in \p hEvent. + * See ::cuEventRecord() for details on what is captured by an event. * * Waiting for an event that was created with the ::CU_EVENT_BLOCKING_SYNC * flag will cause the calling CPU thread to block until the event has @@ -8489,7 +9641,8 @@ CUresult CUDAAPI cuEventQuery(CUevent hEvent); * ::cuEventRecord, * ::cuEventQuery, * ::cuEventDestroy, - * ::cuEventElapsedTime + * ::cuEventElapsedTime, + * ::cudaEventSynchronize */ CUresult CUDAAPI cuEventSynchronize(CUevent hEvent); @@ -8499,10 +9652,10 @@ CUresult CUDAAPI cuEventSynchronize(CUevent hEvent); * * Destroys the event specified by \p hEvent. * - * In case \p hEvent has been recorded but has not yet been completed - * when ::cuEventDestroy() is called, the function will return immediately and - * the resources associated with \p hEvent will be released automatically once - * the device has completed \p hEvent. + * An event may be destroyed before it is complete (i.e., while + * ::cuEventQuery() would return ::CUDA_ERROR_NOT_READY). In this case, the + * call does not block on completion of the event, and any associated + * resources will automatically be released asynchronously at completion. * * \param hEvent - Event to destroy * @@ -8518,7 +9671,8 @@ CUresult CUDAAPI cuEventSynchronize(CUevent hEvent); * ::cuEventRecord, * ::cuEventQuery, * ::cuEventSynchronize, - * ::cuEventElapsedTime + * ::cuEventElapsedTime, + * ::cudaEventDestroy */ CUresult CUDAAPI cuEventDestroy(CUevent hEvent); #endif /* __CUDA_API_VERSION >= 4000 */ @@ -8562,112 +9716,701 @@ CUresult CUDAAPI cuEventDestroy(CUevent hEvent); * ::cuEventRecord, * ::cuEventQuery, * ::cuEventSynchronize, - * ::cuEventDestroy + * ::cuEventDestroy, + * ::cudaEventElapsedTime */ CUresult CUDAAPI cuEventElapsedTime(float *pMilliseconds, CUevent hStart, CUevent hEnd); -#if __CUDA_API_VERSION >= 8000 +/** @} */ /* END CUDA_EVENT */ + /** - * \brief Wait on a memory location + * \defgroup CUDA_EXTRES_INTEROP External Resource Interoperability * - * Enqueues a synchronization of the stream on the given memory location. Work - * ordered after the operation will block until the given condition on the - * memory is satisfied. By default, the condition is to wait for - * (int32_t)(*addr - value) >= 0, a cyclic greater-or-equal. - * Other condition types can be specified via \p flags. + * ___MANBRIEF___ External resource interoperability functions of the low-level CUDA driver API + * (___CURRENT_FILE___) ___ENDMANBRIEF___ * - * If the memory was registered via ::cuMemHostRegister(), the device pointer - * should be obtained with ::cuMemHostGetDevicePointer(). This function cannot - * be used with managed memory (::cuMemAllocManaged). + * This section describes the external resource interoperability functions of the low-level CUDA + * driver application programming interface. + * + * @{ + */ + +#if __CUDA_API_VERSION >= 10000 + + /** + * \brief Imports an external memory object * - * On Windows, the device must be using TCC, or the operation is not supported. - * See ::cuDeviceGetAttributes(). + * Imports an externally allocated memory object and returns + * a handle to that in \p extMem_out. * - * \param stream The stream to synchronize on the memory location. - * \param addr The memory location to wait on. - * \param value The value to compare with the memory location. - * \param flags See ::CUstreamWaitValue_flags. + * The properties of the handle being imported must be described in + * \p memHandleDesc. The ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC structure + * is defined as follows: + * + * \code + typedef struct CUDA_EXTERNAL_MEMORY_HANDLE_DESC_st { + CUexternalMemoryHandleType type; + union { + int fd; + struct { + void *handle; + const void *name; + } win32; + } handle; + unsigned long long size; + unsigned int flags; + } CUDA_EXTERNAL_MEMORY_HANDLE_DESC; + * \endcode + * + * where ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type specifies the type + * of handle being imported. ::CUexternalMemoryHandleType is + * defined as: + * + * \code + typedef enum CUexternalMemoryHandleType_enum { + CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD = 1, + CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 = 2, + CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3, + CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP = 4, + CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE = 5 + } CUexternalMemoryHandleType; + * \endcode + * + * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is + * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD, then + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::fd must be a valid + * file descriptor referencing a memory object. Ownership of + * the file descriptor is transferred to the CUDA driver when the + * handle is imported successfully. Performing any operations on the + * file descriptor after it is imported results in undefined behavior. + * + * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is + * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32, then exactly one + * of ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle and + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name must not be + * NULL. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle + * is not NULL, then it must represent a valid shared NT handle that + * references a memory object. Ownership of this handle is + * not transferred to CUDA after the import operation, so the + * application must release the handle using the appropriate system + * call. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name + * is not NULL, then it must point to a NULL-terminated array of + * UTF-16 characters that refers to a memory object. + * + * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is + * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT, then + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle must + * be non-NULL and + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name + * must be NULL. The handle specified must be a globally shared KMT + * handle. This handle does not hold a reference to the underlying + * object, and thus will be invalid when all references to the + * memory object are destroyed. + * + * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is + * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP, then exactly one + * of ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle and + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name must not be + * NULL. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle + * is not NULL, then it must represent a valid shared NT handle that + * is returned by ID3DDevice::CreateSharedHandle when referring to a + * ID3D12Heap object. This handle holds a reference to the underlying + * object. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name + * is not NULL, then it must point to a NULL-terminated array of + * UTF-16 characters that refers to a ID3D12Heap object. + * + * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is + * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE, then exactly one + * of ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle and + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name must not be + * NULL. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle + * is not NULL, then it must represent a valid shared NT handle that + * is returned by ID3DDevice::CreateSharedHandle when referring to a + * ID3D12Resource object. This handle holds a reference to the + * underlying object. If + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name + * is not NULL, then it must point to a NULL-terminated array of + * UTF-16 characters that refers to a ID3D12Resource object. + * + * The size of the memory object must be specified in + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::size. + * + * Specifying the flag ::CUDA_EXTERNAL_MEMORY_DEDICATED in + * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::flags indicates that the + * resource is a dedicated resource. The definition of what a + * dedicated resource is outside the scope of this extension. + * + * \param extMem_out - Returned handle to an external memory object + * \param memHandleDesc - Memory import handle descriptor * * \return * ::CUDA_SUCCESS, - * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_NOT_SUPPORTED + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * - * \sa ::cuStreamWriteValue32, - * ::cuStreamBatchMemOp, - * ::cuMemHostRegister, - * ::cuStreamWaitEvent + * \note If the Vulkan memory imported into CUDA is mapped on the CPU then the + * application must use vkInvalidateMappedMemoryRanges/vkFlushMappedMemoryRanges + * as well as appropriate Vulkan pipeline barriers to maintain coherence between + * CPU and GPU. For more information on these APIs, please refer to "Synchronization + * and Cache Control" chapter from Vulkan specification. + * + * \sa ::cuDestroyExternalMemory, + * ::cuExternalMemoryGetMappedBuffer, + * ::cuExternalMemoryGetMappedMipmappedArray */ -CUresult CUDAAPI cuStreamWaitValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags); +CUresult CUDAAPI cuImportExternalMemory(CUexternalMemory *extMem_out, const CUDA_EXTERNAL_MEMORY_HANDLE_DESC *memHandleDesc); /** - * \brief Write a value to memory + * \brief Maps a buffer onto an imported memory object * - * Write a value to memory. Unless the ::CU_STREAM_WRITE_VALUE_NO_MEMORY_BARRIER - * flag is passed, the write is preceded by a system-wide memory fence, - * equivalent to a __threadfence_system() but scoped to the stream - * rather than a CUDA thread. + * Maps a buffer onto an imported memory object and returns a device + * pointer in \p devPtr. * - * If the memory was registered via ::cuMemHostRegister(), the device pointer - * should be obtained with ::cuMemHostGetDevicePointer(). This function cannot - * be used with managed memory (::cuMemAllocManaged). + * The properties of the buffer being mapped must be described in + * \p bufferDesc. The ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC structure is + * defined as follows: + * + * \code + typedef struct CUDA_EXTERNAL_MEMORY_BUFFER_DESC_st { + unsigned long long offset; + unsigned long long size; + unsigned int flags; + } CUDA_EXTERNAL_MEMORY_BUFFER_DESC; + * \endcode * - * On Windows, the device must be using TCC, or the operation is not supported. - * See ::cuDeviceGetAttribute(). + * where ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC::offset is the offset in + * the memory object where the buffer's base address is. + * ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC::size is the size of the buffer. + * ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC::flags must be zero. * - * \param stream The stream to do the write in. - * \param addr The device address to write to. - * \param value The value to write. - * \param flags See ::CUstreamWriteValue_flags. + * The offset and size have to be suitably aligned to match the + * requirements of the external API. Mapping two buffers whose ranges + * overlap may or may not result in the same virtual address being + * returned for the overlapped portion. In such cases, the application + * must ensure that all accesses to that region from the GPU are + * volatile. Otherwise writes made via one address are not guaranteed + * to be visible via the other address, even if they're issued by the + * same thread. It is recommended that applications map the combined + * range instead of mapping separate buffers and then apply the + * appropriate offsets to the returned pointer to derive the + * individual buffers. + * + * The returned pointer \p devPtr must be freed using ::cuMemFree. + * + * \param devPtr - Returned device pointer to buffer + * \param extMem - Handle to external memory object + * \param bufferDesc - Buffer descriptor * * \return * ::CUDA_SUCCESS, - * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_NOT_SUPPORTED + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE * \notefnerr * - * \sa ::cuStreamWaitValue32, - * ::cuStreamBatchMemOp, - * ::cuMemHostRegister, - * ::cuEventRecord + * \sa ::cuImportExternalMemory + * ::cuDestroyExternalMemory, + * ::cuExternalMemoryGetMappedMipmappedArray */ -CUresult CUDAAPI cuStreamWriteValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags); +CUresult CUDAAPI cuExternalMemoryGetMappedBuffer(CUdeviceptr *devPtr, CUexternalMemory extMem, const CUDA_EXTERNAL_MEMORY_BUFFER_DESC *bufferDesc); /** - * \brief Batch operations to synchronize the stream via memory operations + * \brief Maps a CUDA mipmapped array onto an external memory object * - * This is a batch version of ::cuStreamWaitValue32() and ::cuStreamWriteValue32(). - * Batching operations may avoid some performance overhead in both the API call - * and the device execution versus adding them to the stream in separate API - * calls. The operations are enqueued in the order they appear in the array. + * Maps a CUDA mipmapped array onto an external object and returns a + * handle to it in \p mipmap. * - * See ::CUstreamBatchMemOpType for the full set of supported operations, and - * ::cuStreamWaitValue32() and ::cuStreamWriteValue32() for details of specific - * operations. + * The properties of the CUDA mipmapped array being mapped must be + * described in \p mipmapDesc. The structure + * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC is defined as follows: + * + * \code + typedef struct CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_st { + unsigned long long offset; + CUDA_ARRAY3D_DESCRIPTOR arrayDesc; + unsigned int numLevels; + } CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC; + * \endcode * - * On Windows, the device must be using TCC, or this call is not supported. See - * ::cuDeviceGetAttribute(). + * where ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::offset is the + * offset in the memory object where the base level of the mipmap + * chain is. + * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::arrayDesc describes + * the format, dimensions and type of the base level of the mipmap + * chain. For further details on these parameters, please refer to the + * documentation for ::cuMipmappedArrayCreate. Note that if the mipmapped + * array is bound as a color target in the graphics API, then the flag + * ::CUDA_ARRAY3D_COLOR_ATTACHMENT must be specified in + * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::arrayDesc::Flags. + * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::numLevels specifies + * the total number of levels in the mipmap chain. * - * \param stream The stream to enqueue the operations in. - * \param count The number of operations in the array. Must be less than 256. - * \param paramArray The types and parameters of the individual operations. - * \param flags Reserved for future expansion; must be 0. + * The returned CUDA mipmapped array must be freed using ::cuMipmappedArrayDestroy. + * + * \param mipmap - Returned CUDA mipmapped array + * \param extMem - Handle to external memory object + * \param mipmapDesc - CUDA array descriptor * * \return * ::CUDA_SUCCESS, - * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_NOT_SUPPORTED + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE + * \notefnerr + * + * \sa ::cuImportExternalMemory + * ::cuDestroyExternalMemory, + * ::cuExternalMemoryGetMappedBuffer + */ +CUresult CUDAAPI cuExternalMemoryGetMappedMipmappedArray(CUmipmappedArray *mipmap, CUexternalMemory extMem, const CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC *mipmapDesc); + +/** + * \brief Destroys an external memory object. + * + * Destroys the specified external memory object. Any existing buffers + * and CUDA mipmapped arrays mapped onto this object must no longer be + * used and must be explicitly freed using ::cuMemFree and + * ::cuMipmappedArrayDestroy respectively. + * + * \param extMem - External memory object to be destroyed + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE + * \notefnerr + * + * \sa ::cuImportExternalMemory + * ::cuExternalMemoryGetMappedBuffer, + * ::cuExternalMemoryGetMappedMipmappedArray + */ +CUresult CUDAAPI cuDestroyExternalMemory(CUexternalMemory extMem); + +/** + * \brief Imports an external semaphore + * + * Imports an externally allocated synchronization object and returns + * a handle to that in \p extSem_out. + * + * The properties of the handle being imported must be described in + * \p semHandleDesc. The ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC is + * defined as follows: + * + * \code + typedef struct CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_st { + CUexternalSemaphoreHandleType type; + union { + int fd; + struct { + void *handle; + const void *name; + } win32; + } handle; + unsigned int flags; + } CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC; + * \endcode + * + * where ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type specifies the type of + * handle being imported. ::CUexternalSemaphoreHandleType is defined + * as: + * + * \code + typedef enum CUexternalSemaphoreHandleType_enum { + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD = 1, + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32 = 2, + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3, + CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE = 4 + } CUexternalSemaphoreHandleType; + * \endcode + * + * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD, then + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::fd must be a valid + * file descriptor referencing a synchronization object. Ownership of + * the file descriptor is transferred to the CUDA driver when the + * handle is imported successfully. Performing any operations on the + * file descriptor after it is imported results in undefined behavior. + * + * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32, then exactly one + * of ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle and + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name must not be + * NULL. If + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle + * is not NULL, then it must represent a valid shared NT handle that + * references a synchronization object. Ownership of this handle is + * not transferred to CUDA after the import operation, so the + * application must release the handle using the appropriate system + * call. If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name + * is not NULL, then it must name a valid synchronization object. + * + * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT, then + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle must + * be non-NULL and + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name + * must be NULL. The handle specified must be a globally shared KMT + * handle. This handle does not hold a reference to the underlying + * object, and thus will be invalid when all references to the + * synchronization object are destroyed. + * + * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE, then exactly one + * of ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle and + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name must not be + * NULL. If + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle + * is not NULL, then it must represent a valid shared NT handle that + * is returned by ID3DDevice::CreateSharedHandle when referring to a + * ID3D12Fence object. This handle holds a reference to the underlying + * object. If + * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name + * is not NULL, then it must name a valid synchronization object that + * refers to a valid ID3D12Fence object. + * + * \param extSem_out - Returned handle to an external semaphore + * \param semHandleDesc - Semaphore import handle descriptor + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE + * \notefnerr + * + * \sa ::cuDestroyExternalSemaphore, + * ::cuSignalExternalSemaphoresAsync, + * ::cuWaitExternalSemaphoresAsync + */ +CUresult CUDAAPI cuImportExternalSemaphore(CUexternalSemaphore *extSem_out, const CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC *semHandleDesc); + +/** + * \brief Signals a set of external semaphore objects + * + * Enqueues a signal operation on a set of externally allocated + * semaphore object in the specified stream. The operations will be + * executed when all prior operations in the stream complete. + * + * The exact semantics of signaling a semaphore depends on the type of + * the object. + * + * If the semaphore object is any one of the following types: + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD, + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32, + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT + * then signaling the semaphore will set it to the signaled state. + * + * If the semaphore object is of the type + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE, then the + * semaphore will be set to the value specified in + * ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS::params::fence::value. + * + * \param extSemArray - Set of external semaphores to be signaled + * \param paramsArray - Array of semaphore parameters + * \param numExtSems - Number of semaphores to signal + * \param stream - Stream to enqueue the signal operations in + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE + * \notefnerr + * + * \sa ::cuImportExternalSemaphore, + * ::cuDestroyExternalSemaphore, + * ::cuWaitExternalSemaphoresAsync + */ +CUresult CUDAAPI cuSignalExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream); + +/** + * \brief Waits on a set of external semaphore objects + * + * Enqueues a wait operation on a set of externally allocated + * semaphore object in the specified stream. The operations will be + * executed when all prior operations in the stream complete. + * + * The exact semantics of waiting on a semaphore depends on the type + * of the object. + * + * If the semaphore object is any one of the following types: + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD, + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32, + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT + * then waiting on the semaphore will wait until the semaphore reaches + * the signaled state. The semaphore will then be reset to the + * unsignaled state. Therefore for every signal operation, there can + * only be one wait operation. + * + * If the semaphore object is of the type + * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE, then waiting on + * the semaphore will wait until the value of the semaphore is + * greater than or equal to + * ::CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS::params::fence::value. + * + * \param extSemArray - External semaphores to be waited on + * \param paramsArray - Array of semaphore parameters + * \param numExtSems - Number of semaphores to wait on + * \param stream - Stream to enqueue the wait operations in + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE + * \notefnerr + * + * \sa ::cuImportExternalSemaphore, + * ::cuDestroyExternalSemaphore, + * ::cuSignalExternalSemaphoresAsync + */ +CUresult CUDAAPI cuWaitExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream); + +/** + * \brief Destroys an external semaphore + * + * Destroys an external semaphore object and releases any references + * to the underlying resource. Any outstanding signals or waits must + * have completed before the semaphore is destroyed. + * + * \param extSem - External semaphore to be destroyed + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_HANDLE + * \notefnerr + * + * \sa ::cuImportExternalSemaphore, + * ::cuSignalExternalSemaphoresAsync, + * ::cuWaitExternalSemaphoresAsync + */ +CUresult CUDAAPI cuDestroyExternalSemaphore(CUexternalSemaphore extSem); + +#endif /* __CUDA_API_VERSION >= 10000 */ + +/** @} */ /* END CUDA_EXTRES_INTEROP */ + +/** + * \defgroup CUDA_MEMOP Stream memory operations + * + * ___MANBRIEF___ Stream memory operations of the low-level CUDA driver API + * (___CURRENT_FILE___) ___ENDMANBRIEF___ + * + * This section describes the stream memory operations of the low-level CUDA + * driver application programming interface. + * + * The whole set of operations is disabled by default. Users are required + * to explicitly enable them, e.g. on Linux by passing the kernel module + * parameter shown below: + * modprobe nvidia NVreg_EnableStreamMemOPs=1 + * There is currently no way to enable these operations on other operating + * systems. + * + * Users can programmatically query whether the device supports these + * operations with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS. + * + * Support for the ::CU_STREAM_WAIT_VALUE_NOR flag can be queried with + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR. + * + * Support for the ::cuStreamWriteValue64() and ::cuStreamWaitValue64() + * functions, as well as for the ::CU_STREAM_MEM_OP_WAIT_VALUE_64 and + * ::CU_STREAM_MEM_OP_WRITE_VALUE_64 flags, can be queried with + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS. + * + * Support for both ::CU_STREAM_WAIT_VALUE_FLUSH and + * ::CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES requires dedicated platform + * hardware features and can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_FLUSH_REMOTE_WRITES. + * + * Note that all memory pointers passed as parameters to these operations + * are device pointers. Where necessary a device pointer should be + * obtained, for example with ::cuMemHostGetDevicePointer(). + * + * None of the operations accepts pointers to managed memory buffers + * (::cuMemAllocManaged). + * + * @{ + */ + +#if __CUDA_API_VERSION >= 8000 +/** + * \brief Wait on a memory location + * + * Enqueues a synchronization of the stream on the given memory location. Work + * ordered after the operation will block until the given condition on the + * memory is satisfied. By default, the condition is to wait for + * (int32_t)(*addr - value) >= 0, a cyclic greater-or-equal. + * Other condition types can be specified via \p flags. + * + * If the memory was registered via ::cuMemHostRegister(), the device pointer + * should be obtained with ::cuMemHostGetDevicePointer(). This function cannot + * be used with managed memory (::cuMemAllocManaged). + * + * Support for this can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS. + * + * Support for CU_STREAM_WAIT_VALUE_NOR can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR. + * + * \param stream The stream to synchronize on the memory location. + * \param addr The memory location to wait on. + * \param value The value to compare with the memory location. + * \param flags See ::CUstreamWaitValue_flags. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_NOT_SUPPORTED + * \notefnerr + * + * \sa ::cuStreamWaitValue64, + * ::cuStreamWriteValue32, + * ::cuStreamWriteValue64 + * ::cuStreamBatchMemOp, + * ::cuMemHostRegister, + * ::cuStreamWaitEvent + */ +CUresult CUDAAPI cuStreamWaitValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags); + +/** + * \brief Wait on a memory location + * + * Enqueues a synchronization of the stream on the given memory location. Work + * ordered after the operation will block until the given condition on the + * memory is satisfied. By default, the condition is to wait for + * (int64_t)(*addr - value) >= 0, a cyclic greater-or-equal. + * Other condition types can be specified via \p flags. + * + * If the memory was registered via ::cuMemHostRegister(), the device pointer + * should be obtained with ::cuMemHostGetDevicePointer(). + * + * Support for this can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS. + * + * \param stream The stream to synchronize on the memory location. + * \param addr The memory location to wait on. + * \param value The value to compare with the memory location. + * \param flags See ::CUstreamWaitValue_flags. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_NOT_SUPPORTED + * \notefnerr + * + * \sa ::cuStreamWaitValue32, + * ::cuStreamWriteValue32, + * ::cuStreamWriteValue64, + * ::cuStreamBatchMemOp, + * ::cuMemHostRegister, + * ::cuStreamWaitEvent + */ +CUresult CUDAAPI cuStreamWaitValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags); + +/** + * \brief Write a value to memory + * + * Write a value to memory. Unless the ::CU_STREAM_WRITE_VALUE_NO_MEMORY_BARRIER + * flag is passed, the write is preceded by a system-wide memory fence, + * equivalent to a __threadfence_system() but scoped to the stream + * rather than a CUDA thread. + * + * If the memory was registered via ::cuMemHostRegister(), the device pointer + * should be obtained with ::cuMemHostGetDevicePointer(). This function cannot + * be used with managed memory (::cuMemAllocManaged). + * + * Support for this can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS. + * + * \param stream The stream to do the write in. + * \param addr The device address to write to. + * \param value The value to write. + * \param flags See ::CUstreamWriteValue_flags. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_NOT_SUPPORTED + * \notefnerr + * + * \sa ::cuStreamWriteValue64, + * ::cuStreamWaitValue32, + * ::cuStreamWaitValue64, + * ::cuStreamBatchMemOp, + * ::cuMemHostRegister, + * ::cuEventRecord + */ +CUresult CUDAAPI cuStreamWriteValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags); + +/** + * \brief Write a value to memory + * + * Write a value to memory. Unless the ::CU_STREAM_WRITE_VALUE_NO_MEMORY_BARRIER + * flag is passed, the write is preceded by a system-wide memory fence, + * equivalent to a __threadfence_system() but scoped to the stream + * rather than a CUDA thread. + * + * If the memory was registered via ::cuMemHostRegister(), the device pointer + * should be obtained with ::cuMemHostGetDevicePointer(). + * + * Support for this can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS. + * + * \param stream The stream to do the write in. + * \param addr The device address to write to. + * \param value The value to write. + * \param flags See ::CUstreamWriteValue_flags. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_NOT_SUPPORTED + * \notefnerr + * + * \sa ::cuStreamWriteValue32, + * ::cuStreamWaitValue32, + * ::cuStreamWaitValue64, + * ::cuStreamBatchMemOp, + * ::cuMemHostRegister, + * ::cuEventRecord + */ +CUresult CUDAAPI cuStreamWriteValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags); + +/** + * \brief Batch operations to synchronize the stream via memory operations + * + * This is a batch version of ::cuStreamWaitValue32() and ::cuStreamWriteValue32(). + * Batching operations may avoid some performance overhead in both the API call + * and the device execution versus adding them to the stream in separate API + * calls. The operations are enqueued in the order they appear in the array. + * + * See ::CUstreamBatchMemOpType for the full set of supported operations, and + * ::cuStreamWaitValue32(), ::cuStreamWaitValue64(), ::cuStreamWriteValue32(), + * and ::cuStreamWriteValue64() for details of specific operations. + * + * Basic support for this can be queried with ::cuDeviceGetAttribute() and + * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS. See related APIs for details + * on querying support for specific operations. + * + * \param stream The stream to enqueue the operations in. + * \param count The number of operations in the array. Must be less than 256. + * \param paramArray The types and parameters of the individual operations. + * \param flags Reserved for future expansion; must be 0. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_NOT_SUPPORTED * \notefnerr * * \sa ::cuStreamWaitValue32, + * ::cuStreamWaitValue64, * ::cuStreamWriteValue32, + * ::cuStreamWriteValue64, * ::cuMemHostRegister */ CUresult CUDAAPI cuStreamBatchMemOp(CUstream stream, unsigned int count, CUstreamBatchMemOpParams *paramArray, unsigned int flags); #endif /* __CUDA_API_VERSION >= 8000 */ -/** @} */ /* END CUDA_EVENT */ +/** @} */ /* END CUDA_MEMOP */ /** * \defgroup CUDA_EXEC Execution Control @@ -8711,8 +10454,12 @@ CUresult CUDAAPI cuStreamBatchMemOp(CUstream stream, unsigned int count, CUstrea * would return the value 13. Note that this will return a value of 10 for * legacy cubins that do not have a properly-encoded binary architecture * version. - * - ::CU_FUNC_CACHE_MODE_CA: The attribute to indicate whether the function has + * - ::CU_FUNC_CACHE_MODE_CA: The attribute to indicate whether the function has * been compiled with user specified option "-Xptxas --dlcm=ca" set . + * - ::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES: The maximum size in bytes of + * dynamically-allocated shared memory. + * - ::CU_FUNC_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT: Preferred shared memory-L1 + * cache split ratio in percent of total shared memory. * * \param pi - Returned attribute value * \param attrib - Attribute requested @@ -8730,23 +10477,74 @@ CUresult CUDAAPI cuStreamBatchMemOp(CUstream stream, unsigned int count, CUstrea * \sa ::cuCtxGetCacheConfig, * ::cuCtxSetCacheConfig, * ::cuFuncSetCacheConfig, - * ::cuLaunchKernel + * ::cuLaunchKernel, + * ::cudaFuncGetAttributes + * ::cudaFuncSetAttribute */ CUresult CUDAAPI cuFuncGetAttribute(int *pi, CUfunction_attribute attrib, CUfunction hfunc); +#if __CUDA_API_VERSION >= 9000 + /** - * \brief Sets the preferred cache configuration for a device function + * \brief Sets information about a function * - * On devices where the L1 cache and shared memory use the same hardware - * resources, this sets through \p config the preferred cache configuration for - * the device function \p hfunc. This is only a preference. The driver will use - * the requested configuration if possible, but it is free to choose a different - * configuration if required to execute \p hfunc. Any context-wide preference - * set via ::cuCtxSetCacheConfig() will be overridden by this per-function - * setting unless the per-function setting is ::CU_FUNC_CACHE_PREFER_NONE. In - * that case, the current context-wide setting will be used. + * This call sets the value of a specified attribute \p attrib on the kernel given + * by \p hfunc to an integer value specified by \p val + * This function returns CUDA_SUCCESS if the new value of the attribute could be + * successfully set. If the set fails, this call will return an error. + * Not all attributes can have values set. Attempting to set a value on a read-only + * attribute will result in an error (CUDA_ERROR_INVALID_VALUE) * - * This setting does nothing on devices where the size of the L1 cache and + * Supported attributes for the cuFuncSetAttribute call are: + * - ::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES: This maximum size in bytes of + * dynamically-allocated shared memory. The value should contain the requested + * maximum size of dynamically-allocated shared memory. The sum of this value and + * the function attribute ::CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES cannot exceed the + * device attribute ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN. + * The maximal size of requestable dynamic shared memory may differ by GPU + * architecture. + * - ::CU_FUNC_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT: On devices where the L1 + * cache and shared memory use the same hardware resources, this sets the shared memory + * carveout preference, in percent of the total shared memory. + * See ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR + * This is only a hint, and the driver can choose a different ratio if required to execute the function. + * + * \param hfunc - Function to query attribute of + * \param attrib - Attribute requested + * \param value - The value to set + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuCtxGetCacheConfig, + * ::cuCtxSetCacheConfig, + * ::cuFuncSetCacheConfig, + * ::cuLaunchKernel, + * ::cudaFuncGetAttributes + * ::cudaFuncSetAttribute + */ +CUresult CUDAAPI cuFuncSetAttribute(CUfunction hfunc, CUfunction_attribute attrib, int value); +#endif // __CUDA_API_VERSION >= 9000 + +/** + * \brief Sets the preferred cache configuration for a device function + * + * On devices where the L1 cache and shared memory use the same hardware + * resources, this sets through \p config the preferred cache configuration for + * the device function \p hfunc. This is only a preference. The driver will use + * the requested configuration if possible, but it is free to choose a different + * configuration if required to execute \p hfunc. Any context-wide preference + * set via ::cuCtxSetCacheConfig() will be overridden by this per-function + * setting unless the per-function setting is ::CU_FUNC_CACHE_PREFER_NONE. In + * that case, the current context-wide setting will be used. + * + * This setting does nothing on devices where the size of the L1 cache and * shared memory are fixed. * * Launching a kernel with a different preference than the most recent @@ -8773,7 +10571,8 @@ CUresult CUDAAPI cuFuncGetAttribute(int *pi, CUfunction_attribute attrib, CUfunc * \sa ::cuCtxGetCacheConfig, * ::cuCtxSetCacheConfig, * ::cuFuncGetAttribute, - * ::cuLaunchKernel + * ::cuLaunchKernel, + * ::cudaFuncSetCacheConfig */ CUresult CUDAAPI cuFuncSetCacheConfig(CUfunction hfunc, CUfunc_cache config); @@ -8781,28 +10580,28 @@ CUresult CUDAAPI cuFuncSetCacheConfig(CUfunction hfunc, CUfunc_cache config); /** * \brief Sets the shared memory configuration for a device function. * - * On devices with configurable shared memory banks, this function will + * On devices with configurable shared memory banks, this function will * force all subsequent launches of the specified device function to have * the given shared memory bank size configuration. On any given launch of the * function, the shared memory configuration of the device will be temporarily * changed if needed to suit the function's preferred configuration. Changes in - * shared memory configuration between subsequent launches of functions, + * shared memory configuration between subsequent launches of functions, * may introduce a device side synchronization point. * - * Any per-function setting of shared memory bank size set via + * Any per-function setting of shared memory bank size set via * ::cuFuncSetSharedMemConfig will override the context wide setting set with * ::cuCtxSetSharedMemConfig. * * Changing the shared memory bank size will not increase shared memory usage - * or affect occupancy of kernels, but may have major effects on performance. + * or affect occupancy of kernels, but may have major effects on performance. * Larger bank sizes will allow for greater potential bandwidth to shared memory, - * but will change what kinds of accesses to shared memory will result in bank + * but will change what kinds of accesses to shared memory will result in bank * conflicts. * * This function will do nothing on devices with fixed shared memory bank size. * * The supported bank configurations are: - * - ::CU_SHARED_MEM_CONFIG_DEFAULT_BANK_SIZE: use the context's shared memory + * - ::CU_SHARED_MEM_CONFIG_DEFAULT_BANK_SIZE: use the context's shared memory * configuration when launching this function. * - ::CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE: set shared memory bank width to * be natively four bytes when launching this function. @@ -8825,7 +10624,8 @@ CUresult CUDAAPI cuFuncSetCacheConfig(CUfunction hfunc, CUfunc_cache config); * ::cuCtxGetSharedMemConfig, * ::cuCtxSetSharedMemConfig, * ::cuFuncGetAttribute, - * ::cuLaunchKernel + * ::cuLaunchKernel, + * ::cudaFuncSetSharedMemConfig */ CUresult CUDAAPI cuFuncSetSharedMemConfig(CUfunction hfunc, CUsharedconfig config); #endif @@ -8902,424 +10702,1752 @@ CUresult CUDAAPI cuFuncSetSharedMemConfig(CUfunction hfunc, CUsharedconfig confi * block shape, shared size and parameter info associated with \p f * is overwritten. * - * Note that to use ::cuLaunchKernel(), the kernel \p f must either have - * been compiled with toolchain version 3.2 or later so that it will - * contain kernel parameter information, or have no kernel parameters. - * If either of these conditions is not met, then ::cuLaunchKernel() will - * return ::CUDA_ERROR_INVALID_IMAGE. + * Note that to use ::cuLaunchKernel(), the kernel \p f must either have + * been compiled with toolchain version 3.2 or later so that it will + * contain kernel parameter information, or have no kernel parameters. + * If either of these conditions is not met, then ::cuLaunchKernel() will + * return ::CUDA_ERROR_INVALID_IMAGE. + * + * \param f - Kernel to launch + * \param gridDimX - Width of grid in blocks + * \param gridDimY - Height of grid in blocks + * \param gridDimZ - Depth of grid in blocks + * \param blockDimX - X dimension of each thread block + * \param blockDimY - Y dimension of each thread block + * \param blockDimZ - Z dimension of each thread block + * \param sharedMemBytes - Dynamic shared-memory size per thread block in bytes + * \param hStream - Stream identifier + * \param kernelParams - Array of pointers to kernel parameters + * \param extra - Extra options + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_IMAGE, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_LAUNCH_FAILED, + * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, + * ::CUDA_ERROR_LAUNCH_TIMEOUT, + * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * \note_null_stream + * \notefnerr + * + * \sa ::cuCtxGetCacheConfig, + * ::cuCtxSetCacheConfig, + * ::cuFuncSetCacheConfig, + * ::cuFuncGetAttribute, + * ::cudaLaunchKernel + */ +CUresult CUDAAPI cuLaunchKernel(CUfunction f, + unsigned int gridDimX, + unsigned int gridDimY, + unsigned int gridDimZ, + unsigned int blockDimX, + unsigned int blockDimY, + unsigned int blockDimZ, + unsigned int sharedMemBytes, + CUstream hStream, + void **kernelParams, + void **extra); +#endif /* __CUDA_API_VERSION >= 4000 */ +#if __CUDA_API_VERSION >= 9000 +/** + * \brief Launches a CUDA function where thread blocks can cooperate and synchronize as they execute + * + * Invokes the kernel \p f on a \p gridDimX x \p gridDimY x \p gridDimZ + * grid of blocks. Each block contains \p blockDimX x \p blockDimY x + * \p blockDimZ threads. + * + * \p sharedMemBytes sets the amount of dynamic shared memory that will be + * available to each thread block. + * + * The device on which this kernel is invoked must have a non-zero value for + * the device attribute ::CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH. + * + * The total number of blocks launched cannot exceed the maximum number of blocks per + * multiprocessor as returned by ::cuOccupancyMaxActiveBlocksPerMultiprocessor (or + * ::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags) times the number of multiprocessors + * as specified by the device attribute ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT. + * + * The kernel cannot make use of CUDA dynamic parallelism. + * + * Kernel parameters must be specified via \p kernelParams. If \p f + * has N parameters, then \p kernelParams needs to be an array of N + * pointers. Each of \p kernelParams[0] through \p kernelParams[N-1] + * must point to a region of memory from which the actual kernel + * parameter will be copied. The number of kernel parameters and their + * offsets and sizes do not need to be specified as that information is + * retrieved directly from the kernel's image. + * + * Calling ::cuLaunchCooperativeKernel() sets persistent function state that is + * the same as function state set through ::cuLaunchKernel API + * + * When the kernel \p f is launched via ::cuLaunchCooperativeKernel(), the previous + * block shape, shared size and parameter info associated with \p f + * is overwritten. + * + * Note that to use ::cuLaunchCooperativeKernel(), the kernel \p f must either have + * been compiled with toolchain version 3.2 or later so that it will + * contain kernel parameter information, or have no kernel parameters. + * If either of these conditions is not met, then ::cuLaunchCooperativeKernel() will + * return ::CUDA_ERROR_INVALID_IMAGE. + * + * \param f - Kernel to launch + * \param gridDimX - Width of grid in blocks + * \param gridDimY - Height of grid in blocks + * \param gridDimZ - Depth of grid in blocks + * \param blockDimX - X dimension of each thread block + * \param blockDimY - Y dimension of each thread block + * \param blockDimZ - Z dimension of each thread block + * \param sharedMemBytes - Dynamic shared-memory size per thread block in bytes + * \param hStream - Stream identifier + * \param kernelParams - Array of pointers to kernel parameters + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_IMAGE, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_LAUNCH_FAILED, + * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, + * ::CUDA_ERROR_LAUNCH_TIMEOUT, + * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, + * ::CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE, + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * \note_null_stream + * \notefnerr + * + * \sa ::cuCtxGetCacheConfig, + * ::cuCtxSetCacheConfig, + * ::cuFuncSetCacheConfig, + * ::cuFuncGetAttribute, + * ::cuLaunchCooperativeKernelMultiDevice, + * ::cudaLaunchCooperativeKernel + */ +CUresult CUDAAPI cuLaunchCooperativeKernel(CUfunction f, + unsigned int gridDimX, + unsigned int gridDimY, + unsigned int gridDimZ, + unsigned int blockDimX, + unsigned int blockDimY, + unsigned int blockDimZ, + unsigned int sharedMemBytes, + CUstream hStream, + void **kernelParams); + +/** + * \brief Launches CUDA functions on multiple devices where thread blocks can cooperate and synchronize as they execute + * + * Invokes kernels as specified in the \p launchParamsList array where each element + * of the array specifies all the parameters required to perform a single kernel launch. + * These kernels can cooperate and synchronize as they execute. The size of the array is + * specified by \p numDevices. + * + * No two kernels can be launched on the same device. All the devices targeted by this + * multi-device launch must be identical. All devices must have a non-zero value for the + * device attribute ::CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH. + * + * All kernels launched must be identical with respect to the compiled code. Note that + * any __device__, __constant__ or __managed__ variables present in the module that owns + * the kernel launched on each device, are independently instantiated on every device. + * It is the application's responsiblity to ensure these variables are initialized and + * used appropriately. + * + * The size of the grids as specified in blocks, the size of the blocks themselves + * and the amount of shared memory used by each thread block must also match across + * all launched kernels. + * + * The streams used to launch these kernels must have been created via either ::cuStreamCreate + * or ::cuStreamCreateWithPriority. The NULL stream or ::CU_STREAM_LEGACY or ::CU_STREAM_PER_THREAD + * cannot be used. + * + * The total number of blocks launched per kernel cannot exceed the maximum number of blocks + * per multiprocessor as returned by ::cuOccupancyMaxActiveBlocksPerMultiprocessor (or + * ::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags) times the number of multiprocessors + * as specified by the device attribute ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT. Since the + * total number of blocks launched per device has to match across all devices, the maximum + * number of blocks that can be launched per device will be limited by the device with the + * least number of multiprocessors. + * + * The kernels cannot make use of CUDA dynamic parallelism. + * + * The ::CUDA_LAUNCH_PARAMS structure is defined as: + * \code + typedef struct CUDA_LAUNCH_PARAMS_st + { + CUfunction function; + unsigned int gridDimX; + unsigned int gridDimY; + unsigned int gridDimZ; + unsigned int blockDimX; + unsigned int blockDimY; + unsigned int blockDimZ; + unsigned int sharedMemBytes; + CUstream hStream; + void **kernelParams; + } CUDA_LAUNCH_PARAMS; + * \endcode + * where: + * - ::CUDA_LAUNCH_PARAMS::function specifies the kernel to be launched. All functions must + * be identical with respect to the compiled code. + * - ::CUDA_LAUNCH_PARAMS::gridDimX is the width of the grid in blocks. This must match across + * all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::gridDimY is the height of the grid in blocks. This must match across + * all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::gridDimZ is the depth of the grid in blocks. This must match across + * all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::blockDimX is the X dimension of each thread block. This must match across + * all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::blockDimX is the Y dimension of each thread block. This must match across + * all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::blockDimZ is the Z dimension of each thread block. This must match across + * all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::sharedMemBytes is the dynamic shared-memory size per thread block in bytes. + * This must match across all kernels launched. + * - ::CUDA_LAUNCH_PARAMS::hStream is the handle to the stream to perform the launch in. This cannot + * be the NULL stream or ::CU_STREAM_LEGACY or ::CU_STREAM_PER_THREAD. The CUDA context associated + * with this stream must match that associated with ::CUDA_LAUNCH_PARAMS::function. + * - ::CUDA_LAUNCH_PARAMS::kernelParams is an array of pointers to kernel parameters. If + * ::CUDA_LAUNCH_PARAMS::function has N parameters, then ::CUDA_LAUNCH_PARAMS::kernelParams + * needs to be an array of N pointers. Each of ::CUDA_LAUNCH_PARAMS::kernelParams[0] through + * ::CUDA_LAUNCH_PARAMS::kernelParams[N-1] must point to a region of memory from which the actual + * kernel parameter will be copied. The number of kernel parameters and their offsets and sizes + * do not need to be specified as that information is retrieved directly from the kernel's image. + * + * By default, the kernel won't begin execution on any GPU until all prior work in all the specified + * streams has completed. This behavior can be overridden by specifying the flag + * ::CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC. When this flag is specified, each kernel + * will only wait for prior work in the stream corresponding to that GPU to complete before it begins + * execution. + * + * Similarly, by default, any subsequent work pushed in any of the specified streams will not begin + * execution until the kernels on all GPUs have completed. This behavior can be overridden by specifying + * the flag ::CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC. When this flag is specified, + * any subsequent work pushed in any of the specified streams will only wait for the kernel launched + * on the GPU corresponding to that stream to complete before it begins execution. + * + * Calling ::cuLaunchCooperativeKernelMultiDevice() sets persistent function state that is + * the same as function state set through ::cuLaunchKernel API when called individually for each + * element in \p launchParamsList. + * + * When kernels are launched via ::cuLaunchCooperativeKernelMultiDevice(), the previous + * block shape, shared size and parameter info associated with each ::CUDA_LAUNCH_PARAMS::function + * in \p launchParamsList is overwritten. + * + * Note that to use ::cuLaunchCooperativeKernelMultiDevice(), the kernels must either have + * been compiled with toolchain version 3.2 or later so that it will + * contain kernel parameter information, or have no kernel parameters. + * If either of these conditions is not met, then ::cuLaunchCooperativeKernelMultiDevice() will + * return ::CUDA_ERROR_INVALID_IMAGE. + * + * \param launchParamsList - List of launch parameters, one per device + * \param numDevices - Size of the \p launchParamsList array + * \param flags - Flags to control launch behavior + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_IMAGE, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_LAUNCH_FAILED, + * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, + * ::CUDA_ERROR_LAUNCH_TIMEOUT, + * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, + * ::CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE, + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * \note_null_stream + * \notefnerr + * + * \sa ::cuCtxGetCacheConfig, + * ::cuCtxSetCacheConfig, + * ::cuFuncSetCacheConfig, + * ::cuFuncGetAttribute, + * ::cuLaunchCooperativeKernel, + * ::cudaLaunchCooperativeKernelMultiDevice + */ +CUresult CUDAAPI cuLaunchCooperativeKernelMultiDevice(CUDA_LAUNCH_PARAMS *launchParamsList, unsigned int numDevices, unsigned int flags); + +#endif /* __CUDA_API_VERSION >= 9000 */ + +#if __CUDA_API_VERSION >= 10000 + +/** + * \brief Enqueues a host function call in a stream + * + * Enqueues a host function to run in a stream. The function will be called + * after currently enqueued work and will block work added after it. + * + * The host function must not make any CUDA API calls. Attempting to use a + * CUDA API may result in ::CUDA_ERROR_NOT_PERMITTED, but this is not required. + * The host function must not perform any synchronization that may depend on + * outstanding CUDA work not mandated to run earlier. Host functions without a + * mandated order (such as in independent streams) execute in undefined order + * and may be serialized. + * + * For the purposes of Unified Memory, execution makes a number of guarantees: + *
      + *
    • The stream is considered idle for the duration of the function's + * execution. Thus, for example, the function may always use memory attached + * to the stream it was enqueued in.
      • + *
    • The start of execution of the function has the same effect as + * synchronizing an event recorded in the same stream immediately prior to + * the function. It thus synchronizes streams which have been "joined" + * prior to the function.
      • + *
    • Adding device work to any stream does not have the effect of making + * the stream active until all preceding host functions and stream callbacks + * have executed. Thus, for + * example, a function might use global attached memory even if work has + * been added to another stream, if the work has been ordered behind the + * function call with an event.
      • + *
    • Completion of the function does not cause a stream to become + * active except as described above. The stream will remain idle + * if no device work follows the function, and will remain idle across + * consecutive host functions or stream callbacks without device work in + * between. Thus, for example, + * stream synchronization can be done by signaling from a host function at the + * end of the stream.
