aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/Makefile2
-rw-r--r--src/abstract_hardware_model.h40
-rw-r--r--src/cuda-sim/Makefile8
-rw-r--r--src/cuda-sim/cuda-math.h16
-rw-r--r--src/cuda-sim/instructions.cc10
-rw-r--r--src/cuda-sim/ptx.y23
-rw-r--r--src/cuda-sim/ptx_parser.cc26
-rw-r--r--src/cuda-sim/ptx_parser.h38
-rw-r--r--src/gpgpu-sim/Makefile2
-rw-r--r--src/intersim2/Makefile1
10 files changed, 73 insertions, 93 deletions
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..68cb693 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -173,9 +173,7 @@ enum _memory_op_t {
#include <algorithm>
#if !defined(__VECTOR_TYPES_H__)
-struct dim3 {
- unsigned int x, y, z;
-};
+#include "vector_types.h"
#endif
struct dim3comp {
bool operator() (const dim3 & a, const dim3 & b) const
@@ -454,19 +452,7 @@ protected:
#if !defined(__CUDA_RUNTIME_API_H__)
-enum cudaChannelFormatKind {
- cudaChannelFormatKindSigned,
- cudaChannelFormatKindUnsigned,
- cudaChannelFormatKindFloat
-};
-
-struct cudaChannelFormatDesc {
- int x;
- int y;
- int z;
- int w;
- enum cudaChannelFormatKind f;
-};
+#include "builtin_types.h"
struct cudaArray {
void *devPtr;
@@ -478,28 +464,6 @@ 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;
-};
-
#endif
// Struct that record other attributes in the textureReference declaration
diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile
index 999dad7..85d1c8c 100644
--- a/src/cuda-sim/Makefile
+++ b/src/cuda-sim/Makefile
@@ -79,16 +79,16 @@ libgpgpu_ptx_sim.a: $(OBJS)
ar rcs $(OUTPUT_DIR)/libgpgpu_ptx_sim.a $(OUTPUT_DIR)/ptx.tab.o $(OUTPUT_DIR)/lex.ptx_.o $(OUTPUT_DIR)/ptxinfo.tab.o $(OUTPUT_DIR)/lex.ptxinfo_.o $(OBJS)
$(OUTPUT_DIR)/ptx.tab.o: $(OUTPUT_DIR)/ptx.tab.c
- $(CPP) -c $(OPT) -DYYDEBUG $(OUTPUT_DIR)/ptx.tab.c -o $(OUTPUT_DIR)/ptx.tab.o
+ $(CPP) -c $(CXX_OPT) -DYYDEBUG $(OUTPUT_DIR)/ptx.tab.c -o $(OUTPUT_DIR)/ptx.tab.o
$(OUTPUT_DIR)/lex.ptx_.o: $(OUTPUT_DIR)/lex.ptx_.c
- $(CPP) -c $(OPT) $(OUTPUT_DIR)/lex.ptx_.c -o $(OUTPUT_DIR)/lex.ptx_.o
+ $(CPP) -c $(CXX_OPT) $(OUTPUT_DIR)/lex.ptx_.c -o $(OUTPUT_DIR)/lex.ptx_.o
$(OUTPUT_DIR)/ptxinfo.tab.o: $(OUTPUT_DIR)/ptxinfo.tab.c
- $(CPP) -c $(OPT) -DYYDEBUG $(