      • + *
    + * + * Note that, in contrast to ::cuStreamAddCallback, the function will not be + * called in the event of an error in the CUDA context. + * + * \param hStream - Stream to enqueue function call in + * \param fn - The function to call once preceding stream operations are complete + * \param userData - User-specified data to be passed to the function + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_NOT_SUPPORTED + * \note_null_stream + * \notefnerr + * + * \sa ::cuStreamCreate, + * ::cuStreamQuery, + * ::cuStreamSynchronize, + * ::cuStreamWaitEvent, + * ::cuStreamDestroy, + * ::cuMemAllocManaged, + * ::cuStreamAttachMemAsync, + * ::cuStreamAddCallback + */ +CUresult CUDAAPI cuLaunchHostFunc(CUstream hStream, CUhostFn fn, void *userData); + +#endif /* __CUDA_API_VERSION >= 10000 */ + +/** @} */ /* END CUDA_EXEC */ + +/** + * \defgroup CUDA_EXEC_DEPRECATED Execution Control [DEPRECATED] + * + * ___MANBRIEF___ deprecated execution control functions of the low-level CUDA + * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ + * + * This section describes the deprecated execution control functions of the + * low-level CUDA driver application programming interface. + * + * @{ + */ + +/** + * \brief Sets the block-dimensions for the function + * + * \deprecated + * + * Specifies the \p x, \p y, and \p z dimensions of the thread blocks that are + * created when the kernel given by \p hfunc is launched. + * + * \param hfunc - Kernel to specify dimensions of + * \param x - X dimension + * \param y - Y dimension + * \param z - Z dimension + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuFuncSetSharedSize, + * ::cuFuncSetCacheConfig, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSeti, + * ::cuParamSetf, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuFuncSetBlockShape(CUfunction hfunc, int x, int y, int z); + +/** + * \brief Sets the dynamic shared-memory size for the function + * + * \deprecated + * + * Sets through \p bytes the amount of dynamic shared memory that will be + * available to each thread block when the kernel given by \p hfunc is launched. + * + * \param hfunc - Kernel to specify dynamic shared-memory size for + * \param bytes - Dynamic shared-memory size per thread in bytes + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetCacheConfig, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSeti, + * ::cuParamSetf, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuFuncSetSharedSize(CUfunction hfunc, unsigned int bytes); + +/** + * \brief Sets the parameter size for the function + * + * \deprecated + * + * Sets through \p numbytes the total size in bytes needed by the function + * parameters of the kernel corresponding to \p hfunc. + * + * \param hfunc - Kernel to set parameter size for + * \param numbytes - Size of parameter list in bytes + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetf, + * ::cuParamSeti, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetSize(CUfunction hfunc, unsigned int numbytes); + +/** + * \brief Adds an integer parameter to the function's argument list + * + * \deprecated + * + * Sets an integer parameter that will be specified the next time the + * kernel corresponding to \p hfunc will be invoked. \p offset is a byte offset. + * + * \param hfunc - Kernel to add parameter to + * \param offset - Offset to add parameter to argument list + * \param value - Value of parameter + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSetf, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuParamSeti(CUfunction hfunc, int offset, unsigned int value); + +/** + * \brief Adds a floating-point parameter to the function's argument list + * + * \deprecated + * + * Sets a floating-point parameter that will be specified the next time the + * kernel corresponding to \p hfunc will be invoked. \p offset is a byte offset. + * + * \param hfunc - Kernel to add parameter to + * \param offset - Offset to add parameter to argument list + * \param value - Value of parameter + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSeti, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetf(CUfunction hfunc, int offset, float value); + +/** + * \brief Adds arbitrary data to the function's argument list + * + * \deprecated + * + * Copies an arbitrary amount of data (specified in \p numbytes) from \p ptr + * into the parameter space of the kernel corresponding to \p hfunc. \p offset + * is a byte offset. + * + * \param hfunc - Kernel to add data to + * \param offset - Offset to add data to argument list + * \param ptr - Pointer to arbitrary data + * \param numbytes - Size of data to copy in bytes + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSetf, + * ::cuParamSeti, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetv(CUfunction hfunc, int offset, void *ptr, unsigned int numbytes); + +/** + * \brief Launches a CUDA function + * + * \deprecated + * + * Invokes the kernel \p f on a 1 x 1 x 1 grid of blocks. The block + * contains the number of threads specified by a previous call to + * ::cuFuncSetBlockShape(). + * + * \param f - Kernel to launch + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_LAUNCH_FAILED, + * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, + * ::CUDA_ERROR_LAUNCH_TIMEOUT, + * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSetf, + * ::cuParamSeti, + * ::cuParamSetv, + * ::cuLaunchGrid, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuLaunch(CUfunction f); + +/** + * \brief Launches a CUDA function + * + * \deprecated + * + * Invokes the kernel \p f on a \p grid_width x \p grid_height grid of + * blocks. Each block contains the number of threads specified by a previous + * call to ::cuFuncSetBlockShape(). + * + * \param f - Kernel to launch + * \param grid_width - Width of grid in blocks + * \param grid_height - Height of grid in blocks + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_LAUNCH_FAILED, + * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, + * ::CUDA_ERROR_LAUNCH_TIMEOUT, + * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSetf, + * ::cuParamSeti, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGridAsync, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuLaunchGrid(CUfunction f, int grid_width, int grid_height); + +/** + * \brief Launches a CUDA function + * + * \deprecated + * + * Invokes the kernel \p f on a \p grid_width x \p grid_height grid of + * blocks. Each block contains the number of threads specified by a previous + * call to ::cuFuncSetBlockShape(). + * + * \param f - Kernel to launch + * \param grid_width - Width of grid in blocks + * \param grid_height - Height of grid in blocks + * \param hStream - Stream identifier + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_LAUNCH_FAILED, + * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, + * ::CUDA_ERROR_LAUNCH_TIMEOUT, + * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, + * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * + * \note In certain cases where cubins are created with no ABI (i.e., using \p ptxas \p --abi-compile \p no), + * this function may serialize kernel launches. In order to force the CUDA driver to retain + * asynchronous behavior, set the ::CU_CTX_LMEM_RESIZE_TO_MAX flag during context creation (see ::cuCtxCreate). + * + * \note_null_stream + * \notefnerr + * + * \sa ::cuFuncSetBlockShape, + * ::cuFuncSetSharedSize, + * ::cuFuncGetAttribute, + * ::cuParamSetSize, + * ::cuParamSetf, + * ::cuParamSeti, + * ::cuParamSetv, + * ::cuLaunch, + * ::cuLaunchGrid, + * ::cuLaunchKernel + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuLaunchGridAsync(CUfunction f, int grid_width, int grid_height, CUstream hStream); + + +/** + * \brief Adds a texture-reference to the function's argument list + * + * \deprecated + * + * Makes the CUDA array or linear memory bound to the texture reference + * \p hTexRef available to a device program as a texture. In this version of + * CUDA, the texture-reference must be obtained via ::cuModuleGetTexRef() and + * the \p texunit parameter must be set to ::CU_PARAM_TR_DEFAULT. + * + * \param hfunc - Kernel to add texture-reference to + * \param texunit - Texture unit (must be ::CU_PARAM_TR_DEFAULT) + * \param hTexRef - Texture-reference to add to argument list + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_CONTEXT, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + */ +__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetTexRef(CUfunction hfunc, int texunit, CUtexref hTexRef); +/** @} */ /* END CUDA_EXEC_DEPRECATED */ + +#if __CUDA_API_VERSION >= 10000 +/** + * \defgroup CUDA_GRAPH Graph Management + * + * ___MANBRIEF___ graph management functions of the low-level CUDA driver API + * (___CURRENT_FILE___) ___ENDMANBRIEF___ + * + * This section describes the graph management functions of the low-level CUDA + * driver application programming interface. + * + * @{ + */ + +/** + * \brief Creates a graph + * + * Creates an empty graph, which is returned via \p phGraph. + * + * \param phGraph - Returns newly created graph + * \param flags - Graph creation flags, must be 0 + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_OUT_OF_MEMORY + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemcpyNode, + * ::cuGraphAddMemsetNode, + * ::cuGraphInstantiate, + * ::cuGraphDestroy, + * ::cuGraphGetNodes, + * ::cuGraphGetRootNodes, + * ::cuGraphGetEdges, + * ::cuGraphClone + */ +CUresult CUDAAPI cuGraphCreate(CUgraph *phGraph, unsigned int flags); + +/** + * \brief Creates a kernel execution node and adds it to a graph + * + * Creates a new kernel execution node and adds it to \p hGraph with \p numDependencies + * dependencies specified via \p dependencies and arguments specified in \p nodeParams. + * It is possible for \p numDependencies to be 0, in which case the node will be placed + * at the root of the graph. \p dependencies may not have any duplicate entries. + * A handle to the new node will be returned in \p phGraphNode. + * + * The CUDA_KERNEL_NODE_PARAMS structure is defined as: + * + * \code + * typedef struct CUDA_KERNEL_NODE_PARAMS_st { + * CUfunction func; + * unsigned int gridDimX; + * unsigned int gridDimY; + * unsigned int gridDimZ; + * unsigned int blockDimX; + * unsigned int blockDimY; + * unsigned int blockDimZ; + * unsigned int sharedMemBytes; + * void **kernelParams; + * void **extra; + * } CUDA_KERNEL_NODE_PARAMS; + * \endcode + * + * When the graph is launched, the node will invoke kernel \p func on a (\p gridDimX x + * \p gridDimY x \p gridDimZ) grid of blocks. Each block contains + * (\p blockDimX x \p blockDimY x \p blockDimZ) threads. + * + * \p sharedMemBytes sets the amount of dynamic shared memory that will be + * available to each thread block. + * + * Kernel parameters to \p func can be specified in one of two ways: + * + * 1) Kernel parameters can be specified via \p kernelParams. If the kernel has N + * parameters, then \p kernelParams needs to be an array of N pointers. Each pointer, + * from \p kernelParams[0] to \p kernelParams[N-1], points to the region of memory from which the actual + * parameter will be copied. The number of kernel parameters and their offsets and sizes do not need + * to be specified as that information is retrieved directly from the kernel's image. + * + * 2) Kernel parameters can also be packaged by the application into a single buffer that is passed in + * via \p extra. This places the burden on the application of knowing each kernel + * parameter's size and alignment/padding within the buffer. The \p extra parameter exists + * to allow this function to take additional less commonly used arguments. \p extra specifies + * a list of names of extra settings and their corresponding values. Each extra setting name is + * immediately followed by the corresponding value. The list must be terminated with either NULL or + * CU_LAUNCH_PARAM_END. + * + * - ::CU_LAUNCH_PARAM_END, which indicates the end of the \p extra + * array; + * - ::CU_LAUNCH_PARAM_BUFFER_POINTER, which specifies that the next + * value in \p extra will be a pointer to a buffer + * containing all the kernel parameters for launching kernel + * \p func; + * - ::CU_LAUNCH_PARAM_BUFFER_SIZE, which specifies that the next + * value in \p extra will be a pointer to a size_t + * containing the size of the buffer specified with + * ::CU_LAUNCH_PARAM_BUFFER_POINTER; + * + * The error ::CUDA_ERROR_INVALID_VALUE will be returned if kernel parameters are specified with both + * \p kernelParams and \p extra (i.e. both \p kernelParams and + * \p extra are non-NULL). + * + * The \p kernelParams or \p extra array, as well as the argument values it points to, + * are copied during this call. + * + * \note Kernels launched using graphs must not use texture and surface references. Reading or + * writing through any texture or surface reference is undefined behavior. + * This restriction does not apply to texture and surface objects. + * + * \param phGraphNode - Returns newly created node + * \param hGraph - Graph to which to add the node + * \param dependencies - Dependencies of the node + * \param numDependencies - Number of dependencies + * \param nodeParams - Parameters for the GPU execution node + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuLaunchKernel, + * ::cuGraphKernelNodeGetParams, + * ::cuGraphKernelNodeSetParams, + * ::cuGraphCreate, + * ::cuGraphDestroyNode, + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemcpyNode, + * ::cuGraphAddMemsetNode + */ +CUresult CUDAAPI cuGraphAddKernelNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_KERNEL_NODE_PARAMS *nodeParams); + +/** + * \brief Returns a kernel node's parameters + * + * Returns the parameters of kernel node \p hNode in \p nodeParams. + * The \p kernelParams or \p extra array returned in \p nodeParams, + * as well as the argument values it points to, are owned by the node. + * This memory remains valid until the node is destroyed or its + * parameters are modified, and should not be modified + * directly. Use ::cuGraphKernelNodeSetParams to update the + * parameters of this node. + * + * The params will contain either \p kernelParams or \p extra, + * according to which of these was most recently set on the node. + * + * \param hNode - Node to get the parameters for + * \param nodeParams - Pointer to return the parameters + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuLaunchKernel, + * ::cuGraphAddKernelNode, + * ::cuGraphKernelNodeSetParams + */ +CUresult CUDAAPI cuGraphKernelNodeGetParams(CUgraphNode hNode, CUDA_KERNEL_NODE_PARAMS *nodeParams); + +/** + * \brief Sets a kernel node's parameters + * + * Sets the parameters of kernel node \p hNode to \p nodeParams. + * + * \param hNode - Node to set the parameters for + * \param nodeParams - Parameters to copy + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_HANDLE, + * ::CUDA_ERROR_OUT_OF_MEMORY + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuLaunchKernel, + * ::cuGraphAddKernelNode, + * ::cuGraphKernelNodeGetParams + */ +CUresult CUDAAPI cuGraphKernelNodeSetParams(CUgraphNode hNode, const CUDA_KERNEL_NODE_PARAMS *nodeParams); + +/** + * \brief Creates a memcpy node and adds it to a graph + * + * Creates a new memcpy node and adds it to \p hGraph with \p numDependencies + * dependencies specified via \p dependencies. + * It is possible for \p numDependencies to be 0, in which case the node will be placed + * at the root of the graph. \p dependencies may not have any duplicate entries. + * A handle to the new node will be returned in \p phGraphNode. + * + * When the graph is launched, the node will perform the memcpy described by \p copyParams. + * See ::cuMemcpy3D() for a description of the structure and its restrictions. + * + * Memcpy nodes have some additional restrictions with regards to managed memory, if the + * system contains at least one device which has a zero value for the device attribute + * ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS. If one or more of the operands refer + * to managed memory, then using the memory type ::CU_MEMORYTYPE_UNIFIED is disallowed + * for those operand(s). The managed memory will be treated as residing on either the + * host or the device, depending on which memory type is specified. + * + * \param phGraphNode - Returns newly created node + * \param hGraph - Graph to which to add the node + * \param dependencies - Dependencies of the node + * \param numDependencies - Number of dependencies + * \param copyParams - Parameters for the memory copy + * \param ctx - Context on which to run the node + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuMemcpy3D, + * ::cuGraphMemcpyNodeGetParams, + * ::cuGraphMemcpyNodeSetParams, + * ::cuGraphCreate, + * ::cuGraphDestroyNode, + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemsetNode + */ +CUresult CUDAAPI cuGraphAddMemcpyNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_MEMCPY3D *copyParams, CUcontext ctx); + +/** + * \brief Returns a memcpy node's parameters + * + * Returns the parameters of memcpy node \p hNode in \p nodeParams. + * + * \param hNode - Node to get the parameters for + * \param nodeParams - Pointer to return the parameters + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuMemcpy3D, + * ::cuGraphAddMemcpyNode, + * ::cuGraphMemcpyNodeSetParams + */ +CUresult CUDAAPI cuGraphMemcpyNodeGetParams(CUgraphNode hNode, CUDA_MEMCPY3D *nodeParams); + +/** + * \brief Sets a memcpy node's parameters + * + * Sets the parameters of memcpy node \p hNode to \p nodeParams. + * + * \param hNode - Node to set the parameters for + * \param nodeParams - Parameters to copy + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuMemcpy3D, + * ::cuGraphAddMemcpyNode, + * ::cuGraphMemcpyNodeGetParams + */ +CUresult CUDAAPI cuGraphMemcpyNodeSetParams(CUgraphNode hNode, const CUDA_MEMCPY3D *nodeParams); + +/** + * \brief Creates a memset node and adds it to a graph + * + * Creates a new memset node and adds it to \p hGraph with \p numDependencies + * dependencies specified via \p dependencies. + * It is possible for \p numDependencies to be 0, in which case the node will be placed + * at the root of the graph. \p dependencies may not have any duplicate entries. + * A handle to the new node will be returned in \p phGraphNode. + * + * The element size must be 1, 2, or 4 bytes. + * When the graph is launched, the node will perform the memset described by \p memsetParams. + * + * \param phGraphNode - Returns newly created node + * \param hGraph - Graph to which to add the node + * \param dependencies - Dependencies of the node + * \param numDependencies - Number of dependencies + * \param memsetParams - Parameters for the memory set + * \param ctx - Context on which to run the node + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_INVALID_CONTEXT + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuMemsetD2D32, + * ::cuGraphMemsetNodeGetParams, + * ::cuGraphMemsetNodeSetParams, + * ::cuGraphCreate, + * ::cuGraphDestroyNode, + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemcpyNode + */ +CUresult CUDAAPI cuGraphAddMemsetNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_MEMSET_NODE_PARAMS *memsetParams, CUcontext ctx); + +/** + * \brief Returns a memset node's parameters + * + * Returns the parameters of memset node \p hNode in \p nodeParams. + * + * \param hNode - Node to get the parameters for + * \param nodeParams - Pointer to return the parameters + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuMemsetD2D32, + * ::cuGraphAddMemsetNode, + * ::cuGraphMemsetNodeSetParams + */ +CUresult CUDAAPI cuGraphMemsetNodeGetParams(CUgraphNode hNode, CUDA_MEMSET_NODE_PARAMS *nodeParams); + +/** + * \brief Sets a memset node's parameters + * + * Sets the parameters of memset node \p hNode to \p nodeParams. + * + * \param hNode - Node to set the parameters for + * \param nodeParams - Parameters to copy + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuMemsetD2D32, + * ::cuGraphAddMemsetNode, + * ::cuGraphMemsetNodeGetParams + */ +CUresult CUDAAPI cuGraphMemsetNodeSetParams(CUgraphNode hNode, const CUDA_MEMSET_NODE_PARAMS *nodeParams); + +/** + * \brief Creates a host execution node and adds it to a graph + * + * Creates a new CPU execution node and adds it to \p hGraph with \p numDependencies + * dependencies specified via \p dependencies and arguments specified in \p nodeParams. + * It is possible for \p numDependencies to be 0, in which case the node will be placed + * at the root of the graph. \p dependencies may not have any duplicate entries. + * A handle to the new node will be returned in \p phGraphNode. + * + * When the graph is launched, the node will invoke the specified CPU function. + * Host nodes are not supported under MPS with pre-Volta GPUs. + * + * \param phGraphNode - Returns newly created node + * \param hGraph - Graph to which to add the node + * \param dependencies - Dependencies of the node + * \param numDependencies - Number of dependencies + * \param nodeParams - Parameters for the host node + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_NOT_SUPPORTED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuLaunchHostFunc, + * ::cuGraphHostNodeGetParams, + * ::cuGraphHostNodeSetParams, + * ::cuGraphCreate, + * ::cuGraphDestroyNode, + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddMemcpyNode, + * ::cuGraphAddMemsetNode + */ +CUresult CUDAAPI cuGraphAddHostNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_HOST_NODE_PARAMS *nodeParams); + +/** + * \brief Returns a host node's parameters + * + * Returns the parameters of host node \p hNode in \p nodeParams. + * + * \param hNode - Node to get the parameters for + * \param nodeParams - Pointer to return the parameters + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuLaunchHostFunc, + * ::cuGraphAddHostNode, + * ::cuGraphHostNodeSetParams + */ +CUresult CUDAAPI cuGraphHostNodeGetParams(CUgraphNode hNode, CUDA_HOST_NODE_PARAMS *nodeParams); + +/** + * \brief Sets a host node's parameters + * + * Sets the parameters of host node \p hNode to \p nodeParams. + * + * \param hNode - Node to set the parameters for + * \param nodeParams - Parameters to copy + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuLaunchHostFunc, + * ::cuGraphAddHostNode, + * ::cuGraphHostNodeGetParams + */ +CUresult CUDAAPI cuGraphHostNodeSetParams(CUgraphNode hNode, const CUDA_HOST_NODE_PARAMS *nodeParams); + +/** + * \brief Creates a child graph node and adds it to a graph + * + * Creates a new node which executes an embedded graph, and adds it to \p hGraph with + * \p numDependencies dependencies specified via \p dependencies. + * It is possible for \p numDependencies to be 0, in which case the node will be placed + * at the root of the graph. \p dependencies may not have any duplicate entries. + * A handle to the new node will be returned in \p phGraphNode. + * + * The node executes an embedded child graph. The child graph is cloned in this call. + * + * \param phGraphNode - Returns newly created node + * \param hGraph - Graph to which to add the node + * \param dependencies - Dependencies of the node + * \param numDependencies - Number of dependencies + * \param childGraph - The graph to clone into this node + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphChildGraphNodeGetGraph, + * ::cuGraphCreate, + * ::cuGraphDestroyNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemcpyNode, + * ::cuGraphAddMemsetNode, + * ::cuGraphClone + */ +CUresult CUDAAPI cuGraphAddChildGraphNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, CUgraph childGraph); + +/** + * \brief Gets a handle to the embedded graph of a child graph node + * + * Gets a handle to the embedded graph in a child graph node. This call + * does not clone the graph. Changes to the graph will be reflected in + * the node, and the node retains ownership of the graph. + * + * \param hNode - Node to get the embedded graph for + * \param phGraph - Location to store a handle to the graph + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphAddChildGraphNode, + * ::cuGraphNodeFindInClone + */ +CUresult CUDAAPI cuGraphChildGraphNodeGetGraph(CUgraphNode hNode, CUgraph *phGraph); + +/** + * \brief Creates an empty node and adds it to a graph + * + * Creates a new node which performs no operation, and adds it to \p hGraph with + * \p numDependencies dependencies specified via \p dependencies. + * It is possible for \p numDependencies to be 0, in which case the node will be placed + * at the root of the graph. \p dependencies may not have any duplicate entries. + * A handle to the new node will be returned in \p phGraphNode. + * + * An empty node performs no operation during execution, but can be used for + * transitive ordering. For example, a phased execution graph with 2 groups of n + * nodes with a barrier between them can be represented using an empty node and + * 2*n dependency edges, rather than no empty node and n^2 dependency edges. + * + * \param phGraphNode - Returns newly created node + * \param hGraph - Graph to which to add the node + * \param dependencies - Dependencies of the node + * \param numDependencies - Number of dependencies + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE, + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphCreate, + * ::cuGraphDestroyNode, + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemcpyNode, + * ::cuGraphAddMemsetNode + */ +CUresult CUDAAPI cuGraphAddEmptyNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies); + +/** + * \brief Clones a graph + * + * This function creates a copy of \p originalGraph and returns it in \p * phGraphClone. + * All parameters are copied into the cloned graph. The original graph may be modified + * after this call without affecting the clone. + * + * Child graph nodes in the original graph are recursively copied into the clone. + * + * \param phGraphClone - Returns newly created cloned graph + * \param originalGraph - Graph to clone + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * ::CUDA_ERROR_OUT_OF_MEMORY + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphCreate, + * ::cuGraphNodeFindInClone + */ +CUresult CUDAAPI cuGraphClone(CUgraph *phGraphClone, CUgraph originalGraph); + +/** + * \brief Finds a cloned version of a node + * + * This function returns the node in \p hClonedGraph corresponding to \p hOriginalNode + * in the original graph. + * + * \p hClonedGraph must have been cloned from \p hOriginalGraph via ::cuGraphClone. + * \p hOriginalNode must have been in \p hOriginalGraph at the time of the call to + * ::cuGraphClone, and the corresponding cloned node in \p hClonedGraph must not have + * been removed. The cloned node is then returned via \p phClonedNode. + * + * \param phNode - Returns handle to the cloned node + * \param hOriginalNode - Handle to the original node + * \param hClonedGraph - Cloned graph to query + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE, + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphClone + */ +CUresult CUDAAPI cuGraphNodeFindInClone(CUgraphNode *phNode, CUgraphNode hOriginalNode, CUgraph hClonedGraph); + +/** + * \brief Returns a node's type + * + * Returns the node type of \p hNode in \p type. + * + * \param hNode - Node to query + * \param type - Pointer to return the node type + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphGetNodes, + * ::cuGraphGetRootNodes, + * ::cuGraphChildGraphNodeGetGraph, + * ::cuGraphKernelNodeGetParams, + * ::cuGraphKernelNodeSetParams, + * ::cuGraphHostNodeGetParams, + * ::cuGraphHostNodeSetParams, + * ::cuGraphMemcpyNodeGetParams, + * ::cuGraphMemcpyNodeSetParams, + * ::cuGraphMemsetNodeGetParams, + * ::cuGraphMemsetNodeSetParams + */ +CUresult CUDAAPI cuGraphNodeGetType(CUgraphNode hNode, CUgraphNodeType *type); + +/** + * \brief Returns a graph's nodes + * + * Returns a list of \p hGraph's nodes. \p nodes may be NULL, in which case this + * function will return the number of nodes in \p numNodes. Otherwise, + * \p numNodes entries will be filled in. If \p numNodes is higher than the actual + * number of nodes, the remaining entries in \p nodes will be set to NULL, and the + * number of nodes actually obtained will be returned in \p numNodes. * - * \param f - Kernel to launch - * \param gridDimX - Width of grid in blocks - * \param gridDimY - Height of grid in blocks - * \param gridDimZ - Depth of grid in blocks - * \param blockDimX - X dimension of each thread block - * \param blockDimY - Y dimension of each thread block - * \param blockDimZ - Z dimension of each thread block - * \param sharedMemBytes - Dynamic shared-memory size per thread block in bytes - * \param hStream - Stream identifier - * \param kernelParams - Array of pointers to kernel parameters - * \param extra - Extra options + * \param hGraph - Graph to query + * \param nodes - Pointer to return the nodes + * \param numNodes - See description * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_HANDLE, - * ::CUDA_ERROR_INVALID_IMAGE, - * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_LAUNCH_FAILED, - * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, - * ::CUDA_ERROR_LAUNCH_TIMEOUT, - * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED - * \note_null_stream + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuCtxGetCacheConfig, - * ::cuCtxSetCacheConfig, - * ::cuFuncSetCacheConfig, - * ::cuFuncGetAttribute + * \sa + * ::cuGraphCreate, + * ::cuGraphGetRootNodes, + * ::cuGraphGetEdges, + * ::cuGraphNodeGetType, + * ::cuGraphNodeGetDependencies, + * ::cuGraphNodeGetDependentNodes */ -CUresult CUDAAPI cuLaunchKernel(CUfunction f, - unsigned int gridDimX, - unsigned int gridDimY, - unsigned int gridDimZ, - unsigned int blockDimX, - unsigned int blockDimY, - unsigned int blockDimZ, - unsigned int sharedMemBytes, - CUstream hStream, - void **kernelParams, - void **extra); -#endif /* __CUDA_API_VERSION >= 4000 */ - -/** @} */ /* END CUDA_EXEC */ +CUresult CUDAAPI cuGraphGetNodes(CUgraph hGraph, CUgraphNode *nodes, size_t *numNodes); /** - * \defgroup CUDA_EXEC_DEPRECATED Execution Control [DEPRECATED] + * \brief Returns a graph's root nodes * - * ___MANBRIEF___ deprecated execution control functions of the low-level CUDA - * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ + * Returns a list of \p hGraph's root nodes. \p rootNodes may be NULL, in which case this + * function will return the number of root nodes in \p numRootNodes. Otherwise, + * \p numRootNodes entries will be filled in. If \p numRootNodes is higher than the actual + * number of root nodes, the remaining entries in \p rootNodes will be set to NULL, and the + * number of nodes actually obtained will be returned in \p numRootNodes. * - * This section describes the deprecated execution control functions of the - * low-level CUDA driver application programming interface. + * \param hGraph - Graph to query + * \param rootNodes - Pointer to return the root nodes + * \param numRootNodes - See description * - * @{ + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphCreate, + * ::cuGraphGetNodes, + * ::cuGraphGetEdges, + * ::cuGraphNodeGetType, + * ::cuGraphNodeGetDependencies, + * ::cuGraphNodeGetDependentNodes */ +CUresult CUDAAPI cuGraphGetRootNodes(CUgraph hGraph, CUgraphNode *rootNodes, size_t *numRootNodes); /** - * \brief Sets the block-dimensions for the function - * - * \deprecated + * \brief Returns a graph's dependency edges * - * Specifies the \p x, \p y, and \p z dimensions of the thread blocks that are - * created when the kernel given by \p hfunc is launched. + * Returns a list of \p hGraph's dependency edges. Edges are returned via corresponding + * indices in \p from and \p to; that is, the node in \p to[i] has a dependency on the + * node in \p from[i]. \p from and \p to may both be NULL, in which + * case this function only returns the number of edges in \p numEdges. Otherwise, + * \p numEdges entries will be filled in. If \p numEdges is higher than the actual + * number of edges, the remaining entries in \p from and \p to will be set to NULL, and + * the number of edges actually returned will be written to \p numEdges. * - * \param hfunc - Kernel to specify dimensions of - * \param x - X dimension - * \param y - Y dimension - * \param z - Z dimension + * \param hGraph - Graph to get the edges from + * \param from - Location to return edge endpoints + * \param to - Location to return edge endpoints + * \param numEdges - See description * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_HANDLE, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetSharedSize, - * ::cuFuncSetCacheConfig, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSeti, - * ::cuParamSetf, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphGetNodes, + * ::cuGraphGetRootNodes, + * ::cuGraphAddDependencies, + * ::cuGraphRemoveDependencies, + * ::cuGraphNodeGetDependencies, + * ::cuGraphNodeGetDependentNodes */ -CUresult CUDAAPI cuFuncSetBlockShape(CUfunction hfunc, int x, int y, int z); +CUresult CUDAAPI cuGraphGetEdges(CUgraph hGraph, CUgraphNode *from, CUgraphNode *to, size_t *numEdges); /** - * \brief Sets the dynamic shared-memory size for the function - * - * \deprecated + * \brief Returns a node's dependencies * - * Sets through \p bytes the amount of dynamic shared memory that will be - * available to each thread block when the kernel given by \p hfunc is launched. + * Returns a list of \p node's dependencies. \p dependencies may be NULL, in which case this + * function will return the number of dependencies in \p numDependencies. Otherwise, + * \p numDependencies entries will be filled in. If \p numDependencies is higher than the actual + * number of dependencies, the remaining entries in \p dependencies will be set to NULL, and the + * number of nodes actually obtained will be returned in \p numDependencies. * - * \param hfunc - Kernel to specify dynamic shared-memory size for - * \param bytes - Dynamic shared-memory size per thread in bytes + * \param hNode - Node to query + * \param dependencies - Pointer to return the dependencies + * \param numDependencies - See description * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_HANDLE, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetCacheConfig, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSeti, - * ::cuParamSetf, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphNodeGetDependentNodes, + * ::cuGraphGetNodes, + * ::cuGraphGetRootNodes, + * ::cuGraphGetEdges, + * ::cuGraphAddDependencies, + * ::cuGraphRemoveDependencies */ -CUresult CUDAAPI cuFuncSetSharedSize(CUfunction hfunc, unsigned int bytes); +CUresult CUDAAPI cuGraphNodeGetDependencies(CUgraphNode hNode, CUgraphNode *dependencies, size_t *numDependencies); /** - * \brief Sets the parameter size for the function - * - * \deprecated + * \brief Returns a node's dependent nodes * - * Sets through \p numbytes the total size in bytes needed by the function - * parameters of the kernel corresponding to \p hfunc. + * Returns a list of \p node's dependent nodes. \p dependentNodes may be NULL, in which + * case this function will return the number of dependent nodes in \p numDependentNodes. + * Otherwise, \p numDependentNodes entries will be filled in. If \p numDependentNodes is + * higher than the actual number of dependent nodes, the remaining entries in + * \p dependentNodes will be set to NULL, and the number of nodes actually obtained will + * be returned in \p numDependentNodes. * - * \param hfunc - Kernel to set parameter size for - * \param numbytes - Size of parameter list in bytes + * \param hNode - Node to query + * \param dependentNodes - Pointer to return the dependent nodes + * \param numDependentNodes - See description * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetf, - * ::cuParamSeti, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphNodeGetDependencies, + * ::cuGraphGetNodes, + * ::cuGraphGetRootNodes, + * ::cuGraphGetEdges, + * ::cuGraphAddDependencies, + * ::cuGraphRemoveDependencies */ -CUresult CUDAAPI cuParamSetSize(CUfunction hfunc, unsigned int numbytes); +CUresult CUDAAPI cuGraphNodeGetDependentNodes(CUgraphNode hNode, CUgraphNode *dependentNodes, size_t *numDependentNodes); /** - * \brief Adds an integer parameter to the function's argument list + * \brief Adds dependency edges to a graph * - * \deprecated + * The number of dependencies to be added is defined by \p numDependencies + * Elements in \p from and \p to at corresponding indices define a dependency. + * Each node in \p from and \p to must belong to \p hGraph. * - * Sets an integer parameter that will be specified the next time the - * kernel corresponding to \p hfunc will be invoked. \p offset is a byte offset. + * If \p numDependencies is 0, elements in \p from and \p to will be ignored. + * Specifying an existing dependency will return an error. * - * \param hfunc - Kernel to add parameter to - * \param offset - Offset to add parameter to argument list - * \param value - Value of parameter + * \param hGraph - Graph to which dependencies are added + * \param from - Array of nodes that provide the dependencies + * \param to - Array of dependent nodes + * \param numDependencies - Number of dependencies to be added * * \return * ::CUDA_SUCCESS, - * ::CUDA_ERROR_DEINITIALIZED, - * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSetf, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphRemoveDependencies, + * ::cuGraphGetEdges, + * ::cuGraphNodeGetDependencies, + * ::cuGraphNodeGetDependentNodes */ -CUresult CUDAAPI cuParamSeti(CUfunction hfunc, int offset, unsigned int value); +CUresult CUDAAPI cuGraphAddDependencies(CUgraph hGraph, const CUgraphNode *from, const CUgraphNode *to, size_t numDependencies); /** - * \brief Adds a floating-point parameter to the function's argument list + * \brief Removes dependency edges from a graph * - * \deprecated + * The number of \p dependencies to be removed is defined by \p numDependencies. + * Elements in \p from and \p to at corresponding indices define a dependency. + * Each node in \p from and \p to must belong to \p hGraph. * - * Sets a floating-point parameter that will be specified the next time the - * kernel corresponding to \p hfunc will be invoked. \p offset is a byte offset. + * If \p numDependencies is 0, elements in \p from and \p to will be ignored. + * Specifying a non-existing dependency will return an error. * - * \param hfunc - Kernel to add parameter to - * \param offset - Offset to add parameter to argument list - * \param value - Value of parameter + * \param hGraph - Graph from which to remove dependencies + * \param from - Array of nodes that provide the dependencies + * \param to - Array of dependent nodes + * \param numDependencies - Number of dependencies to be removed * * \return * ::CUDA_SUCCESS, - * ::CUDA_ERROR_DEINITIALIZED, - * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSeti, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphAddDependencies, + * ::cuGraphGetEdges, + * ::cuGraphNodeGetDependencies, + * ::cuGraphNodeGetDependentNodes */ -CUresult CUDAAPI cuParamSetf(CUfunction hfunc, int offset, float value); +CUresult CUDAAPI cuGraphRemoveDependencies(CUgraph hGraph, const CUgraphNode *from, const CUgraphNode *to, size_t numDependencies); /** - * \brief Adds arbitrary data to the function's argument list + * \brief Remove a node from the graph * - * \deprecated + * Removes \p hNode from its graph. This operation also severs any dependencies of other nodes + * on \p hNode and vice versa. * - * Copies an arbitrary amount of data (specified in \p numbytes) from \p ptr - * into the parameter space of the kernel corresponding to \p hfunc. \p offset - * is a byte offset. + * \param hNode - Node to remove * - * \param hfunc - Kernel to add data to - * \param offset - Offset to add data to argument list - * \param ptr - Pointer to arbitrary data - * \param numbytes - Size of data to copy in bytes + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety + * \notefnerr + * + * \sa + * ::cuGraphAddChildGraphNode, + * ::cuGraphAddEmptyNode, + * ::cuGraphAddKernelNode, + * ::cuGraphAddHostNode, + * ::cuGraphAddMemcpyNode, + * ::cuGraphAddMemsetNode + */ +CUresult CUDAAPI cuGraphDestroyNode(CUgraphNode hNode); + +/** + * \brief Creates an executable graph from a graph + * + * Instantiates \p hGraph as an executable graph. The graph is validated for any + * structural constraints or intra-node constraints which were not previously + * validated. If instantiation is successful, a handle to the instantiated graph + * is returned in \p graphExec. + * + * If there are any errors, diagnostic information may be returned in \p errorNode and + * \p logBuffer. This is the primary way to inspect instantiation errors. The output + * will be null terminated unless the diagnostics overflow + * the buffer. In this case, they will be truncated, and the last byte can be + * inspected to determine if truncation occurred. + * + * \param phGraphExec - Returns instantiated graph + * \param hGraph - Graph to instantiate + * \param phErrorNode - In case of an instantiation error, this may be modified to + * indicate a node contributing to the error + * \param logBuffer - A character buffer to store diagnostic messages + * \param bufferSize - Size of the log buffer in bytes * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSetf, - * ::cuParamSeti, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphCreate, + * ::cuGraphLaunch, + * ::cuGraphExecDestroy */ -CUresult CUDAAPI cuParamSetv(CUfunction hfunc, int offset, void *ptr, unsigned int numbytes); +CUresult CUDAAPI cuGraphInstantiate(CUgraphExec *phGraphExec, CUgraph hGraph, CUgraphNode *phErrorNode, char *logBuffer, size_t bufferSize); + +#if __CUDA_API_VERSION >= 10010 /** - * \brief Launches a CUDA function + * \brief Sets the parameters for a kernel node in the given graphExec * - * \deprecated + * Sets the parameters of a kernel node in an executable graph \p hGraphExec. + * The node is identified by the corresponding node \p hNode in the + * non-executable graph, from which the executable graph was instantiated. * - * Invokes the kernel \p f on a 1 x 1 x 1 grid of blocks. The block - * contains the number of threads specified by a previous call to - * ::cuFuncSetBlockShape(). + * \p hNode must not have been removed from the original graph. The \p func field + * of \p nodeParams cannot be modified and must match the original value. + * All other values can be modified. * - * \param f - Kernel to launch + * The modifications take effect at the next launch of \p hGraphExec. Already + * enqueued or running launches of \p hGraphExec are not affected by this call. + * \p hNode is also not modified by this call. + * + * \param hGraphExec - The executable graph in which to set the specified node + * \param hNode - kernel node from the graph from which graphExec was instantiated + * \param nodeParams - Updated Parameters to set * * \return * ::CUDA_SUCCESS, - * ::CUDA_ERROR_DEINITIALIZED, - * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_LAUNCH_FAILED, - * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, - * ::CUDA_ERROR_LAUNCH_TIMEOUT, - * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSetf, - * ::cuParamSeti, - * ::cuParamSetv, - * ::cuLaunchGrid, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphAddKernelNode, + * ::cuGraphKernelNodeSetParams, + * ::cuGraphInstantiate */ -CUresult CUDAAPI cuLaunch(CUfunction f); + CUresult CUDAAPI cuGraphExecKernelNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_KERNEL_NODE_PARAMS *nodeParams); + +#endif /* __CUDA_API_VERSION >= 10010 */ /** - * \brief Launches a CUDA function - * - * \deprecated + * \brief Launches an executable graph in a stream * - * Invokes the kernel \p f on a \p grid_width x \p grid_height grid of - * blocks. Each block contains the number of threads specified by a previous - * call to ::cuFuncSetBlockShape(). + * Executes \p hGraphExec in \p hStream. Only one instance of \p hGraphExec may be executing + * at a time. Each launch is ordered behind both any previous work in \p hStream + * and any previous launches of \p hGraphExec. To execute a graph concurrently, it must be + * instantiated multiple times into multiple executable graphs. * - * \param f - Kernel to launch - * \param grid_width - Width of grid in blocks - * \param grid_height - Height of grid in blocks + * \param hGraphExec - Executable graph to launch + * \param hStream - Stream in which to launch the graph * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_LAUNCH_FAILED, - * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, - * ::CUDA_ERROR_LAUNCH_TIMEOUT, - * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSetf, - * ::cuParamSeti, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGridAsync, - * ::cuLaunchKernel + * \sa + * ::cuGraphInstantiate, + * ::cuGraphExecDestroy */ -CUresult CUDAAPI cuLaunchGrid(CUfunction f, int grid_width, int grid_height); +CUresult CUDAAPI cuGraphLaunch(CUgraphExec hGraphExec, CUstream hStream); /** - * \brief Launches a CUDA function - * - * \deprecated + * \brief Destroys an executable graph * - * Invokes the kernel \p f on a \p grid_width x \p grid_height grid of - * blocks. Each block contains the number of threads specified by a previous - * call to ::cuFuncSetBlockShape(). + * Destroys the executable graph specified by \p hGraphExec, as well + * as all of its executable nodes. If the executable graph is + * in-flight, it will not be terminated, but rather freed + * asynchronously on completion. * - * \param f - Kernel to launch - * \param grid_width - Width of grid in blocks - * \param grid_height - Height of grid in blocks - * \param hStream - Stream identifier + * \param hGraphExec - Executable graph to destroy * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, - * ::CUDA_ERROR_INVALID_HANDLE, - * ::CUDA_ERROR_INVALID_VALUE, - * ::CUDA_ERROR_LAUNCH_FAILED, - * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES, - * ::CUDA_ERROR_LAUNCH_TIMEOUT, - * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING, - * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED - * - * \note In certain cases where cubins are created with no ABI (i.e., using \p ptxas \p --abi-compile \p no), - * this function may serialize kernel launches. In order to force the CUDA driver to retain - * asynchronous behavior, set the ::CU_CTX_LMEM_RESIZE_TO_MAX flag during context creation (see ::cuCtxCreate). - * - * \note_null_stream + * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr * - * \sa ::cuFuncSetBlockShape, - * ::cuFuncSetSharedSize, - * ::cuFuncGetAttribute, - * ::cuParamSetSize, - * ::cuParamSetf, - * ::cuParamSeti, - * ::cuParamSetv, - * ::cuLaunch, - * ::cuLaunchGrid, - * ::cuLaunchKernel + * \sa + * ::cuGraphInstantiate, + * ::cuGraphLaunch */ -CUresult CUDAAPI cuLaunchGridAsync(CUfunction f, int grid_width, int grid_height, CUstream hStream); - +CUresult CUDAAPI cuGraphExecDestroy(CUgraphExec hGraphExec); /** - * \brief Adds a texture-reference to the function's argument list - * - * \deprecated + * \brief Destroys a graph * - * Makes the CUDA array or linear memory bound to the texture reference - * \p hTexRef available to a device program as a texture. In this version of - * CUDA, the texture-reference must be obtained via ::cuModuleGetTexRef() and - * the \p texunit parameter must be set to ::CU_PARAM_TR_DEFAULT. + * Destroys the graph specified by \p hGraph, as well as all of its nodes. * - * \param hfunc - Kernel to add texture-reference to - * \param texunit - Texture unit (must be ::CU_PARAM_TR_DEFAULT) - * \param hTexRef - Texture-reference to add to argument list + * \param hGraph - Graph to destroy * * \return * ::CUDA_SUCCESS, * ::CUDA_ERROR_DEINITIALIZED, * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE + * \note_graph_thread_safety * \notefnerr + * + * \sa + * ::cuGraphCreate */ -CUresult CUDAAPI cuParamSetTexRef(CUfunction hfunc, int texunit, CUtexref hTexRef); -/** @} */ /* END CUDA_EXEC_DEPRECATED */ - +CUresult CUDAAPI cuGraphDestroy(CUgraph hGraph); +/** @} */ /* END CUDA_GRAPH */ +#endif /* __CUDA_API_VERSION >= 10000 */ #if __CUDA_API_VERSION >= 6050 /** @@ -9354,6 +12482,8 @@ CUresult CUDAAPI cuParamSetTexRef(CUfunction hfunc, int texunit, CUtexref hTexRe * ::CUDA_ERROR_UNKNOWN * \notefnerr * + * \sa + * ::cudaOccupancyMaxActiveBlocksPerMultiprocessor */ CUresult CUDAAPI cuOccupancyMaxActiveBlocksPerMultiprocessor(int *numBlocks, CUfunction func, int blockSize, size_t dynamicSMemSize); @@ -9394,9 +12524,11 @@ CUresult CUDAAPI cuOccupancyMaxActiveBlocksPerMultiprocessor(int *numBlocks, CUf * ::CUDA_ERROR_UNKNOWN * \notefnerr * + * \sa + * ::cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags */ CUresult CUDAAPI cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(int *numBlocks, CUfunction func, int blockSize, size_t dynamicSMemSize, unsigned int flags); - + /** * \brief Suggest a launch configuration with reasonable occupancy * @@ -9444,6 +12576,8 @@ CUresult CUDAAPI cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(int *numBl * ::CUDA_ERROR_UNKNOWN * \notefnerr * + * \sa + * ::cudaOccupancyMaxPotentialBlockSize */ CUresult CUDAAPI cuOccupancyMaxPotentialBlockSize(int *minGridSize, int *blockSize, CUfunction func, CUoccupancyB2DSize blockSizeToDynamicSMemSize, size_t dynamicSMemSize, int blockSizeLimit); @@ -9488,6 +12622,8 @@ CUresult CUDAAPI cuOccupancyMaxPotentialBlockSize(int *minGridSize, int *blockSi * ::CUDA_ERROR_UNKNOWN * \notefnerr * + * \sa + * ::cudaOccupancyMaxPotentialBlockSizeWithFlags */ CUresult CUDAAPI cuOccupancyMaxPotentialBlockSizeWithFlags(int *minGridSize, int *blockSize, CUfunction func, CUoccupancyB2DSize blockSizeToDynamicSMemSize, size_t dynamicSMemSize, int blockSizeLimit, unsigned int flags); @@ -9495,13 +12631,13 @@ CUresult CUDAAPI cuOccupancyMaxPotentialBlockSizeWithFlags(int *minGridSize, int #endif /* __CUDA_API_VERSION >= 6050 */ /** - * \defgroup CUDA_TEXREF Texture Reference Management + * \defgroup CUDA_TEXREF_DEPRECATED Texture Reference Management [DEPRECATED] * - * ___MANBRIEF___ texture reference management functions of the low-level CUDA - * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ + * ___MANBRIEF___ deprecated texture reference management functions of the + * low-level CUDA driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ * - * This section describes the texture reference management functions of the - * low-level CUDA driver application programming interface. + * This section describes the deprecated texture reference management + * functions of the low-level CUDA driver application programming interface. * * @{ */ @@ -9509,6 +12645,8 @@ CUresult CUDAAPI cuOccupancyMaxPotentialBlockSizeWithFlags(int *minGridSize, int /** * \brief Binds an array as a texture reference * + * \deprecated + * * Binds the CUDA array \p hArray to the texture reference \p hTexRef. Any * previous address or CUDA array state associated with the texture reference * is superseded by this function. \p Flags must be set to @@ -9530,16 +12668,19 @@ CUresult CUDAAPI cuOccupancyMaxPotentialBlockSizeWithFlags(int *minGridSize, int * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToArray */ CUresult CUDAAPI cuTexRefSetArray(CUtexref hTexRef, CUarray hArray, unsigned int Flags); /** * \brief Binds a mipmapped array to a texture reference * + * \deprecated + * * Binds the CUDA mipmapped array \p hMipmappedArray to the texture reference \p hTexRef. * Any previous address or CUDA array state associated with the texture reference - * is superseded by this function. \p Flags must be set to ::CU_TRSA_OVERRIDE_FORMAT. + * is superseded by this function. \p Flags must be set to ::CU_TRSA_OVERRIDE_FORMAT. * Any CUDA array previously bound to \p hTexRef is unbound. * * \param hTexRef - Texture reference to bind @@ -9557,7 +12698,8 @@ CUresult CUDAAPI cuTexRefSetArray(CUtexref hTexRef, CUarray hArray, unsigned int * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetMipmappedArray(CUtexref hTexRef, CUmipmappedArray hMipmappedArray, unsigned int Flags); @@ -9565,6 +12707,8 @@ CUresult CUDAAPI cuTexRefSetMipmappedArray(CUtexref hTexRef, CUmipmappedArray hM /** * \brief Binds an address as a texture reference * + * \deprecated + * * Binds a linear address range to the texture reference \p hTexRef. Any * previous address or CUDA array state associated with the texture reference * is superseded by this function. Any memory previously bound to \p hTexRef @@ -9583,7 +12727,7 @@ CUresult CUDAAPI cuTexRefSetMipmappedArray(CUtexref hTexRef, CUmipmappedArray hM * The total number of elements (or texels) in the linear address range * cannot exceed ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH. * The number of elements is computed as (\p bytes / bytesPerElement), - * where bytesPerElement is determined from the data format and number of + * where bytesPerElement is determined from the data format and number of * components set using ::cuTexRefSetFormat(). * * \param ByteOffset - Returned byte offset @@ -9601,13 +12745,16 @@ CUresult CUDAAPI cuTexRefSetMipmappedArray(CUtexref hTexRef, CUmipmappedArray hM * \sa ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTexture */ CUresult CUDAAPI cuTexRefSetAddress(size_t *ByteOffset, CUtexref hTexRef, CUdeviceptr dptr, size_t bytes); /** * \brief Binds an address as a 2D texture reference * + * \deprecated + * * Binds a linear address range to the texture reference \p hTexRef. Any * previous address or CUDA array state associated with the texture reference * is superseded by this function. Any memory previously bound to \p hTexRef @@ -9627,14 +12774,14 @@ CUresult CUDAAPI cuTexRefSetAddress(size_t *ByteOffset, CUtexref hTexRef, CUdevi * supplied, ::CUDA_ERROR_INVALID_VALUE is returned. * * \p Pitch has to be aligned to the hardware-specific texture pitch alignment. - * This value can be queried using the device attribute - * ::CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT. If an unaligned \p Pitch is + * This value can be queried using the device attribute + * ::CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT. If an unaligned \p Pitch is * supplied, ::CUDA_ERROR_INVALID_VALUE is returned. * * Width and Height, which are specified in elements (or texels), cannot exceed * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH and * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT respectively. - * \p Pitch, which is specified in bytes, cannot exceed + * \p Pitch, which is specified in bytes, cannot exceed * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH. * * \param hTexRef - Texture reference to bind @@ -9653,7 +12800,8 @@ CUresult CUDAAPI cuTexRefSetAddress(size_t *ByteOffset, CUtexref hTexRef, CUdevi * ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTexture2D */ CUresult CUDAAPI cuTexRefSetAddress2D(CUtexref hTexRef, const CUDA_ARRAY_DESCRIPTOR *desc, CUdeviceptr dptr, size_t Pitch); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -9661,6 +12809,8 @@ CUresult CUDAAPI cuTexRefSetAddress2D(CUtexref hTexRef, const CUDA_ARRAY_DESCRIP /** * \brief Sets the format for a texture reference * + * \deprecated + * * Specifies the format of the data to be read by the texture reference * \p hTexRef. \p fmt and \p NumPackedComponents are exactly analogous to the * ::Format and ::NumChannels members of the ::CUDA_ARRAY_DESCRIPTOR structure: @@ -9682,13 +12832,20 @@ CUresult CUDAAPI cuTexRefSetAddress2D(CUtexref hTexRef, const CUDA_ARRAY_DESCRIP * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaCreateChannelDesc, + * ::cudaBindTexture, + * ::cudaBindTexture2D, + * ::cudaBindTextureToArray, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetFormat(CUtexref hTexRef, CUarray_format fmt, int NumPackedComponents); /** * \brief Sets the addressing mode for a texture reference * + * \deprecated + * * Specifies the addressing mode \p am for the given dimension \p dim of the * texture reference \p hTexRef. If \p dim is zero, the addressing mode is * applied to the first parameter of the functions used to fetch from the @@ -9704,7 +12861,7 @@ CUresult CUDAAPI cuTexRefSetFormat(CUtexref hTexRef, CUarray_format fmt, int Num * \endcode * * Note that this call has no effect if \p hTexRef is bound to linear memory. - * Also, if the flag, ::CU_TRSF_NORMALIZED_COORDINATES, is not set, the only + * Also, if the flag, ::CU_TRSF_NORMALIZED_COORDINATES, is not set, the only * supported address mode is ::CU_TR_ADDRESS_MODE_CLAMP. * * \param hTexRef - Texture reference @@ -9722,13 +12879,19 @@ CUresult CUDAAPI cuTexRefSetFormat(CUtexref hTexRef, CUarray_format fmt, int Num * ::cuTexRefSetAddress2D, ::cuTexRefSetArray, * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTexture, + * ::cudaBindTexture2D, + * ::cudaBindTextureToArray, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetAddressMode(CUtexref hTexRef, int dim, CUaddress_mode am); /** * \brief Sets the filtering mode for a texture reference * + * \deprecated + * * Specifies the filtering mode \p fm to be used when reading memory through * the texture reference \p hTexRef. ::CUfilter_mode_enum is defined as: * @@ -9755,13 +12918,16 @@ CUresult CUDAAPI cuTexRefSetAddressMode(CUtexref hTexRef, int dim, CUaddress_mod * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToArray */ CUresult CUDAAPI cuTexRefSetFilterMode(CUtexref hTexRef, CUfilter_mode fm); /** * \brief Sets the mipmap filtering mode for a texture reference * + * \deprecated + * * Specifies the mipmap filtering mode \p fm to be used when reading memory through * the texture reference \p hTexRef. ::CUfilter_mode_enum is defined as: * @@ -9788,14 +12954,17 @@ CUresult CUDAAPI cuTexRefSetFilterMode(CUtexref hTexRef, CUfilter_mode fm); * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetMipmapFilterMode(CUtexref hTexRef, CUfilter_mode fm); /** * \brief Sets the mipmap level bias for a texture reference * - * Specifies the mipmap level bias \p bias to be added to the specified mipmap level when + * \deprecated + * + * Specifies the mipmap level bias \p bias to be added to the specified mipmap level when * reading memory through the texture reference \p hTexRef. * * Note that this call has no effect if \p hTexRef is not bound to a mipmapped array. @@ -9814,15 +12983,18 @@ CUresult CUDAAPI cuTexRefSetMipmapFilterMode(CUtexref hTexRef, CUfilter_mode fm) * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetMipmapLevelBias(CUtexref hTexRef, float bias); /** * \brief Sets the mipmap min/max mipmap level clamps for a texture reference * + * \deprecated + * * Specifies the min/max mipmap level clamps, \p minMipmapLevelClamp and \p maxMipmapLevelClamp - * respectively, to be used when reading memory through the texture reference + * respectively, to be used when reading memory through the texture reference * \p hTexRef. * * Note that this call has no effect if \p hTexRef is not bound to a mipmapped array. @@ -9842,15 +13014,18 @@ CUresult CUDAAPI cuTexRefSetMipmapLevelBias(CUtexref hTexRef, float bias); * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetMipmapLevelClamp(CUtexref hTexRef, float minMipmapLevelClamp, float maxMipmapLevelClamp); /** * \brief Sets the maximum anisotropy for a texture reference * + * \deprecated + * * Specifies the maximum anisotropy \p maxAniso to be used when reading memory through - * the texture reference \p hTexRef. + * the texture reference \p hTexRef. * * Note that this call has no effect if \p hTexRef is bound to linear memory. * @@ -9868,13 +13043,17 @@ CUresult CUDAAPI cuTexRefSetMipmapLevelClamp(CUtexref hTexRef, float minMipmapLe * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFlags, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTextureToArray, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetMaxAnisotropy(CUtexref hTexRef, unsigned int maxAniso); /** * \brief Sets the border color for a texture reference * + * \deprecated + * * Specifies the value of the RGBA color via the \p pBorderColor to the texture reference * \p hTexRef. The color value supports only float type and holds color components in * the following sequence: @@ -9898,13 +13077,19 @@ CUresult CUDAAPI cuTexRefSetMaxAnisotropy(CUtexref hTexRef, unsigned int maxAnis * ::CUDA_ERROR_INVALID_VALUE * * \sa ::cuTexRefSetAddressMode, - * ::cuTexRefGetAddressMode, ::cuTexRefGetBorderColor + * ::cuTexRefGetAddressMode, ::cuTexRefGetBorderColor, + * ::cudaBindTexture, + * ::cudaBindTexture2D, + * ::cudaBindTextureToArray, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetBorderColor(CUtexref hTexRef, float *pBorderColor); /** * \brief Sets the flags for a texture reference * + * \deprecated + * * Specifies optional flags via \p Flags to specify the behavior of data * returned through the texture reference \p hTexRef. The valid flags are: * @@ -9913,7 +13098,7 @@ CUresult CUDAAPI cuTexRefSetBorderColor(CUtexref hTexRef, float *pBorderColor); * range [0, 1]. Note that texture with 32-bit integer format * would not be promoted, regardless of whether or not this * flag is specified; - * - ::CU_TRSF_NORMALIZED_COORDINATES, which suppresses the + * - ::CU_TRSF_NORMALIZED_COORDINATES, which suppresses the * default behavior of having the texture coordinates range * from [0, Dim) where Dim is the width or height of the CUDA * array. Instead, the texture coordinates [0, 1.0) reference @@ -9933,7 +13118,11 @@ CUresult CUDAAPI cuTexRefSetBorderColor(CUtexref hTexRef, float *pBorderColor); * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray, * ::cuTexRefSetFilterMode, ::cuTexRefSetFormat, * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray, - * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat + * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat, + * ::cudaBindTexture, + * ::cudaBindTexture2D, + * ::cudaBindTextureToArray, + * ::cudaBindTextureToMipmappedArray */ CUresult CUDAAPI cuTexRefSetFlags(CUtexref hTexRef, unsigned int Flags); @@ -9941,6 +13130,8 @@ CUresult CUDAAPI cuTexRefSetFlags(CUtexref hTexRef, unsigned int Flags); /** * \brief Gets the address associated with a texture reference * + * \deprecated + * * Returns in \p *pdptr the base address bound to the texture reference * \p hTexRef, or returns ::CUDA_ERROR_INVALID_VALUE if the texture reference * is not bound to any device memory range. @@ -9967,6 +13158,8 @@ CUresult CUDAAPI cuTexRefGetAddress(CUdeviceptr *pdptr, CUtexref hTexRef); /** * \brief Gets the array bound to a texture reference * + * \deprecated + * * Returns in \p *phArray the CUDA array bound to the texture reference * \p hTexRef, or returns ::CUDA_ERROR_INVALID_VALUE if the texture reference * is not bound to any CUDA array. @@ -9992,7 +13185,9 @@ CUresult CUDAAPI cuTexRefGetArray(CUarray *phArray, CUtexref hTexRef); /** * \brief Gets the mipmapped array bound to a texture reference * - * Returns in \p *phMipmappedArray the CUDA mipmapped array bound to the texture + * \deprecated + * + * Returns in \p *phMipmappedArray the CUDA mipmapped array bound to the texture * reference \p hTexRef, or returns ::CUDA_ERROR_INVALID_VALUE if the texture reference * is not bound to any CUDA mipmapped array. * @@ -10017,6 +13212,8 @@ CUresult CUDAAPI cuTexRefGetMipmappedArray(CUmipmappedArray *phMipmappedArray, C /** * \brief Gets the addressing mode used by a texture reference * + * \deprecated + * * Returns in \p *pam the addressing mode corresponding to the * dimension \p dim of the texture reference \p hTexRef. Currently, the only * valid value for \p dim are 0 and 1. @@ -10043,6 +13240,8 @@ CUresult CUDAAPI cuTexRefGetAddressMode(CUaddress_mode *pam, CUtexref hTexRef, i /** * \brief Gets the filter-mode used by a texture reference * + * \deprecated + * * Returns in \p *pfm the filtering mode of the texture reference * \p hTexRef. * @@ -10067,6 +13266,8 @@ CUresult CUDAAPI cuTexRefGetFilterMode(CUfilter_mode *pfm, CUtexref hTexRef); /** * \brief Gets the format used by a texture reference * + * \deprecated + * * Returns in \p *pFormat and \p *pNumChannels the format and number * of components of the CUDA array bound to the texture reference \p hTexRef. * If \p pFormat or \p pNumChannels is NULL, it will be ignored. @@ -10093,6 +13294,8 @@ CUresult CUDAAPI cuTexRefGetFormat(CUarray_format *pFormat, int *pNumChannels, C /** * \brief Gets the mipmap filtering mode for a texture reference * + * \deprecated + * * Returns the mipmap filtering mode in \p pfm that's used when reading memory through * the texture reference \p hTexRef. * @@ -10117,6 +13320,8 @@ CUresult CUDAAPI cuTexRefGetMipmapFilterMode(CUfilter_mode *pfm, CUtexref hTexRe /** * \brief Gets the mipmap level bias for a texture reference * + * \deprecated + * * Returns the mipmap level bias in \p pBias that's added to the specified mipmap * level when reading memory through the texture reference \p hTexRef. * @@ -10141,8 +13346,10 @@ CUresult CUDAAPI cuTexRefGetMipmapLevelBias(float *pbias, CUtexref hTexRef); /** * \brief Gets the min/max mipmap level clamps for a texture reference * + * \deprecated + * * Returns the min/max mipmap level clamps in \p pminMipmapLevelClamp and \p pmaxMipmapLevelClamp - * that's used when reading memory through the texture reference \p hTexRef. + * that's used when reading memory through the texture reference \p hTexRef. * * \param pminMipmapLevelClamp - Returned mipmap min level clamp * \param pmaxMipmapLevelClamp - Returned mipmap max level clamp @@ -10166,8 +13373,10 @@ CUresult CUDAAPI cuTexRefGetMipmapLevelClamp(float *pminMipmapLevelClamp, float /** * \brief Gets the maximum anisotropy for a texture reference * + * \deprecated + * * Returns the maximum anisotropy in \p pmaxAniso that's used when reading memory through - * the texture reference \p hTexRef. + * the texture reference \p hTexRef. * * \param pmaxAniso - Returned maximum anisotropy * \param hTexRef - Texture reference @@ -10190,6 +13399,8 @@ CUresult CUDAAPI cuTexRefGetMaxAnisotropy(int *pmaxAniso, CUtexref hTexRef); /** * \brief Gets the border color used by a texture reference * + * \deprecated + * * Returns in \p pBorderColor, values of the RGBA color used by * the texture reference \p hTexRef. * The color value is of type float and holds color components in @@ -10212,11 +13423,13 @@ CUresult CUDAAPI cuTexRefGetMaxAnisotropy(int *pmaxAniso, CUtexref hTexRef); * \sa ::cuTexRefSetAddressMode, * ::cuTexRefSetAddressMode, ::cuTexRefSetBorderColor */ -CUresult CUDAAPI cuTexRefGetBorderColor(float *pBorderColor, CUtexref hTexRef); +CUresult CUDAAPI cuTexRefGetBorderColor(float *pBorderColor, CUtexref hTexRef); /** * \brief Gets the flags used by a texture reference * + * \deprecated + * * Returns in \p *pFlags the flags of the texture reference \p hTexRef. * * \param pFlags - Returned flags @@ -10237,20 +13450,6 @@ CUresult CUDAAPI cuTexRefGetBorderColor(float *pBorderColor, CUtexref hTexRef); */ CUresult CUDAAPI cuTexRefGetFlags(unsigned int *pFlags, CUtexref hTexRef); -/** @} */ /* END CUDA_TEXREF */ - -/** - * \defgroup CUDA_TEXREF_DEPRECATED Texture Reference Management [DEPRECATED] - * - * ___MANBRIEF___ deprecated texture reference management functions of the - * low-level CUDA driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ - * - * This section describes the deprecated texture reference management - * functions of the low-level CUDA driver application programming interface. - * - * @{ - */ - /** * \brief Creates a texture reference * @@ -10300,7 +13499,7 @@ CUresult CUDAAPI cuTexRefDestroy(CUtexref hTexRef); /** - * \defgroup CUDA_SURFREF Surface Reference Management + * \defgroup CUDA_SURFREF_DEPRECATED Surface Reference Management [DEPRECATED] * * ___MANBRIEF___ surface reference management functions of the low-level CUDA * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ @@ -10314,6 +13513,8 @@ CUresult CUDAAPI cuTexRefDestroy(CUtexref hTexRef); /** * \brief Sets the CUDA array for a surface reference. * + * \deprecated + * * Sets the CUDA array \p hArray to be read and written by the surface reference * \p hSurfRef. Any previous CUDA array state associated with the surface * reference is superseded by this function. \p Flags must be set to 0. @@ -10331,13 +13532,18 @@ CUresult CUDAAPI cuTexRefDestroy(CUtexref hTexRef); * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuModuleGetSurfRef, ::cuSurfRefGetArray + * \sa + * ::cuModuleGetSurfRef, + * ::cuSurfRefGetArray, + * ::cudaBindSurfaceToArray */ CUresult CUDAAPI cuSurfRefSetArray(CUsurfref hSurfRef, CUarray hArray, unsigned int Flags); /** * \brief Passes back the CUDA array bound to a surface reference. * + * \deprecated + * * Returns in \p *phArray the CUDA array bound to the surface reference * \p hSurfRef, or returns ::CUDA_ERROR_INVALID_VALUE if the surface reference * is not bound to any CUDA array. @@ -10356,7 +13562,7 @@ CUresult CUDAAPI cuSurfRefSetArray(CUsurfref hSurfRef, CUarray hArray, unsigned */ CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef); -/** @} */ /* END CUDA_SURFREF */ +/** @} */ /* END CUDA_SURFREF_DEPRECATED */ #if __CUDA_API_VERSION >= 5000 /** @@ -10444,7 +13650,7 @@ CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef); * must be set to a valid device pointer, that is aligned to ::CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT. * ::CUDA_RESOURCE_DESC::res::linear::format and ::CUDA_RESOURCE_DESC::res::linear::numChannels * describe the format of each component and the number of components per array element. ::CUDA_RESOURCE_DESC::res::linear::sizeInBytes - * specifies the size of the array in bytes. The total number of elements in the linear address range cannot exceed + * specifies the size of the array in bytes. The total number of elements in the linear address range cannot exceed * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH. The number of elements is computed as (sizeInBytes / (sizeof(format) * numChannels)). * * \par @@ -10454,7 +13660,7 @@ CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef); * describe the format of each component and the number of components per array element. ::CUDA_RESOURCE_DESC::res::pitch2D::width * and ::CUDA_RESOURCE_DESC::res::pitch2D::height specify the width and height of the array in elements, and cannot exceed * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH and ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT respectively. - * ::CUDA_RESOURCE_DESC::res::pitch2D::pitchInBytes specifies the pitch between two rows in bytes and has to be aligned to + * ::CUDA_RESOURCE_DESC::res::pitch2D::pitchInBytes specifies the pitch between two rows in bytes and has to be aligned to * ::CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT. Pitch cannot exceed ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH. * * - ::flags must be set to zero. @@ -10483,7 +13689,7 @@ CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef); CU_TR_ADDRESS_MODE_BORDER = 3 } CUaddress_mode; * \endcode - * This is ignored if ::CUDA_RESOURCE_DESC::resType is ::CU_RESOURCE_TYPE_LINEAR. Also, if the flag, ::CU_TRSF_NORMALIZED_COORDINATES + * This is ignored if ::CUDA_RESOURCE_DESC::resType is ::CU_RESOURCE_TYPE_LINEAR. Also, if the flag, ::CU_TRSF_NORMALIZED_COORDINATES * is not set, the only supported address mode is ::CU_TR_ADDRESS_MODE_CLAMP. * * - ::CUDA_TEXTURE_DESC::filterMode specifies the filtering mode to be used when fetching from the texture. CUfilter_mode is defined as: @@ -10558,14 +13764,14 @@ CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef); * - ::CUDA_RESOURCE_VIEW_DESC::firstLayer specifies the first layer index for layered textures. This will be the new layer zero. * For non-layered resources, this value has to be zero. * - * - ::CUDA_RESOURCE_VIEW_DESC::lastLayer specifies the last layer index for layered textures. For non-layered resources, + * - ::CUDA_RESOURCE_VIEW_DESC::lastLayer specifies the last layer index for layered textures. For non-layered resources, * this value has to be zero. * * * \param pTexObject - Texture object to create * \param pResDesc - Resource descriptor * \param pTexDesc - Texture descriptor - * \param pResViewDesc - Resource view descriptor + * \param pResViewDesc - Resource view descriptor * * \return * ::CUDA_SUCCESS, @@ -10574,7 +13780,9 @@ CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef); * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuTexObjectDestroy + * \sa + * ::cuTexObjectDestroy, + * ::cudaCreateTextureObject */ CUresult CUDAAPI cuTexObjectCreate(CUtexObject *pTexObject, const CUDA_RESOURCE_DESC *pResDesc, const CUDA_TEXTURE_DESC *pTexDesc, const CUDA_RESOURCE_VIEW_DESC *pResViewDesc); @@ -10592,7 +13800,9 @@ CUresult CUDAAPI cuTexObjectCreate(CUtexObject *pTexObject, const CUDA_RESOURCE_ * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuTexObjectCreate + * \sa + * ::cuTexObjectCreate, + * ::cudaDestroyTextureObject */ CUresult CUDAAPI cuTexObjectDestroy(CUtexObject texObject); @@ -10611,7 +13821,9 @@ CUresult CUDAAPI cuTexObjectDestroy(CUtexObject texObject); * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuTexObjectCreate + * \sa + * ::cuTexObjectCreate, + * ::cudaGetTextureObjectResourceDesc, */ CUresult CUDAAPI cuTexObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUtexObject texObject); @@ -10630,7 +13842,9 @@ CUresult CUDAAPI cuTexObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUtexO * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuTexObjectCreate + * \sa + * ::cuTexObjectCreate, + * ::cudaGetTextureObjectTextureDesc */ CUresult CUDAAPI cuTexObjectGetTextureDesc(CUDA_TEXTURE_DESC *pTexDesc, CUtexObject texObject); @@ -10650,7 +13864,9 @@ CUresult CUDAAPI cuTexObjectGetTextureDesc(CUDA_TEXTURE_DESC *pTexDesc, CUtexObj * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuTexObjectCreate + * \sa + * ::cuTexObjectCreate, + * ::cudaGetTextureObjectResourceViewDesc */ CUresult CUDAAPI cuTexObjectGetResourceViewDesc(CUDA_RESOURCE_VIEW_DESC *pResViewDesc, CUtexObject texObject); @@ -10673,7 +13889,7 @@ CUresult CUDAAPI cuTexObjectGetResourceViewDesc(CUDA_RESOURCE_VIEW_DESC *pResVie * \brief Creates a surface object * * Creates a surface object and returns it in \p pSurfObject. \p pResDesc describes - * the data to perform surface load/stores on. ::CUDA_RESOURCE_DESC::resType must be + * the data to perform surface load/stores on. ::CUDA_RESOURCE_DESC::resType must be * ::CU_RESOURCE_TYPE_ARRAY and ::CUDA_RESOURCE_DESC::res::array::hArray * must be set to a valid CUDA array handle. ::CUDA_RESOURCE_DESC::flags must be set to zero. * @@ -10691,7 +13907,9 @@ CUresult CUDAAPI cuTexObjectGetResourceViewDesc(CUDA_RESOURCE_VIEW_DESC *pResVie * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuSurfObjectDestroy + * \sa + * ::cuSurfObjectDestroy, + * ::cudaCreateSurfaceObject */ CUresult CUDAAPI cuSurfObjectCreate(CUsurfObject *pSurfObject, const CUDA_RESOURCE_DESC *pResDesc); @@ -10709,7 +13927,9 @@ CUresult CUDAAPI cuSurfObjectCreate(CUsurfObject *pSurfObject, const CUDA_RESOUR * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuSurfObjectCreate + * \sa + * ::cuSurfObjectCreate, + * ::cudaDestroySurfaceObject */ CUresult CUDAAPI cuSurfObjectDestroy(CUsurfObject surfObject); @@ -10728,26 +13948,29 @@ CUresult CUDAAPI cuSurfObjectDestroy(CUsurfObject surfObject); * ::CUDA_ERROR_INVALID_CONTEXT, * ::CUDA_ERROR_INVALID_VALUE * - * \sa ::cuSurfObjectCreate + * \sa + * ::cuSurfObjectCreate, + * ::cudaGetSurfaceObjectResourceDesc */ CUresult CUDAAPI cuSurfObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUsurfObject surfObject); /** @} */ /* END CUDA_SURFOBJECT */ #endif /* __CUDA_API_VERSION >= 5000 */ -#if __CUDA_API_VERSION >= 4000 /** * \defgroup CUDA_PEER_ACCESS Peer Context Memory Access * * ___MANBRIEF___ direct peer context memory access functions of the low-level * CUDA driver API (___CURRENT_FILE___) ___ENDMANBRIEF___ * - * This section describes the direct peer context memory access functions + * This section describes the direct peer context memory access functions * of the low-level CUDA driver application programming interface. * * @{ */ - + +#if __CUDA_API_VERSION >= 4000 + /** * \brief Queries if a device may directly access a peer device's memory. * @@ -10759,7 +13982,7 @@ CUresult CUDAAPI cuSurfObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUsur * \param canAccessPeer - Returned access capability * \param dev - Device from which allocations on \p peerDev are to * be directly accessed. - * \param peerDev - Device on which the allocations to be directly accessed + * \param peerDev - Device on which the allocations to be directly accessed * by \p dev reside. * * \return @@ -10769,47 +13992,12 @@ CUresult CUDAAPI cuSurfObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUsur * ::CUDA_ERROR_INVALID_DEVICE * \notefnerr * - * \sa ::cuCtxEnablePeerAccess, - * ::cuCtxDisablePeerAccess - */ -CUresult CUDAAPI cuDeviceCanAccessPeer(int *canAccessPeer, CUdevice dev, CUdevice peerDev); - - -/** - * \brief Queries attributes of the link between two devices. - * - * Returns in \p *value the value of the requested attribute \p attrib of the - * link between \p srcDevice and \p dstDevice. The supported attributes are: - * - ::CU_DEVICE_P2P_ATTRIBUTE_PERFORMANCE_RANK: A relative value indicating the - * performance of the link between two devices. - * - ::CU_DEVICE_P2P_ATTRIBUTE_ACCESS_SUPPORTED P2P: 1 if P2P Access is enable. - * - ::CU_DEVICE_P2P_ATTRIBUTE_NATIVE_ATOMIC_SUPPORTED: 1 if Atomic operations over - * the link are supported. - * - * Returns ::CUDA_ERROR_INVALID_DEVICE if \p srcDevice or \p dstDevice are not valid - * or if they represent the same device. - * - * Returns ::CUDA_ERROR_INVALID_VALUE if \p attrib is not valid or if \p value is - * a null pointer. - * - * \param value - Returned value of the requested attribute - * \param attrib - The requested attribute of the link between \p srcDevice and \p dstDevice. - * \param srcDevice - The source device of the target link. - * \param dstDevice - The destination device of the target link. - * - * \return - * ::CUDA_SUCCESS, - * ::CUDA_ERROR_DEINITIALIZED, - * ::CUDA_ERROR_NOT_INITIALIZED, - * ::CUDA_ERROR_INVALID_DEVICE, - * ::CUDA_ERROR_INVALID_VALUE - * \notefnerr - * - * \sa ::cuCtxEnablePeerAccess, + * \sa + * ::cuCtxEnablePeerAccess, * ::cuCtxDisablePeerAccess, - * ::cuCtxCanAccessPeer + * ::cudaDeviceCanAccessPeer */ -CUresult CUDAAPI cuDeviceGetP2PAttribute(int* value, CUdevice_P2PAttribute attrib, CUdevice srcDevice, CUdevice dstDevice); +CUresult CUDAAPI cuDeviceCanAccessPeer(int *canAccessPeer, CUdevice dev, CUdevice peerDev); /** * \brief Enables direct access to memory allocations in a peer context. @@ -10821,7 +14009,7 @@ CUresult CUDAAPI cuDeviceGetP2PAttribute(int* value, CUdevice_P2PAttribute attri * details. * * Note that access granted by this call is unidirectional and that in order to access - * memory from the current context in \p peerContext, a separate symmetric call + * memory from the current context in \p peerContext, a separate symmetric call * to ::cuCtxEnablePeerAccess() is required. * * There is a system-wide maximum of eight peer connections per device. @@ -10833,7 +14021,7 @@ CUresult CUDAAPI cuDeviceGetP2PAttribute(int* value, CUdevice_P2PAttribute attri * Returns ::CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED if direct access of * \p peerContext from the current context has already been enabled. * - * Returns ::CUDA_ERROR_TOO_MANY_PEERS if direct peer access is not possible + * Returns ::CUDA_ERROR_TOO_MANY_PEERS if direct peer access is not possible * because hardware resources required for peer access have been exhausted. * * Returns ::CUDA_ERROR_INVALID_CONTEXT if there is no current context, \p peerContext @@ -10855,16 +14043,18 @@ CUresult CUDAAPI cuDeviceGetP2PAttribute(int* value, CUdevice_P2PAttribute attri * ::CUDA_ERROR_INVALID_VALUE * \notefnerr * - * \sa ::cuDeviceCanAccessPeer, - * ::cuCtxDisablePeerAccess + * \sa + * ::cuDeviceCanAccessPeer, + * ::cuCtxDisablePeerAccess, + * ::cudaDeviceEnablePeerAccess */ CUresult CUDAAPI cuCtxEnablePeerAccess(CUcontext peerContext, unsigned int Flags); /** - * \brief Disables direct access to memory allocations in a peer context and + * \brief Disables direct access to memory allocations in a peer context and * unregisters any registered allocations. * - Returns ::CUDA_ERROR_PEER_ACCESS_NOT_ENABLED if direct peer access has + Returns ::CUDA_ERROR_PEER_ACCESS_NOT_ENABLED if direct peer access has * not yet been enabled from \p peerContext to the current context. * * Returns ::CUDA_ERROR_INVALID_CONTEXT if there is no current context, or if @@ -10880,14 +14070,61 @@ CUresult CUDAAPI cuCtxEnablePeerAccess(CUcontext peerContext, unsigned int Flags * ::CUDA_ERROR_INVALID_CONTEXT, * \notefnerr * - * \sa ::cuDeviceCanAccessPeer, - * ::cuCtxEnablePeerAccess + * \sa + * ::cuDeviceCanAccessPeer, + * ::cuCtxEnablePeerAccess, + * ::cudaDeviceDisablePeerAccess */ CUresult CUDAAPI cuCtxDisablePeerAccess(CUcontext peerContext); -/** @} */ /* END CUDA_PEER_ACCESS */ #endif /* __CUDA_API_VERSION >= 4000 */ +#if __CUDA_API_VERSION >= 8000 + +/** + * \brief Queries attributes of the link between two devices. + * + * Returns in \p *value the value of the requested attribute \p attrib of the + * link between \p srcDevice and \p dstDevice. The supported attributes are: + * - ::CU_DEVICE_P2P_ATTRIBUTE_PERFORMANCE_RANK: A relative value indicating the + * performance of the link between two devices. + * - ::CU_DEVICE_P2P_ATTRIBUTE_ACCESS_SUPPORTED P2P: 1 if P2P Access is enable. + * - ::CU_DEVICE_P2P_ATTRIBUTE_NATIVE_ATOMIC_SUPPORTED: 1 if Atomic operations over + * the link are supported. + * - ::CU_DEVICE_P2P_ATTRIBUTE_CUDA_ARRAY_ACCESS_SUPPORTED: 1 if cudaArray can + * be accessed over the link. + * + * Returns ::CUDA_ERROR_INVALID_DEVICE if \p srcDevice or \p dstDevice are not valid + * or if they represent the same device. + * + * Returns ::CUDA_ERROR_INVALID_VALUE if \p attrib is not valid or if \p value is + * a null pointer. + * + * \param value - Returned value of the requested attribute + * \param attrib - The requested attribute of the link between \p srcDevice and \p dstDevice. + * \param srcDevice - The source device of the target link. + * \param dstDevice - The destination device of the target link. + * + * \return + * ::CUDA_SUCCESS, + * ::CUDA_ERROR_DEINITIALIZED, + * ::CUDA_ERROR_NOT_INITIALIZED, + * ::CUDA_ERROR_INVALID_DEVICE, + * ::CUDA_ERROR_INVALID_VALUE + * \notefnerr + * + * \sa + * ::cuCtxEnablePeerAccess, + * ::cuCtxDisablePeerAccess, + * ::cuDeviceCanAccessPeer, + * ::cudaDeviceGetP2PAttribute + */ +CUresult CUDAAPI cuDeviceGetP2PAttribute(int* value, CUdevice_P2PAttribute attrib, CUdevice srcDevice, CUdevice dstDevice); + +#endif /* __CUDA_API_VERSION >= 8000 */ + +/** @} */ /* END CUDA_PEER_ACCESS */ + /** * \defgroup CUDA_GRAPHICS Graphics Interoperability * @@ -10925,7 +14162,8 @@ CUresult CUDAAPI cuCtxDisablePeerAccess(CUcontext peerContext); * ::cuGraphicsD3D10RegisterResource, * ::cuGraphicsD3D11RegisterResource, * ::cuGraphicsGLRegisterBuffer, - * ::cuGraphicsGLRegisterImage + * ::cuGraphicsGLRegisterImage, + * ::cudaGraphicsUnregisterResource */ CUresult CUDAAPI cuGraphicsUnregisterResource(CUgraphicsResource resource); @@ -10963,7 +14201,9 @@ CUresult CUDAAPI cuGraphicsUnregisterResource(CUgraphicsResource resource); * ::CUDA_ERROR_NOT_MAPPED_AS_ARRAY * \notefnerr * - * \sa ::cuGraphicsResourceGetMappedPointer + * \sa + * ::cuGraphicsResourceGetMappedPointer, + * ::cudaGraphicsSubResourceGetMappedArray */ CUresult CUDAAPI cuGraphicsSubResourceGetMappedArray(CUarray *pArray, CUgraphicsResource resource, unsigned int arrayIndex, unsigned int mipLevel); @@ -10972,8 +14212,8 @@ CUresult CUDAAPI cuGraphicsSubResourceGetMappedArray(CUarray *pArray, CUgraphics /** * \brief Get a mipmapped array through which to access a mapped graphics resource. * - * Returns in \p *pMipmappedArray a mipmapped array through which the mapped graphics - * resource \p resource. The value set in \p *pMipmappedArray may change every time + * Returns in \p *pMipmappedArray a mipmapped array through which the mapped graphics + * resource \p resource. The value set in \p *pMipmappedArray may change every time * that \p resource is mapped. * * If \p resource is not a texture then it cannot be accessed via a mipmapped array and @@ -10994,7 +14234,9 @@ CUresult CUDAAPI cuGraphicsSubResourceGetMappedArray(CUarray *pArray, CUgraphics * ::CUDA_ERROR_NOT_MAPPED_AS_ARRAY * \notefnerr * - * \sa ::cuGraphicsResourceGetMappedPointer + * \sa + * ::cuGraphicsResourceGetMappedPointer, + * ::cudaGraphicsResourceGetMappedMipmappedArray */ CUresult CUDAAPI cuGraphicsResourceGetMappedMipmappedArray(CUmipmappedArray *pMipmappedArray, CUgraphicsResource resource); @@ -11030,7 +14272,8 @@ CUresult CUDAAPI cuGraphicsResourceGetMappedMipmappedArray(CUmipmappedArray *pMi * * \sa * ::cuGraphicsMapResources, - * ::cuGraphicsSubResourceGetMappedArray + * ::cuGraphicsSubResourceGetMappedArray, + * ::cudaGraphicsResourceGetMappedPointer */ CUresult CUDAAPI cuGraphicsResourceGetMappedPointer(CUdeviceptr *pDevPtr, size_t *pSize, CUgraphicsResource resource); #endif /* __CUDA_API_VERSION >= 3020 */ @@ -11071,7 +14314,8 @@ CUresult CUDAAPI cuGraphicsResourceGetMappedPointer(CUdeviceptr *pDevPtr, size_t * \notefnerr * * \sa - * ::cuGraphicsMapResources + * ::cuGraphicsMapResources, + * ::cudaGraphicsResourceSetMapFlags */ CUresult CUDAAPI cuGraphicsResourceSetMapFlags(CUgraphicsResource resource, unsigned int flags); @@ -11110,7 +14354,8 @@ CUresult CUDAAPI cuGraphicsResourceSetMapFlags(CUgraphicsResource resource, unsi * \sa * ::cuGraphicsResourceGetMappedPointer, * ::cuGraphicsSubResourceGetMappedArray, - * ::cuGraphicsUnmapResources + * ::cuGraphicsUnmapResources, + * ::cudaGraphicsMapResources */ CUresult CUDAAPI cuGraphicsMapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream); @@ -11146,7 +14391,8 @@ CUresult CUDAAPI cuGraphicsMapResources(unsigned int count, CUgraphicsResource * * \notefnerr * * \sa - * ::cuGraphicsMapResources + * ::cuGraphicsMapResources, + * ::cudaGraphicsUnmapResources */ CUresult CUDAAPI cuGraphicsUnmapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream); @@ -11225,6 +14471,7 @@ CUresult CUDAAPI cuGetExportTable(const void **ppExportTable, const CUuuid *pExp #undef cuMemsetD2D32Async #undef cuStreamGetPriority #undef cuStreamGetFlags + #undef cuStreamGetCtx #undef cuStreamWaitEvent #undef cuStreamAddCallback #undef cuStreamAttachMemAsync @@ -11232,12 +14479,23 @@ CUresult CUDAAPI cuGetExportTable(const void **ppExportTable, const CUuuid *pExp #undef cuStreamSynchronize #undef cuEventRecord #undef cuLaunchKernel + #undef cuLaunchHostFunc #undef cuGraphicsMapResources #undef cuGraphicsUnmapResources - #undef cuMemPrefetchAsync #undef cuStreamWriteValue32 #undef cuStreamWaitValue32 + #undef cuStreamWriteValue64 + #undef cuStreamWaitValue64 #undef cuStreamBatchMemOp + #undef cuMemPrefetchAsync + #undef cuLaunchCooperativeKernel + #undef cuSignalExternalSemaphoresAsync + #undef cuWaitExternalSemaphoresAsync + #undef cuStreamBeginCapture + #undef cuStreamEndCapture + #undef cuStreamIsCapturing + #undef cuStreamGetCaptureInfo + #undef cuGraphLaunch #endif /* __CUDA_API_VERSION_INTERNAL */ #if defined(__CUDA_API_VERSION_INTERNAL) || (__CUDA_API_VERSION >= 4000 && __CUDA_API_VERSION < 6050) @@ -11455,6 +14713,7 @@ CUresult CUDAAPI cuEventDestroy(CUevent hEvent); CUresult CUDAAPI cuStreamGetPriority(CUstream hStream, int *priority); CUresult CUDAAPI cuStreamGetFlags(CUstream hStream, unsigned int *flags); + CUresult CUDAAPI cuStreamGetCtx(CUstream hStream, CUcontext *pctx); CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int Flags); CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback, void *userData, unsigned int flags); CUresult CUDAAPI cuStreamAttachMemAsync(CUstream hStream, CUdeviceptr dptr, size_t length, unsigned int flags); @@ -11462,12 +14721,25 @@ CUresult CUDAAPI cuEventDestroy(CUevent hEvent); CUresult CUDAAPI cuStreamSynchronize(CUstream hStream); CUresult CUDAAPI cuEventRecord(CUevent hEvent, CUstream hStream); CUresult CUDAAPI cuLaunchKernel(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams, void **extra); + CUresult CUDAAPI cuLaunchHostFunc(CUstream hStream, CUhostFn fn, void *userData); CUresult CUDAAPI cuGraphicsMapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream); CUresult CUDAAPI cuGraphicsUnmapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream); - CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream); CUresult CUDAAPI cuStreamWriteValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags); CUresult CUDAAPI cuStreamWaitValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags); + CUresult CUDAAPI cuStreamWriteValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags); + CUresult CUDAAPI cuStreamWaitValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags); CUresult CUDAAPI cuStreamBatchMemOp(CUstream stream, unsigned int count, CUstreamBatchMemOpParams *paramArray, unsigned int flags); + CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream); + CUresult CUDAAPI cuLaunchCooperativeKernel(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams); + CUresult CUDAAPI cuSignalExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream); + CUresult CUDAAPI cuWaitExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream); + CUresult CUDAAPI cuStreamBeginCapture(CUstream hStream); + CUresult CUDAAPI cuStreamBeginCapture_ptsz(CUstream hStream); + CUresult CUDAAPI cuStreamBeginCapture_v2(CUstream hStream, CUstreamCaptureMode mode); + CUresult CUDAAPI cuStreamEndCapture(CUstream hStream, CUgraph *phGraph); + CUresult CUDAAPI cuStreamIsCapturing(CUstream hStream, CUstreamCaptureStatus *captureStatus); + CUresult CUDAAPI cuStreamGetCaptureInfo(CUstream hStream, CUstreamCaptureStatus *captureStatus, cuuint64_t *id); + CUresult CUDAAPI cuGraphLaunch(CUgraphExec hGraph, CUstream hStream); #endif #ifdef __cplusplus @@ -11475,6 +14747,6 @@ CUresult CUDAAPI cuEventDestroy(CUevent hEvent); #endif #undef __CUDA_API_VERSION +#undef __CUDA_DEPRECATED #endif /* __cuda_cuda_h__ */ - -- cgit v1.3 From 2ea18072618e7fe4e541f84de7d8575998299a1c Mon Sep 17 00:00:00 2001 From: tgrogers Date: Mon, 10 Jun 2019 09:19:30 -0400 Subject: Code to get CUDA 10 to work - looks like the gridDim/blockDim args given to cudaLaunchKernel are not complete garbage. We must rely on the PushConfig to get the proper sizing info --- libcuda/cuda_runtime_api.cc | 43 +++++++++++++++++++++++++++---------------- 1 file changed, 27 insertions(+), 16 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 718db49..500b5f2 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -2108,20 +2108,21 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 gridDim, dim3 blockDim, const void** args, size_t sharedMem, cudaStream_t stream ) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st *context = GPGPUSim_Context(); - function_info *entry = context->get_kernel(hostFun); - - cudaConfigureCallInternal(gridDim, blockDim, sharedMem, stream); - for(unsigned i = 0; i < entry->num_args(); i++){ - std::pair p = entry->get_param_config(i); - cudaSetupArgumentInternal(args[i], p.first, p.second); - } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + CUctx_st *context = GPGPUSim_Context(); + function_info *entry = context->get_kernel(hostFun); +#if CUDART_VERSION < 10000 + cudaConfigureCallInternal(gridDim, blockDim, sharedMem, stream); +#endif + for(unsigned i = 0; i < entry->num_args(); i++){ + std::pair p = entry->get_param_config(i); + cudaSetupArgumentInternal(args[i], p.first, p.second); + } - cudaLaunchInternal(hostFun); - return g_last_cudaError = cudaSuccess; + cudaLaunchInternal(hostFun); + return g_last_cudaError = cudaSuccess; } @@ -2974,12 +2975,17 @@ extern "C" { void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) { + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } return cudaRegisterFatBinaryInternal(fatCubin); } void CUDARTAPI __cudaRegisterFatBinaryEnd( void **fatCubinHandle ) { - + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } } unsigned CUDARTAPI __cudaPushCallConfiguration(dim3 gridDim, @@ -2987,7 +2993,10 @@ unsigned CUDARTAPI __cudaPushCallConfiguration(dim3 gridDim, size_t sharedMem = 0, struct CUstream_st *stream = 0) { - + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + cudaConfigureCallInternal(gridDim, blockDim, sharedMem, stream); } cudaError_t CUDARTAPI __cudaPopCallConfiguration( @@ -2997,6 +3006,9 @@ cudaError_t CUDARTAPI __cudaPopCallConfiguration( void *stream ) { + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } return g_last_cudaError = cudaSuccess; } @@ -3056,7 +3068,6 @@ void __cudaUnregisterFatBinary(void **fatCubinHandle) if(g_debug_execution >= 3){ announce_call(__my_func__); } - ; } cudaError_t cudaDeviceReset ( void ) { -- cgit v1.3 From c29246408c963ece65515fae92540e76ac71b72b Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 10 Jun 2019 11:17:23 -0400 Subject: Move some struct upward in cuda_runtime_api.cc Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 149 +++++++++++++++++++++++++++++++++++++++++--- libcuda/cuda_runtime_api.cc | 141 ----------------------------------------- libcuda/cuobjdump.h | 1 + 3 files changed, 143 insertions(+), 148 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index 41337c6..73c077e 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -6,13 +6,10 @@ #include #include -class cuobjdumpSection; -class cuobjdumpELFSection; -class cuobjdumpPTXSection; -class symbol_table; -class gpgpu_ptx_sim_arg; -class kernel_config; -class kernel_info_t; +#include "../src/gpgpu-sim/gpu-sim.h" +#include "../src/cuda-sim/ptx_ir.h" +#include "../src/abstract_hardware_model.h" +#include "cuobjdump.h" typedef std::list gpgpu_ptx_sim_arg_list_t; @@ -29,6 +26,144 @@ struct glbmap_entry { typedef struct glbmap_entry glbmap_entry_t; +struct _cuda_device_id { + _cuda_device_id(gpgpu_sim* gpu) {m_id = 0; m_next = NULL; m_gpgpu=gpu;} + struct _cuda_device_id *next() { return m_next; } + unsigned num_shader() const { return m_gpgpu->get_config().num_shader(); } + int num_devices() const { + if( m_next == NULL ) return 1; + else return 1 + m_next->num_devices(); + } + struct _cuda_device_id *get_device( unsigned n ) + { + assert( n < (unsigned)num_devices() ); + struct _cuda_device_id *p=this; + for(unsigned i=0; im_next; + return p; + } + const struct cudaDeviceProp *get_prop() const + { + return m_gpgpu->get_prop(); + } + unsigned get_id() const { return m_id; } + + gpgpu_sim *get_gpgpu() { return m_gpgpu; } +private: + unsigned m_id; + class gpgpu_sim *m_gpgpu; + struct _cuda_device_id *m_next; +}; + +struct CUctx_st { + CUctx_st( _cuda_device_id *gpu ) + { + m_gpu = gpu; + m_binary_info.cmem = 0; + m_binary_info.gmem = 0; + no_of_ptx=0; + } + + _cuda_device_id *get_device() { return m_gpu; } + + void add_binary( symbol_table *symtab, unsigned fat_cubin_handle ) + { + m_code[fat_cubin_handle] = symtab; + m_last_fat_cubin_handle = fat_cubin_handle; + } + + void add_ptxinfo( const char *deviceFun, const struct gpgpu_ptx_sim_info &info ) + { + symbol *s = m_code[m_last_fat_cubin_handle]->lookup(deviceFun); + assert( s != NULL ); + function_info *f = s->get_pc(); + assert( f != NULL ); + f->set_kernel_info(info); + } + + void add_ptxinfo( const struct gpgpu_ptx_sim_info &info ) + { + m_binary_info = info; + } + + void register_function( unsigned fat_cubin_handle, const char *hostFun, const char *deviceFun ) + { + if( m_code.find(fat_cubin_handle) != m_code.end() ) { + symbol *s = m_code[fat_cubin_handle]->lookup(deviceFun); + if(s != NULL) { + function_info *f = s->get_pc(); + assert( f != NULL ); + m_kernel_lookup[hostFun] = f; + } + else { + printf("Warning: cannot find deviceFun %s\n", deviceFun); + m_kernel_lookup[hostFun] = NULL; + } + // assert( s != NULL ); + // function_info *f = s->get_pc(); + // assert( f != NULL ); + // m_kernel_lookup[hostFun] = f; + } else { + m_kernel_lookup[hostFun] = NULL; + } + } + + void register_hostFun_function( const char*hostFun, function_info* f){ + m_kernel_lookup[hostFun] = f; + } + + function_info *get_kernel(const char *hostFun) + { + std::map::iterator i=m_kernel_lookup.find(hostFun); + assert( i != m_kernel_lookup.end() ); + return i->second; + } + + int no_of_ptx; + +private: + _cuda_device_id *m_gpu; // selected gpu + std::map m_code; // fat binary handle => global symbol table + unsigned m_last_fat_cubin_handle; + std::map m_kernel_lookup; // unique id (CUDA app function address) => kernel entry point + struct gpgpu_ptx_sim_info m_binary_info; + +}; + +class kernel_config { +public: + kernel_config( dim3 GridDim, dim3 BlockDim, size_t sharedMem, struct CUstream_st *stream ) + { + m_GridDim=GridDim; + m_BlockDim=BlockDim; + m_sharedMem=sharedMem; + m_stream = stream; + } + kernel_config() + { + m_GridDim=dim3(-1,-1,-1); + m_BlockDim=dim3(-1,-1,-1); + m_sharedMem=0; + m_stream =NULL; + } + void set_arg( const void *arg, size_t size, size_t offset ) + { + m_args.push_front( gpgpu_ptx_sim_arg(arg,size,offset) ); + } + dim3 grid_dim() const { return m_GridDim; } + dim3 block_dim() const { return m_BlockDim; } + void set_grid_dim(dim3 *d) { m_GridDim = *d; } + void set_block_dim(dim3 *d) { m_BlockDim = *d; } + gpgpu_ptx_sim_arg_list_t get_args() { return m_args; } + struct CUstream_st *get_stream() { return m_stream; } + +private: + dim3 m_GridDim; + dim3 m_BlockDim; + size_t m_sharedMem; + struct CUstream_st *m_stream; + gpgpu_ptx_sim_arg_list_t m_args; +}; class cuda_runtime_api { public: diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 72ab002..2e2b50b 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -141,8 +141,6 @@ #include "../src/gpgpusim_entrypoint.h" #include "../src/stream_manager.h" #include "../src/abstract_hardware_model.h" -typedef void * yyscan_t; -#include "cuobjdump.h" #include #include @@ -208,145 +206,6 @@ void register_ptx_function( const char *name, function_info *impl ) # endif #endif -struct _cuda_device_id { - _cuda_device_id(gpgpu_sim* gpu) {m_id = 0; m_next = NULL; m_gpgpu=gpu;} - struct _cuda_device_id *next() { return m_next; } - unsigned num_shader() const { return m_gpgpu->get_config().num_shader(); } - int num_devices() const { - if( m_next == NULL ) return 1; - else return 1 + m_next->num_devices(); - } - struct _cuda_device_id *get_device( unsigned n ) - { - assert( n < (unsigned)num_devices() ); - struct _cuda_device_id *p=this; - for(unsigned i=0; im_next; - return p; - } - const struct cudaDeviceProp *get_prop() const - { - return m_gpgpu->get_prop(); - } - unsigned get_id() const { return m_id; } - - gpgpu_sim *get_gpgpu() { return m_gpgpu; } -private: - unsigned m_id; - class gpgpu_sim *m_gpgpu; - struct _cuda_device_id *m_next; -}; - -struct CUctx_st { - CUctx_st( _cuda_device_id *gpu ) - { - m_gpu = gpu; - m_binary_info.cmem = 0; - m_binary_info.gmem = 0; - no_of_ptx=0; - } - - _cuda_device_id *get_device() { return m_gpu; } - - void add_binary( symbol_table *symtab, unsigned fat_cubin_handle ) - { - m_code[fat_cubin_handle] = symtab; - m_last_fat_cubin_handle = fat_cubin_handle; - } - - void add_ptxinfo( const char *deviceFun, const struct gpgpu_ptx_sim_info &info ) - { - symbol *s = m_code[m_last_fat_cubin_handle]->lookup(deviceFun); - assert( s != NULL ); - function_info *f = s->get_pc(); - assert( f != NULL ); - f->set_kernel_info(info); - } - - void add_ptxinfo( const struct gpgpu_ptx_sim_info &info ) - { - m_binary_info = info; - } - - void register_function( unsigned fat_cubin_handle, const char *hostFun, const char *deviceFun ) - { - if( m_code.find(fat_cubin_handle) != m_code.end() ) { - symbol *s = m_code[fat_cubin_handle]->lookup(deviceFun); - if(s != NULL) { - function_info *f = s->get_pc(); - assert( f != NULL ); - m_kernel_lookup[hostFun] = f; - } - else { - printf("Warning: cannot find deviceFun %s\n", deviceFun); - m_kernel_lookup[hostFun] = NULL; - } - // assert( s != NULL ); - // function_info *f = s->get_pc(); - // assert( f != NULL ); - // m_kernel_lookup[hostFun] = f; - } else { - m_kernel_lookup[hostFun] = NULL; - } - } - - void register_hostFun_function( const char*hostFun, function_info* f){ - m_kernel_lookup[hostFun] = f; - } - - function_info *get_kernel(const char *hostFun) - { - std::map::iterator i=m_kernel_lookup.find(hostFun); - assert( i != m_kernel_lookup.end() ); - return i->second; - } - - int no_of_ptx; - -private: - _cuda_device_id *m_gpu; // selected gpu - std::map m_code; // fat binary handle => global symbol table - unsigned m_last_fat_cubin_handle; - std::map m_kernel_lookup; // unique id (CUDA app function address) => kernel entry point - struct gpgpu_ptx_sim_info m_binary_info; - -}; - -class kernel_config { -public: - kernel_config( dim3 GridDim, dim3 BlockDim, size_t sharedMem, struct CUstream_st *stream ) - { - m_GridDim=GridDim; - m_BlockDim=BlockDim; - m_sharedMem=sharedMem; - m_stream = stream; - } - kernel_config() - { - m_GridDim=dim3(-1,-1,-1); - m_BlockDim=dim3(-1,-1,-1); - m_sharedMem=0; - m_stream =NULL; - } - void set_arg( const void *arg, size_t size, size_t offset ) - { - m_args.push_front( gpgpu_ptx_sim_arg(arg,size,offset) ); - } - dim3 grid_dim() const { return m_GridDim; } - dim3 block_dim() const { return m_BlockDim; } - void set_grid_dim(dim3 *d) { m_GridDim = *d; } - void set_block_dim(dim3 *d) { m_BlockDim = *d; } - gpgpu_ptx_sim_arg_list_t get_args() { return m_args; } - struct CUstream_st *get_stream() { return m_stream; } - -private: - dim3 m_GridDim; - dim3 m_BlockDim; - size_t m_sharedMem; - struct CUstream_st *m_stream; - gpgpu_ptx_sim_arg_list_t m_args; -}; - struct _cuda_device_id *GPGPUSim_Init() { //static _cuda_device_id *the_device = NULL; diff --git a/libcuda/cuobjdump.h b/libcuda/cuobjdump.h index 49af3e2..6ab6778 100644 --- a/libcuda/cuobjdump.h +++ b/libcuda/cuobjdump.h @@ -4,6 +4,7 @@ #include #include +typedef void * yyscan_t; struct cuobjdump_parser { yyscan_t scanner; int elfserial; -- cgit v1.3 From c68377b8eb2ce322e195a23062f5ba635dba546a Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 10 Jun 2019 19:43:36 -0400 Subject: Remove those duplicated defined Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 2 ++ src/Makefile | 2 ++ src/abstract_hardware_model.h | 75 ++++++++++++++++++++++--------------------- src/cuda-sim/cuda-math.h | 16 ++++----- src/cuda-sim/instructions.cc | 10 +++--- src/gpgpu-sim/Makefile | 2 ++ src/intersim2/Makefile | 1 + 7 files changed, 59 insertions(+), 49 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index 73c077e..d931fd5 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -6,6 +6,8 @@ #include #include +#include "builtin_types.h" + #include "../src/gpgpu-sim/gpu-sim.h" #include "../src/cuda-sim/ptx_ir.h" #include "../src/abstract_hardware_model.h" diff --git a/src/Makefile b/src/Makefile index 6001669..3ad511e 100644 --- a/src/Makefile +++ b/src/Makefile @@ -51,6 +51,8 @@ else CXXFLAGS += endif +CXXFLAGS += -I$(CUDA_INSTALL_PATH)/include + OPTFLAGS += -g3 -fPIC CPP = g++ $(SNOW) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 77d5f58..7455f25 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -173,9 +173,10 @@ enum _memory_op_t { #include #if !defined(__VECTOR_TYPES_H__) -struct dim3 { - unsigned int x, y, z; -}; +#include "vector_types.h" +//struct dim3 { +// unsigned int x, y, z; +//}; #endif struct dim3comp { bool operator() (const dim3 & a, const dim3 & b) const @@ -454,19 +455,21 @@ protected: #if !defined(__CUDA_RUNTIME_API_H__) -enum cudaChannelFormatKind { - cudaChannelFormatKindSigned, - cudaChannelFormatKindUnsigned, - cudaChannelFormatKindFloat -}; +#include "builtin_types.h" -struct cudaChannelFormatDesc { - int x; - int y; - int z; - int w; - enum cudaChannelFormatKind f; -}; +//enum cudaChannelFormatKind { +// cudaChannelFormatKindSigned, +// cudaChannelFormatKindUnsigned, +// cudaChannelFormatKindFloat +//}; + +//struct cudaChannelFormatDesc { +// int x; +// int y; +// int z; +// int w; +// enum cudaChannelFormatKind f; +//}; struct cudaArray { void *devPtr; @@ -478,27 +481,27 @@ struct cudaArray { unsigned dimensions; }; -enum cudaTextureAddressMode { - cudaAddressModeWrap, - cudaAddressModeClamp -}; - -enum cudaTextureFilterMode { - cudaFilterModePoint, - cudaFilterModeLinear -}; - -enum cudaTextureReadMode { - cudaReadModeElementType, - cudaReadModeNormalizedFloat -}; - -struct textureReference { - int normalized; - enum cudaTextureFilterMode filterMode; - enum cudaTextureAddressMode addressMode[3]; - struct cudaChannelFormatDesc channelDesc; -}; +//enum cudaTextureAddressMode { +// cudaAddressModeWrap, +// cudaAddressModeClamp +//}; + +//enum cudaTextureFilterMode { +// cudaFilterModePoint, +// cudaFilterModeLinear +//}; + +//enum cudaTextureReadMode { +// cudaReadModeElementType, +// cudaReadModeNormalizedFloat +//}; + +//struct textureReference { +// int normalized; +// enum cudaTextureFilterMode filterMode; +// enum cudaTextureAddressMode addressMode[3]; +// struct cudaChannelFormatDesc channelDesc; +//}; #endif diff --git a/src/cuda-sim/cuda-math.h b/src/cuda-sim/cuda-math.h index a5db337..9a5468c 100644 --- a/src/cuda-sim/cuda-math.h +++ b/src/cuda-sim/cuda-math.h @@ -277,10 +277,10 @@ int float2int(float a, enum cudaRoundMode mode) { int tmp; switch (mode) { - case cuda_math::cudaRoundZero: tmp = truncf(a); break; - case cuda_math::cudaRoundNearest: tmp = nearbyintf(a); break; - case cuda_math::cudaRoundMinInf: tmp = floorf(a); break; - case cuda_math::cudaRoundPosInf: tmp = ceilf(a); break; + case cudaRoundZero: tmp = truncf(a); break; + case cudaRoundNearest: tmp = nearbyintf(a); break; + case cudaRoundMinInf: tmp = floorf(a); break; + case cudaRoundPosInf: tmp = ceilf(a); break; default: abort(); } return tmp; @@ -296,10 +296,10 @@ unsigned int float2uint(float a, enum cudaRoundMode mode) { unsigned int tmp; switch (mode) { - case cuda_math::cudaRoundZero: tmp = truncf(a); break; - case cuda_math::cudaRoundNearest: tmp = nearbyintf(a); break; - case cuda_math::cudaRoundMinInf: tmp = floorf(a); break; - case cuda_math::cudaRoundPosInf: tmp = ceilf(a); break; + case cudaRoundZero: tmp = truncf(a); break; + case cudaRoundNearest: tmp = nearbyintf(a); break; + case cudaRoundMinInf: tmp = floorf(a); break; + case cudaRoundPosInf: tmp = ceilf(a); break; default: abort(); } return tmp; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 69a97b6..186bb1e 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -2302,12 +2302,12 @@ ptx_reg_t f2x( ptx_reg_t x, unsigned from_width, unsigned to_width, int to_sign, half_float::half tmp_h; //assert( from_width == 32); - enum cuda_math::cudaRoundMode mode = cuda_math::cudaRoundZero; + enum cudaRoundMode mode = cudaRoundZero; switch (rounding_mode) { - case RZI_OPTION: mode = cuda_math::cudaRoundZero; break; - case RNI_OPTION: mode = cuda_math::cudaRoundNearest; break; - case RMI_OPTION: mode = cuda_math::cudaRoundMinInf; break; - case RPI_OPTION: mode = cuda_math::cudaRoundPosInf; break; + case RZI_OPTION: mode = cudaRoundZero; break; + case RNI_OPTION: mode = cudaRoundNearest; break; + case RMI_OPTION: mode = cudaRoundMinInf; break; + case RPI_OPTION: mode = cudaRoundPosInf; break; default: break; } diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile index f10a8a4..4994577 100644 --- a/src/gpgpu-sim/Makefile +++ b/src/gpgpu-sim/Makefile @@ -53,6 +53,8 @@ else CXXFLAGS += endif +CXXFLAGS += -I$(CUDA_INSTALL_PATH)/include + POWER_FLAGS= ifneq ($(GPGPUSIM_POWER_MODEL),) POWER_FLAGS = -I$(GPGPUSIM_POWER_MODEL) -DGPGPUSIM_POWER_MODEL diff --git a/src/intersim2/Makefile b/src/intersim2/Makefile index 7d10b3f..3eeeb70 100644 --- a/src/intersim2/Makefile +++ b/src/intersim2/Makefile @@ -48,6 +48,7 @@ CPPFLAGS += -O3 endif CPPFLAGS += -g CPPFLAGS += -fPIC +CPPFLAGS += -I$(CUDA_INSTALL_PATH)/include LFLAGS += -- cgit v1.3 From 269fbbdb72dacdf4dc3c482b213174336c3eeb1f Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 10 Jun 2019 22:42:27 -0400 Subject: Move getters for gpgpu_context to globals Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- libcuda/gpgpu_context.h | 2 ++ 2 files changed, 3 insertions(+), 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 2e2b50b..c5ebc20 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -273,7 +273,7 @@ static CUctx_st* GPGPUSim_Context() return the_context; } -static gpgpu_context* GPGPU_Context() +gpgpu_context* GPGPU_Context() { static gpgpu_context *gpgpu_ctx = NULL; if( gpgpu_ctx == NULL ) { diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 4622f00..6c8a293 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -12,4 +12,6 @@ class gpgpu_context { cuda_runtime_api* api; // member function list }; +gpgpu_context* GPGPU_Context(); + #endif /* __gpgpu_context_h__ */ -- cgit v1.3 From ba1ed2941753ae8406bc9ec4a0de1eddc6454a1c Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 00:42:37 -0400 Subject: Move some vars Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 5 +++ libcuda/cuda_runtime_api.cc | 74 ++++++++++++++++++++++++++------------------- 2 files changed, 48 insertions(+), 31 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index d931fd5..f9a4fde 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -171,6 +171,7 @@ class cuda_runtime_api { public: cuda_runtime_api() { g_glbmap = NULL; + g_active_device = 0; //active gpu that runs the code } // global list std::list cuobjdumpSectionList; @@ -185,6 +186,7 @@ class cuda_runtime_api { std::map pinned_memory; //support for pinned memories added std::map pinned_memory_size; glbmap_entry_t* g_glbmap; + int g_active_device; //active gpu that runs the code // member function list void cuobjdumpInit(); void extract_code_using_cuobjdump(); @@ -201,5 +203,8 @@ class cuda_runtime_api { struct dim3 gridDim, struct dim3 blockDim, struct CUctx_st* context ); + int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr, gpgpu_t *gpu ); + int load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu ); + }; #endif /* __cuda_api_object_h__ */ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index c5ebc20..09a13a7 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -153,15 +153,6 @@ extern void synchronize(); extern void exit_simulation(); -static int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr, gpgpu_t *gpu ); -static int load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu ); - -//static kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, -// gpgpu_ptx_sim_arg_list_t args, -// struct dim3 gridDim, -// struct dim3 blockDim, -// struct CUctx_st* context ); - /*DEVICE_BUILTIN*/ struct cudaArray { @@ -353,7 +344,6 @@ typedef std::map event_tracker_t; int CUevent_st::m_next_event_uid; event_tracker_t g_timer_events; -int g_active_device = 0; //active gpu that runs the code extern int cuobjdump_lex_init(yyscan_t* scanner); extern void cuobjdump_set_in (FILE * _in_str ,yyscan_t yyscanner ); @@ -492,6 +482,41 @@ void cuda_runtime_api::cuobjdumpRegisterFatBinary(unsigned int handle, const cha /******************************************************************************* * Add internal cuda runtime API call to accept gpgpu_context * *******************************************************************************/ +cudaError_t cudaSetDeviceInternal(int device, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + //set the active device to run cuda + if ( device <= GPGPUSim_Init()->num_devices() ) { + ctx->api->g_active_device = device; + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorInvalidDevice; + } +} + +cudaError_t cudaGetDeviceInternal(int *device, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + *device = ctx->api->g_active_device; + return g_last_cudaError = cudaSuccess; +} + void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = NULL) { @@ -618,8 +643,8 @@ void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); } source_num++; - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + ctx->api->load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + ctx->api->load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); } else { printf("GPGPU-Sim PTX: warning -- did not find an appropriate PTX in cubin\n"); } @@ -988,8 +1013,8 @@ cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, std::string fname(path); ctx->api->name_symtab[fname] = symtab; context->add_binary(symtab, 1); - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + ctx->api->load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + ctx->api->load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); addedFile = true; return CUDA_SUCCESS; } @@ -1692,25 +1717,12 @@ __host__ cudaError_t CUDARTAPI cudaChooseDevice(int *device, const struct cudaDe __host__ cudaError_t CUDARTAPI cudaSetDevice(int device) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - //set the active device to run cuda - if ( device <= GPGPUSim_Init()->num_devices() ) { - g_active_device = device; - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorInvalidDevice; - } + return cudaSetDeviceInternal(device); } __host__ cudaError_t CUDARTAPI cudaGetDevice(int *device) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - *device = g_active_device; - return g_last_cudaError = cudaSuccess; + return cudaGetDeviceInternal(device); } __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit limit ) @@ -3269,7 +3281,7 @@ int CUDARTAPI __cudaSynchronizeThreads(void**, void*) /// static functions -static int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr, gpgpu_t *gpu ) +int cuda_runtime_api::load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr, gpgpu_t *gpu ) { if(g_debug_execution >= 3){ announce_call(__my_func__); @@ -3308,7 +3320,7 @@ static int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsign return ng_bytes; } -static int load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu ) +int cuda_runtime_api::load_constants( symbol_table *symtab, addr_t min_gaddr, gpgpu_t *gpu ) { if(g_debug_execution >= 3){ announce_call(__my_func__); -- cgit v1.3 From 9e52c143a883f682c02d81149748cdf8aa5508f7 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 01:42:52 -0400 Subject: Move some function from ptx_loader to gpgpu_context Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 1 - libcuda/cuda_runtime_api.cc | 42 ++++++++++++++++++++--------------------- libcuda/gpgpu_context.h | 3 +++ libopencl/opencl_runtime_api.cc | 5 ++++- src/cuda-sim/ptx_loader.cc | 5 +++-- src/cuda-sim/ptx_loader.h | 2 -- 6 files changed, 31 insertions(+), 27 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index f9a4fde..0054697 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -191,7 +191,6 @@ class cuda_runtime_api { void cuobjdumpInit(); void extract_code_using_cuobjdump(); void extract_ptx_files_using_cuobjdump(CUctx_st *context); - void cuobjdumpParseBinary(unsigned int handle); std::list pruneSectionList(CUctx_st *context); std::list mergeMatchingSections(std::string identifier); std::list mergeSections(); diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 09a13a7..25642a7 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -685,7 +685,7 @@ void cudaRegisterFunctionInternal( printf("GPGPU-Sim PTX: __cudaRegisterFunction %s : hostFun 0x%p, fat_cubin_handle = %u\n", deviceFun, hostFun, fat_cubin_handle); if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - ctx->api->cuobjdumpParseBinary(fat_cubin_handle); + ctx->cuobjdumpParseBinary(fat_cubin_handle); context->register_function( fat_cubin_handle, hostFun, deviceFun ); } @@ -712,7 +712,7 @@ void cudaRegisterVarInternal( printf("GPGPU-Sim PTX: __cudaRegisterVar: hostVar = %p; deviceAddress = %s; deviceName = %s\n", hostVar, deviceAddress, deviceName); printf("GPGPU-Sim PTX: __cudaRegisterVar: Registering const memory space of %d bytes\n", size); if(GPGPUSim_Context()->get_device()->get_gpgpu()->get_config().use_cuobjdump()) - ctx->api->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); + ctx->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); fflush(stdout); if ( constant && !global && !ext ) { gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); @@ -1009,7 +1009,7 @@ cuLinkAddFileInternal(CUlinkState state, CUjitInputType type, const char *path, } strcat(file,"/"); strcat(file,path); - symbol_table *symtab = gpgpu_ptx_sim_load_ptx_from_filename( file ); + symbol_table *symtab = ctx->gpgpu_ptx_sim_load_ptx_from_filename( file ); std::string fname(path); ctx->api->name_symtab[fname] = symtab; context->add_binary(symtab, 1); @@ -2750,15 +2750,15 @@ void cuda_runtime_api::cuobjdumpInit(){ //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it -void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ +void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ CUctx_st *context = GPGPUSim_Context(); - if(fatbin_registered[handle]) return; - fatbin_registered[handle] = true; - std::string fname = fatbinmap[handle]; + if(api->fatbin_registered[handle]) return; + api->fatbin_registered[handle] = true; + std::string fname = api->fatbinmap[handle]; - if (name_symtab.find(fname) != name_symtab.end()) { - symbol_table *symtab = name_symtab[fname]; + if (api->name_symtab.find(fname) != api->name_symtab.end()) { + symbol_table *symtab = api->name_symtab[fname]; context->add_binary(symtab, handle); return; } @@ -2767,7 +2767,7 @@ void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ #if (CUDART_VERSION >= 6000) //loops through all ptx files from smallest sm version to largest std::map >::iterator itr_m; - for (itr_m = version_filename.begin(); itr_m!=version_filename.end(); itr_m++){ + for (itr_m = api->version_filename.begin(); itr_m!=api->version_filename.end(); itr_m++){ std::set::iterator itr_s; for (itr_s = itr_m->second.begin(); itr_s!=itr_m->second.end(); itr_s++){ std::string ptx_filename = *itr_s; @@ -2775,11 +2775,11 @@ void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ symtab = gpgpu_ptx_sim_load_ptx_from_filename( ptx_filename.c_str() ); } } - name_symtab[fname] = symtab; + api->name_symtab[fname] = symtab; context->add_binary(symtab, handle); - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - for (itr_m = version_filename.begin(); itr_m!=version_filename.end(); itr_m++){ + api->load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + api->load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + for (itr_m = api->version_filename.begin(); itr_m!=api->version_filename.end(); itr_m++){ std::set::iterator itr_s; for (itr_s = itr_m->second.begin(); itr_s!=itr_m->second.end(); itr_s++){ std::string ptx_filename = *itr_s; @@ -2791,8 +2791,8 @@ void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ #endif unsigned max_capability = 0; - for ( std::list::iterator iter = cuobjdumpSectionList.begin(); - iter != cuobjdumpSectionList.end(); + for ( std::list::iterator iter = api->cuobjdumpSectionList.begin(); + iter != api->cuobjdumpSectionList.end(); iter++){ unsigned capability = (*iter)->getArch(); if (capability > max_capability) max_capability = capability; @@ -2803,7 +2803,7 @@ void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ cuobjdumpPTXSection* ptx = NULL; const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); if(pre_load==NULL || strlen(pre_load)==0) - ptx = findPTXSection(fname); + ptx = api->findPTXSection(fname); char *ptxcode; const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) { @@ -2813,7 +2813,7 @@ void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ ptxcode = readfile(override_ptx_name); } if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier()); + cuobjdumpELFSection* elfsection = api->findELFSection(ptx->getIdentifier()); assert (elfsection!= NULL); char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( ptx->getPTXfilename(), @@ -2831,9 +2831,9 @@ void cuda_runtime_api::cuobjdumpParseBinary(unsigned int handle){ context->add_binary(symtab, handle); gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability, context->no_of_ptx ); } - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - name_symtab[fname] = symtab; + api->load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + api->load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + api->name_symtab[fname] = symtab; //TODO: Remove temporarily files as per configurations } diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 6c8a293..16626eb 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -11,6 +11,9 @@ class gpgpu_context { // objects pointers for each file cuda_runtime_api* api; // member function list + void cuobjdumpParseBinary(unsigned int handle); + class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); + class symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ); }; gpgpu_context* GPGPU_Context(); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index 97a54d8..0ec635e 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -84,6 +84,7 @@ #include "../src/gpgpusim_entrypoint.h" #include "../src/gpgpu-sim/gpu-sim.h" #include "../src/gpgpu-sim/shader.h" +#include "../libcuda/gpgpu_context.h" //# define __my_func__ __PRETTY_FUNCTION__ # if defined __cplusplus ? __GNUC_PREREQ (2, 6) : __GNUC_PREREQ (2, 4) @@ -424,6 +425,8 @@ void register_ptx_function( const char *name, function_info *impl ) void _cl_program::Build(const char *options) { + gpgpu_context *ctx; + ctx = GPGPU_Context(); printf("GPGPU-Sim OpenCL API: compiling OpenCL kernels...\n"); std::map::iterator i; for( i = m_pgm.begin(); i!= m_pgm.end(); i++ ) { @@ -576,7 +579,7 @@ void _cl_program::Build(const char *options) } } info.m_asm = tmp; - info.m_symtab = gpgpu_ptx_sim_load_ptx_from_string( tmp, source_num ); + info.m_symtab = ctx->gpgpu_ptx_sim_load_ptx_from_string( tmp, source_num ); gpgpu_ptxinfo_load_from_string( tmp, source_num ); free(tmp); } diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index f037c34..d7e9b71 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -33,6 +33,7 @@ #include #include #include +#include "../../libcuda/gpgpu_context.h" /// globals @@ -165,7 +166,7 @@ char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilenam } -symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ) +symbol_table *gpgpu_context::gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ) { char buf[1024]; snprintf(buf,1024,"_%u.ptx", source_num ); @@ -200,7 +201,7 @@ symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source return symtab; } -symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ) +symbol_table *gpgpu_context::gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ) { symbol_table *symtab=init_parser(filename); printf("GPGPU-Sim PTX: finished parsing EMBEDDED .ptx file %s\n",filename); diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index c4d8292..2af611a 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -42,8 +42,6 @@ class ptxinfo_data{ extern bool g_override_embedded_ptx; extern int no_of_ptx; //counter to track number of ptx files to be extracted in an application. -class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); -class symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ); void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version ); char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); -- cgit v1.3 From c7d713104df5bab5583b3a0e96323cbe346f9759 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 09:42:25 -0400 Subject: Fix for 4.2 Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 25642a7..14e3329 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -638,7 +638,7 @@ void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = "\tEither enable cuobjdump or disable PTXPlus in your configuration file\n"); exit(1); } else { - symtab=gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); + symtab=ctx->gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); context->add_binary(symtab,fat_cubin_handle); gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); } -- cgit v1.3 From 7d02cbb061485db38ed8e5f6bf06c9b2fa40eed2 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 13:19:08 -0400 Subject: Integrate ptxinfo into gpgpu_context Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- libcuda/gpgpu_context.h | 5 +++++ libopencl/opencl_runtime_api.cc | 2 +- src/cuda-sim/ptx_loader.cc | 43 +++++++++++++++++++---------------------- src/cuda-sim/ptx_loader.h | 2 -- 5 files changed, 27 insertions(+), 27 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 14e3329..0a6eabb 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -640,7 +640,7 @@ void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = } else { symtab=ctx->gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); context->add_binary(symtab,fat_cubin_handle); - gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); + ctx->gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability, context->no_of_ptx ); } source_num++; ctx->api->load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 16626eb..d3db1ad 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -1,19 +1,24 @@ #ifndef __gpgpu_context_h__ #define __gpgpu_context_h__ #include "cuda_api_object.h" +#include "../src/cuda-sim/ptx_loader.h" class gpgpu_context { public: gpgpu_context() { api = new cuda_runtime_api(); + ptxinfo = new ptxinfo_data(); } // global list // objects pointers for each file cuda_runtime_api* api; + ptxinfo_data* ptxinfo; // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); class symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ); + void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version); + void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); }; gpgpu_context* GPGPU_Context(); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index 0ec635e..50a02fa 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -580,7 +580,7 @@ void _cl_program::Build(const char *options) } info.m_asm = tmp; info.m_symtab = ctx->gpgpu_ptx_sim_load_ptx_from_string( tmp, source_num ); - gpgpu_ptxinfo_load_from_string( tmp, source_num ); + ctx->gpgpu_ptxinfo_load_from_string( tmp, source_num ); free(tmp); } printf("GPGPU-Sim OpenCL API: finished compiling OpenCL kernels.\n"); diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 9857741..2a6d930 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -332,7 +332,7 @@ char* get_app_binary_name(){ return self_exe_path; } -void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version) +void gpgpu_context::gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version) { std::string ptxas_filename(std::string(filename) + "as"); char buff[1024], extra_flags[1024]; @@ -352,17 +352,16 @@ void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_versio } FILE *ptxinfo_in; - ptxinfo_data ptxinfo; - ptxinfo.g_ptxinfo_filename = strdup(ptxas_filename.c_str()); - ptxinfo_in = fopen(ptxinfo.g_ptxinfo_filename,"r"); - ptxinfo_lex_init(&(ptxinfo.scanner)); - ptxinfo_set_in(ptxinfo_in, ptxinfo.scanner); - ptxinfo_parse(ptxinfo.scanner, &ptxinfo); - ptxinfo_lex_destroy(ptxinfo.scanner); + ptxinfo->g_ptxinfo_filename = strdup(ptxas_filename.c_str()); + ptxinfo_in = fopen(ptxinfo->g_ptxinfo_filename,"r"); + ptxinfo_lex_init(&(ptxinfo->scanner)); + ptxinfo_set_in(ptxinfo_in, ptxinfo->scanner); + ptxinfo_parse(ptxinfo->scanner, ptxinfo); + ptxinfo_lex_destroy(ptxinfo->scanner); fclose(ptxinfo_in); } -void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version, int no_of_ptx ) +void gpgpu_context::gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version, int no_of_ptx ) { //do ptxas for individual files instead of one big embedded ptx. This prevents the duplicate defs and declarations. char ptx_file[1000]; @@ -420,14 +419,13 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num if( result != 0 ) { // 65280 = duplicate errors if (result == 65280) { - ptxinfo_data ptxinfo; FILE *ptxinfo_in; ptxinfo_in = fopen(tempfile_ptxinfo,"r"); - ptxinfo.g_ptxinfo_filename = tempfile_ptxinfo; - ptxinfo_lex_init(&(ptxinfo.scanner)); - ptxinfo_set_in(ptxinfo_in, ptxinfo.scanner); - ptxinfo_parse(ptxinfo.scanner, &ptxinfo); - ptxinfo_lex_destroy(ptxinfo.scanner); + ptxinfo->g_ptxinfo_filename = tempfile_ptxinfo; + ptxinfo_lex_init(&(ptxinfo->scanner)); + ptxinfo_set_in(ptxinfo_in, ptxinfo->scanner); + ptxinfo_parse(ptxinfo->scanner, ptxinfo); + ptxinfo_lex_destroy(ptxinfo->scanner); fclose(ptxinfo_in); fix_duplicate_errors(fname2); @@ -507,18 +505,17 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num } } - ptxinfo_data ptxinfo; if(no_of_ptx>0) - ptxinfo.g_ptxinfo_filename = final_tempfile_ptxinfo; + ptxinfo->g_ptxinfo_filename = final_tempfile_ptxinfo; else - ptxinfo.g_ptxinfo_filename = tempfile_ptxinfo; + ptxinfo->g_ptxinfo_filename = tempfile_ptxinfo; FILE *ptxinfo_in; - ptxinfo_in = fopen(ptxinfo.g_ptxinfo_filename,"r"); + ptxinfo_in = fopen(ptxinfo->g_ptxinfo_filename,"r"); - ptxinfo_lex_init(&(ptxinfo.scanner)); - ptxinfo_set_in(ptxinfo_in, ptxinfo.scanner); - ptxinfo_parse(ptxinfo.scanner, &ptxinfo); - ptxinfo_lex_destroy(ptxinfo.scanner); + ptxinfo_lex_init(&(ptxinfo->scanner)); + ptxinfo_set_in(ptxinfo_in, ptxinfo->scanner); + ptxinfo_parse(ptxinfo->scanner, ptxinfo); + ptxinfo_lex_destroy(ptxinfo->scanner); fclose(ptxinfo_in); snprintf(commandline,1024,"rm -f *info"); diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index c3ce888..3637bff 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -44,8 +44,6 @@ class ptxinfo_data{ extern bool g_override_embedded_ptx; extern int no_of_ptx; //counter to track number of ptx files to be extracted in an application. -void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); -void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version ); char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); bool keep_intermediate_files(); -- cgit v1.3 From 7ac6034a99b52d09db7ef07bc008b8f6b039f929 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 18:03:35 -0400 Subject: Move print_ptx_file Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 1 + src/cuda-sim/ptx_loader.cc | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index d3db1ad..dbc3140 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -19,6 +19,7 @@ class gpgpu_context { class symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename ); void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version); void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); + void print_ptx_file( const char *p, unsigned source_num, const char *filename ); }; gpgpu_context* GPGPU_Context(); diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 2a6d930..38597e0 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -81,7 +81,7 @@ void ptx_reg_options(option_parser_t opp) "0"); } -void print_ptx_file( const char *p, unsigned source_num, const char *filename ) +void gpgpu_context::print_ptx_file( const char *p, unsigned source_num, const char *filename ) { printf("\nGPGPU-Sim PTX: file _%u.ptx contents:\n\n", source_num ); char *s = strdup(p); -- cgit v1.3 From c146d3707216f235a1dde3bd978239fd55b111ed Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 18:17:45 -0400 Subject: Move all ptx_parser inside gpgpu_context Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 4 ++++ src/cuda-sim/ptx_loader.cc | 9 ++++----- src/cuda-sim/ptx_parser.cc | 18 +++++++++--------- src/cuda-sim/ptx_parser.h | 1 - 4 files changed, 17 insertions(+), 15 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index dbc3140..b586e47 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -2,17 +2,20 @@ #define __gpgpu_context_h__ #include "cuda_api_object.h" #include "../src/cuda-sim/ptx_loader.h" +#include "../src/cuda-sim/ptx_parser.h" class gpgpu_context { public: gpgpu_context() { api = new cuda_runtime_api(); ptxinfo = new ptxinfo_data(); + ptx_parser = new ptx_recognizer(); } // global list // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; + ptx_recognizer* ptx_parser; // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); @@ -20,6 +23,7 @@ class gpgpu_context { void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version); void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); void print_ptx_file( const char *p, unsigned source_num, const char *filename ); + class symbol_table* init_parser(const char*); }; gpgpu_context* GPGPU_Context(); diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 38597e0..921d1e6 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -170,10 +170,9 @@ symbol_table *gpgpu_context::gpgpu_ptx_sim_load_ptx_from_string( const char *p, fclose(fp); } symbol_table *symtab=init_parser(buf); - ptx_recognizer recognizer; - ptx_lex_init(&(recognizer.scanner)); - ptx__scan_string(p, recognizer.scanner); - int errors = ptx_parse (recognizer.scanner, &recognizer); + ptx_lex_init(&(ptx_parser->scanner)); + ptx__scan_string(p, ptx_parser->scanner); + int errors = ptx_parse (ptx_parser->scanner, ptx_parser); if ( errors ) { char fname[1024]; snprintf(fname,1024,"_ptx_errors_XXXXXX"); @@ -186,7 +185,7 @@ symbol_table *gpgpu_context::gpgpu_ptx_sim_load_ptx_from_string( const char *p, abort(); exit(40); } - ptx_lex_destroy(recognizer.scanner); + ptx_lex_destroy(ptx_parser->scanner); if ( g_debug_execution >= 100 ) print_ptx_file(p,source_num,buf); diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 897beaf..7080ee9 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -27,6 +27,7 @@ #include "ptx_parser.h" #include "ptx_ir.h" +#include "../../libcuda/gpgpu_context.h" typedef void * yyscan_t; #include "ptx.tab.h" @@ -122,7 +123,7 @@ void ptx_recognizer::init_instruction_state() init_directive_state(); } -symbol_table *init_parser( const char *ptx_filename ) +symbol_table * gpgpu_context::init_parser( const char *ptx_filename ) { g_filename = strdup(ptx_filename); if (g_global_allfiles_symbol_table == NULL) { @@ -151,17 +152,16 @@ symbol_table *init_parser( const char *ptx_filename ) g_ptx_token_decode[generic_space] = "generic_space"; g_ptx_token_decode[instruction_space] = "instruction_space"; - ptx_recognizer recognizer; - ptx_lex_init(&(recognizer.scanner)); - recognizer.init_directive_state(); - recognizer.init_instruction_state(); + ptx_lex_init(&(ptx_parser->scanner)); + ptx_parser->init_directive_state(); + ptx_parser->init_instruction_state(); FILE *ptx_in; ptx_in = fopen(ptx_filename, "r"); - ptx_set_in(ptx_in, recognizer.scanner); - ptx_parse(recognizer.scanner, &recognizer); - ptx_in = ptx_get_in(recognizer.scanner); - ptx_lex_destroy(recognizer.scanner); + ptx_set_in(ptx_in, ptx_parser->scanner); + ptx_parse(ptx_parser->scanner, ptx_parser); + ptx_in = ptx_get_in(ptx_parser->scanner); + ptx_lex_destroy(ptx_parser->scanner); fclose(ptx_in); return g_global_symbol_table; } diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index c52fe20..5293c2e 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -34,7 +34,6 @@ extern const char *g_filename; extern int g_error_detected; #ifdef __cplusplus -class symbol_table* init_parser(const char*); const class ptx_instruction *ptx_instruction_lookup( const char *filename, unsigned linenumber ); #endif -- cgit v1.3 From 5530b51c7c572d25717ad7f5d81196011dfcb540 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 12 Jun 2019 18:49:58 -0400 Subject: Move symbol table globals Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/cuda-sim/ptx_parser.cc | 10 +++------- src/cuda-sim/ptx_parser.h | 7 +++++++ 3 files changed, 12 insertions(+), 7 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index b586e47..fa5d02b 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -7,11 +7,13 @@ class gpgpu_context { public: gpgpu_context() { + g_global_allfiles_symbol_table = NULL; api = new cuda_runtime_api(); ptxinfo = new ptxinfo_data(); ptx_parser = new ptx_recognizer(); } // global list + symbol_table *g_global_allfiles_symbol_table; // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 7080ee9..32e12a1 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -52,12 +52,8 @@ static bool g_debug_ir_generation=false; const char *g_filename; // the program intermediate representation... -static symbol_table *g_global_allfiles_symbol_table = NULL; -static symbol_table *g_global_symbol_table = NULL; std::map g_sym_name_to_symbol_table; -static symbol_table *g_current_symbol_table = NULL; static std::list g_instructions; -static symbol *g_last_symbol = NULL; int g_error_detected = 0; @@ -128,7 +124,7 @@ symbol_table * gpgpu_context::init_parser( const char *ptx_filename ) g_filename = strdup(ptx_filename); if (g_global_allfiles_symbol_table == NULL) { g_global_allfiles_symbol_table = new symbol_table("global_allfiles", 0, NULL); - g_global_symbol_table = g_current_symbol_table = g_global_allfiles_symbol_table; + ptx_parser->g_global_symbol_table = ptx_parser->g_current_symbol_table = g_global_allfiles_symbol_table; } /*else { g_global_symbol_table = g_current_symbol_table = new symbol_table("global",0,g_global_allfiles_symbol_table); @@ -163,7 +159,7 @@ symbol_table * gpgpu_context::init_parser( const char *ptx_filename ) ptx_in = ptx_get_in(ptx_parser->scanner); ptx_lex_destroy(ptx_parser->scanner); fclose(ptx_in); - return g_global_symbol_table; + return ptx_parser->g_global_symbol_table; } static int g_entry_point; @@ -331,7 +327,7 @@ void ptx_recognizer::set_variable_type() g_extern_spec ); } -bool check_for_duplicates( const char *identifier ) +bool ptx_recognizer::check_for_duplicates( const char *identifier ) { const symbol *s = g_current_symbol_table->lookup(identifier); return ( s != NULL ); diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index 5293c2e..cc8fe18 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -56,6 +56,9 @@ class ptx_recognizer { g_ident_add_uid = 0; g_const_alloc = 1; g_max_regs_per_thread = 0; + g_global_symbol_table = NULL; + g_current_symbol_table = NULL; + symbol *g_last_symbol = NULL; } // global list yyscan_t scanner; @@ -89,6 +92,9 @@ class ptx_recognizer { int g_ident_add_uid; unsigned g_const_alloc; unsigned g_max_regs_per_thread; + symbol_table *g_global_symbol_table; + symbol_table *g_current_symbol_table; + symbol *g_last_symbol; // member function list void init_directive_state(); @@ -154,6 +160,7 @@ class ptx_recognizer { //Jin: handle instructino group for cdp void start_inst_group(); void end_inst_group(); + bool check_for_duplicates( const char *identifier ); }; -- cgit v1.3 From f63324eda157f742d06c889b1be73717771e60a5 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Tue, 18 Jun 2019 22:34:07 -0400 Subject: Some stuff to get mnist to run with CUDA 9.1 --- libcuda/cuda_runtime_api.cc | 49 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 49 insertions(+) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index f46d9f1..8ba2b0f 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1694,6 +1694,12 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic break; case 88: case 89: + case 90: + case 91: + case 92: + case 93: + case 94: + case 95: *value= 0; break; default: @@ -3246,6 +3252,49 @@ __host__ cudaError_t CUDARTAPI cudaDeviceSetLimit(enum cudaLimit limit, size_t v #endif +/** + * \brief Set attributes for a given function + * + * This function sets the attributes of a function specified via \p entry. + * The parameter \p entry must be a pointer to a function that executes + * on the device. The parameter specified by \p entry must be declared as a \p __global__ + * function. The enumeration defined by \p attr is set to the value defined by \p value + * If the specified function does not exist, then ::cudaErrorInvalidDeviceFunction is returned. + * If the specified attribute cannot be written, or if the value is incorrect, + * then ::cudaErrorInvalidValue is returned. + * + * Valid values for \p attr are: + * ::cuFuncAttrMaxDynamicSharedMem - Maximum size of dynamic shared memory per block + * ::cudaFuncAttributePreferredSharedMemoryCarveout - Preferred shared memory-L1 cache split ratio + * + * \param entry - Function to get attributes of + * \param attr - Attribute to set + * \param value - Value to set + * + * \return + * ::cudaSuccess, + * ::cudaErrorInitializationError, + * ::cudaErrorInvalidDeviceFunction, + * ::cudaErrorInvalidValue + * \notefnerr + * + * \ref ::cudaLaunchKernel(const T *func, dim3 gridDim, dim3 blockDim, void **args, size_t sharedMem, cudaStream_t stream) "cudaLaunchKernel (C++ API)", + * \ref ::cudaFuncSetCacheConfig(T*, enum cudaFuncCache) "cudaFuncSetCacheConfig (C++ API)", + * \ref ::cudaFuncGetAttributes(struct cudaFuncAttributes*, const void*) "cudaFuncGetAttributes (C API)", + * ::cudaSetDoubleForDevice, + * ::cudaSetDoubleForHost, + * \ref ::cudaSetupArgument(T, size_t) "cudaSetupArgument (C++ API)" + */ +cudaError_t CUDARTAPI cudaFuncSetAttribute(const void *func, enum cudaFuncAttribute attr, int value) +{ + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + printf("GPGPU-Sim PTX: Execution warning: ignoring call to \"%s ( func=%p, attr=%d, value=%d )\"\n", + __my_func__, func, attr, value ); + return g_last_cudaError = cudaSuccess; +} + cudaError_t CUDARTAPI cudaGLSetGLDevice(int device) { if(g_debug_execution >= 3){ -- cgit v1.3 From 48830687ede62b3acaebeba93633255b4d8ec9c8 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 19 Jun 2019 10:22:41 -0400 Subject: Fixing 4.2 build --- libcuda/cuda_runtime_api.cc | 3 +++ 1 file changed, 3 insertions(+) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 8ba2b0f..015fbc0 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -3252,6 +3252,8 @@ __host__ cudaError_t CUDARTAPI cudaDeviceSetLimit(enum cudaLimit limit, size_t v #endif + +#if CUDART_VERSION >= 9000 /** * \brief Set attributes for a given function * @@ -3294,6 +3296,7 @@ cudaError_t CUDARTAPI cudaFuncSetAttribute(const void *func, enum cudaFuncAttrib __my_func__, func, attr, value ); return g_last_cudaError = cudaSuccess; } +#endif cudaError_t CUDARTAPI cudaGLSetGLDevice(int device) { -- cgit v1.3 From cb678c3670de4a435a3260ed80dc476da3860082 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 1 Jul 2019 16:23:33 -0400 Subject: Move g_debug_ir_generation and GPGPUSim_Init Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 674 ++++++++++++++++++++++------------------ libcuda/gpgpu_context.h | 2 + libopencl/opencl_runtime_api.cc | 4 +- src/cuda-sim/ptx_parser.cc | 3 +- src/cuda-sim/ptx_parser.h | 3 + src/gpgpusim_entrypoint.cc | 5 +- src/gpgpusim_entrypoint.h | 1 - 7 files changed, 375 insertions(+), 317 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 61f8415..a34727f 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -197,7 +197,7 @@ void register_ptx_function( const char *name, function_info *impl ) # endif #endif -struct _cuda_device_id *GPGPUSim_Init() +struct _cuda_device_id *gpgpu_context::GPGPUSim_Init() { //static _cuda_device_id *the_device = NULL; _cuda_device_id *the_device = GPGPUsim_ctx_ptr()->the_cude_device; @@ -255,9 +255,10 @@ struct _cuda_device_id *GPGPUSim_Init() static CUctx_st* GPGPUSim_Context() { //static CUctx_st *the_context = NULL; + gpgpu_context *cur_ctx = GPGPU_Context(); CUctx_st *the_context = GPGPUsim_ctx_ptr()->the_context; if( the_context == NULL ) { - _cuda_device_id *the_gpu = GPGPUSim_Init(); + _cuda_device_id *the_gpu = cur_ctx->GPGPUSim_Init(); GPGPUsim_ctx_ptr()->the_context = new CUctx_st(the_gpu); the_context = GPGPUsim_ctx_ptr()->the_context; } @@ -496,7 +497,7 @@ cudaError_t cudaSetDeviceInternal(int device, gpgpu_context* gpgpu_ctx = NULL) announce_call(__my_func__); } //set the active device to run cuda - if ( device <= GPGPUSim_Init()->num_devices() ) { + if ( device <= ctx->GPGPUSim_Init()->num_devices() ) { ctx->api->g_active_device = device; return g_last_cudaError = cudaSuccess; } else { @@ -519,6 +520,55 @@ cudaError_t cudaGetDeviceInternal(int *device, gpgpu_context* gpgpu_ctx = NULL) return g_last_cudaError = cudaSuccess; } +__host__ cudaError_t CUDARTAPI cudaDeviceGetLimitInternal( size_t* pValue, cudaLimit limit, gpgpu_context* gpgpu_ctx = NULL ) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + _cuda_device_id *dev = ctx->GPGPUSim_Init(); + const struct cudaDeviceProp *prop = dev->get_prop(); + const gpgpu_sim_config& config=dev->get_gpgpu()->get_config(); + switch(limit) { + case 0: // cudaLimitStackSize + *pValue=config.stack_limit(); + break; + case 2: // cudaLimitMallocHeapSize + *pValue=config.heap_limit(); + break; +#if (CUDART_VERSION > 5050) + case 3: // cudaLimitDevRuntimeSyncDepth + if(prop->major > 2){ + *pValue=config.sync_depth_limit(); + break; + } + else{ + printf("ERROR:Limit %s is not supported on this architecture \n",limit); + abort(); + } + case 4: // cudaLimitDevRuntimePendingLaunchCount + if(prop->major > 2){ + *pValue=config.pending_launch_count_limit(); + break; + } + else{ + printf("ERROR:Limit %s is not supported on this architecture \n",limit); + abort(); + } +#endif + default: + printf("ERROR:Limit %s unimplemented \n",limit); + abort(); + } + return g_last_cudaError = cudaSuccess; + +} + void** cudaRegisterFatBinaryInternal( void *fatCubin, gpgpu_context* gpgpu_ctx = NULL) { @@ -739,6 +789,260 @@ cudaError_t cudaConfigureCallInternal(dim3 gridDim, dim3 blockDim, size_t shared return g_last_cudaError = cudaSuccess; } +__host__ cudaError_t CUDARTAPI cudaGetDeviceCountInternal(int *count, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + _cuda_device_id *dev = ctx->GPGPUSim_Init(); + *count = dev->num_devices(); + return g_last_cudaError = cudaSuccess; +} + +__host__ cudaError_t CUDARTAPI cudaGetDevicePropertiesInternal(struct cudaDeviceProp *prop, int device, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + _cuda_device_id *dev = ctx->GPGPUSim_Init(); + if (device <= dev->num_devices() ) { + *prop= *dev->get_prop(); + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorInvalidDevice; + } +} + +#if (CUDART_VERSION > 5000) +__host__ cudaError_t CUDARTAPI cudaDeviceGetAttributeInternal(int *value, enum cudaDeviceAttr attr, int device, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + const struct cudaDeviceProp *prop; + _cuda_device_id *dev = ctx->GPGPUSim_Init(); + if (device <= dev->num_devices() ) { + prop = dev->get_prop(); + switch (attr) { + case 1: + *value= prop->maxThreadsDim[0] * prop->maxThreadsDim[1] * prop->maxThreadsDim[2] * prop->maxGridSize[0] * prop->maxGridSize[1] * prop->maxGridSize[2]; + break; + case 2: + *value= prop->maxThreadsDim[0]; + break; + case 3: + *value= prop->maxThreadsDim[1]; + break; + case 4: + *value= prop->maxThreadsDim[2]; + break; + case 5: + *value= prop->maxGridSize[0]; + break; + case 6: + *value= prop->maxGridSize[1]; + break; + case 7: + *value= prop->maxGridSize[2]; + break; + case 8: + *value= prop->sharedMemPerBlock; + break; + case 9: + *value= prop->totalConstMem; + break; + case 10: + *value= prop->warpSize; + break; + case 11: + *value= 16;//dummy value + break; + case 12: + *value= prop->regsPerBlock; + break; + case 13: + *value= 1480000;//for 1080ti + break; + case 14: + *value= prop->textureAlignment ; + break; + case 15: + *value = 0; + break; + case 16: + *value= prop->multiProcessorCount ; + break; + case 17: + case 18: + case 19: + *value = 0; + break; + case 21: + case 22: + case 23: + case 24: + case 25: + case 26: + case 27: + case 28: + case 42: + case 45: + case 46: + case 47: + case 48: + case 49: + case 52: + case 53: + case 55: + case 56: + case 57: + case 58: + case 59: + case 60: + case 61: + case 62: + case 63: + case 64: + case 66: + case 67: + case 69: + case 70: + case 71: + case 73: + case 74: + case 77: + *value = 1000;//dummy value + break; + case 29: + case 43: + case 54: + case 65: + case 68: + case 72: + *value = 10;//dummy value + break; + case 30: + case 51: + *value = 128;//dummy value + break; + case 31: + *value = 1; + break; + case 32: + *value = 0; + break; + case 33: + case 50: + *value = 0;//dummy value + break; + case 34: + *value= 0; + break; + case 35: + *value = 0; + break; + case 36: + *value = 1250000;//CK value for 1080ti + break; + case 37: + *value = 352;//value for 1080ti + break; + case 38: + *value = 3000000;//value for 1080ti + break; + case 39: + *value= dev->get_gpgpu()->threads_per_core(); + break; + case 40: + *value= 0; + break; + case 41: + *value= 0; + break; + case 75://cudaDevAttrComputeCapabilityMajor + *value= prop->major ; + break; + case 76://cudaDevAttrComputeCapabilityMinor + *value= prop->minor ; + break; + case 78: + *value= 0 ; //TODO: as of now, we dont support stream priorities. + break; + case 79: + *value= 0; + break; + case 80: + *value= 0; + break; + #if (CUDART_VERSION > 5050) + case 81: + *value= prop->sharedMemPerMultiprocessor; + break; + case 82: + *value= prop->regsPerMultiprocessor; + break; + #endif + case 83: + case 84: + case 85: + case 86: + *value= 0; + break; + case 87: + *value= 4;//dummy value + break; + case 88: + case 89: + case 90: + case 91: + case 95: + *value= 0; + break; + default: + printf("ERROR: Attribute number %d unimplemented \n",attr); + abort(); + } + return g_last_cudaError = cudaSuccess; + } else { + return g_last_cudaError = cudaErrorInvalidDevice; + } +} +#endif + +__host__ cudaError_t CUDARTAPI cudaChooseDeviceInternal(int *device, const struct cudaDeviceProp *prop, gpgpu_context* gpgpu_ctx = NULL) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + _cuda_device_id *dev = ctx->GPGPUSim_Init(); + *device = dev->get_id(); + return g_last_cudaError = cudaSuccess; +} + cudaError_t cudaSetupArgumentInternal(const void *arg, size_t size, size_t offset, gpgpu_context* gpgpu_ctx = NULL) { gpgpu_context *ctx; @@ -1058,6 +1362,52 @@ cudaError_t cudaHostAllocInternal(void **pHost, size_t bytes, unsigned int flag #endif +size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr, gpgpu_context *ctx) { + _cuda_device_id *dev = ctx->GPGPUSim_Init(); + struct cudaDeviceProp prop; + + prop = *dev->get_prop(); + + size_t max = prop.maxThreadsPerBlock; + + if ((prop.regsPerBlock / attr->numRegs) < max) + max = prop.regsPerBlock / attr->numRegs; + + return max; +} + +cudaError_t CUDARTAPI cudaFuncGetAttributesInternal(struct cudaFuncAttributes *attr, const char *hostFun, gpgpu_context* gpgpu_ctx = NULL ) +{ + gpgpu_context *ctx; + if (gpgpu_ctx){ + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if(g_debug_execution >= 3){ + announce_call(__my_func__); + } + CUctx_st *context = GPGPUSim_Context(); + function_info *entry = context->get_kernel(hostFun); + if( entry ) { + const struct gpgpu_ptx_sim_info *kinfo = entry->get_kernel_info(); + attr->sharedSizeBytes = kinfo->smem; + attr->constSizeBytes = kinfo->cmem; + attr->localSizeBytes = kinfo->lmem; + attr->numRegs = kinfo->regs; + if(kinfo->maxthreads > 0) + attr->maxThreadsPerBlock = kinfo->maxthreads; + else + attr->maxThreadsPerBlock = getMaxThreadsPerBlock(attr, ctx); +#if CUDART_VERSION >= 3000 + attr->ptxVersion = kinfo->ptx_version; + attr->binaryVersion = kinfo->sm_target; +#endif + } + return g_last_cudaError = cudaSuccess; +} + + /******************************************************************************* * * * * @@ -1503,235 +1853,24 @@ __host__ cudaError_t CUDARTAPI cudaGetSymbolSize(size_t *size, const char *symbo *******************************************************************************/ __host__ cudaError_t CUDARTAPI cudaGetDeviceCount(int *count) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - _cuda_device_id *dev = GPGPUSim_Init(); - *count = dev->num_devices(); - return g_last_cudaError = cudaSuccess; + return cudaGetDeviceCountInternal(count); } - -__host__ cudaError_t CUDARTAPI cudaGetDeviceProperties(struct cudaDeviceProp *prop, int device) -{ - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - _cuda_device_id *dev = GPGPUSim_Init(); - if (device <= dev->num_devices() ) { - *prop= *dev->get_prop(); - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorInvalidDevice; - } + +__host__ cudaError_t CUDARTAPI cudaGetDeviceProperties(struct cudaDeviceProp *prop, int device) +{ + return cudaGetDevicePropertiesInternal(prop, device); } #if (CUDART_VERSION > 5000) __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDeviceAttr attr, int device) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - const struct cudaDeviceProp *prop; - _cuda_device_id *dev = GPGPUSim_Init(); - if (device <= dev->num_devices() ) { - prop = dev->get_prop(); - switch (attr) { - case 1: - *value= prop->maxThreadsDim[0] * prop->maxThreadsDim[1] * prop->maxThreadsDim[2] * prop->maxGridSize[0] * prop->maxGridSize[1] * prop->maxGridSize[2]; - break; - case 2: - *value= prop->maxThreadsDim[0]; - break; - case 3: - *value= prop->maxThreadsDim[1]; - break; - case 4: - *value= prop->maxThreadsDim[2]; - break; - case 5: - *value= prop->maxGridSize[0]; - break; - case 6: - *value= prop->maxGridSize[1]; - break; - case 7: - *value= prop->maxGridSize[2]; - break; - case 8: - *value= prop->sharedMemPerBlock; - break; - case 9: - *value= prop->totalConstMem; - break; - case 10: - *value= prop->warpSize; - break; - case 11: - *value= 16;//dummy value - break; - case 12: - *value= prop->regsPerBlock; - break; - case 13: - *value= 1480000;//for 1080ti - break; - case 14: - *value= prop->textureAlignment ; - break; - case 15: - *value = 0; - break; - case 16: - *value= prop->multiProcessorCount ; - break; - case 17: - case 18: - case 19: - *value = 0; - break; - case 21: - case 22: - case 23: - case 24: - case 25: - case 26: - case 27: - case 28: - case 42: - case 45: - case 46: - case 47: - case 48: - case 49: - case 52: - case 53: - case 55: - case 56: - case 57: - case 58: - case 59: - case 60: - case 61: - case 62: - case 63: - case 64: - case 66: - case 67: - case 69: - case 70: - case 71: - case 73: - case 74: - case 77: - *value = 1000;//dummy value - break; - case 29: - case 43: - case 54: - case 65: - case 68: - case 72: - *value = 10;//dummy value - break; - case 30: - case 51: - *value = 128;//dummy value - break; - case 31: - *value = 1; - break; - case 32: - *value = 0; - break; - case 33: - case 50: - *value = 0;//dummy value - break; - case 34: - *value= 0; - break; - case 35: - *value = 0; - break; - case 36: - *value = 1250000;//CK value for 1080ti - break; - case 37: - *value = 352;//value for 1080ti - break; - case 38: - *value = 3000000;//value for 1080ti - break; - case 39: - *value= dev->get_gpgpu()->threads_per_core(); - break; - case 40: - *value= 0; - break; - case 41: - *value= 0; - break; - case 75://cudaDevAttrComputeCapabilityMajor - *value= prop->major ; - break; - case 76://cudaDevAttrComputeCapabilityMinor - *value= prop->minor ; - break; - case 78: - *value= 0 ; //TODO: as of now, we dont support stream priorities. - break; - case 79: - *value= 0; - break; - case 80: - *value= 0; - break; - #if (CUDART_VERSION > 5050) - case 81: - *value= prop->sharedMemPerMultiprocessor; - break; - case 82: - *value= prop->regsPerMultiprocessor; - break; - #endif - case 83: - case 84: - case 85: - case 86: - *value= 0; - break; - case 87: - *value= 4;//dummy value - break; - case 88: - case 89: - case 90: - case 91: - case 92: - case 93: - case 94: - case 95: - *value= 0; - break; - default: - printf("ERROR: Attribute number %d unimplemented \n",attr); - abort(); - } - return g_last_cudaError = cudaSuccess; - } else { - return g_last_cudaError = cudaErrorInvalidDevice; - } + return cudaDeviceGetAttributeInternal(value, attr, device); } #endif __host__ cudaError_t CUDARTAPI cudaChooseDevice(int *device, const struct cudaDeviceProp *prop) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - _cuda_device_id *dev = GPGPUSim_Init(); - *device = dev->get_id(); - return g_last_cudaError = cudaSuccess; + return cudaChooseDeviceInternal(device, prop); } __host__ cudaError_t CUDARTAPI cudaSetDevice(int device) @@ -1744,47 +1883,9 @@ __host__ cudaError_t CUDARTAPI cudaGetDevice(int *device) return cudaGetDeviceInternal(device); } -__host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit limit ) +__host__ cudaError_t CUDARTAPI cudaDeviceGetLimit( size_t* pValue, cudaLimit limit ) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - _cuda_device_id *dev = GPGPUSim_Init(); - const struct cudaDeviceProp *prop = dev->get_prop(); - const gpgpu_sim_config& config=dev->get_gpgpu()->get_config(); - switch(limit) { - case 0: // cudaLimitStackSize - *pValue=config.stack_limit(); - break; - case 2: // cudaLimitMallocHeapSize - *pValue=config.heap_limit(); - break; -#if (CUDART_VERSION > 5050) - case 3: // cudaLimitDevRuntimeSyncDepth - if(prop->major > 2){ - *pValue=config.sync_depth_limit(); - break; - } - else{ - printf("ERROR:Limit %s is not supported on this architecture \n",limit); - abort(); - } - case 4: // cudaLimitDevRuntimePendingLaunchCount - if(prop->major > 2){ - *pValue=config.pending_launch_count_limit(); - break; - } - else{ - printf("ERROR:Limit %s is not supported on this architecture \n",limit); - abort(); - } -#endif - default: - printf("ERROR:Limit %s unimplemented \n",limit); - abort(); - } - return g_last_cudaError = cudaSuccess; - + return cudaDeviceGetLimitInternal( pValue, limit ); } __host__ cudaError_t CUDARTAPI cudaStreamGetPriority ( cudaStream_t hStream, int* priority ) @@ -3167,58 +3268,9 @@ cudaError_t CUDARTAPI cudaSetDeviceFlags( int flags ) } } -size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr) { - _cuda_device_id *dev = GPGPUSim_Init(); - struct cudaDeviceProp prop; - - prop = *dev->get_prop(); - - size_t max = prop.maxThreadsPerBlock; - - if ((prop.regsPerBlock / attr->numRegs) < max) - max = prop.regsPerBlock / attr->numRegs; - - return max; -} - cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, const char *hostFun ) { - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUctx_st *context = GPGPUSim_Context(); - function_info *entry = context->get_kernel(hostFun); - if( entry ) { - const struct gpgpu_ptx_sim_info *kinfo = entry->get_kernel_info(); - attr->sharedSizeBytes = kinfo->smem; - attr->constSizeBytes = kinfo->cmem; - attr->localSizeBytes = kinfo->lmem; - attr->numRegs = kinfo->regs; - if(kinfo->maxthreads > 0) - attr->maxThreadsPerBlock = kinfo->maxthreads; - else - attr->maxThreadsPerBlock = getMaxThreadsPerBlock(attr); -#if CUDART_VERSION >= 3000 - attr->ptxVersion = kinfo->ptx_version; - attr->binaryVersion = kinfo->sm_target; -#endif - } - return g_last_cudaError = cudaSuccess; -} - -cudaError_t CUDARTAPI cudaEventCreateWithFlags(cudaEvent_t *event, int flags) -{ - if(g_debug_execution >= 3){ - announce_call(__my_func__); - } - CUevent_st *e = new CUevent_st(flags==cudaEventBlockingSync); - g_timer_events[e->get_uid()] = e; -#if CUDART_VERSION >= 3000 - *event = e; -#else - *event = e->get_uid(); -#endif - return g_last_cudaError = cudaSuccess; + return cudaFuncGetAttributesInternal(attr, hostFun ); } cudaError_t CUDARTAPI cudaDriverGetVersion(int *driverVersion) diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index fa5d02b..dd8b51d 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -26,6 +26,8 @@ class gpgpu_context { void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20, int no_of_ptx=0 ); void print_ptx_file( const char *p, unsigned source_num, const char *filename ); class symbol_table* init_parser(const char*); + class gpgpu_sim *gpgpu_ptx_sim_init_perf(); + struct _cuda_device_id *GPGPUSim_Init(); }; gpgpu_context* GPGPU_Context(); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index 50a02fa..e91e2e0 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -648,7 +648,9 @@ class _cl_device_id *GPGPUSim_Init() { static _cl_device_id *the_device = NULL; if( !the_device ) { - gpgpu_sim *the_gpu = gpgpu_ptx_sim_init_perf(); + gpgpu_context *ctx; + ctx = GPGPU_Context(); + gpgpu_sim *the_gpu = ctx->gpgpu_ptx_sim_init_perf(); the_device = new _cl_device_id(the_gpu); } start_sim_thread(2); diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 2872b84..9094ec3 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -48,7 +48,6 @@ void set_ptx_warp_size(const struct core_config * warp_size) g_shader_core_config=warp_size; } -static bool g_debug_ir_generation=false; const char *g_filename; // the program intermediate representation... @@ -70,7 +69,7 @@ const char *decode_token( int type ) return g_ptx_token_decode[type].c_str(); } -void read_parser_environment_variables() +void ptx_recognizer::read_parser_environment_variables() { g_filename = getenv("PTX_SIM_KERNELFILE"); char *dbg_level = getenv("PTX_SIM_DEBUG"); diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index 25c01fe..bc9a872 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -62,6 +62,7 @@ class ptx_recognizer { g_error_detected = 0; g_entry_func_param_index=0; g_func_info = NULL; + g_debug_ir_generation=false; } // global list yyscan_t scanner; @@ -103,6 +104,7 @@ class ptx_recognizer { unsigned g_entry_func_param_index; function_info *g_func_info; operand_info g_return_var; + bool g_debug_ir_generation; // member function list void init_directive_state(); @@ -169,6 +171,7 @@ class ptx_recognizer { void start_inst_group(); void end_inst_group(); bool check_for_duplicates( const char *identifier ); + void read_parser_environment_variables(); }; diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index b09674a..476a9d4 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -35,6 +35,7 @@ #include "gpgpu-sim/gpu-sim.h" #include "gpgpu-sim/icnt_wrapper.h" #include "stream_manager.h" +#include "../libcuda/gpgpu_context.h" #define MAX(a,b) (((a)>(b))?(a):(b)) @@ -208,12 +209,12 @@ void exit_simulation() extern bool g_cuda_launch_blocking; -gpgpu_sim *gpgpu_ptx_sim_init_perf() +gpgpu_sim *gpgpu_context::gpgpu_ptx_sim_init_perf() { srand(1); print_splash(); read_sim_environment_variables(); - read_parser_environment_variables(); + ptx_parser->read_parser_environment_variables(); option_parser_t opp = option_parser_create(); ptx_reg_options(opp); diff --git a/src/gpgpusim_entrypoint.h b/src/gpgpusim_entrypoint.h index 9ce7fef..a443151 100644 --- a/src/gpgpusim_entrypoint.h +++ b/src/gpgpusim_entrypoint.h @@ -74,7 +74,6 @@ struct GPGPUsim_ctx { }; -class gpgpu_sim *gpgpu_ptx_sim_init_perf(); void start_sim_thread(int api); class gpgpu_sim* g_the_gpu(); -- cgit v1.3 From 3600ed5c59adafe40840524d19f622aa25e60dd6 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 2 Jul 2019 00:22:15 -0400 Subject: Move ptxinfo_addinfo Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- src/cuda-sim/ptx_loader.h | 2 +- src/cuda-sim/ptxinfo.y | 5 ++--- 3 files changed, 4 insertions(+), 5 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index a34727f..eb645ce 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -274,7 +274,7 @@ gpgpu_context* GPGPU_Context() return gpgpu_ctx; } - void ptxinfo_addinfo() + void ptxinfo_data::ptxinfo_addinfo() { if(!get_ptxinfo_kname()){ /* This info is not per kernel (since CUDA 5.0 some info (e.g. gmem, and cmem) is added at the beginning for the whole binary ) */ diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index ee09e16..77f27c8 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -37,7 +37,7 @@ class ptxinfo_data{ char linebuf[PTXINFO_LINEBUF_SIZE]; unsigned col; const char *g_ptxinfo_filename; - + void ptxinfo_addinfo(); }; diff --git a/src/cuda-sim/ptxinfo.y b/src/cuda-sim/ptxinfo.y index 00c81e0..b303958 100644 --- a/src/cuda-sim/ptxinfo.y +++ b/src/cuda-sim/ptxinfo.y @@ -29,7 +29,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %{ typedef void * yyscan_t; -class ptxinfo_data; +#include "ptx_loader.h" %} %define api.pure full @@ -79,7 +79,6 @@ class ptxinfo_data; static unsigned g_system; int ptxinfo_lex(YYSTYPE * yylval_param, yyscan_t yyscanner, ptxinfo_data* ptxinfo); void yyerror(yyscan_t yyscanner, ptxinfo_data* ptxinfo, const char* msg); - void ptxinfo_addinfo(); void ptxinfo_function(const char *fname ); void ptxinfo_regs( unsigned nregs ); void ptxinfo_lmem( unsigned declared, unsigned system ); @@ -104,7 +103,7 @@ line: HEADER INFO COLON line_info ; line_info: function_name - | function_info { ptxinfo_addinfo(); } + | function_info { ptxinfo->ptxinfo_addinfo(); } | gmem_info ; -- cgit v1.3 From 71eaa4e7b4aba49efef39391f3e1bb3a7e4e1529 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 3 Jul 2019 02:44:08 -0400 Subject: Connect gpgpu_context and GPGPUsim_ctx Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 3 +++ src/gpgpusim_entrypoint.cc | 6 +++--- src/gpgpusim_entrypoint.h | 10 ++++++---- 3 files changed, 12 insertions(+), 7 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index dd8b51d..69341be 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -3,6 +3,7 @@ #include "cuda_api_object.h" #include "../src/cuda-sim/ptx_loader.h" #include "../src/cuda-sim/ptx_parser.h" +#include "../src/gpgpusim_entrypoint.h" class gpgpu_context { public: @@ -11,6 +12,7 @@ class gpgpu_context { api = new cuda_runtime_api(); ptxinfo = new ptxinfo_data(); ptx_parser = new ptx_recognizer(); + the_gpgpusim = new GPGPUsim_ctx(this); } // global list symbol_table *g_global_allfiles_symbol_table; @@ -18,6 +20,7 @@ class gpgpu_context { cuda_runtime_api* api; ptxinfo_data* ptxinfo; ptx_recognizer* ptx_parser; + GPGPUsim_ctx* the_gpgpusim; // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index 476a9d4..c91a9d1 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -43,11 +43,11 @@ static int sg_argc = 3; static const char *sg_argv[] = {"", "-config","gpgpusim.config"}; -struct GPGPUsim_ctx* the_gpgpusim = NULL; +GPGPUsim_ctx* the_gpgpusim = NULL; -struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(){ +GPGPUsim_ctx* GPGPUsim_ctx_ptr(){ if(the_gpgpusim == NULL) - the_gpgpusim = new GPGPUsim_ctx(); + the_gpgpusim = GPGPU_Context()->the_gpgpusim; return the_gpgpusim; } diff --git a/src/gpgpusim_entrypoint.h b/src/gpgpusim_entrypoint.h index a443151..dfb82d0 100644 --- a/src/gpgpusim_entrypoint.h +++ b/src/gpgpusim_entrypoint.h @@ -34,11 +34,11 @@ #include //extern time_t g_simulation_starttime; +class gpgpu_context; - -struct GPGPUsim_ctx { - - GPGPUsim_ctx() { +class GPGPUsim_ctx { + public: + GPGPUsim_ctx(gpgpu_context* ctx) { g_sim_active = false; g_sim_done = true; break_limit = false; @@ -49,6 +49,7 @@ struct GPGPUsim_ctx { g_stream_manager=NULL; the_cude_device=NULL; the_context=NULL; + gpgpu_ctx = ctx; } //struct gpgpu_ptx_sim_arg *grid_params; @@ -65,6 +66,7 @@ struct GPGPUsim_ctx { struct _cuda_device_id *the_cude_device; struct CUctx_st* the_context; + gpgpu_context* gpgpu_ctx; pthread_mutex_t g_sim_lock; -- cgit v1.3 From 60c38b0f77378eb111e88f632702d19dc3746cc7 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Wed, 3 Jul 2019 14:54:16 -0400 Subject: Remove g_filename Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 5 +++-- src/cuda-sim/ptx.l | 5 ++--- src/cuda-sim/ptx.y | 5 ++--- src/cuda-sim/ptx_loader.h | 5 +++++ src/cuda-sim/ptx_parser.cc | 26 ++++++++++++-------------- src/cuda-sim/ptx_parser.h | 7 +++++-- src/cuda-sim/ptxinfo.l | 4 ++-- src/gpgpu-sim/gpu-sim.h | 1 + 8 files changed, 32 insertions(+), 26 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 69341be..d819559 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -10,12 +10,13 @@ class gpgpu_context { gpgpu_context() { g_global_allfiles_symbol_table = NULL; api = new cuda_runtime_api(); - ptxinfo = new ptxinfo_data(); - ptx_parser = new ptx_recognizer(); + ptxinfo = new ptxinfo_data(this); + ptx_parser = new ptx_recognizer(this); the_gpgpusim = new GPGPUsim_ctx(this); } // global list symbol_table *g_global_allfiles_symbol_table; + const char *g_filename; // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index 18952a9..3a2a839 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -40,6 +40,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include "ptx_parser.h" #include "ptx.tab.h" #include +#include "../../libcuda/gpgpu_context.h" #define LINEBUF_SIZE (4*1024) #define TC recognizer->col+=strlen(yytext); @@ -452,8 +453,6 @@ breakaddr TC; yylval->int_value = BREAKADDR_OP; return OPCODE; . TC; ptx_error(yyscanner, recognizer, (const char*)NULL); %% -extern const char *g_filename; - int ptx_error( yyscan_t yyscanner, ptx_recognizer* recognizer, const char *s ) { struct yyguts_t * yyg = (struct yyguts_t*)yyscanner; @@ -461,7 +460,7 @@ int ptx_error( yyscan_t yyscanner, ptx_recognizer* recognizer, const char *s ) recognizer->g_error_detected = 1; fflush(stdout); if( s != NULL ) - printf("%s:%u Syntax error:\n\n", g_filename, yylineno ); + printf("%s:%u Syntax error:\n\n", recognizer->gpgpu_ctx->g_filename, yylineno ); printf(" %s\n", recognizer->linebuf ); printf(" "); for( i=0; i < recognizer->col-1; i++ ) { diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index 837fbe9..a01c3c6 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -30,6 +30,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %{ typedef void * yyscan_t; class ptx_recognizer; +#include "../../libcuda/gpgpu_context.h" %} %define api.pure full @@ -628,11 +629,9 @@ address_expression: IDENTIFIER { recognizer->add_address_operand($1,0); } %% -extern const char *g_filename; - void syntax_not_implemented(yyscan_t yyscanner, ptx_recognizer* recognizer) { - printf("Parse error (%s): this syntax is not (yet) implemented:\n",g_filename); + printf("Parse error (%s): this syntax is not (yet) implemented:\n", recognizer->gpgpu_ctx->g_filename); ptx_error(yyscanner, recognizer, NULL); abort(); } diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index 77f27c8..c214b95 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -30,13 +30,18 @@ #include #define PTXINFO_LINEBUF_SIZE 1024 +class gpgpu_context; typedef void * yyscan_t; class ptxinfo_data{ public: + ptxinfo_data(gpgpu_context* ctx) { + gpgpu_ctx = ctx; + } yyscan_t scanner; char linebuf[PTXINFO_LINEBUF_SIZE]; unsigned col; const char *g_ptxinfo_filename; + class gpgpu_context* gpgpu_ctx; void ptxinfo_addinfo(); }; diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 09bc15d..05fc618 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -47,14 +47,12 @@ void ptx_recognizer::set_ptx_warp_size(const struct core_config * warp_size) g_shader_core_config=warp_size; } -const char *g_filename; - // the program intermediate representation... std::map g_sym_name_to_symbol_table; #define PTX_PARSE_DPRINTF(...) \ if( g_debug_ir_generation ) { \ - printf(" %s:%u => ",g_filename,ptx_get_lineno(scanner)); \ + printf(" %s:%u => ",gpgpu_ctx->g_filename,ptx_get_lineno(scanner)); \ printf(" (%s:%u) ", __FILE__, __LINE__); \ printf(__VA_ARGS__); \ printf("\n"); \ @@ -70,7 +68,7 @@ const char *decode_token( int type ) void ptx_recognizer::read_parser_environment_variables() { - g_filename = getenv("PTX_SIM_KERNELFILE"); + gpgpu_ctx->g_filename = getenv("PTX_SIM_KERNELFILE"); char *dbg_level = getenv("PTX_SIM_DEBUG"); if ( dbg_level && strlen(dbg_level) ) { int debug_execution=0; @@ -180,7 +178,7 @@ void ptx_recognizer::add_function_name( const char *name ) if( prior_decl ) { g_func_info->remove_args(); } - g_global_symbol_table->add_function( g_func_info, g_filename, ptx_get_lineno(scanner) ); + g_global_symbol_table->add_function( g_func_info, gpgpu_ctx->g_filename, ptx_get_lineno(scanner) ); } //Jin: handle instruction group for cdp @@ -229,7 +227,7 @@ void ptx_recognizer::parse_error_impl( const char *file, unsigned line, const ch va_end(ap); g_error_detected = 1; - printf("%s:%u: Parse error: %s (%s:%u)\n\n", g_filename, ptx_get_lineno(scanner), buf, file, line); + printf("%s:%u: Parse error: %s (%s:%u)\n\n", gpgpu_ctx->g_filename, ptx_get_lineno(scanner), buf, file, line); ptx_error(scanner, NULL); abort(); exit(1); @@ -284,12 +282,12 @@ void ptx_recognizer::add_instruction() g_wmma_options, g_scalar_type, g_space_spec, - g_filename, + gpgpu_ctx->g_filename, ptx_get_lineno(scanner), linebuf, g_shader_core_config ); g_instructions.push_back(i); - g_inst_lookup[g_filename][ptx_get_lineno(scanner)] = i; + g_inst_lookup[gpgpu_ctx->g_filename][ptx_get_lineno(scanner)] = i; init_instruction_state(); } @@ -391,7 +389,7 @@ void ptx_recognizer::add_identifier( const char *identifier, int array_dim, unsi default: break; } - g_last_symbol = g_current_symbol_table->add_variable(identifier,type,num_bits/8,g_filename,ptx_get_lineno(scanner)); + g_last_symbol = g_current_symbol_table->add_variable(identifier,type,num_bits/8,gpgpu_ctx->g_filename,ptx_get_lineno(scanner)); switch ( ti.get_memory_space().get_type() ) { case reg_space: { regnum = g_current_symbol_table->next_reg_num(); @@ -654,7 +652,7 @@ void ptx_recognizer::add_label( const char *identifier ) if ( s != NULL ) { g_label = s; } else { - g_label = g_current_symbol_table->add_variable(identifier,NULL,0,g_filename,ptx_get_lineno(scanner)); + g_label = g_current_symbol_table->add_variable(identifier,NULL,0,gpgpu_ctx->g_filename,ptx_get_lineno(scanner)); } } @@ -912,7 +910,7 @@ void ptx_recognizer::add_scalar_operand( const char *identifier ) if ( s == NULL ) { if ( g_opcode == BRA_OP || g_opcode == CALLP_OP) { // forward branch target... - s = g_current_symbol_table->add_variable(identifier,NULL,0,g_filename,ptx_get_lineno(scanner)); + s = g_current_symbol_table->add_variable(identifier,NULL,0,gpgpu_ctx->g_filename,ptx_get_lineno(scanner)); } else { std::string msg = std::string("operand \"") + identifier + "\" has no declaration."; parse_error( msg.c_str() ); @@ -926,7 +924,7 @@ void ptx_recognizer::add_neg_pred_operand( const char *identifier ) PTX_PARSE_DPRINTF("add_neg_pred_operand"); const symbol *s = g_current_symbol_table->lookup(identifier); if ( s == NULL ) { - s = g_current_symbol_table->add_variable(identifier,NULL,1,g_filename,ptx_get_lineno(scanner)); + s = g_current_symbol_table->add_variable(identifier,NULL,1,gpgpu_ctx->g_filename,ptx_get_lineno(scanner)); } operand_info op(s); op.set_neg_pred(); @@ -962,7 +960,7 @@ void ptx_recognizer::add_version_info( float ver, unsigned ext ) void ptx_recognizer::add_file( unsigned num, const char *filename ) { - if( g_filename == NULL ) { + if( gpgpu_ctx->g_filename == NULL ) { char *b = strdup(filename); char *l=b; char *n=b; @@ -978,7 +976,7 @@ void ptx_recognizer::add_file( unsigned num, const char *filename ) char *q = strtok(NULL,"."); if( q && !strcmp(q,"cu") ) { - g_filename = strdup(buf); + gpgpu_ctx->g_filename = strdup(buf); } free( b ); diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index ec300a1..e09aab1 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -30,17 +30,17 @@ #include "../abstract_hardware_model.h" #include "ptx_ir.h" -extern const char *g_filename; extern int g_error_detected; #ifdef __cplusplus const class ptx_instruction *ptx_instruction_lookup( const char *filename, unsigned linenumber ); #endif +class gpgpu_context; typedef void * yyscan_t; class ptx_recognizer { public: - ptx_recognizer() { + ptx_recognizer( gpgpu_context* ctx ) { scanner = NULL; g_size = -1; g_add_identifier_cached__identifier = NULL; @@ -63,6 +63,7 @@ class ptx_recognizer { g_entry_func_param_index=0; g_func_info = NULL; g_debug_ir_generation=false; + gpgpu_ctx = ctx; } // global list yyscan_t scanner; @@ -107,6 +108,8 @@ class ptx_recognizer { bool g_debug_ir_generation; int g_entry_point; const struct core_config *g_shader_core_config; + // backward pointer + class gpgpu_context* gpgpu_ctx; // member function list void init_directive_state(); diff --git a/src/cuda-sim/ptxinfo.l b/src/cuda-sim/ptxinfo.l index b0ada09..3a152b0 100644 --- a/src/cuda-sim/ptxinfo.l +++ b/src/cuda-sim/ptxinfo.l @@ -39,6 +39,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include "ptx_loader.h" #include "ptxinfo.tab.h" #include +#include "../../libcuda/gpgpu_context.h" #define LINEBUF_SIZE 1024 #define TC if( (yylineno == 1) && (ptxinfo->col + strlen(yytext) < LINEBUF_SIZE) ) { \ @@ -88,7 +89,6 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %% extern int g_ptxinfo_error_detected; -extern const char *g_filename; int ptxinfo_error(yyscan_t yyscanner, ptxinfo_data* ptxinfo, const char* msg) { @@ -98,7 +98,7 @@ int ptxinfo_error(yyscan_t yyscanner, ptxinfo_data* ptxinfo, const char* msg) fflush(stdout); printf("GPGPU-Sim: ERROR while parsing output of ptxas (used to capture resource usage information)\n"); if( msg != NULL ) - printf("GPGPU-Sim: %s (%s:%u) Syntax error:\n\n", g_filename, ptxinfo->g_ptxinfo_filename, yylineno ); + printf("GPGPU-Sim: %s (%s:%u) Syntax error:\n\n", ptxinfo->gpgpu_ctx->g_filename, ptxinfo->g_ptxinfo_filename, yylineno ); printf(" %s\n", ptxinfo->linebuf ); printf(" "); for( i=0; i < ptxinfo->col-1; i++ ) { diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index b1d2e45..e2c913a 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -506,6 +506,7 @@ private: void gpgpu_debug(); ///// data ///// +// backward pointer class gpgpu_context* gpgpu_ctx; class simt_core_cluster **m_cluster; -- cgit v1.3 From 7e9591a24ffbe645cfaa7a84b33a82edb9fca0bc Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 7 Jul 2019 23:56:21 -0400 Subject: g_keep_intermediate_files Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 1 + libopencl/opencl_runtime_api.cc | 2 +- src/cuda-sim/ptx_ir.h | 3 --- src/cuda-sim/ptx_loader.cc | 7 +++---- src/cuda-sim/ptx_loader.h | 3 ++- 5 files changed, 7 insertions(+), 9 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index d819559..a8b60f4 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -32,6 +32,7 @@ class gpgpu_context { class symbol_table* init_parser(const char*); class gpgpu_sim *gpgpu_ptx_sim_init_perf(); struct _cuda_device_id *GPGPUSim_Init(); + void ptx_reg_options(option_parser_t opp); }; gpgpu_context* GPGPU_Context(); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index e91e2e0..03ec80c 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -537,7 +537,7 @@ void _cl_program::Build(const char *options) exit(1); } } - if( !g_keep_intermediate_files ) { + if( !ctx->ptxinfo->g_keep_intermediate_files ) { // clean up files... snprintf(commandline,1024,"rm -f %s", cl_fname ); int result = system(commandline); diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 1af85de..fd869c6 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1575,11 +1575,8 @@ struct textureInfo { extern std::map g_sym_name_to_symbol_table; -extern bool g_keep_intermediate_files; - void gpgpu_ptx_assemble( std::string kname, void *kinfo ); #include "../option_parser.h" -void ptx_reg_options(option_parser_t opp); unsigned ptx_kernel_shmem_size( void *kernel_impl ); unsigned ptx_kernel_nregs( void *kernel_impl ); diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 921d1e6..e0d9a11 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -58,17 +58,16 @@ extern int ptxinfo_lex_destroy(yyscan_t scanner); static bool g_save_embedded_ptx; static int g_occupancy_sm_number; -bool g_keep_intermediate_files; bool m_ptx_save_converted_ptxplus; -bool keep_intermediate_files() {return g_keep_intermediate_files;} +bool ptxinfo_data::keep_intermediate_files() {return g_keep_intermediate_files;} -void ptx_reg_options(option_parser_t opp) +void gpgpu_context::ptx_reg_options(option_parser_t opp) { option_parser_register(opp, "-save_embedded_ptx", OPT_BOOL, &g_save_embedded_ptx, "saves ptx files embedded in binary as .ptx", "0"); - option_parser_register(opp, "-keep", OPT_BOOL, &g_keep_intermediate_files, + option_parser_register(opp, "-keep", OPT_BOOL, &(ptxinfo->g_keep_intermediate_files), "keep intermediate files created by GPGPU-Sim when interfacing with external programs", "0"); option_parser_register(opp, "-gpgpu_ptx_save_converted_ptxplus", OPT_BOOL, diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index c214b95..aa3e2f3 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -42,7 +42,9 @@ class ptxinfo_data{ unsigned col; const char *g_ptxinfo_filename; class gpgpu_context* gpgpu_ctx; + bool g_keep_intermediate_files; void ptxinfo_addinfo(); + bool keep_intermediate_files(); }; @@ -50,6 +52,5 @@ extern bool g_override_embedded_ptx; extern int no_of_ptx; //counter to track number of ptx files to be extracted in an application. char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); -bool keep_intermediate_files(); #endif -- cgit v1.3 From b3e786e3d8d720217f36a214e9b5be9a19ab9dd2 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 8 Jul 2019 12:37:52 -0400 Subject: Move opcode_latency_int thus pass gpgpu_context into many classes Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 3 +++ src/cuda-sim/cuda-sim.cc | 7 ++++--- src/cuda-sim/cuda-sim.h | 11 +++++++++-- src/cuda-sim/ptx_ir.cc | 4 +++- src/cuda-sim/ptx_ir.h | 7 ++++++- src/cuda-sim/ptx_parser.cc | 3 ++- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 11 ++++++++--- src/gpgpu-sim/mem_latency_stat.cc | 2 +- src/gpgpu-sim/mem_latency_stat.h | 4 ++-- src/gpgpu-sim/power_interface.cc | 2 +- src/gpgpu-sim/power_interface.h | 2 +- src/gpgpu-sim/power_stat.cc | 6 +++--- src/gpgpu-sim/power_stat.h | 6 +++--- src/gpgpu-sim/shader.cc | 7 ++++--- src/gpgpu-sim/shader.h | 18 ++++++++++++------ src/gpgpusim_entrypoint.cc | 4 ++-- 17 files changed, 65 insertions(+), 34 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index a8b60f4..2e21009 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -4,6 +4,7 @@ #include "../src/cuda-sim/ptx_loader.h" #include "../src/cuda-sim/ptx_parser.h" #include "../src/gpgpusim_entrypoint.h" +#include "../src/cuda-sim/cuda-sim.h" class gpgpu_context { public: @@ -13,6 +14,7 @@ class gpgpu_context { ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); the_gpgpusim = new GPGPUsim_ctx(this); + func_sim = new cuda_sim(); } // global list symbol_table *g_global_allfiles_symbol_table; @@ -22,6 +24,7 @@ class gpgpu_context { ptxinfo_data* ptxinfo; ptx_recognizer* ptx_parser; GPGPUsim_ctx* the_gpgpusim; + cuda_sim* func_sim; // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 261d605..df0bbd7 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -51,6 +51,7 @@ typedef void * yyscan_t; #include "decuda_pred_table/decuda_pred_table.h" #include "../stream_manager.h" #include "cuda_device_runtime.h" +#include "../../libcuda/gpgpu_context.h" int gpgpu_ptx_instruction_classification; void ** g_inst_classification_stat = NULL; @@ -66,12 +67,12 @@ int cp_cta_resume; unsigned g_ptx_sim_num_insn = 0; unsigned gpgpu_param_num_shaders = 0; -char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu,*opcode_latency_tensor; +char *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu,*opcode_latency_tensor; char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp,*opcode_initiation_sfu,*opcode_initiation_tensor; char *cdp_latency_str; unsigned cdp_latency[5]; -void ptx_opcocde_latency_options (option_parser_t opp) { +void cuda_sim::ptx_opcocde_latency_options (option_parser_t opp) { option_parser_register(opp, "-ptx_opcode_latency_int", OPT_CSTR, &opcode_latency_int, "Opcode latencies for integers " "Default 1,1,19,25,145", @@ -667,7 +668,7 @@ void ptx_instruction::set_opcode_and_latency() * [3] MAD * [4] DIV */ - sscanf(opcode_latency_int, "%u,%u,%u,%u,%u", + sscanf(gpgpu_ctx->func_sim->opcode_latency_int, "%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], &int_latency[3],&int_latency[4]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index e690356..96d34f5 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -47,10 +47,9 @@ extern int g_debug_thread_uid; extern void ** g_inst_classification_stat; extern void ** g_inst_op_classification_stat; extern int g_ptx_kernel_count; // used for classification stat collection purposes -extern char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu,*opcode_latency_tensor; +extern char *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu,*opcode_latency_tensor; -void ptx_opcocde_latency_options (option_parser_t opp); extern class kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, gpgpu_ptx_sim_arg_list_t args, struct dim3 gridDim, @@ -134,4 +133,12 @@ void print_ptxinfo(); void clear_ptxinfo(); struct gpgpu_ptx_sim_info get_ptxinfo(); +class cuda_sim { + public: + //global variables + char *opcode_latency_int; + //global functions + void ptx_opcocde_latency_options (option_parser_t opp); +}; + #endif diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 8cedf79..1bd409e 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -1095,8 +1095,10 @@ ptx_instruction::ptx_instruction( int opcode, const char *file, unsigned line, const char *source, - const core_config *config ) : warp_inst_t(config) + const core_config *config, + gpgpu_context* ctx ) : warp_inst_t(config) { + gpgpu_ctx = ctx; m_uid = ++g_num_ptx_inst_uid; m_PC = 0; m_opcode = opcode; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index fd869c6..1604551 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -43,6 +43,8 @@ #include "memory.h" +class gpgpu_context; + class type_info_key { public: type_info_key() @@ -931,7 +933,8 @@ public: const char *file, unsigned line, const char *source, - const core_config *config ); + const core_config *config, + gpgpu_context* ctx); void print_insn() const; virtual void print_insn( FILE *fp ) const; @@ -1187,6 +1190,8 @@ private: virtual void pre_decode(); friend class function_info; static unsigned g_num_ptx_inst_uid; + // backward pointer + class gpgpu_context* gpgpu_ctx; }; class param_info { diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 05fc618..0139534 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -285,7 +285,8 @@ void ptx_recognizer::add_instruction() gpgpu_ctx->g_filename, ptx_get_lineno(scanner), linebuf, - g_shader_core_config ); + g_shader_core_config, + gpgpu_ctx ); g_instructions.push_back(i); g_inst_lookup[gpgpu_ctx->g_filename][ptx_get_lineno(scanner)] = i; init_instruction_state(); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 2ff37d1..39acdd9 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1826,7 +1826,7 @@ void gpgpu_sim::dump_pipeline( int mask, int s, int m ) const fflush(stdout); } -const struct shader_core_config * gpgpu_sim::getShaderCoreConfig() +const shader_core_config * gpgpu_sim::getShaderCoreConfig() { return m_shader_config; } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index e2c913a..7eeb7dd 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -295,7 +295,10 @@ extern bool g_interactive_debugger_enabled; class gpgpu_sim_config : public power_config, public gpgpu_functional_sim_config { public: - gpgpu_sim_config() { m_valid = false; } + gpgpu_sim_config(gpgpu_context* ctx): m_shader_config(ctx) { + m_valid = false; + gpgpu_ctx = ctx; + } void reg_options(class OptionParser * opp); void init() { @@ -341,6 +344,8 @@ private: void init_clock_domains(void ); + // backward pointer + class gpgpu_context* gpgpu_ctx; bool m_valid; shader_core_config m_shader_config; memory_config m_memory_config; @@ -473,7 +478,7 @@ public: /*! * Returning the configuration of the shader core, used by the functional simulation only so far */ - const struct shader_core_config * getShaderCoreConfig(); + const shader_core_config * getShaderCoreConfig(); //! Get shader core Memory Configuration @@ -537,7 +542,7 @@ private: const gpgpu_sim_config &m_config; const struct cudaDeviceProp *m_cuda_properties; - const struct shader_core_config *m_shader_config; + const shader_core_config *m_shader_config; const struct memory_config *m_memory_config; // stats diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index 04dc75b..d08ba39 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -42,7 +42,7 @@ #include #include -memory_stats_t::memory_stats_t( unsigned n_shader, const struct shader_core_config *shader_config, const struct memory_config *mem_config, const class gpgpu_sim* gpu ) +memory_stats_t::memory_stats_t( unsigned n_shader, const shader_core_config *shader_config, const struct memory_config *mem_config, const class gpgpu_sim* gpu ) { assert( mem_config->m_valid ); assert( shader_config->m_valid ); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index b9285c1..6ce568d 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -35,7 +35,7 @@ class memory_stats_t { public: memory_stats_t( unsigned n_shader, - const struct shader_core_config *shader_config, + const class shader_core_config *shader_config, const struct memory_config *mem_config, const class gpgpu_sim* gpu); @@ -53,7 +53,7 @@ public: unsigned m_n_shader; - const struct shader_core_config *m_shader_config; + const shader_core_config *m_shader_config; const struct memory_config *m_memory_config; const class gpgpu_sim* m_gpu; diff --git a/src/gpgpu-sim/power_interface.cc b/src/gpgpu-sim/power_interface.cc index 3861b6a..0272aa6 100644 --- a/src/gpgpu-sim/power_interface.cc +++ b/src/gpgpu-sim/power_interface.cc @@ -38,7 +38,7 @@ void init_mcpat(const gpgpu_sim_config &config, class gpgpu_sim_wrapper *wrapper } -void mcpat_cycle(const gpgpu_sim_config &config, const struct shader_core_config *shdr_config, class gpgpu_sim_wrapper *wrapper, class power_stat_t *power_stats, unsigned stat_sample_freq, unsigned tot_cycle, unsigned cycle, unsigned tot_inst, unsigned inst){ +void mcpat_cycle(const gpgpu_sim_config &config, const shader_core_config *shdr_config, class gpgpu_sim_wrapper *wrapper, class power_stat_t *power_stats, unsigned stat_sample_freq, unsigned tot_cycle, unsigned cycle, unsigned tot_inst, unsigned inst){ static bool mcpat_init=true; diff --git a/src/gpgpu-sim/power_interface.h b/src/gpgpu-sim/power_interface.h index afac22b..a388c23 100644 --- a/src/gpgpu-sim/power_interface.h +++ b/src/gpgpu-sim/power_interface.h @@ -36,7 +36,7 @@ #include "gpgpu_sim_wrapper.h" void init_mcpat(const gpgpu_sim_config &config, class gpgpu_sim_wrapper *wrapper, unsigned stat_sample_freq, unsigned tot_inst, unsigned inst); -void mcpat_cycle(const gpgpu_sim_config &config, const struct shader_core_config *shdr_config, class gpgpu_sim_wrapper *wrapper, class power_stat_t *power_stats, +void mcpat_cycle(const gpgpu_sim_config &config, const shader_core_config *shdr_config, class gpgpu_sim_wrapper *wrapper, class power_stat_t *power_stats, unsigned stat_sample_freq, unsigned tot_cycle, unsigned cycle, unsigned tot_inst, unsigned inst); void mcpat_reset_perf_count(class gpgpu_sim_wrapper *wrapper); diff --git a/src/gpgpu-sim/power_stat.cc b/src/gpgpu-sim/power_stat.cc index 4c995e9..007b4c6 100644 --- a/src/gpgpu-sim/power_stat.cc +++ b/src/gpgpu-sim/power_stat.cc @@ -42,7 +42,7 @@ -power_mem_stat_t::power_mem_stat_t(const struct memory_config *mem_config, const struct shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats){ +power_mem_stat_t::power_mem_stat_t(const struct memory_config *mem_config, const shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats){ assert( mem_config->m_valid ); m_mem_stats = mem_stats; m_config = mem_config; @@ -125,7 +125,7 @@ void power_mem_stat_t::print (FILE *fout) const { } -power_core_stat_t::power_core_stat_t( const struct shader_core_config *shader_config, shader_core_stats *core_stats ) +power_core_stat_t::power_core_stat_t( const shader_core_config *shader_config, shader_core_stats *core_stats ) { assert( shader_config->m_valid ); m_config = shader_config; @@ -266,7 +266,7 @@ for(unsigned i=0; inum_shader(); ++i){ } } -power_stat_t::power_stat_t( const struct shader_core_config *shader_config,float * average_pipeline_duty_cycle,float *active_sms,shader_core_stats * shader_stats, const struct memory_config *mem_config,memory_stats_t * memory_stats) +power_stat_t::power_stat_t( const shader_core_config *shader_config,float * average_pipeline_duty_cycle,float *active_sms,shader_core_stats * shader_stats, const struct memory_config *mem_config,memory_stats_t * memory_stats) { assert( shader_config->m_valid ); assert( mem_config->m_valid ); diff --git a/src/gpgpu-sim/power_stat.h b/src/gpgpu-sim/power_stat.h index 20af2e5..91fade9 100644 --- a/src/gpgpu-sim/power_stat.h +++ b/src/gpgpu-sim/power_stat.h @@ -73,7 +73,7 @@ struct shader_core_power_stats_pod { class power_core_stat_t : public shader_core_power_stats_pod { public: - power_core_stat_t(const struct shader_core_config *shader_config, shader_core_stats *core_stats); + power_core_stat_t(const shader_core_config *shader_config, shader_core_stats *core_stats); void visualizer_print( gzFile visualizer_file ); void print (FILE *fout); void init(); @@ -113,7 +113,7 @@ struct mem_power_stats_pod{ class power_mem_stat_t : public mem_power_stats_pod{ public: - power_mem_stat_t(const struct memory_config *mem_config, const struct shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats); + power_mem_stat_t(const struct memory_config *mem_config, const shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats); void visualizer_print( gzFile visualizer_file ); void print (FILE *fout) const; void init(); @@ -128,7 +128,7 @@ private: class power_stat_t { public: - power_stat_t( const struct shader_core_config *shader_config,float * average_pipeline_duty_cycle,float * active_sms,shader_core_stats * shader_stats, const struct memory_config *mem_config,memory_stats_t * memory_stats); + power_stat_t( const shader_core_config *shader_config,float * average_pipeline_duty_cycle,float * active_sms,shader_core_stats * shader_stats, const struct memory_config *mem_config,memory_stats_t * memory_stats); void visualizer_print( gzFile visualizer_file ); void print (FILE *fout) const; void save_stats(){ diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 69b619a..4d12068 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -46,6 +46,7 @@ #include #include "traffic_breakdown.h" #include "shader_trace.h" +#include "../../libcuda/gpgpu_context.h" #define PRIORITIZE_MSHR_OVER_WB 1 #define MAX(a,b) (((a)>(b))?(a):(b)) @@ -69,7 +70,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, class simt_core_cluster *cluster, unsigned shader_id, unsigned tpc_id, - const struct shader_core_config *config, + const shader_core_config *config, const struct memory_config *mem_config, shader_core_stats *stats ) : core_t( gpu, NULL, config->warp_size, config->n_thread_per_shader ), @@ -3018,7 +3019,7 @@ void shader_core_config::set_pipeline_latency() { * [3] MAD * [4] DIV */ - sscanf(opcode_latency_int, "%u,%u,%u,%u,%u", + sscanf(gpgpu_ctx->func_sim->opcode_latency_int, "%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], &int_latency[3],&int_latency[4]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", @@ -3786,7 +3787,7 @@ void opndcoll_rfu_t::collector_unit_t::dispatch() simt_core_cluster::simt_core_cluster( class gpgpu_sim *gpu, unsigned cluster_id, - const struct shader_core_config *config, + const shader_core_config *config, const struct memory_config *mem_config, shader_core_stats *stats, class memory_stats_t *mstats ) diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 25b9607..e0cefac 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -69,6 +69,8 @@ #define WRITE_MASK_SIZE 8 +class gpgpu_context; + enum exec_unit_type_t { NONE = 0, @@ -294,7 +296,7 @@ typedef std::bitset warp_set_t; int register_bank(int regnum, int wid, unsigned num_banks, unsigned bank_warp_shift, bool sub_core_model, unsigned banks_per_sched, unsigned sched_id ); class shader_core_ctx; -struct shader_core_config; +class shader_core_config; class shader_core_stats; enum scheduler_prioritization_type @@ -1032,7 +1034,7 @@ struct ifetch_buffer_t { unsigned m_warp_id; }; -struct shader_core_config; +class shader_core_config; class simd_function_unit { public: @@ -1362,10 +1364,12 @@ const char* const pipeline_stage_name_decode[] = { "N_PIPELINE_STAGES" }; -struct shader_core_config : public core_config +class shader_core_config : public core_config { - shader_core_config(){ + public: + shader_core_config(gpgpu_context* ctx){ pipeline_widths_string = NULL; + gpgpu_ctx = ctx; } void init() @@ -1425,6 +1429,8 @@ struct shader_core_config : public core_config unsigned cid_to_sid( unsigned cid, unsigned cluster_id ) const { return cluster_id*n_simt_cores_per_cluster + cid; } void set_pipeline_latency(); + // backward pointer + class gpgpu_context* gpgpu_ctx; // data char *gpgpu_shader_core_pipeline_opt; bool gpgpu_perfect_mem; @@ -1770,7 +1776,7 @@ public: class simt_core_cluster *cluster, unsigned shader_id, unsigned tpc_id, - const struct shader_core_config *config, + const shader_core_config *config, const struct memory_config *mem_config, shader_core_stats *stats ); @@ -2065,7 +2071,7 @@ class simt_core_cluster { public: simt_core_cluster( class gpgpu_sim *gpu, unsigned cluster_id, - const struct shader_core_config *config, + const shader_core_config *config, const struct memory_config *mem_config, shader_core_stats *stats, memory_stats_t *mstats ); diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index b54f20c..5018305 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -218,10 +218,10 @@ gpgpu_sim *gpgpu_context::gpgpu_ptx_sim_init_perf() option_parser_t opp = option_parser_create(); ptx_reg_options(opp); - ptx_opcocde_latency_options(opp); + func_sim->ptx_opcocde_latency_options(opp); icnt_reg_options(opp); - GPGPUsim_ctx_ptr()->g_the_gpu_config = new gpgpu_sim_config(); + GPGPUsim_ctx_ptr()->g_the_gpu_config = new gpgpu_sim_config(this); GPGPUsim_ctx_ptr()->g_the_gpu_config->reg_options(opp); // register GPU microrachitecture options option_parser_cmdline(opp, sg_argc, sg_argv); // parse configuration options -- cgit v1.3 From 6e051c42b8c9c820af50a10025d182a96fdd12a6 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 9 Jul 2019 10:06:31 -0400 Subject: Move m_ptx_save_converted_ptxplus Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 2 +- src/cuda-sim/ptx_loader.cc | 5 ++--- src/cuda-sim/ptx_loader.h | 4 ++-- 3 files changed, 5 insertions(+), 6 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index eb645ce..43c5e1f 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -2936,7 +2936,7 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle){ if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { cuobjdumpELFSection* elfsection = api->findELFSection(ptx->getIdentifier()); assert (elfsection!= NULL); - char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( + char *ptxplus_str = ptxinfo->gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( ptx->getPTXfilename(), elfsection->getELFfilename(), elfsection->getSASSfilename()); diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index e0d9a11..2bf464c 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -58,7 +58,6 @@ extern int ptxinfo_lex_destroy(yyscan_t scanner); static bool g_save_embedded_ptx; static int g_occupancy_sm_number; -bool m_ptx_save_converted_ptxplus; bool ptxinfo_data::keep_intermediate_files() {return g_keep_intermediate_files;} @@ -71,7 +70,7 @@ void gpgpu_context::ptx_reg_options(option_parser_t opp) "keep intermediate files created by GPGPU-Sim when interfacing with external programs", "0"); option_parser_register(opp, "-gpgpu_ptx_save_converted_ptxplus", OPT_BOOL, - &m_ptx_save_converted_ptxplus, + &(ptxinfo->m_ptx_save_converted_ptxplus), "Saved converted ptxplus to a file", "0"); option_parser_register(opp, "-gpgpu_occupancy_sm_number", OPT_INT32, &g_occupancy_sm_number, @@ -106,7 +105,7 @@ void gpgpu_context::print_ptx_file( const char *p, unsigned source_num, const ch fflush(stdout); } -char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilename, const std::string elffilename, const std::string sassfilename) +char* ptxinfo_data::gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilename, const std::string elffilename, const std::string sassfilename) { printf("GPGPU-Sim PTX: converting EMBEDDED .ptx file to ptxplus \n"); diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index aa3e2f3..decfdc8 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -43,14 +43,14 @@ class ptxinfo_data{ const char *g_ptxinfo_filename; class gpgpu_context* gpgpu_ctx; bool g_keep_intermediate_files; + bool m_ptx_save_converted_ptxplus; void ptxinfo_addinfo(); bool keep_intermediate_files(); + char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); }; extern bool g_override_embedded_ptx; extern int no_of_ptx; //counter to track number of ptx files to be extracted in an application. -char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); - #endif -- cgit v1.3 From cda7a145b9e28eff0f3e9ac8197c2b6215755fc8 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 9 Jul 2019 14:32:24 -0400 Subject: Move g_ptx_kernel_count Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 5 ++++- libcuda/cuda_runtime_api.cc | 2 +- libcuda/gpgpu_context.h | 2 +- libopencl/opencl_runtime_api.cc | 2 +- src/cuda-sim/cuda-sim.cc | 13 ++++++------- src/cuda-sim/cuda-sim.h | 14 ++++++++------ src/gpgpu-sim/gpu-sim.cc | 4 ++-- 7 files changed, 23 insertions(+), 19 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index 0054697..db5e6a4 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -169,9 +169,10 @@ private: class cuda_runtime_api { public: - cuda_runtime_api() { + cuda_runtime_api( gpgpu_context* ctx ) { g_glbmap = NULL; g_active_device = 0; //active gpu that runs the code + gpgpu_ctx = ctx; } // global list std::list cuobjdumpSectionList; @@ -187,6 +188,8 @@ class cuda_runtime_api { std::map pinned_memory_size; glbmap_entry_t* g_glbmap; int g_active_device; //active gpu that runs the code + // backward pointer + class gpgpu_context* gpgpu_ctx; // member function list void cuobjdumpInit(); void extract_code_using_cuobjdump(); diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 43c5e1f..43c8bae 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -3545,7 +3545,7 @@ kernel_info_t * cuda_runtime_api::gpgpu_cuda_ptx_sim_init_grid( const char *host } entry->finalize(result->get_param_memory()); - g_ptx_kernel_count++; + gpgpu_ctx->func_sim->g_ptx_kernel_count++; fflush(stdout); if(g_debug_execution >= 4){ diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 2e21009..a2ae7b6 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -10,7 +10,7 @@ class gpgpu_context { public: gpgpu_context() { g_global_allfiles_symbol_table = NULL; - api = new cuda_runtime_api(); + api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); the_gpgpusim = new GPGPUsim_ctx(this); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index d302ff8..0a6eb3e 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -956,7 +956,7 @@ clEnqueueNDRangeKernel(cl_command_queue command_queue, gpgpu_ptx_sim_memcpy_symbol( "%_global_launch_offset", zeros, 3 * sizeof(int), 0, 1, gpu ); gpgpu_ptx_sim_memcpy_symbol( "%_global_block_offset", zeros, 3 * sizeof(int), 0, 1, gpu ); } - kernel_info_t *grid = gpgpu_opencl_ptx_sim_init_grid(kernel->get_implementation(),params,GridDim,BlockDim,gpu); + kernel_info_t *grid = ctx->func_sim->gpgpu_opencl_ptx_sim_init_grid(kernel->get_implementation(),params,GridDim,BlockDim,gpu); if ( g_ptx_sim_mode ) ctx->func_sim->gpgpu_opencl_ptx_sim_main_func( grid ); else diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index e86395d..0ed125a 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -56,7 +56,6 @@ typedef void * yyscan_t; int gpgpu_ptx_instruction_classification; void ** g_inst_classification_stat = NULL; void ** g_inst_op_classification_stat= NULL; -int g_ptx_kernel_count = -1; // used for classification stat collection purposes int g_debug_execution = 0; int g_debug_thread_uid = 0; addr_t g_debug_pc = 0xBEEF1518; @@ -1480,7 +1479,7 @@ bool ptx_debug_exec_dump_cond(int thd_uid, addr_t pc) return false; } -void init_inst_classification_stat() +void cuda_sim::init_inst_classification_stat() { static std::set init; if( init.find(g_ptx_kernel_count) != init.end() ) @@ -1690,7 +1689,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) ptx_file_line_stats_add_exec_count(pI); if ( gpgpu_ptx_instruction_classification ) { - init_inst_classification_stat(); + m_gpu->gpgpu_ctx->func_sim->init_inst_classification_stat(); unsigned space_type=0; switch ( pI->get_space().get_type() ) { case global_space: space_type = 10; break; @@ -1706,9 +1705,9 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) space_type = 0 ; break; } - StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], op_classification); - if (space_type) StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], ( int )space_type); - StatAddSample( g_inst_op_classification_stat[g_ptx_kernel_count], (int) pI->get_opcode() ); + StatAddSample( g_inst_classification_stat[m_gpu->gpgpu_ctx->func_sim->g_ptx_kernel_count], op_classification); + if (space_type) StatAddSample( g_inst_classification_stat[m_gpu->gpgpu_ctx->func_sim->g_ptx_kernel_count], ( int )space_type); + StatAddSample( g_inst_op_classification_stat[m_gpu->gpgpu_ctx->func_sim->g_ptx_kernel_count], (int) pI->get_opcode() ); } if ( (m_gpu->gpgpu_ctx->func_sim->g_ptx_sim_num_insn % 100000) == 0 ) { dim3 ctaid = get_ctaid(); @@ -1917,7 +1916,7 @@ size_t get_kernel_code_size( class function_info *entry ) } -kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, +kernel_info_t *cuda_sim::gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, gpgpu_ptx_sim_arg_list_t args, struct dim3 gridDim, struct dim3 blockDim, diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index 16ee46e..e259f1f 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -46,13 +46,7 @@ extern int g_debug_execution; extern int g_debug_thread_uid; extern void ** g_inst_classification_stat; extern void ** g_inst_op_classification_stat; -extern int g_ptx_kernel_count; // used for classification stat collection purposes -extern class kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, - gpgpu_ptx_sim_arg_list_t args, - struct dim3 gridDim, - struct dim3 blockDim, - class gpgpu_t *gpu ); extern void print_splash(); extern void gpgpu_ptx_sim_register_const_variable(void*, const char *deviceName, size_t size ); extern void gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ); @@ -134,6 +128,7 @@ class cuda_sim { public: cuda_sim() { g_ptx_sim_num_insn = 0; + g_ptx_kernel_count = -1; // used for classification stat collection purposes } //global variables char *opcode_latency_int; @@ -151,10 +146,17 @@ class cuda_sim { int g_ptxinfo_error_detected; unsigned g_ptx_sim_num_insn; char *cdp_latency_str; + int g_ptx_kernel_count; // used for classification stat collection purposes //global functions void ptx_opcocde_latency_options (option_parser_t opp); void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL = false ); int gpgpu_opencl_ptx_sim_main_func( kernel_info_t *grid ); + void init_inst_classification_stat(); + kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, + gpgpu_ptx_sim_arg_list_t args, + struct dim3 gridDim, + struct dim3 blockDim, + gpgpu_t *gpu ); }; #endif diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 4f9ccbf..9f47067 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1201,8 +1201,8 @@ void gpgpu_sim::gpu_print_stat() insn_warp_occ_print(stdout); } if ( gpgpu_ptx_instruction_classification ) { - StatDisp( g_inst_classification_stat[g_ptx_kernel_count]); - StatDisp( g_inst_op_classification_stat[g_ptx_kernel_count]); + StatDisp( g_inst_classification_stat[gpgpu_ctx->func_sim->g_ptx_kernel_count]); + StatDisp( g_inst_op_classification_stat[gpgpu_ctx->func_sim->g_ptx_kernel_count]); } #ifdef GPGPUSIM_POWER_MODEL -- cgit v1.3 From b3655bf28a7402db347f9d7f87049806b9315a05 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 9 Jul 2019 15:03:18 -0400 Subject: Move g_global_name_lookup and g_constant_name_lookup Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 4 ++-- src/cuda-sim/cuda-sim.cc | 14 ++++++-------- src/cuda-sim/cuda-sim.h | 6 ++++-- 3 files changed, 12 insertions(+), 12 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 43c8bae..bbbaf23 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -767,9 +767,9 @@ void cudaRegisterVarInternal( ctx->cuobjdumpParseBinary((unsigned)(unsigned long long)fatCubinHandle); fflush(stdout); if ( constant && !global && !ext ) { - gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); + ctx->func_sim->gpgpu_ptx_sim_register_const_variable(hostVar,deviceName,size); } else if ( !constant && !global && !ext ) { - gpgpu_ptx_sim_register_global_variable(hostVar,deviceName,size); + ctx->func_sim->gpgpu_ptx_sim_register_global_variable(hostVar,deviceName,size); } else cuda_not_implemented(__my_func__,__LINE__); } diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 0ed125a..b3e2965 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1948,18 +1948,16 @@ void print_splash() } } -std::map g_const_name_lookup; // indexed by hostVar -std::map g_global_name_lookup; // indexed by hostVar std::set g_globals; std::set g_constants; -void gpgpu_ptx_sim_register_const_variable(void *hostVar, const char *deviceName, size_t size ) +void cuda_sim::gpgpu_ptx_sim_register_const_variable(void *hostVar, const char *deviceName, size_t size ) { printf("GPGPU-Sim PTX registering constant %s (%zu bytes) to name mapping\n", deviceName, size ); g_const_name_lookup[hostVar] = deviceName; } -void gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ) +void cuda_sim::gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ) { printf("GPGPU-Sim PTX registering global %s hostVar to name mapping\n", deviceName ); g_global_name_lookup[hostVar] = deviceName; @@ -1972,14 +1970,14 @@ void gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, size_t co memory_space_t mem_region = undefined_space; std::string sym_name; - std::map::iterator c=g_const_name_lookup.find(hostVar); - if ( c!=g_const_name_lookup.end() ) { + std::map::iterator c=gpu->gpgpu_ctx->func_sim->g_const_name_lookup.find(hostVar); + if ( c!=gpu->gpgpu_ctx->func_sim->g_const_name_lookup.end() ) { found_sym = true; sym_name = c->second; mem_region = const_space; } - std::map::iterator g=g_global_name_lookup.find(hostVar); - if ( g!=g_global_name_lookup.end() ) { + std::map::iterator g=gpu->gpgpu_ctx->func_sim->g_global_name_lookup.find(hostVar); + if ( g!=gpu->gpgpu_ctx->func_sim->g_global_name_lookup.end() ) { if ( found_sym ) { printf("Execution error: PTX symbol \"%s\" w/ hostVar=0x%Lx is declared both const and global?\n", sym_name.c_str(), (unsigned long long)hostVar ); diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index e259f1f..25ebf7b 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -48,8 +48,6 @@ extern void ** g_inst_classification_stat; extern void ** g_inst_op_classification_stat; extern void print_splash(); -extern void gpgpu_ptx_sim_register_const_variable(void*, const char *deviceName, size_t size ); -extern void gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ); extern void gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, size_t count, size_t offset, int to, gpgpu_t *gpu ); extern void read_sim_environment_variables(); @@ -147,6 +145,8 @@ class cuda_sim { unsigned g_ptx_sim_num_insn; char *cdp_latency_str; int g_ptx_kernel_count; // used for classification stat collection purposes + std::map g_global_name_lookup; // indexed by hostVar + std::map g_const_name_lookup; // indexed by hostVar //global functions void ptx_opcocde_latency_options (option_parser_t opp); void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL = false ); @@ -157,6 +157,8 @@ class cuda_sim { struct dim3 gridDim, struct dim3 blockDim, gpgpu_t *gpu ); + void gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ); + void gpgpu_ptx_sim_register_const_variable(void*, const char *deviceName, size_t size ); }; #endif -- cgit v1.3 From 4d4d5938d715d2b79a617c32583184426b4a642d Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 9 Jul 2019 23:16:17 -0400 Subject: Move g_ptx_sim_mode Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 6 +++--- libcuda/gpgpu_context.h | 9 ++++++++- libopencl/opencl_runtime_api.cc | 6 +++--- src/cuda-sim/cuda-sim.cc | 6 ++---- src/cuda-sim/cuda-sim.h | 10 +++++++--- src/cuda-sim/cuda_device_runtime.cc | 7 ++++--- src/cuda-sim/cuda_device_runtime.h | 18 ++++++++++++++++-- src/gpgpu-sim/gpu-sim.cc | 4 ++-- src/gpgpusim_entrypoint.cc | 2 +- 9 files changed, 46 insertions(+), 22 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index bbbaf23..59d2a60 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1076,7 +1076,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = CUctx_st* context = GPGPUSim_Context(); char *mode = getenv("PTX_SIM_MODE_FUNC"); if( mode ) - sscanf(mode,"%u", &g_ptx_sim_mode); + sscanf(mode,"%u", &(ctx->func_sim->g_ptx_sim_mode)); gpgpusim_ptx_assert( !ctx->api->g_cuda_launch_stack.empty(), "empty launch stack" ); kernel_config config = ctx->api->g_cuda_launch_stack.back(); { @@ -1092,7 +1092,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = } struct CUstream_st *stream = config.get_stream(); printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, - g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); + (ctx->func_sim->g_ptx_sim_mode)?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); kernel_info_t *grid = ctx->api->gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); //do dynamic PDOM analysis for performance simulation scenario std::string kname = grid->name(); @@ -1143,7 +1143,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = } printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); - stream_operation op(grid,g_ptx_sim_mode,stream); + stream_operation op(grid,ctx->func_sim->g_ptx_sim_mode,stream); g_stream_manager()->push(op); ctx->api->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index a2ae7b6..3c9f87c 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -5,6 +5,7 @@ #include "../src/cuda-sim/ptx_parser.h" #include "../src/gpgpusim_entrypoint.h" #include "../src/cuda-sim/cuda-sim.h" +#include "../src/cuda-sim/cuda_device_runtime.h" class gpgpu_context { public: @@ -14,7 +15,10 @@ class gpgpu_context { ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); the_gpgpusim = new GPGPUsim_ctx(this); - func_sim = new cuda_sim(); + func_sim = new cuda_sim(this); +#if (CUDART_VERSION >= 5000) + device_runtime = new cuda_device_runtime(this); +#endif } // global list symbol_table *g_global_allfiles_symbol_table; @@ -25,6 +29,9 @@ class gpgpu_context { ptx_recognizer* ptx_parser; GPGPUsim_ctx* the_gpgpusim; cuda_sim* func_sim; +#if (CUDART_VERSION >= 5000) + cuda_device_runtime* device_runtime; +#endif // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index 0a6eb3e..aaaec4f 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -884,9 +884,9 @@ clEnqueueNDRangeKernel(cl_command_queue command_queue, printf("\n\n\n"); char *mode = getenv("PTX_SIM_MODE_FUNC"); if ( mode ) - sscanf(mode,"%u", &g_ptx_sim_mode); + sscanf(mode,"%u", &(ctx->func_sim->g_ptx_sim_mode)); printf("GPGPU-Sim OpenCL API: clEnqueueNDRangeKernel '%s' (mode=%s)\n", kernel->name().c_str(), - g_ptx_sim_mode?"functional simulation":"performance simulation"); + (ctx->func_sim->g_ptx_sim_mode)?"functional simulation":"performance simulation"); if ( !work_dim || work_dim > 3 ) return CL_INVALID_WORK_DIMENSION; size_t _local_size[3]; if( local_work_size != NULL ) { @@ -957,7 +957,7 @@ clEnqueueNDRangeKernel(cl_command_queue command_queue, gpgpu_ptx_sim_memcpy_symbol( "%_global_block_offset", zeros, 3 * sizeof(int), 0, 1, gpu ); } kernel_info_t *grid = ctx->func_sim->gpgpu_opencl_ptx_sim_init_grid(kernel->get_implementation(),params,GridDim,BlockDim,gpu); - if ( g_ptx_sim_mode ) + if ( ctx->func_sim->g_ptx_sim_mode ) ctx->func_sim->gpgpu_opencl_ptx_sim_main_func( grid ); else gpgpu_opencl_ptx_sim_main_perf( grid ); diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index b3e2965..7a7d205 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -2033,13 +2033,11 @@ void gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, size_t co fflush(stdout); } -int g_ptx_sim_mode; // if non-zero run functional simulation only (i.e., no notion of a clock cycle) - extern int ptx_debug; bool g_cuda_launch_blocking = false; -void read_sim_environment_variables() +void cuda_sim::read_sim_environment_variables() { ptx_debug = 0; g_debug_execution = 0; @@ -2185,7 +2183,7 @@ void cuda_sim::gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL cta.execute(cp_count,temp); #if (CUDART_VERSION >= 5000) - launch_all_device_kernels(); + gpgpu_ctx->device_runtime->launch_all_device_kernels(); #endif } else diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index 25ebf7b..3c4336d 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -36,12 +36,12 @@ #include #include"ptx_sim.h" +class gpgpu_context; class memory_space; class function_info; class symbol_table; extern const char *g_gpgpusim_version_string; -extern int g_ptx_sim_mode; extern int g_debug_execution; extern int g_debug_thread_uid; extern void ** g_inst_classification_stat; @@ -50,7 +50,6 @@ extern void ** g_inst_op_classification_stat; extern void print_splash(); extern void gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, size_t count, size_t offset, int to, gpgpu_t *gpu ); -extern void read_sim_environment_variables(); extern void ptxinfo_opencl_addinfo( std::map &kernels ); unsigned ptx_sim_init_thread( kernel_info_t &kernel, class ptx_thread_info** thread_info, @@ -124,9 +123,10 @@ struct gpgpu_ptx_sim_info get_ptxinfo(); class cuda_sim { public: - cuda_sim() { + cuda_sim( gpgpu_context* ctx ) { g_ptx_sim_num_insn = 0; g_ptx_kernel_count = -1; // used for classification stat collection purposes + gpgpu_ctx = ctx; } //global variables char *opcode_latency_int; @@ -147,6 +147,9 @@ class cuda_sim { int g_ptx_kernel_count; // used for classification stat collection purposes std::map g_global_name_lookup; // indexed by hostVar std::map g_const_name_lookup; // indexed by hostVar + int g_ptx_sim_mode; // if non-zero run functional simulation only (i.e., no notion of a clock cycle) + // backward pointer + class gpgpu_context* gpgpu_ctx; //global functions void ptx_opcocde_latency_options (option_parser_t opp); void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL = false ); @@ -159,6 +162,7 @@ class cuda_sim { gpgpu_t *gpu ); void gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ); void gpgpu_ptx_sim_register_const_variable(void*, const char *deviceName, size_t size ); + void read_sim_environment_variables(); }; #endif diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index be8369f..354fa79 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -20,6 +20,7 @@ unsigned long long g_max_total_param_size = 0; #include "../stream_manager.h" #include "../gpgpusim_entrypoint.h" #include "cuda_device_runtime.h" +#include "../../libcuda/gpgpu_context.h" #define DEV_RUNTIME_REPORT(a) \ if( g_debug_execution ) { \ @@ -318,17 +319,17 @@ void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_ } -void launch_one_device_kernel() { +void cuda_device_runtime::launch_one_device_kernel() { if(!g_cuda_device_launch_op.empty()) { device_launch_operation_t &op = g_cuda_device_launch_op.front(); - stream_operation stream_op = stream_operation(op.grid, g_ptx_sim_mode, op.stream); + stream_operation stream_op = stream_operation(op.grid, gpgpu_ctx->func_sim->g_ptx_sim_mode, op.stream); g_stream_manager()->push(stream_op); g_cuda_device_launch_op.pop_front(); } } -void launch_all_device_kernels() { +void cuda_device_runtime::launch_all_device_kernels() { while(!g_cuda_device_launch_op.empty()) { launch_one_device_kernel(); } diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index 6dbcd71..851fed2 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -6,6 +6,20 @@ void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); -void launch_all_device_kernels(); -void launch_one_device_kernel(); +#endif +#if (CUDART_VERSION >= 5000) + +class gpgpu_context; + +class cuda_device_runtime { + public: + cuda_device_runtime( gpgpu_context* ctx ) { + gpgpu_ctx = ctx; + } + // backward pointer + class gpgpu_context* gpgpu_ctx; + void launch_all_device_kernels(); + void launch_one_device_kernel(); +}; + #endif diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 9f47067..bbcc078 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -503,7 +503,7 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) &gpgpu_ptx_instruction_classification, "if enabled will classify ptx instruction types per kernel (Max 255 kernels now)", "0"); - option_parser_register(opp, "-gpgpu_ptx_sim_mode", OPT_INT32, &g_ptx_sim_mode, + option_parser_register(opp, "-gpgpu_ptx_sim_mode", OPT_INT32, &(gpgpu_ctx->func_sim->g_ptx_sim_mode), "Select between Performance (default) or Functional simulation (1)", "0"); option_parser_register(opp, "-gpgpu_clock_domains", OPT_CSTR, &gpgpu_clock_domains, @@ -1753,7 +1753,7 @@ void gpgpu_sim::cycle() #if (CUDART_VERSION >= 5000) //launch device kernel - launch_one_device_kernel(); + gpgpu_ctx->device_runtime->launch_one_device_kernel(); #endif } } diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index d9d1023..683a695 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -213,7 +213,7 @@ gpgpu_sim *gpgpu_context::gpgpu_ptx_sim_init_perf() { srand(1); print_splash(); - read_sim_environment_variables(); + func_sim->read_sim_environment_variables(); ptx_parser->read_parser_environment_variables(); option_parser_t opp = option_parser_create(); -- cgit v1.3 From 2f22ddb07752bc956063f6548d4bc26a6ae066f2 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 11 Jul 2019 16:56:29 -0400 Subject: Fix symbols Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 4 ---- src/cuda-sim/cuda_device_runtime.h | 10 ++++++---- 2 files changed, 6 insertions(+), 8 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 3c9f87c..c0b250a 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -16,9 +16,7 @@ class gpgpu_context { ptx_parser = new ptx_recognizer(this); the_gpgpusim = new GPGPUsim_ctx(this); func_sim = new cuda_sim(this); -#if (CUDART_VERSION >= 5000) device_runtime = new cuda_device_runtime(this); -#endif } // global list symbol_table *g_global_allfiles_symbol_table; @@ -29,9 +27,7 @@ class gpgpu_context { ptx_recognizer* ptx_parser; GPGPUsim_ctx* the_gpgpusim; cuda_sim* func_sim; -#if (CUDART_VERSION >= 5000) cuda_device_runtime* device_runtime; -#endif // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index b49d0eb..a6303b9 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -1,8 +1,7 @@ +#ifndef __cuda_device_runtime_h__ +#define __cuda_device_runtime_h__ //Jin: cuda_device_runtime.h //Defines CUDA device runtime APIs for CDP support -#if (CUDART_VERSION >= 5000) -#pragma once - class device_launch_config_t { public: @@ -51,11 +50,14 @@ class cuda_device_runtime { std::list g_cuda_device_launch_op; // backward pointer class gpgpu_context* gpgpu_ctx; +#if (CUDART_VERSION >= 5000) +#pragma once void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void launch_all_device_kernels(); void launch_one_device_kernel(); +#endif }; -#endif +#endif /* __cuda_device_runtime_h__ */ -- cgit v1.3 From 4671280cfe7252bf875d071e667f57064250a1b7 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Fri, 12 Jul 2019 00:09:29 -0400 Subject: Move g_cdp_enabled Signed-off-by: Mengchi Zhang --- libcuda/cuda_api_object.h | 1 + libcuda/cuda_runtime_api.cc | 11 ++++------- src/abstract_hardware_model.cc | 3 --- src/cuda-sim/cuda_device_runtime.h | 2 ++ src/cuda-sim/ptx_loader.cc | 9 +++------ src/gpgpu-sim/gpu-sim.cc | 3 +-- 6 files changed, 11 insertions(+), 18 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index db5e6a4..51416f2 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -199,6 +199,7 @@ class cuda_runtime_api { std::list mergeSections(); cuobjdumpELFSection* findELFSection(const std::string identifier); cuobjdumpPTXSection* findPTXSection(const std::string identifier); + cuobjdumpPTXSection* findPTXSectionInList(std::list §ionlist, const std::string identifier); void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context); kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key, gpgpu_ptx_sim_arg_list_t args, diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 59d2a60..10a651a 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -2421,7 +2421,6 @@ __host__ cudaError_t CUDARTAPI cudaGetExportTable(const void **ppExportTable, co //extracts all ptx files from binary and dumps into prog_name.unique_no.sm_<>.ptx files void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context){ - extern bool g_cdp_enabled; char command[1000]; char *pytorch_bin = getenv("PYTORCH_BIN"); std::string app_binary = get_app_binary(); @@ -2442,7 +2441,7 @@ void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context){ printf("WARNING: Failed to execute cuobjdump to get list of ptx files \n"); exit(0); } - if(!g_cdp_enabled) { + if(!gpgpu_ctx->device_runtime->g_cdp_enabled) { //based on the list above, dump ptx files individually. Format of dumped ptx file is prog_name.unique_no.sm_<>.ptx std::ifstream infile(ptx_list_file_name); @@ -2515,7 +2514,6 @@ void cuda_runtime_api::extract_code_using_cuobjdump(){ } // Running cuobjdump using dynamic link to current process // Needs the option '-all' to extract PTX from CDP-enabled binary - extern bool g_cdp_enabled; //dump ptx for all individial ptx files into sepearte files which is later used by ptxas. int result=0; @@ -2530,7 +2528,7 @@ void cuda_runtime_api::extract_code_using_cuobjdump(){ snprintf(fname,1024,"_cuobjdump_complete_output_XXXXXX"); int fd=mkstemp(fname); close(fd); - if(!g_cdp_enabled) + if(!gpgpu_ctx->device_runtime->g_cdp_enabled) snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname); else snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname); @@ -2822,7 +2820,7 @@ cuobjdumpELFSection* cuda_runtime_api::findELFSection(const std::string identifi } //! Within the section list, find the PTX section corresponding to a given identifier -cuobjdumpPTXSection* findPTXSectionInList(std::list §ionlist, const std::string identifier){ +cuobjdumpPTXSection* cuda_runtime_api::findPTXSectionInList(std::list §ionlist, const std::string identifier){ std::list::iterator iter; for ( iter = sectionlist.begin(); iter != sectionlist.end(); @@ -2833,8 +2831,7 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list §ionl if(ptxsection->getIdentifier() == identifier) return ptxsection; else { - extern bool g_cdp_enabled; - if(g_cdp_enabled) { + if(gpgpu_ctx->device_runtime->g_cdp_enabled) { printf("Warning: __cudaRegisterFatBinary needs %s, but find PTX section with %s\n", identifier.c_str(), ptxsection->getIdentifier().c_str()); return ptxsection; diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 450087b..fde7874 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -691,9 +691,6 @@ void warp_inst_t::completed( unsigned long long cycle ) const ptx_file_line_stats_add_latency(pc, latency * active_count()); } -//Jin: CDP support -bool g_cdp_enabled; - unsigned kernel_info_t::m_next_uid = 1; /*A snapshot of the texture mappings needs to be stored in the kernel's info as diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index f37849b..7f7a0ca 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -51,6 +51,8 @@ class cuda_device_runtime { std::list g_cuda_device_launch_op; unsigned g_kernel_launch_latency; unsigned long long g_max_total_param_size; + bool g_cdp_enabled; + // backward pointer class gpgpu_context* gpgpu_ctx; #if (CUDART_VERSION >= 5000) diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 6e36a62..dca3cec 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -330,8 +330,7 @@ void gpgpu_context::gpgpu_ptx_info_load_from_filename( const char *filename, uns std::string ptxas_filename(std::string(filename) + "as"); char buff[1024], extra_flags[1024]; extra_flags[0]=0; - extern bool g_cdp_enabled; - if(!g_cdp_enabled) + if(!device_runtime->g_cdp_enabled) snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version); else snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version); @@ -398,8 +397,7 @@ void gpgpu_context::gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsi "A register size/SM mismatch may result in occupancy differences." ); exit(1); } - extern bool g_cdp_enabled; - if(!g_cdp_enabled) + if(!device_runtime->g_cdp_enabled) snprintf(extra_flags,1024,"--gpu-name=sm_%u", g_occupancy_sm_number); else snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",g_occupancy_sm_number); @@ -467,8 +465,7 @@ void gpgpu_context::gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsi #if CUDART_VERSION >= 3000 if (sm_version == 0) sm_version = 20; - extern bool g_cdp_enabled; - if(!g_cdp_enabled) + if(!device_runtime->g_cdp_enabled) snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version); else snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index a3d6a8a..0481259 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -551,9 +551,8 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-gpgpu_kernel_launch_latency", OPT_INT32, &(gpgpu_ctx->device_runtime->g_kernel_launch_latency), "Kernel launch latency in cycles. Default: 0", "0"); - extern bool g_cdp_enabled; option_parser_register(opp, "-gpgpu_cdp_enabled", OPT_BOOL, - &g_cdp_enabled, "Turn on CDP", + &(gpgpu_ctx->device_runtime->g_cdp_enabled), "Turn on CDP", "0"); } -- cgit v1.3 From 23c0bb224295dde9651fd915d854e4f7eafdf88f Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 14 Jul 2019 11:25:42 -0400 Subject: Move sm_next_access_uid Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/abstract_hardware_model.cc | 13 ++++++++++--- src/abstract_hardware_model.h | 22 +++++++++------------- src/gpgpu-sim/dram.cc | 2 +- src/gpgpu-sim/dram.h | 5 +++-- src/gpgpu-sim/gpu-cache.cc | 6 ++++-- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 15 +++++++++------ src/gpgpu-sim/l2cache.cc | 11 ++++++----- src/gpgpu-sim/l2cache.h | 12 ++++++------ src/gpgpu-sim/mem_fetch.cc | 4 ++-- src/gpgpu-sim/mem_fetch.h | 5 +++-- src/gpgpu-sim/mem_latency_stat.cc | 2 +- src/gpgpu-sim/mem_latency_stat.h | 5 +++-- src/gpgpu-sim/power_stat.cc | 4 ++-- src/gpgpu-sim/power_stat.h | 6 +++--- src/gpgpu-sim/shader.cc | 20 ++++++++++++++++---- src/gpgpu-sim/shader.h | 20 ++++---------------- 18 files changed, 85 insertions(+), 71 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index c0b250a..07473be 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -11,6 +11,7 @@ class gpgpu_context { public: gpgpu_context() { g_global_allfiles_symbol_table = NULL; + sm_next_access_uid=0; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -21,6 +22,7 @@ class gpgpu_context { // global list symbol_table *g_global_allfiles_symbol_table; const char *g_filename; + unsigned sm_next_access_uid; // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index fde7874..fe10daa 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -41,9 +41,16 @@ #include #include "../libcuda/gpgpu_context.h" -unsigned mem_access_t::sm_next_access_uid = 0; unsigned warp_inst_t::sm_next_uid = 0; +void mem_access_t::init(gpgpu_context* ctx) +{ + gpgpu_ctx = ctx; + m_uid=++(gpgpu_ctx->sm_next_access_uid); + m_addr=0; + m_req_size=0; +} + checkpoint::checkpoint() { @@ -449,7 +456,7 @@ void warp_inst_t::generate_mem_accesses() byte_mask.set(idx+i); } for( a=accesses.begin(); a != accesses.end(); ++a ) - m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second, byte_mask, mem_access_sector_mask_t())); + m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second, byte_mask, mem_access_sector_mask_t(), m_config->gpgpu_ctx)); } if ( space.get_type() == global_space ) { @@ -681,7 +688,7 @@ void warp_inst_t::memory_coalescing_arch_reduce_and_send( bool is_write, mem_acc assert(lower_half_used && upper_half_used); } } - m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes, info.chunks) ); + m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes, info.chunks,m_config->gpgpu_ctx) ); } void warp_inst_t::completed( unsigned long long cycle ) const diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 8ef8376..a22c8c3 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -732,13 +732,14 @@ enum cache_operator_type { class mem_access_t { public: - mem_access_t() { init(); } + mem_access_t(gpgpu_context* ctx) { init(ctx); } mem_access_t( mem_access_type type, new_addr_type address, unsigned size, - bool wr ) + bool wr, + gpgpu_context* ctx) { - init(); + init(ctx); m_type = type; m_addr = address; m_req_size = size; @@ -750,10 +751,11 @@ public: bool wr, const active_mask_t &active_mask, const mem_access_byte_mask_t &byte_mask, - const mem_access_sector_mask_t §or_mask) + const mem_access_sector_mask_t §or_mask, + gpgpu_context* ctx) : m_warp_mask(active_mask), m_byte_mask(byte_mask), m_sector_mask(sector_mask) { - init(); + init(ctx); m_type = type; m_addr = address; m_req_size = size; @@ -786,13 +788,9 @@ public: } } + gpgpu_context* gpgpu_ctx; private: - void init() - { - m_uid=++sm_next_access_uid; - m_addr=0; - m_req_size=0; - } + void init(gpgpu_context* ctx); unsigned m_uid; new_addr_type m_addr; // request address @@ -802,8 +800,6 @@ private: active_mask_t m_warp_mask; mem_access_byte_mask_t m_byte_mask; mem_access_sector_mask_t m_sector_mask; - - static unsigned sm_next_access_uid; }; class mem_fetch; diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 5e36d4b..d443d79 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -41,7 +41,7 @@ int PRINT_CYCLE = 0; template class fifo_pipeline; template class fifo_pipeline; -dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, memory_stats_t *stats, +dram_t::dram_t( unsigned int partition_id, const memory_config *config, memory_stats_t *stats, memory_partition_unit *mp, gpgpu_sim* gpu ) { id = partition_id; diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 7a3a2da..0bd9725 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -106,11 +106,12 @@ enum bank_grp_bits_position{ }; class mem_fetch; +class memory_config; class dram_t { public: - dram_t( unsigned int parition_id, const struct memory_config *config, class memory_stats_t *stats, + dram_t( unsigned int parition_id, const memory_config *config, class memory_stats_t *stats, class memory_partition_unit *mp, class gpgpu_sim* gpu ); bool full(bool is_write) const; @@ -145,7 +146,7 @@ public: - const struct memory_config *m_config; + const memory_config *m_config; private: bankgrp_t **bkgrp; diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index f1f6e19..1705821 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -1251,7 +1251,8 @@ data_cache::wr_miss_wa_naive( new_addr_type addr, false, // Now performing a read mf->get_access_warp_mask(), mf->get_access_byte_mask(), - mf->get_access_sector_mask()); + mf->get_access_sector_mask(), + m_gpu->gpgpu_ctx); mem_fetch *n_mf = new mem_fetch( *ma, NULL, @@ -1365,7 +1366,8 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, false, // Now performing a read mf->get_access_warp_mask(), mf->get_access_byte_mask(), - mf->get_access_sector_mask()); + mf->get_access_sector_mask(), + m_gpu->gpgpu_ctx); mem_fetch *n_mf = new mem_fetch( *ma, NULL, diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 0481259..0644b44 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1828,7 +1828,7 @@ const shader_core_config * gpgpu_sim::getShaderCoreConfig() return m_shader_config; } -const struct memory_config * gpgpu_sim::getMemoryConfig() +const memory_config * gpgpu_sim::getMemoryConfig() { return m_memory_config; } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 78f0505..19e1eb3 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -33,6 +33,7 @@ #include "../trace.h" #include "addrdec.h" #include "shader.h" +#include "gpu-cache.h" #include #include #include @@ -143,13 +144,14 @@ struct power_config { }; - -struct memory_config { - memory_config() +class memory_config { + public: + memory_config(gpgpu_context* ctx) { m_valid = false; gpgpu_dram_timing_opt=NULL; gpgpu_L2_queue_config=NULL; + gpgpu_ctx = ctx; } void init() { @@ -291,13 +293,14 @@ struct memory_config { unsigned write_high_watermark; unsigned write_low_watermark; bool m_perf_sim_memcpy; + gpgpu_context* gpgpu_ctx; }; extern bool g_interactive_debugger_enabled; class gpgpu_sim_config : public power_config, public gpgpu_functional_sim_config { public: - gpgpu_sim_config(gpgpu_context* ctx): m_shader_config(ctx) { + gpgpu_sim_config(gpgpu_context* ctx): m_shader_config(ctx), m_memory_config(ctx) { m_valid = false; gpgpu_ctx = ctx; } @@ -507,7 +510,7 @@ public: /*! * Returning the memory configuration of the shader core, used by the functional simulation only so far */ - const struct memory_config * getMemoryConfig(); + const memory_config * getMemoryConfig(); //! Get shader core SIMT cluster @@ -567,7 +570,7 @@ private: const struct cudaDeviceProp *m_cuda_properties; const shader_core_config *m_shader_config; - const struct memory_config *m_memory_config; + const memory_config *m_memory_config; // stats class shader_core_stats *m_shader_stats; diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 62e70a7..6540b52 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -49,7 +49,7 @@ mem_fetch * partition_mf_allocator::alloc(new_addr_type addr, mem_access_type type, unsigned size, bool wr, unsigned long long cycle ) const { assert( wr ); - mem_access_t access( type, addr, size, wr ); + mem_access_t access( type, addr, size, wr, m_memory_config->gpgpu_ctx ); mem_fetch *mf = new mem_fetch( access, NULL, WRITE_PACKET_SIZE, @@ -62,7 +62,7 @@ mem_fetch * partition_mf_allocator::alloc(new_addr_type addr, mem_access_type ty } memory_partition_unit::memory_partition_unit( unsigned partition_id, - const struct memory_config *config, + const memory_config *config, class memory_stats_t *stats, class gpgpu_sim* gpu) : m_id(partition_id), m_config(config), m_stats(stats), m_arbitration_metadata(config), m_gpu(gpu) @@ -95,7 +95,7 @@ memory_partition_unit::~memory_partition_unit() delete[] m_sub_partition; } -memory_partition_unit::arbitration_metadata::arbitration_metadata(const struct memory_config *config) +memory_partition_unit::arbitration_metadata::arbitration_metadata(const memory_config *config) : m_last_borrower(config->m_n_sub_partition_per_memory_channel - 1), m_private_credit(config->m_n_sub_partition_per_memory_channel, 0), m_shared_credit(0) @@ -312,7 +312,7 @@ void memory_partition_unit::print( FILE *fp ) const } memory_sub_partition::memory_sub_partition( unsigned sub_partition_id, - const struct memory_config *config, + const memory_config *config, class memory_stats_t *stats, class gpgpu_sim* gpu) { @@ -640,7 +640,8 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques mf->is_write(), mf->get_access_warp_mask(), mf->get_access_byte_mask() & byte_sector_mask, - std::bitset().set(j)); + std::bitset().set(j), + m_gpu->gpgpu_ctx); mem_fetch *n_mf = new mem_fetch( *ma, NULL, diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 8ff2666..1f74c47 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -58,7 +58,7 @@ private: class memory_partition_unit { public: - memory_partition_unit( unsigned partition_id, const struct memory_config *config, class memory_stats_t *stats, class gpgpu_sim* gpu ); + memory_partition_unit( unsigned partition_id, const memory_config *config, class memory_stats_t *stats, class gpgpu_sim* gpu ); ~memory_partition_unit(); bool busy() const; @@ -98,7 +98,7 @@ public: private: unsigned m_id; - const struct memory_config *m_config; + const memory_config *m_config; class memory_stats_t *m_stats; class memory_sub_partition **m_sub_partition; class dram_t *m_dram; @@ -106,7 +106,7 @@ private: class arbitration_metadata { public: - arbitration_metadata(const struct memory_config *config); + arbitration_metadata(const memory_config *config); // check if a subpartition still has credit bool has_credits(int inner_sub_partition_id) const; @@ -130,7 +130,7 @@ private: std::vector m_private_credit; int m_shared_credit; }; - arbitration_metadata m_arbitration_metadata; + arbitration_metadata m_arbitration_metadata; // determine wheither a given subpartition can issue to DRAM bool can_issue_to_dram(int inner_sub_partition_id); @@ -149,7 +149,7 @@ private: class memory_sub_partition { public: - memory_sub_partition( unsigned sub_partition_id, const struct memory_config *config, class memory_stats_t *stats, class gpgpu_sim* gpu ); + memory_sub_partition( unsigned sub_partition_id, const memory_config *config, class memory_stats_t *stats, class gpgpu_sim* gpu ); ~memory_sub_partition(); unsigned get_id() const { return m_id; } @@ -197,7 +197,7 @@ public: private: // data unsigned m_id; //< the global sub partition ID - const struct memory_config *m_config; + const memory_config *m_config; class l2_cache *m_L2cache; class L2interface *m_L2interface; class gpgpu_sim* m_gpu; diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index c9b0484..6a00889 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -39,10 +39,10 @@ mem_fetch::mem_fetch( const mem_access_t &access, unsigned wid, unsigned sid, unsigned tpc, - const struct memory_config *config, + const memory_config *config, unsigned long long cycle, mem_fetch *m_original_mf, - mem_fetch *m_original_wr_mf) + mem_fetch *m_original_wr_mf):m_access(access) { m_request_uid = sm_next_mf_request_uid++; diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index 4eb3a52..1cab9f2 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -47,6 +47,7 @@ enum mf_type { #undef MF_TUP #undef MF_TUP_END +class memory_config; class mem_fetch { public: mem_fetch( const mem_access_t &access, @@ -55,7 +56,7 @@ public: unsigned wid, unsigned sid, unsigned tpc, - const struct memory_config *config, + const memory_config *config, unsigned long long cycle, mem_fetch *original_mf = NULL, mem_fetch *original_wr_mf = NULL); @@ -149,7 +150,7 @@ private: static unsigned sm_next_mf_request_uid; - const struct memory_config *m_mem_config; + const memory_config *m_mem_config; unsigned icnt_flit_size; mem_fetch* original_mf; //this pointer is set up when a request is divided into sector requests at L2 cache (if the req size > L2 sector size), so the pointer refers to the original request diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index d08ba39..4e94991 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -42,7 +42,7 @@ #include #include -memory_stats_t::memory_stats_t( unsigned n_shader, const shader_core_config *shader_config, const struct memory_config *mem_config, const class gpgpu_sim* gpu ) +memory_stats_t::memory_stats_t( unsigned n_shader, const shader_core_config *shader_config, const memory_config *mem_config, const class gpgpu_sim* gpu ) { assert( mem_config->m_valid ); assert( shader_config->m_valid ); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index 6ce568d..0c84972 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -32,11 +32,12 @@ #include #include +class memory_config; class memory_stats_t { public: memory_stats_t( unsigned n_shader, const class shader_core_config *shader_config, - const struct memory_config *mem_config, + const memory_config *mem_config, const class gpgpu_sim* gpu); unsigned memlatstat_done( class mem_fetch *mf ); @@ -54,7 +55,7 @@ public: unsigned m_n_shader; const shader_core_config *m_shader_config; - const struct memory_config *m_memory_config; + const memory_config *m_memory_config; const class gpgpu_sim* m_gpu; unsigned max_mrq_latency; diff --git a/src/gpgpu-sim/power_stat.cc b/src/gpgpu-sim/power_stat.cc index 007b4c6..2c02082 100644 --- a/src/gpgpu-sim/power_stat.cc +++ b/src/gpgpu-sim/power_stat.cc @@ -42,7 +42,7 @@ -power_mem_stat_t::power_mem_stat_t(const struct memory_config *mem_config, const shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats){ +power_mem_stat_t::power_mem_stat_t(const memory_config *mem_config, const shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats){ assert( mem_config->m_valid ); m_mem_stats = mem_stats; m_config = mem_config; @@ -266,7 +266,7 @@ for(unsigned i=0; inum_shader(); ++i){ } } -power_stat_t::power_stat_t( const shader_core_config *shader_config,float * average_pipeline_duty_cycle,float *active_sms,shader_core_stats * shader_stats, const struct memory_config *mem_config,memory_stats_t * memory_stats) +power_stat_t::power_stat_t( const shader_core_config *shader_config,float * average_pipeline_duty_cycle,float *active_sms,shader_core_stats * shader_stats, const memory_config *mem_config,memory_stats_t * memory_stats) { assert( shader_config->m_valid ); assert( mem_config->m_valid ); diff --git a/src/gpgpu-sim/power_stat.h b/src/gpgpu-sim/power_stat.h index 91fade9..24ade99 100644 --- a/src/gpgpu-sim/power_stat.h +++ b/src/gpgpu-sim/power_stat.h @@ -113,7 +113,7 @@ struct mem_power_stats_pod{ class power_mem_stat_t : public mem_power_stats_pod{ public: - power_mem_stat_t(const struct memory_config *mem_config, const shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats); + power_mem_stat_t(const memory_config *mem_config, const shader_core_config *shdr_config, memory_stats_t *mem_stats, shader_core_stats *shdr_stats); void visualizer_print( gzFile visualizer_file ); void print (FILE *fout) const; void init(); @@ -128,7 +128,7 @@ private: class power_stat_t { public: - power_stat_t( const shader_core_config *shader_config,float * average_pipeline_duty_cycle,float * active_sms,shader_core_stats * shader_stats, const struct memory_config *mem_config,memory_stats_t * memory_stats); + power_stat_t( const shader_core_config *shader_config,float * average_pipeline_duty_cycle,float * active_sms,shader_core_stats * shader_stats, const memory_config *mem_config,memory_stats_t * memory_stats); void visualizer_print( gzFile visualizer_file ); void print (FILE *fout) const; void save_stats(){ @@ -621,7 +621,7 @@ public: float * m_average_pipeline_duty_cycle; float * m_active_sms; const shader_core_config *m_config; - const struct memory_config *m_mem_config; + const memory_config *m_mem_config; }; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index f380560..b7ae95d 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -28,7 +28,6 @@ #include #include "shader.h" -#include "gpu-sim.h" #include "addrdec.h" #include "dram.h" #include "stat-tool.h" @@ -53,6 +52,19 @@ #define MIN(a,b) (((a)<(b))?(a):(b)) +mem_fetch *shader_core_mem_fetch_allocator::alloc( new_addr_type addr, mem_access_type type, unsigned size, bool wr, unsigned long long cycle ) const +{ + mem_access_t access( type, addr, size, wr, m_memory_config->gpgpu_ctx); + mem_fetch *mf = new mem_fetch( access, + NULL, + wr?WRITE_PACKET_SIZE:READ_PACKET_SIZE, + -1, + m_core_id, + m_cluster_id, + m_memory_config, + cycle); + return mf; +} ///////////////////////////////////////////////////////////////////////////// std::list shader_core_ctx::get_regs_written( const inst_t &fvt ) const @@ -71,7 +83,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, unsigned shader_id, unsigned tpc_id, const shader_core_config *config, - const struct memory_config *mem_config, + const memory_config *mem_config, shader_core_stats *stats ) : core_t( gpu, NULL, config->warp_size, config->n_thread_per_shader ), m_barriers( this, config->max_warps_per_shader, config->max_cta_per_core, config->max_barriers_per_cta, config->warp_size ), @@ -809,7 +821,7 @@ void shader_core_ctx::fetch() // TODO: replace with use of allocator // mem_fetch *mf = m_mem_fetch_allocator->alloc() - mem_access_t acc(INST_ACC_R,ppc,nbytes,false); + mem_access_t acc(INST_ACC_R,ppc,nbytes,false, m_gpu->gpgpu_ctx); mem_fetch *mf = new mem_fetch(acc, NULL/*we don't have an instruction yet*/, READ_PACKET_SIZE, @@ -3787,7 +3799,7 @@ void opndcoll_rfu_t::collector_unit_t::dispatch() simt_core_cluster::simt_core_cluster( class gpgpu_sim *gpu, unsigned cluster_id, const shader_core_config *config, - const struct memory_config *mem_config, + const memory_config *mem_config, shader_core_stats *stats, class memory_stats_t *mstats ) { diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 2837f1b..b0d7f7f 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1727,6 +1727,7 @@ private: friend class LooseRoundRobbinScheduler; }; +class memory_config; class shader_core_mem_fetch_allocator : public mem_fetch_allocator { public: shader_core_mem_fetch_allocator( unsigned core_id, unsigned cluster_id, const memory_config *config ) @@ -1735,20 +1736,7 @@ public: m_cluster_id = cluster_id; m_memory_config = config; } - mem_fetch *alloc( new_addr_type addr, mem_access_type type, unsigned size, bool wr, unsigned long long cycle ) const - { - mem_access_t access( type, addr, size, wr ); - mem_fetch *mf = new mem_fetch( access, - NULL, - wr?WRITE_PACKET_SIZE:READ_PACKET_SIZE, - -1, - m_core_id, - m_cluster_id, - m_memory_config, - cycle); - return mf; - } - + mem_fetch *alloc( new_addr_type addr, mem_access_type type, unsigned size, bool wr, unsigned long long cycle ) const; mem_fetch *alloc( const warp_inst_t &inst, const mem_access_t &access, unsigned long long cycle ) const { warp_inst_t inst_copy = inst; @@ -1777,7 +1765,7 @@ public: unsigned shader_id, unsigned tpc_id, const shader_core_config *config, - const struct memory_config *mem_config, + const memory_config *mem_config, shader_core_stats *stats ); // used by simt_core_cluster: @@ -2072,7 +2060,7 @@ public: simt_core_cluster( class gpgpu_sim *gpu, unsigned cluster_id, const shader_core_config *config, - const struct memory_config *mem_config, + const memory_config *mem_config, shader_core_stats *stats, memory_stats_t *mstats ); -- cgit v1.3 From c4782ba3c78e6e97db9f56780991dc5cf827c755 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Sun, 14 Jul 2019 13:06:46 -0400 Subject: Move warp_inst_sm_next_uid Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/abstract_hardware_model.cc | 15 +++++++++++++-- src/abstract_hardware_model.h | 18 +++--------------- 3 files changed, 18 insertions(+), 17 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 07473be..0dd6311 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -12,6 +12,7 @@ class gpgpu_context { gpgpu_context() { g_global_allfiles_symbol_table = NULL; sm_next_access_uid=0; + warp_inst_sm_next_uid=0; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -23,6 +24,7 @@ class gpgpu_context { symbol_table *g_global_allfiles_symbol_table; const char *g_filename; unsigned sm_next_access_uid; + unsigned warp_inst_sm_next_uid; // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index fe10daa..1dddecd 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -41,8 +41,6 @@ #include #include "../libcuda/gpgpu_context.h" -unsigned warp_inst_t::sm_next_uid = 0; - void mem_access_t::init(gpgpu_context* ctx) { gpgpu_ctx = ctx; @@ -50,6 +48,19 @@ void mem_access_t::init(gpgpu_context* ctx) m_addr=0; m_req_size=0; } +void warp_inst_t::issue( const active_mask_t &mask, unsigned warp_id, unsigned long long cycle, int dynamic_warp_id, int sch_id ) +{ + m_warp_active_mask = mask; + m_warp_issued_mask = mask; + m_uid = ++(m_config->gpgpu_ctx->warp_inst_sm_next_uid); + m_warp_id = warp_id; + m_dynamic_warp_id = dynamic_warp_id; + issue_cycle = cycle; + cycles = initiation_interval; + m_cache_hit=false; + m_empty=false; + m_scheduler_id=sch_id; +} checkpoint::checkpoint() { diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index a22c8c3..519416d 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -958,19 +958,9 @@ public: { m_empty=true; } - void issue( const active_mask_t &mask, unsigned warp_id, unsigned long long cycle, int dynamic_warp_id, int sch_id ) - { - m_warp_active_mask = mask; - m_warp_issued_mask = mask; - m_uid = ++sm_next_uid; - m_warp_id = warp_id; - m_dynamic_warp_id = dynamic_warp_id; - issue_cycle = cycle; - cycles = initiation_interval; - m_cache_hit=false; - m_empty=false; - m_scheduler_id=sch_id; - } + + void issue( const active_mask_t &mask, unsigned warp_id, unsigned long long cycle, int dynamic_warp_id, int sch_id ); + const active_mask_t & get_active_mask() const { return m_warp_active_mask; @@ -1133,8 +1123,6 @@ protected: bool m_mem_accesses_created; std::list m_accessq; - static unsigned sm_next_uid; - unsigned m_scheduler_id; //the scheduler that issues this inst //Jin: cdp support -- cgit v1.3 From 67950eb12eec495b1976934acc763db481d955f8 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 11:58:10 -0400 Subject: Move operand_info::sm_next_uid Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/cuda-sim/ptx_ir.cc | 20 ++++++++++------- src/cuda-sim/ptx_ir.h | 54 ++++++++++++++++++++++++---------------------- src/cuda-sim/ptx_parser.cc | 30 +++++++++++++------------- src/cuda-sim/ptx_parser.h | 2 +- src/cuda-sim/ptx_sim.cc | 8 +++---- 6 files changed, 62 insertions(+), 54 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 0dd6311..9c5ae7b 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -13,6 +13,7 @@ class gpgpu_context { g_global_allfiles_symbol_table = NULL; sm_next_access_uid=0; warp_inst_sm_next_uid=0; + operand_info_sm_next_uid = 1; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -25,6 +26,7 @@ class gpgpu_context { const char *g_filename; unsigned sm_next_access_uid; unsigned warp_inst_sm_next_uid; + unsigned operand_info_sm_next_uid;//uid for operand_info // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index c537091..0fa29eb 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -39,6 +39,7 @@ typedef void * yyscan_t; #include "assert.h" #include "cuda-sim.h" +#include "../../libcuda/gpgpu_context.h" #define STR_SIZE 1024 @@ -322,11 +323,9 @@ void symbol_table::dump() printf("\n"); } -unsigned operand_info::sm_next_uid=1; - unsigned operand_info::get_uid() { - unsigned result = sm_next_uid++; + unsigned result = (gpgpu_ctx->operand_info_sm_next_uid)++; return result; } @@ -1015,7 +1014,7 @@ void copy_buffer_to_frame(ptx_thread_info * thread, const arg_buffer_t &a) { if( a.is_reg() ) { ptx_reg_t value = a.get_reg(); - operand_info dst_reg = operand_info(a.get_dst()); + operand_info dst_reg = operand_info(a.get_dst(), thread->get_gpu()->gpgpu_ctx); thread->set_reg(dst_reg.get_symbol(),value); } else { const void *buffer = a.get_param_buffer(); @@ -1039,7 +1038,8 @@ void copy_buffer_list_into_frame(ptx_thread_info * thread, arg_buffer_list_t &ar static std::list check_operands( int opcode, const std::list &scalar_type, - const std::list &operands ) + const std::list &operands, + gpgpu_context* ctx) { static int g_warn_literal_operands_two_type_inst; if( (opcode == CVT_OP) || (opcode == SET_OP) || (opcode == SLCT_OP) || (opcode == TEX_OP) || (opcode==MMA_OP) || (opcode == DP4A_OP)) { @@ -1066,7 +1066,7 @@ static std::list check_operands( int opcode, if( (op.get_type() == double_op_t) && (inst_type == F32_TYPE) ) { ptx_reg_t v = op.get_literal_value(); float u = (float)v.f64; - operand_info n(u); + operand_info n(u, ctx); result.push_back(n); } else { result.push_back(op); @@ -1097,7 +1097,7 @@ ptx_instruction::ptx_instruction( int opcode, unsigned line, const char *source, const core_config *config, - gpgpu_context* ctx ) : warp_inst_t(config) + gpgpu_context* ctx ) : warp_inst_t(config), m_return_var(ctx) { gpgpu_ctx = ctx; m_uid = ++g_num_ptx_inst_uid; @@ -1107,7 +1107,7 @@ ptx_instruction::ptx_instruction( int opcode, m_neg_pred = neg_pred; m_pred_mod = pred_mod; m_label = label; - const std::list checked_operands = check_operands(opcode,scalar_type,operands); + const std::list checked_operands = check_operands(opcode,scalar_type,operands, ctx); m_operands.insert(m_operands.begin(), checked_operands.begin(), checked_operands.end() ); m_return_var = return_var; m_options = options; @@ -1371,6 +1371,10 @@ std::string ptx_instruction::to_string() const m_source.c_str() ); return std::string( buf ); } +operand_info ptx_instruction::get_pred() const +{ + return operand_info( m_pred, gpgpu_ctx); +} unsigned function_info::sm_next_uid = 1; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index babd54b..8ed94d9 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -377,9 +377,9 @@ private: class operand_info { public: - operand_info() + operand_info(gpgpu_context* ctx) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -392,9 +392,9 @@ public: m_addr_offset = 0; m_value.m_symbolic=NULL; } - operand_info( const symbol *addr ) + operand_info( const symbol *addr, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -435,9 +435,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( const symbol *addr1, const symbol *addr2 ) + operand_info( const symbol *addr1, const symbol *addr2, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -462,9 +462,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( int builtin_id, int dim_mod ) + operand_info( int builtin_id, int dim_mod, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -481,9 +481,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( const symbol *addr, int offset ) + operand_info( const symbol *addr, int offset, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -500,9 +500,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( unsigned x ) + operand_info( unsigned x, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -519,9 +519,9 @@ public: m_is_return_var = false; m_immediate_address=true; } - operand_info( int x ) + operand_info( int x, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -538,9 +538,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( float x ) + operand_info( float x, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -557,9 +557,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( double x ) + operand_info( double x, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -576,9 +576,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( const symbol *s1, const symbol *s2, const symbol *s3, const symbol *s4 ) + operand_info( const symbol *s1, const symbol *s2, const symbol *s3, const symbol *s4, gpgpu_context* ctx ) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -603,9 +603,9 @@ public: m_is_return_var = false; m_immediate_address=false; } - operand_info( const symbol *s1, const symbol *s2, const symbol *s3, const symbol *s4 ,const symbol *s5,const symbol *s6,const symbol *s7, const symbol *s8) + operand_info( const symbol *s1, const symbol *s2, const symbol *s3, const symbol *s4 ,const symbol *s5,const symbol *s6,const symbol *s7, const symbol *s8, gpgpu_context* ctx) { - init(); + init(ctx); m_is_non_arch_reg = false; m_addr_space = undefined_space; m_operand_lohi = 0; @@ -631,8 +631,9 @@ public: m_immediate_address=false; } - void init() + void init(gpgpu_context* ctx) { + gpgpu_ctx = ctx; m_uid=(unsigned)-1; m_valid=false; m_vector=false; @@ -842,6 +843,7 @@ public: bool is_non_arch_reg() const { return m_is_non_arch_reg; } private: + gpgpu_context* gpgpu_ctx; unsigned m_uid; bool m_valid; bool m_vector; @@ -957,7 +959,7 @@ public: unsigned source_line() const { return m_source_line;} unsigned get_num_operands() const { return m_operands.size();} bool has_pred() const { return m_pred != NULL;} - operand_info get_pred() const { return operand_info( m_pred );} + operand_info get_pred() const; bool get_pred_neg() const { return m_neg_pred;} int get_pred_mod() const { return m_pred_mod;} const char *get_source() const { return m_source.c_str();} @@ -1448,14 +1450,14 @@ private: class arg_buffer_t { public: - arg_buffer_t() + arg_buffer_t(gpgpu_context* ctx) : m_src_op(ctx) { m_is_reg=false; m_is_param=false; m_param_value=NULL; m_reg_value=ptx_reg_t(); } - arg_buffer_t( const arg_buffer_t &another ) + arg_buffer_t( const arg_buffer_t &another, gpgpu_context* ctx ) : m_src_op(ctx) { make_copy(another); } diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 1d38edf..81b70af 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -102,7 +102,7 @@ void ptx_recognizer::init_instruction_state() g_opcode = -1; g_options.clear(); g_wmma_options.clear(); - g_return_var = operand_info(); + g_return_var = operand_info(gpgpu_ctx); init_directive_state(); } @@ -689,7 +689,7 @@ void ptx_recognizer::add_double_operand( const char *d1, const char *d2 ) const symbol *s1 = g_current_symbol_table->lookup(d1); const symbol *s2 = g_current_symbol_table->lookup(d2); parse_assert( s1 != NULL && s2 != NULL, "component(s) missing declarations."); - g_operands.push_back( operand_info(s1,s2) ); + g_operands.push_back( operand_info(s1,s2,gpgpu_ctx) ); } void ptx_recognizer::add_1vector_operand( const char *d1 ) @@ -698,7 +698,7 @@ void ptx_recognizer::add_1vector_operand( const char *d1 ) PTX_PARSE_DPRINTF("add_1vector_operand"); const symbol *s1 = g_current_symbol_table->lookup(d1); parse_assert( s1 != NULL, "component(s) missing declarations."); - g_operands.push_back( operand_info(s1,NULL,NULL,NULL) ); + g_operands.push_back( operand_info(s1,NULL,NULL,NULL,gpgpu_ctx) ); } void ptx_recognizer::add_2vector_operand( const char *d1, const char *d2 ) @@ -707,7 +707,7 @@ void ptx_recognizer::add_2vector_operand( const char *d1, const char *d2 ) const symbol *s1 = g_current_symbol_table->lookup(d1); const symbol *s2 = g_current_symbol_table->lookup(d2); parse_assert( s1 != NULL && s2 != NULL, "v2 component(s) missing declarations."); - g_operands.push_back( operand_info(s1,s2,NULL,NULL) ); + g_operands.push_back( operand_info(s1,s2,NULL,NULL,gpgpu_ctx) ); } void ptx_recognizer::add_3vector_operand( const char *d1, const char *d2, const char *d3 ) @@ -717,7 +717,7 @@ void ptx_recognizer::add_3vector_operand( const char *d1, const char *d2, const const symbol *s2 = g_current_symbol_table->lookup(d2); const symbol *s3 = g_current_symbol_table->lookup(d3); parse_assert( s1 != NULL && s2 != NULL && s3 != NULL, "v3 component(s) missing declarations."); - g_operands.push_back( operand_info(s1,s2,s3,NULL) ); + g_operands.push_back( operand_info(s1,s2,s3,NULL,gpgpu_ctx) ); } void ptx_recognizer::add_4vector_operand( const char *d1, const char *d2, const char *d3, const char *d4 ) @@ -732,7 +732,7 @@ void ptx_recognizer::add_4vector_operand( const char *d1, const char *d2, const if ( s2 == null_op ) s2 = NULL; if ( s3 == null_op ) s3 = NULL; if ( s4 == null_op ) s4 = NULL; - g_operands.push_back( operand_info(s1,s2,s3,s4) ); + g_operands.push_back( operand_info(s1,s2,s3,s4,gpgpu_ctx) ); } void ptx_recognizer::add_8vector_operand( const char *d1, const char *d2, const char *d3, const char *d4,const char *d5,const char *d6,const char *d7,const char *d8 ) { @@ -754,13 +754,13 @@ void ptx_recognizer::add_8vector_operand( const char *d1, const char *d2, const if ( s6 == null_op ) s6 = NULL; if ( s7 == null_op ) s7 = NULL; if ( s8 == null_op ) s8 = NULL; - g_operands.push_back( operand_info(s1,s2,s3,s4,s5,s6,s7,s8) ); + g_operands.push_back( operand_info(s1,s2,s3,s4,s5,s6,s7,s8,gpgpu_ctx) ); } void ptx_recognizer::add_builtin_operand( int builtin, int dim_modifier ) { PTX_PARSE_DPRINTF("add_builtin_operand"); - g_operands.push_back( operand_info(builtin,dim_modifier) ); + g_operands.push_back( operand_info(builtin,dim_modifier,gpgpu_ctx) ); } void ptx_recognizer::add_memory_operand() @@ -883,19 +883,19 @@ void ptx_recognizer::change_operand_neg( ) void ptx_recognizer::add_literal_int( int value ) { PTX_PARSE_DPRINTF("add_literal_int"); - g_operands.push_back( operand_info(value) ); + g_operands.push_back( operand_info(value,gpgpu_ctx) ); } void ptx_recognizer::add_literal_float( float value ) { PTX_PARSE_DPRINTF("add_literal_float"); - g_operands.push_back( operand_info(value) ); + g_operands.push_back( operand_info(value,gpgpu_ctx) ); } void ptx_recognizer::add_literal_double( double value ) { PTX_PARSE_DPRINTF("add_literal_double"); - g_operands.push_back( operand_info(value) ); + g_operands.push_back( operand_info(value,gpgpu_ctx) ); } void ptx_recognizer::add_scalar_operand( const char *identifier ) @@ -911,7 +911,7 @@ void ptx_recognizer::add_scalar_operand( const char *identifier ) parse_error( msg.c_str() ); } } - g_operands.push_back( operand_info(s) ); + g_operands.push_back( operand_info(s,gpgpu_ctx) ); } void ptx_recognizer::add_neg_pred_operand( const char *identifier ) @@ -921,7 +921,7 @@ void ptx_recognizer::add_neg_pred_operand( const char *identifier ) if ( s == NULL ) { s = g_current_symbol_table->add_variable(identifier,NULL,1,gpgpu_ctx->g_filename,ptx_get_lineno(scanner)); } - operand_info op(s); + operand_info op(s, gpgpu_ctx); op.set_neg_pred(); g_operands.push_back( op ); } @@ -934,13 +934,13 @@ void ptx_recognizer::add_address_operand( const char *identifier, int offset ) std::string msg = std::string("operand \"") + identifier + "\" has no declaration."; parse_error( msg.c_str() ); } - g_operands.push_back( operand_info(s,offset) ); + g_operands.push_back( operand_info(s,offset,gpgpu_ctx) ); } void ptx_recognizer::add_address_operand2( int offset ) { PTX_PARSE_DPRINTF("add_address_operand"); - g_operands.push_back( operand_info((unsigned)offset) ); + g_operands.push_back( operand_info((unsigned)offset,gpgpu_ctx) ); } void ptx_recognizer::add_array_initializer() diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index 1e559a2..11a3d20 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -35,7 +35,7 @@ class gpgpu_context; typedef void * yyscan_t; class ptx_recognizer { public: - ptx_recognizer( gpgpu_context* ctx ) { + ptx_recognizer( gpgpu_context* ctx ) : g_return_var(ctx) { scanner = NULL; g_size = -1; g_add_identifier_cached__identifier = NULL; diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index b6cf4bd..62bb529 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -422,9 +422,9 @@ bool ptx_thread_info::callstack_pop() assert( !((rv_src != NULL) ^ (rv_dst != NULL)) ); // ensure caller and callee agree on whether there is a return value // read return value from callee frame - arg_buffer_t buffer; + arg_buffer_t buffer(m_gpu->gpgpu_ctx); if( rv_src != NULL ) - buffer = copy_arg_to_buffer(this, operand_info(rv_src), rv_dst ); + buffer = copy_arg_to_buffer(this, operand_info(rv_src, m_gpu->gpgpu_ctx), rv_dst ); m_symbol_table = m_callstack.back().m_symbol_table; m_NPC = m_callstack.back().m_PC; @@ -456,9 +456,9 @@ bool ptx_thread_info::callstack_pop_plus() assert( !((rv_src != NULL) ^ (rv_dst != NULL)) ); // ensure caller and callee agree on whether there is a return value // read return value from callee frame - arg_buffer_t buffer; + arg_buffer_t buffer(m_gpu->gpgpu_ctx); if( rv_src != NULL ) - buffer = copy_arg_to_buffer(this, operand_info(rv_src), rv_dst ); + buffer = copy_arg_to_buffer(this, operand_info(rv_src, m_gpu->gpgpu_ctx), rv_dst ); m_symbol_table = m_callstack.back().m_symbol_table; m_NPC = m_callstack.back().m_PC; -- cgit v1.3 From bd170826a00fe1ea1960ff2e23247ed3f980547b Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 12:16:01 -0400 Subject: Move kernel_info_t::m_next_uid Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/abstract_hardware_model.cc | 3 +-- src/abstract_hardware_model.h | 1 - 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 9c5ae7b..1e20c62 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -14,6 +14,7 @@ class gpgpu_context { sm_next_access_uid=0; warp_inst_sm_next_uid=0; operand_info_sm_next_uid = 1; + kernel_info_m_next_uid = 1; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -27,6 +28,7 @@ class gpgpu_context { unsigned sm_next_access_uid; unsigned warp_inst_sm_next_uid; unsigned operand_info_sm_next_uid;//uid for operand_info + unsigned kernel_info_m_next_uid;//uid for kernel_info_t // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 1dddecd..733d602 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -709,7 +709,6 @@ void warp_inst_t::completed( unsigned long long cycle ) const ptx_file_line_stats_add_latency(pc, latency * active_count()); } -unsigned kernel_info_t::m_next_uid = 1; /*A snapshot of the texture mappings needs to be stored in the kernel's info as kernels should use the texture bindings seen at the time of launch and textures @@ -724,7 +723,7 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * m_next_cta.z=0; m_next_tid=m_next_cta; m_num_cores_running=0; - m_uid = m_next_uid++; + m_uid = (entry->gpgpu_ctx->kernel_info_m_next_uid)++; m_param_mem = new memory_space_impl<8192>("param",64*1024); //Jin: parent and child kernel management for CDP diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 519416d..d13b8c6 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -299,7 +299,6 @@ private: class function_info *m_kernel_entry; unsigned m_uid; - static unsigned m_next_uid; //These maps contain the snapshot of the texture mappings at kernel launch std::map m_NameToCudaArray; -- cgit v1.3 From 36aaec0a30445533f04aaf6b55c5ef135e20484a Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 13:31:36 -0400 Subject: Move ptx_line_stats_filename and enable_ptx_file_line_stats Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 3 +++ src/cuda-sim/ptx-stats.cc | 9 +++------ src/cuda-sim/ptx-stats.h | 23 ++++++++++++++++------- src/gpgpu-sim/gpu-sim.cc | 4 ++-- 4 files changed, 24 insertions(+), 15 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 1e20c62..337ebb2 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -6,6 +6,7 @@ #include "../src/gpgpusim_entrypoint.h" #include "../src/cuda-sim/cuda-sim.h" #include "../src/cuda-sim/cuda_device_runtime.h" +#include "../src/cuda-sim/ptx-stats.h" class gpgpu_context { public: @@ -21,6 +22,7 @@ class gpgpu_context { the_gpgpusim = new GPGPUsim_ctx(this); func_sim = new cuda_sim(this); device_runtime = new cuda_device_runtime(this); + stats = new ptx_stats(this); } // global list symbol_table *g_global_allfiles_symbol_table; @@ -36,6 +38,7 @@ class gpgpu_context { GPGPUsim_ctx* the_gpgpusim; cuda_sim* func_sim; cuda_device_runtime* device_runtime; + ptx_stats* stats; // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); diff --git a/src/cuda-sim/ptx-stats.cc b/src/cuda-sim/ptx-stats.cc index 0a9f08d..298729f 100644 --- a/src/cuda-sim/ptx-stats.cc +++ b/src/cuda-sim/ptx-stats.cc @@ -32,12 +32,9 @@ #include #include #include "../tr1_hash_map.h" +#include "../../libcuda/gpgpu_context.h" -// options -bool enable_ptx_file_line_stats; -char * ptx_line_stats_filename = NULL; - -void ptx_file_line_stats_options(option_parser_t opp) +void ptx_stats::ptx_file_line_stats_options(option_parser_t opp) { option_parser_register(opp, "-enable_ptx_file_line_stats", OPT_BOOL, &enable_ptx_file_line_stats, @@ -118,7 +115,7 @@ typedef tr1_hash_map ptx static ptx_file_line_stats_map_t ptx_file_line_stats_tracker; // output statistics to a file -void ptx_file_line_stats_write_file() +void ptx_stats::ptx_file_line_stats_write_file() { // check if stat collection is turned on if (enable_ptx_file_line_stats == 0) return; diff --git a/src/cuda-sim/ptx-stats.h b/src/cuda-sim/ptx-stats.h index 4fc6599..c75fc58 100644 --- a/src/cuda-sim/ptx-stats.h +++ b/src/cuda-sim/ptx-stats.h @@ -29,13 +29,6 @@ #include "../option_parser.h" -extern bool enable_ptx_file_line_stats; - -// set options -void ptx_file_line_stats_options(option_parser_t opp); - -// output stats to a file -void ptx_file_line_stats_write_file(); #ifdef __cplusplus // stat collection interface to cuda-sim @@ -56,3 +49,19 @@ void ptx_file_line_stats_commit_exposed_latency(int sc_id, int exposed_latency); void ptx_file_line_stats_add_warp_divergence(unsigned pc, unsigned n_way_divergence); +class gpgpu_context; +class ptx_stats { + public: + ptx_stats(gpgpu_context* ctx) { + ptx_line_stats_filename = NULL; + gpgpu_ctx = ctx; + } + char * ptx_line_stats_filename; + bool enable_ptx_file_line_stats; + gpgpu_context* gpgpu_ctx; + // set options + void ptx_file_line_stats_options(option_parser_t opp); + + // output stats to a file + void ptx_file_line_stats_write_file(); +}; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 0644b44..bdf989a 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -545,7 +545,7 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-trace_sampling_memory_partition", OPT_INT32, &Trace::sampling_memory_partition, "The memory partition which is printed using MEMPART_DPRINTF. Default -1 (i.e. all)", "-1"); - ptx_file_line_stats_options(opp); + gpgpu_ctx->stats->ptx_file_line_stats_options(opp); //Jin: kernel launch latency option_parser_register(opp, "-gpgpu_kernel_launch_latency", OPT_INT32, @@ -932,7 +932,7 @@ void gpgpu_sim::update_stats() { void gpgpu_sim::print_stats() { - ptx_file_line_stats_write_file(); + gpgpu_ctx->stats->ptx_file_line_stats_write_file(); gpu_print_stat(); if (g_network_mode) { -- cgit v1.3 From 3b88f9db669d8ec9ffc477352dcfd2a3d423781f Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 13:38:38 -0400 Subject: Move g_num_ptx_inst_uid Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/cuda-sim/instructions.cc | 1 - src/cuda-sim/ptx_ir.cc | 2 +- src/cuda-sim/ptx_ir.h | 2 -- 4 files changed, 3 insertions(+), 4 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 337ebb2..909f267 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -16,6 +16,7 @@ class gpgpu_context { warp_inst_sm_next_uid=0; operand_info_sm_next_uid = 1; kernel_info_m_next_uid = 1; + g_num_ptx_inst_uid = 0; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -31,6 +32,7 @@ class gpgpu_context { unsigned warp_inst_sm_next_uid; unsigned operand_info_sm_next_uid;//uid for operand_info unsigned kernel_info_m_next_uid;//uid for kernel_info_t + unsigned g_num_ptx_inst_uid; //uid for ptx inst inside ptx_instruction // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 565340c..4d4a80d 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -60,7 +60,6 @@ class ptx_recognizer; using half_float::half; -unsigned ptx_instruction::g_num_ptx_inst_uid=0; bool debug_tensorcore = 0; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 0fa29eb..849cc5d 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -1100,7 +1100,7 @@ ptx_instruction::ptx_instruction( int opcode, gpgpu_context* ctx ) : warp_inst_t(config), m_return_var(ctx) { gpgpu_ctx = ctx; - m_uid = ++g_num_ptx_inst_uid; + m_uid = ++(ctx->g_num_ptx_inst_uid); m_PC = 0; m_opcode = opcode; m_pred = pred; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 8ed94d9..1ffbebc 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -878,7 +878,6 @@ private: }; extern const char *g_opcode_string[]; -extern unsigned g_num_ptx_inst_uid; struct basic_block_t { basic_block_t( unsigned ID, ptx_instruction *begin, ptx_instruction *end, bool entry, bool ex) { @@ -1194,7 +1193,6 @@ private: virtual void pre_decode(); friend class function_info; - static unsigned g_num_ptx_inst_uid; // backward pointer class gpgpu_context* gpgpu_ctx; }; -- cgit v1.3 From 971722f51189e80034a9c80a4846a4ec045f0d4d Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 13:57:48 -0400 Subject: Move g_ptx_cta_info_uid and symbol::sm_next_uid Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 4 ++++ src/cuda-sim/ptx_ir.cc | 8 +++----- src/cuda-sim/ptx_ir.h | 6 +++--- src/cuda-sim/ptx_sim.cc | 3 +-- 4 files changed, 11 insertions(+), 10 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 909f267..e168850 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -17,6 +17,8 @@ class gpgpu_context { operand_info_sm_next_uid = 1; kernel_info_m_next_uid = 1; g_num_ptx_inst_uid = 0; + g_ptx_cta_info_uid = 1; + symbol_sm_next_uid = 1; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -33,6 +35,8 @@ class gpgpu_context { unsigned operand_info_sm_next_uid;//uid for operand_info unsigned kernel_info_m_next_uid;//uid for kernel_info_t unsigned g_num_ptx_inst_uid; //uid for ptx inst inside ptx_instruction + unsigned long long g_ptx_cta_info_uid; + unsigned symbol_sm_next_uid; //uid for symbol // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 849cc5d..3719fb2 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -43,11 +43,9 @@ typedef void * yyscan_t; #define STR_SIZE 1024 -unsigned symbol::sm_next_uid = 1; - unsigned symbol::get_uid() { - unsigned result = sm_next_uid++; + unsigned result = (gpgpu_ctx->symbol_sm_next_uid)++; return result; } @@ -152,7 +150,7 @@ symbol *symbol_table::add_variable( const char *identifier, const type_info *typ std::string key(identifier); assert( m_symbols.find(key) == m_symbols.end() ); snprintf(buf,1024,"%s:%u",filename,line); - symbol *s = new symbol(identifier,type,buf,size); + symbol *s = new symbol(identifier,type,buf,size,gpgpu_ctx); m_symbols[ key ] = s; if ( type != NULL && type->get_key().is_global() ) { @@ -173,7 +171,7 @@ void symbol_table::add_function( function_info *func, const char *filename, unsi char buf[1024]; snprintf(buf,1024,"%s:%u",filename,linenumber); type_info *type = add_type( func ); - symbol *s = new symbol(func->get_name().c_str(),type,buf,0); + symbol *s = new symbol(func->get_name().c_str(),type,buf,0,gpgpu_ctx); s->set_function(func); m_symbols[ func->get_name() ] = s; } diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 1ffbebc..2fbda1c 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -152,8 +152,9 @@ class operand_info; class symbol { public: - symbol( const char *name, const type_info *type, const char *location, unsigned size ) + symbol( const char *name, const type_info *type, const char *location, unsigned size, gpgpu_context* ctx ) { + gpgpu_ctx = ctx; m_uid = get_uid(); m_name = name; m_decl_location = location; @@ -273,6 +274,7 @@ public: unsigned uid() const { return m_uid; } private: + gpgpu_context* gpgpu_ctx; unsigned get_uid(); unsigned m_uid; const type_info *m_type; @@ -299,8 +301,6 @@ private: bool m_reg_num_valid; std::list m_initializer; - static unsigned sm_next_uid; - }; class symbol_table { diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 62bb529..949ee66 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -37,7 +37,6 @@ typedef void * yyscan_t; void feature_not_implemented( const char *f ); -unsigned long long g_ptx_cta_info_uid = 1; ptx_cta_info::ptx_cta_info( unsigned sm_idx, gpgpu_context* ctx ) { @@ -45,7 +44,7 @@ ptx_cta_info::ptx_cta_info( unsigned sm_idx, gpgpu_context* ctx ) ctx->func_sim->g_ptx_cta_info_sm_idx_used.insert(sm_idx); m_sm_idx = sm_idx; - m_uid = g_ptx_cta_info_uid++; + m_uid = (ctx->g_ptx_cta_info_uid)++; m_bar_threads = 0; gpgpu_ctx = ctx; } -- cgit v1.3 From 6ec4563b11fe99e539eb83412acaadc6f67b05ba Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 14:04:43 -0400 Subject: Move function_info::sm_next_id Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/cuda-sim/ptx_ir.cc | 3 +-- src/cuda-sim/ptx_ir.h | 2 -- 3 files changed, 3 insertions(+), 4 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index e168850..ed4f746 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -19,6 +19,7 @@ class gpgpu_context { g_num_ptx_inst_uid = 0; g_ptx_cta_info_uid = 1; symbol_sm_next_uid = 1; + function_info_sm_next_uid = 1; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -37,6 +38,7 @@ class gpgpu_context { unsigned g_num_ptx_inst_uid; //uid for ptx inst inside ptx_instruction unsigned long long g_ptx_cta_info_uid; unsigned symbol_sm_next_uid; //uid for symbol + unsigned function_info_sm_next_uid; // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 3719fb2..5fa9379 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -1374,12 +1374,11 @@ operand_info ptx_instruction::get_pred() const return operand_info( m_pred, gpgpu_ctx); } -unsigned function_info::sm_next_uid = 1; function_info::function_info(int entry_point, gpgpu_context* ctx ) { gpgpu_ctx = ctx; - m_uid = sm_next_uid++; + m_uid = (gpgpu_ctx->function_info_sm_next_uid)++; m_entry_point = (entry_point==1)?true:false; m_extern = (entry_point==2)?true:false; num_reconvergence_pairs = 0; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 2fbda1c..8fc0a06 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -873,7 +873,6 @@ private: bool m_is_return_var; bool m_is_non_arch_reg; - static unsigned sm_next_uid; unsigned get_uid(); }; @@ -1438,7 +1437,6 @@ private: symbol_table *m_symtab; static std::vector s_g_pc_to_insn; // a direct mapping from PC to instruction - static unsigned sm_next_uid; //parameter size for device kernels int m_args_aligned_size; -- cgit v1.3 From 352d2a3336b1c8e5258ca9d92d214973e98837c0 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 16:17:07 -0400 Subject: Move s_g_pc_to_insn Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 5 +++++ src/abstract_hardware_model.cc | 10 +++++----- src/cuda-sim/cuda-sim.cc | 20 +++++++++----------- src/cuda-sim/cuda-sim.h | 2 -- src/cuda-sim/ptx-stats.cc | 28 ++++++++++++++-------------- src/cuda-sim/ptx-stats.h | 15 ++++++++------- src/cuda-sim/ptx_ir.cc | 8 ++++++++ src/cuda-sim/ptx_ir.h | 9 --------- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/mem_latency_stat.cc | 3 ++- src/gpgpu-sim/shader.cc | 4 ++-- src/gpgpu-sim/stat-tool.cc | 33 +++++++++++++++++---------------- src/gpgpu-sim/stat-tool.h | 12 +++++++----- 13 files changed, 78 insertions(+), 73 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index ed4f746..346a8a4 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -39,6 +39,8 @@ class gpgpu_context { unsigned long long g_ptx_cta_info_uid; unsigned symbol_sm_next_uid; //uid for symbol unsigned function_info_sm_next_uid; + std::vector s_g_pc_to_insn; // a direct mapping from PC to instruction + // objects pointers for each file cuda_runtime_api* api; ptxinfo_data* ptxinfo; @@ -58,6 +60,9 @@ class gpgpu_context { class gpgpu_sim *gpgpu_ptx_sim_init_perf(); struct _cuda_device_id *GPGPUSim_Init(); void ptx_reg_options(option_parser_t opp); + const ptx_instruction* pc_to_instruction(unsigned pc); + const warp_inst_t *ptx_fetch_inst( address_type pc ); + unsigned translate_pc_to_ptxlineno(unsigned pc); }; gpgpu_context* GPGPU_Context(); diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 733d602..d8d5fbd 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -426,7 +426,7 @@ void warp_inst_t::generate_mem_accesses() } assert( total_accesses > 0 && total_accesses <= m_config->warp_size ); cycles = total_accesses; // shared memory conflicts modeled as larger initiation interval - ptx_file_line_stats_add_smem_bank_conflict( pc, total_accesses ); + m_config->gpgpu_ctx->stats->ptx_file_line_stats_add_smem_bank_conflict( pc, total_accesses ); break; } @@ -471,7 +471,7 @@ void warp_inst_t::generate_mem_accesses() } if ( space.get_type() == global_space ) { - ptx_file_line_stats_add_uncoalesced_gmem( pc, m_accessq.size() - starting_queue_size ); + m_config->gpgpu_ctx->stats->ptx_file_line_stats_add_uncoalesced_gmem( pc, m_accessq.size() - starting_queue_size ); } m_mem_accesses_created=true; } @@ -706,7 +706,7 @@ void warp_inst_t::completed( unsigned long long cycle ) const { unsigned long long latency = cycle - issue_cycle; assert(latency <= cycle); // underflow detection - ptx_file_line_stats_add_latency(pc, latency * active_count()); + m_config->gpgpu_ctx->stats->ptx_file_line_stats_add_latency(pc, latency * active_count()); } @@ -1110,7 +1110,7 @@ void simt_stack::update( simt_mask_t &thread_done, addr_vector_t &next_pc, addre if (warp_diverged) { - ptx_file_line_stats_add_warp_divergence(top_pc, 1); + m_gpu->gpgpu_ctx->stats->ptx_file_line_stats_add_warp_divergence(top_pc, 1); } } @@ -1157,7 +1157,7 @@ warp_inst_t core_t::getExecuteWarp(unsigned warpId) { unsigned pc,rpc; m_simt_stack[warpId]->get_pdom_stack_top_info(&pc,&rpc); - warp_inst_t wi= *ptx_fetch_inst(pc); + warp_inst_t wi= *(m_gpu->gpgpu_ctx->ptx_fetch_inst(pc)); wi.set_active(m_simt_stack[warpId]->get_active_mask()); return wi; } diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index c06f093..b9e6552 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -215,8 +215,6 @@ void gpgpu_t::gpgpu_ptx_sim_unbindTexture(const struct textureReference* texref) m_NameToTextureInfo.erase(texname); } -std::vector function_info::s_g_pc_to_insn; - #define MAX_INST_SIZE 8 /*bytes*/ void function_info::ptx_assemble() @@ -237,14 +235,14 @@ void function_info::ptx_assemble() addr_t PC = gpgpu_ctx->func_sim->g_assemble_code_next_pc; // globally unique address (across functions) // start function on an aligned address for( unsigned i=0; i < (PC%MAX_INST_SIZE); i++ ) - s_g_pc_to_insn.push_back((ptx_instruction*)NULL); + gpgpu_ctx->s_g_pc_to_insn.push_back((ptx_instruction*)NULL); PC += PC%MAX_INST_SIZE; m_start_PC = PC; addr_t n=0; // offset in m_instr_mem //Why s_g_pc_to_insn.size() is needed to reserve additional memory for insts? reserve is cumulative. //s_g_pc_to_insn.reserve(s_g_pc_to_insn.size() + MAX_INST_SIZE*m_instructions.size()); - s_g_pc_to_insn.reserve(MAX_INST_SIZE*m_instructions.size()); + gpgpu_ctx->s_g_pc_to_insn.reserve(MAX_INST_SIZE*m_instructions.size()); for ( i=m_instructions.begin(); i != m_instructions.end(); i++ ) { ptx_instruction *pI = *i; if ( pI->is_label() ) { @@ -253,13 +251,13 @@ void function_info::ptx_assemble() } else { gpgpu_ctx->func_sim->g_pc_to_finfo[PC] = this; m_instr_mem[n] = pI; - s_g_pc_to_insn.push_back(pI); - assert(pI == s_g_pc_to_insn[PC]); + gpgpu_ctx->s_g_pc_to_insn.push_back(pI); + assert(pI == gpgpu_ctx->s_g_pc_to_insn[PC]); pI->set_m_instr_mem_index(n); pI->set_PC(PC); assert( pI->inst_size() <= MAX_INST_SIZE ); for( unsigned i=1; i < pI->inst_size(); i++ ) { - s_g_pc_to_insn.push_back((ptx_instruction*)NULL); + gpgpu_ctx->s_g_pc_to_insn.push_back((ptx_instruction*)NULL); m_instr_mem[n+i]=NULL; } n += pI->inst_size(); @@ -1738,9 +1736,9 @@ const struct gpgpu_ptx_sim_info* ptx_sim_kernel_info(const function_info *kernel return kernel->get_kernel_info(); } -const warp_inst_t *ptx_fetch_inst( address_type pc ) +const warp_inst_t *gpgpu_context::ptx_fetch_inst( address_type pc ) { - return function_info::pc_to_instruction(pc); + return pc_to_instruction(pc); } unsigned ptx_sim_init_thread( kernel_info_t &kernel, @@ -2366,11 +2364,11 @@ void functionalCoreSim::executeWarp(unsigned i, bool &allAtBarrier, bool & someO if(!m_warpAtBarrier[i]&& m_liveThreadCount[i]>0) allAtBarrier = false; } -unsigned translate_pc_to_ptxlineno(unsigned pc) +unsigned gpgpu_context::translate_pc_to_ptxlineno(unsigned pc) { // this function assumes that the kernel fits inside a single PTX file // function_info *pFunc = g_func_info; // assume that the current kernel is the one in query - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = pc_to_instruction(pc); unsigned ptx_line_number = pInsn->source_line(); return ptx_line_number; diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index 5bd4cb2..1be3d19 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -58,10 +58,8 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, unsigned hw_warp_id, gpgpu_t *gpu, bool functionalSimulationMode = false); -const warp_inst_t *ptx_fetch_inst( address_type pc ); const struct gpgpu_ptx_sim_info* ptx_sim_kernel_info(const class function_info *kernel); - /*! * This class functionally executes a kernel. It uses the basic data structures and procedures in core_t */ diff --git a/src/cuda-sim/ptx-stats.cc b/src/cuda-sim/ptx-stats.cc index 298729f..22517df 100644 --- a/src/cuda-sim/ptx-stats.cc +++ b/src/cuda-sim/ptx-stats.cc @@ -151,27 +151,27 @@ void ptx_file_line_stats_add_exec_count(const ptx_instruction *pInsn) // attribute pipeline latency to this ptx instruction (specified by the pc) // pipeline latency is the number of cycles a warp with this instruction spent in the pipeline -void ptx_file_line_stats_add_latency(unsigned pc, unsigned latency) +void ptx_stats::ptx_file_line_stats_add_latency(unsigned pc, unsigned latency) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(), pInsn->source_line())].latency += latency; } // attribute dram traffic to this ptx instruction (specified by the pc) // dram traffic is counted in number of requests -void ptx_file_line_stats_add_dram_traffic(unsigned pc, unsigned dram_traffic) +void ptx_stats::ptx_file_line_stats_add_dram_traffic(unsigned pc, unsigned dram_traffic) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(), pInsn->source_line())].dram_traffic += dram_traffic; } // attribute the number of shared memory access cycles to a ptx instruction // counts both the number of warps doing shared memory access and the number of cycles involved -void ptx_file_line_stats_add_smem_bank_conflict(unsigned pc, unsigned n_way_bkconflict) +void ptx_stats::ptx_file_line_stats_add_smem_bank_conflict(unsigned pc, unsigned n_way_bkconflict) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); ptx_file_line_stats& line_stats = ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(), pInsn->source_line())]; line_stats.smem_n_way_bank_conflict_total += n_way_bkconflict; @@ -180,9 +180,9 @@ void ptx_file_line_stats_add_smem_bank_conflict(unsigned pc, unsigned n_way_bkco // attribute a non-coalesced mem access to a ptx instruction // counts both the number of warps causing this and the number of memory requests generated -void ptx_file_line_stats_add_uncoalesced_gmem(unsigned pc, unsigned n_access) +void ptx_stats::ptx_file_line_stats_add_uncoalesced_gmem(unsigned pc, unsigned n_access) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); ptx_file_line_stats& line_stats = ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(), pInsn->source_line())]; line_stats.gmem_n_access_total += n_access; @@ -239,17 +239,17 @@ void ptx_file_line_stats_create_exposed_latency_tracker(int n_shader_cores) } // add an inflight memory instruction -void ptx_file_line_stats_add_inflight_memory_insn(int sc_id, unsigned pc) +void ptx_stats::ptx_file_line_stats_add_inflight_memory_insn(int sc_id, unsigned pc) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); inflight_mem_tracker[sc_id].add_count(pInsn); } // remove an inflight memory instruction -void ptx_file_line_stats_sub_inflight_memory_insn(int sc_id, unsigned pc) +void ptx_stats::ptx_file_line_stats_sub_inflight_memory_insn(int sc_id, unsigned pc) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); inflight_mem_tracker[sc_id].sub_count(pInsn); } @@ -262,9 +262,9 @@ void ptx_file_line_stats_commit_exposed_latency(int sc_id, int exposed_latency) } // attribute the number of warp divergence to a ptx instruction -void ptx_file_line_stats_add_warp_divergence(unsigned pc, unsigned n_way_divergence) +void ptx_stats::ptx_file_line_stats_add_warp_divergence(unsigned pc, unsigned n_way_divergence) { - const ptx_instruction *pInsn = function_info::pc_to_instruction(pc); + const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc); ptx_file_line_stats& line_stats = ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(), pInsn->source_line())]; line_stats.warp_divergence += n_way_divergence; diff --git a/src/cuda-sim/ptx-stats.h b/src/cuda-sim/ptx-stats.h index c75fc58..246b4ce 100644 --- a/src/cuda-sim/ptx-stats.h +++ b/src/cuda-sim/ptx-stats.h @@ -37,17 +37,10 @@ void ptx_file_line_stats_add_exec_count(const ptx_instruction *pInsn); #endif // stat collection interface to gpgpu-sim -void ptx_file_line_stats_add_latency(unsigned pc, unsigned latency); -void ptx_file_line_stats_add_dram_traffic(unsigned pc, unsigned dram_traffic); -void ptx_file_line_stats_add_smem_bank_conflict(unsigned pc, unsigned n_way_bkconflict); -void ptx_file_line_stats_add_uncoalesced_gmem(unsigned pc, unsigned n_access); void ptx_file_line_stats_create_exposed_latency_tracker(int n_shader_cores); -void ptx_file_line_stats_add_inflight_memory_insn(int sc_id, unsigned pc); -void ptx_file_line_stats_sub_inflight_memory_insn(int sc_id, unsigned pc); void ptx_file_line_stats_commit_exposed_latency(int sc_id, int exposed_latency); -void ptx_file_line_stats_add_warp_divergence(unsigned pc, unsigned n_way_divergence); class gpgpu_context; class ptx_stats { @@ -64,4 +57,12 @@ class ptx_stats { // output stats to a file void ptx_file_line_stats_write_file(); + // stat collection interface to gpgpu-sim + void ptx_file_line_stats_add_latency(unsigned pc, unsigned latency); + void ptx_file_line_stats_add_dram_traffic(unsigned pc, unsigned dram_traffic); + void ptx_file_line_stats_add_smem_bank_conflict(unsigned pc, unsigned n_way_bkconflict); + void ptx_file_line_stats_add_uncoalesced_gmem(unsigned pc, unsigned n_access); + void ptx_file_line_stats_add_inflight_memory_insn(int sc_id, unsigned pc); + void ptx_file_line_stats_sub_inflight_memory_insn(int sc_id, unsigned pc); + void ptx_file_line_stats_add_warp_divergence(unsigned pc, unsigned n_way_divergence); }; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 5fa9379..3384d49 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -43,6 +43,14 @@ typedef void * yyscan_t; #define STR_SIZE 1024 +const ptx_instruction* gpgpu_context::pc_to_instruction(unsigned pc) +{ + if( pc < s_g_pc_to_insn.size() ) + return s_g_pc_to_insn[pc]; + else + return NULL; +} + unsigned symbol::get_uid() { unsigned result = (gpgpu_ctx->symbol_sm_next_uid)++; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 8fc0a06..f4c5c37 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1372,13 +1372,6 @@ public: return m_symtab; } - static const ptx_instruction* pc_to_instruction(unsigned pc) - { - if( pc < s_g_pc_to_insn.size() ) - return s_g_pc_to_insn[pc]; - else - return NULL; - } unsigned local_mem_framesize() const { return m_local_mem_framesize; @@ -1436,8 +1429,6 @@ private: symbol_table *m_symtab; - static std::vector s_g_pc_to_insn; // a direct mapping from PC to instruction - //parameter size for device kernels int m_args_aligned_size; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index bdf989a..e4ae04f 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -886,7 +886,7 @@ void gpgpu_sim::init() m_shader_stats->new_grid(); // initialize the control-flow, memory access, memory latency logger if (m_config.g_visualizer_enabled) { - create_thread_CFlogger( m_config.num_shader(), m_shader_config->n_thread_per_shader, 0, m_config.gpgpu_cflog_interval ); + create_thread_CFlogger( gpgpu_ctx, m_config.num_shader(), m_shader_config->n_thread_per_shader, 0, m_config.gpgpu_cflog_interval ); } shader_CTA_count_create( m_config.num_shader(), m_config.gpgpu_cflog_interval); if (m_config.gpgpu_cflog_interval != 0) { diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index 4e94991..a1b43a8 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -41,6 +41,7 @@ #include #include #include +#include "../../libcuda/gpgpu_context.h" memory_stats_t::memory_stats_t( unsigned n_shader, const shader_core_config *shader_config, const memory_config *mem_config, const class gpgpu_sim* gpu ) { @@ -195,7 +196,7 @@ void memory_stats_t::memlatstat_dram_access(mem_fetch *mf) mem_access_type_stats[mf->get_access_type()][dram_id][bank]++; } if (mf->get_pc() != (unsigned)-1) - ptx_file_line_stats_add_dram_traffic(mf->get_pc(), mf->get_data_size()); + m_gpu->gpgpu_ctx->stats->ptx_file_line_stats_add_dram_traffic(mf->get_pc(), mf->get_data_size()); } void memory_stats_t::memlatstat_icnt2mem_pop(mem_fetch *mf) diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index b7ae95d..c697450 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -744,7 +744,7 @@ void shader_core_ctx::decode() if( m_inst_fetch_buffer.m_valid ) { // decode 1 or 2 instructions and place them into ibuffer address_type pc = m_inst_fetch_buffer.m_pc; - const warp_inst_t* pI1 = ptx_fetch_inst(pc); + const warp_inst_t* pI1 = m_gpu->gpgpu_ctx->ptx_fetch_inst(pc); m_warp[m_inst_fetch_buffer.m_warp_id].ibuffer_fill(0,pI1); m_warp[m_inst_fetch_buffer.m_warp_id].inc_inst_in_pipeline(); if( pI1 ) { @@ -754,7 +754,7 @@ void shader_core_ctx::decode() }else if(pI1->oprnd_type==FP_OP) { m_stats->m_num_FPdecoded_insn[m_sid]++; } - const warp_inst_t* pI2 = ptx_fetch_inst(pc+pI1->isize); + const warp_inst_t* pI2 = m_gpu->gpgpu_ctx->ptx_fetch_inst(pc+pI1->isize); if( pI2 ) { m_warp[m_inst_fetch_buffer.m_warp_id].ibuffer_fill(1,pI2); m_warp[m_inst_fetch_buffer.m_warp_id].inc_inst_in_pipeline(); diff --git a/src/gpgpu-sim/stat-tool.cc b/src/gpgpu-sim/stat-tool.cc index 6a4c75b..35a4cc3 100644 --- a/src/gpgpu-sim/stat-tool.cc +++ b/src/gpgpu-sim/stat-tool.cc @@ -37,6 +37,7 @@ #include #include #include +#include "../../libcuda/gpgpu_context.h" //////////////////////////////////////////////////////////////////////////////// @@ -110,12 +111,10 @@ void spill_log_to_file (FILE *fout, int final, unsigned long long current_cycle //////////////////////////////////////////////////////////////////////////////// -unsigned translate_pc_to_ptxlineno(unsigned pc); - static int n_thread_CFloggers = 0; static thread_CFlocality** thread_CFlogger = NULL; -void create_thread_CFlogger( int n_loggers, int n_threads, address_type start_pc, unsigned long long logging_interval) +void create_thread_CFlogger(gpgpu_context* ctx, int n_loggers, int n_threads, address_type start_pc, unsigned long long logging_interval) { destroy_thread_CFlogger(); @@ -126,7 +125,7 @@ void create_thread_CFlogger( int n_loggers, int n_threads, address_type start_pc char buffer[32]; for (int i = 0; i < n_thread_CFloggers; i++) { snprintf(buffer, 32, "%02d", i); - thread_CFlogger[i] = new thread_CFlocality( name_tpl + buffer, logging_interval, n_threads, start_pc); + thread_CFlogger[i] = new thread_CFlocality( ctx, name_tpl + buffer, logging_interval, n_threads, start_pc); if (logging_interval != 0) { add_snap_shot_trigger(thread_CFlogger[i]); add_spill_log(thread_CFlogger[i]); @@ -368,10 +367,10 @@ static int s_cache_access_logger_n_types = 0; static std::vector s_cache_access_logger; enum cache_access_logger_types { - NORMAL, TEXTURE, CONSTANT, INSTRUCTION + NORMALS, TEXTURE, CONSTANT, INSTRUCTION }; -int get_shader_normal_cache_id() { return NORMAL; } +int get_shader_normal_cache_id() { return NORMALS; } int get_shader_texture_cache_id() { return TEXTURE; } int get_shader_constant_cache_id() { return CONSTANT; } int get_shader_instruction_cache_id() { return INSTRUCTION; } @@ -394,7 +393,7 @@ void shader_cache_access_log( int logger_id, int type, int miss) { if (s_cache_access_logger_n_types == 0) return; if (logger_id < 0) return; - assert(type == NORMAL || type == TEXTURE || type == CONSTANT || type == INSTRUCTION); + assert(type == NORMALS || type == TEXTURE || type == CONSTANT || type == INSTRUCTION); assert(miss == 0 || miss == 1); s_cache_access_logger[logger_id].log(2 * type + miss); @@ -404,7 +403,7 @@ void shader_cache_access_unlog( int logger_id, int type, int miss) { if (s_cache_access_logger_n_types == 0) return; if (logger_id < 0) return; - assert(type == NORMAL || type == TEXTURE || type == CONSTANT || type == INSTRUCTION); + assert(type == NORMALS || type == TEXTURE || type == CONSTANT || type == INSTRUCTION); assert(miss == 0 || miss == 1); s_cache_access_logger[logger_id].unlog(2 * type + miss); @@ -477,22 +476,24 @@ void shader_CTA_count_visualizer_gzprint( gzFile fout ) //////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////// -thread_insn_span::thread_insn_span(unsigned long long cycle) +thread_insn_span::thread_insn_span(unsigned long long cycle, gpgpu_context* ctx) : m_cycle(cycle), #if (tr1_hash_map_ismap == 1) m_insn_span_count() #else m_insn_span_count(32*1024) #endif -{ +{ + gpgpu_ctx = ctx; } thread_insn_span::~thread_insn_span() { } -thread_insn_span::thread_insn_span(const thread_insn_span& other) +thread_insn_span::thread_insn_span(const thread_insn_span& other, gpgpu_context* ctx) : m_cycle(other.m_cycle), - m_insn_span_count(other.m_insn_span_count) + m_insn_span_count(other.m_insn_span_count) { + gpgpu_ctx = ctx; } thread_insn_span& thread_insn_span::operator=(const thread_insn_span& other) @@ -551,7 +552,7 @@ void thread_insn_span::print_sparse_histo(FILE *fout) const int n_printed_entries = 0; span_count_map::const_iterator i_sc = m_insn_span_count.begin(); for (; i_sc != m_insn_span_count.end(); ++i_sc) { - unsigned ptx_lineno = translate_pc_to_ptxlineno(i_sc->first); + unsigned ptx_lineno = gpgpu_ctx->translate_pc_to_ptxlineno(i_sc->first); fprintf(fout, "%u %d ", ptx_lineno, i_sc->second); n_printed_entries++; } @@ -566,7 +567,7 @@ void thread_insn_span::print_sparse_histo(gzFile fout) const int n_printed_entries = 0; span_count_map::const_iterator i_sc = m_insn_span_count.begin(); for (; i_sc != m_insn_span_count.end(); ++i_sc) { - unsigned ptx_lineno = translate_pc_to_ptxlineno(i_sc->first); + unsigned ptx_lineno = gpgpu_ctx->translate_pc_to_ptxlineno(i_sc->first); gzprintf(fout, "%u %d ", ptx_lineno, i_sc->second); n_printed_entries++; } @@ -578,14 +579,14 @@ void thread_insn_span::print_sparse_histo(gzFile fout) const //////////////////////////////////////////////////////////////////////////////// -thread_CFlocality::thread_CFlocality(std::string name, +thread_CFlocality::thread_CFlocality( gpgpu_context* ctx, std::string name, unsigned long long snap_shot_interval, int nthreads, address_type start_pc, unsigned long long start_cycle) : snap_shot_trigger(snap_shot_interval), m_name(name), m_nthreads(nthreads), m_thread_pc(nthreads, start_pc), m_cycle(start_cycle), - m_thd_span(start_cycle) + m_thd_span(start_cycle, ctx) { std::fill(m_thread_pc.begin(), m_thread_pc.end(), -1); // so that hw thread with no work assigned will not clobber results } diff --git a/src/gpgpu-sim/stat-tool.h b/src/gpgpu-sim/stat-tool.h index 5646f01..67b3923 100644 --- a/src/gpgpu-sim/stat-tool.h +++ b/src/gpgpu-sim/stat-tool.h @@ -35,6 +35,7 @@ #include #include +class gpgpu_context; ///////////////////////////////////////////////////////////////////////////////////// // logger snapshot trigger: // - automate the snap_shot part of loggers to avoid modifying simulation loop everytime @@ -80,8 +81,8 @@ public: class thread_insn_span { public: - thread_insn_span(unsigned long long cycle); - thread_insn_span(const thread_insn_span& other); + thread_insn_span(unsigned long long cycle, gpgpu_context* ctx); + thread_insn_span(const thread_insn_span& other, gpgpu_context* ctx); ~thread_insn_span(); thread_insn_span& operator=(const thread_insn_span& other); @@ -94,7 +95,8 @@ public: void print_sparse_histo(FILE *fout) const; void print_sparse_histo(gzFile fout) const; -private: +private: + gpgpu_context* gpgpu_ctx; typedef tr1_hash_map span_count_map; unsigned long long m_cycle; span_count_map m_insn_span_count; @@ -102,7 +104,7 @@ private: class thread_CFlocality : public snap_shot_trigger, public spill_log_interface { public: - thread_CFlocality(std::string name, unsigned long long snap_shot_interval, + thread_CFlocality(gpgpu_context* ctx, std::string name, unsigned long long snap_shot_interval, int nthreads, address_type start_pc, unsigned long long start_cycle = 0); ~thread_CFlocality(); @@ -270,7 +272,7 @@ void try_snap_shot (unsigned long long current_cycle); void set_spill_interval (unsigned long long interval); void spill_log_to_file (FILE *fout, int final, unsigned long long current_cycle); -void create_thread_CFlogger( int n_loggers, int n_threads, address_type start_pc, unsigned long long logging_interval); +void create_thread_CFlogger(gpgpu_context* ctx, int n_loggers, int n_threads, address_type start_pc, unsigned long long logging_interval); void destroy_thread_CFlogger( ); void cflog_update_thread_pc( int logger_id, int thread_id, address_type pc ); void cflog_snapshot( int logger_id, unsigned long long cycle ); -- cgit v1.3 From e1fa1a3cc7c509417064a8e4cdab71e3f7feb881 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Mon, 15 Jul 2019 22:16:55 -0400 Subject: Move debug_tensorcore Signed-off-by: Mengchi Zhang --- libcuda/gpgpu_context.h | 2 ++ src/cuda-sim/instructions.cc | 43 +++++++++++++++++++++---------------------- 2 files changed, 23 insertions(+), 22 deletions(-) (limited to 'libcuda') diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index 346a8a4..45c5cdd 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -20,6 +20,7 @@ class gpgpu_context { g_ptx_cta_info_uid = 1; symbol_sm_next_uid = 1; function_info_sm_next_uid = 1; + debug_tensorcore = 0; api = new cuda_runtime_api(this); ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); @@ -40,6 +41,7 @@ class gpgpu_context { unsigned symbol_sm_next_uid; //uid for symbol unsigned function_info_sm_next_uid; std::vector s_g_pc_to_insn; // a direct mapping from PC to instruction + bool debug_tensorcore; // objects pointers for each file cuda_runtime_api* api; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 4d4a80d..58a077e 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -60,7 +60,6 @@ class ptx_recognizer; using half_float::half; -bool debug_tensorcore = 0; const char *g_opcode_string[NUM_OPCODES] = { @@ -1840,14 +1839,14 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) for (thrd=0; thrd < core->get_warp_size(); thrd++){ thread = core->get_thread_info()[tid+thrd]; - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("THREAD=%d\n:",thrd); for(int operand_num=1;operand_num<=3;operand_num++){ const operand_info &src_a= pI->operand_lookup(operand_num); unsigned nelem = src_a.get_vect_nelem(); ptx_reg_t v[8]; thread->get_vector_operand_values( src_a, v, nelem ); - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ printf("Thread%d_Iteration=%d\n:",thrd,operand_num); for(k=0;kget_gpu()->gpgpu_ctx->debug_tensorcore) printf("%.2f ",temp); } - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("\n"); } else{ - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ for(k=0;k<8;k++){ printf("%.2f ",v[k].f32); } @@ -1887,7 +1886,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) case 1 ://operand 1 for(k=0;k<8;k++){ mapping(thrd,LOAD_A,a_layout,F16_TYPE,k,16,row,col,offset); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("A:thread=%d,row=%d,col=%d,offset=%d\n",thrd,row,col,offset); matrix_a[row][col]=nw_v[offset]; } @@ -1895,7 +1894,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) case 2 ://operand 2 for(k=0;k<8;k++){ mapping(thrd,LOAD_B,b_layout,F16_TYPE,k,16,row,col,offset); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("B:thread=%d,row=%d,col=%d,offset=%d\n",thrd,row,col,offset); matrix_b[row][col]=nw_v[offset]; } @@ -1903,7 +1902,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) case 3 ://operand 3 for(k=0;k<8;k++){ mapping(thrd,LOAD_C,ROW,type2,k,16,row,col,offset); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("C:thread=%d,row=%d,col=%d,offset=%d\n",thrd,row,col,offset); if(type2!=F16_TYPE){ matrix_c[row][col]=v[offset]; @@ -1917,10 +1916,10 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) printf("Invalid Operand Index\n" ); } } - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("\n"); } - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ printf("MATRIX_A\n"); for (i=0;i<16;i++){ for(j=0;j<16;j++){ @@ -1980,7 +1979,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) } } } - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ printf("MATRIX_D\n"); for (i=0;i<16;i++){ for(j=0;j<16;j++){ @@ -1999,7 +1998,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) int col_t[8]; for(k=0;k<8;k++){ mapping(thrd,LOAD_C,ROW,type,k,16,row_t[k],col_t[k],offset); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("mma:store:row:%d,col%d\n",row_t[k],col_t[k]); } thread = core->get_thread_info()[tid+thrd]; @@ -2008,7 +2007,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) if(type==F32_TYPE){ thread->set_wmma_vector_operand_values(dst,matrix_d[row_t[0]][col_t[0]],matrix_d[row_t[1]][col_t[1]],matrix_d[row_t[2]][col_t[2]],matrix_d[row_t[3]][col_t[3]],matrix_d[row_t[4]][col_t[4]],matrix_d[row_t[5]][col_t[5]],matrix_d[row_t[6]][col_t[6]],matrix_d[row_t[7]][col_t[7]]); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) { printf("thread%d:",thrd); for(k=0;k<8;k++){ @@ -2018,7 +2017,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) } } else if(type==F16_TYPE){ - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ printf("thread%d:",thrd); for(k=0;k<8;k++){ temp=matrix_d[row_t[k]][col_t[k]].f16; @@ -2038,7 +2037,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) nw_data3.s64=((matrix_d[row_t[4]][col_t[4]].s64 & 0xffff))|((matrix_d[row_t[5]][col_t[5]].s64&0xffff)<<16); nw_data4.s64=((matrix_d[row_t[6]][col_t[6]].s64 & 0xffff))|((matrix_d[row_t[7]][col_t[7]].s64&0xffff)<<16); thread->set_vector_operand_values(dst,nw_data1,nw_data2,nw_data3,nw_data4); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("thread%d=%x,%x,%x,%x",thrd,nw_data1.s64,nw_data2.s64,nw_data3.s64,nw_data4.s64); } @@ -3132,7 +3131,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) decode_space(space,thread,src1,mem,addr); type_info_key::type_decode(type,size,t); - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("mma_st: thrd=%d,addr=%x, fp(size=%d), stride=%d\n",thrd,addr_reg.u32,size,src2_data.u32); addr_t new_addr = addr+thread_group_offset(thrd,wmma_type,wmma_layout,type,stride)*size/8; addr_t push_addr; @@ -3152,7 +3151,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) mem->write(push_addr,size/8,&v[k].s64,thread,pI); mem_txn_addr[num_mem_txn++]=push_addr; - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,v[0].s64,v[1].s64,v[2].s64,v[3].s64,v[4].s64,v[5].s64,v[6].s64,v[7].s64); float temp; int l; @@ -3179,7 +3178,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) mem_txn_addr[num_mem_txn++]=push_addr; } - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,nw_v[0].s64,nw_v[1].s64,nw_v[2].s64,nw_v[3].s64,nw_v[4].s64,nw_v[5].s64,nw_v[6].s64,nw_v[7].s64); } } @@ -3242,7 +3241,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) type_info_key::type_decode(type,size,t); ptx_reg_t data[16]; - if(debug_tensorcore) + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore) printf("mma_ld: thrd=%d,addr=%x, fpsize=%d, stride=%d\n",thrd,src1_data.u32,size,src2_data.u32); addr_t new_addr = addr+thread_group_offset(thrd,wmma_type,wmma_layout,type,stride)*size/8; @@ -3338,7 +3337,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) inst.data_size = 4; // 4 byte transaction assert( inst.memory_op == insn_memory_op ); - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ if(type==F16_TYPE){ printf("\nmma_ld:thread%d= ",thrd); for(i=0;i<16;i++){ @@ -3388,7 +3387,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) thread->set_vector_operand_values(dst,nw_data[0],nw_data[1],nw_data[2],nw_data[3]); else thread->set_wmma_vector_operand_values(dst,nw_data[0],nw_data[1],nw_data[2],nw_data[3],nw_data[4],nw_data[5],nw_data[6],nw_data[7]); - if(debug_tensorcore){ + if(core->get_gpu()->gpgpu_ctx->debug_tensorcore){ printf("mma_ld:data[0].s64=%x,data[1].s64=%x,new_data[0].s64=%x\n",data[0].u64,data[1].u64,nw_data[0].u64); printf("mma_ld:data[2].s64=%x,data[3].s64=%x,new_data[1].s64=%x\n",data[2].u64,data[3].u64,nw_data[1].u64); printf("mma_ld:data[4].s64=%x,data[5].s64=%x,new_data[2].s64=%x\n",data[4].u64,data[5].u64,nw_data[2].u64); -- cgit v1.3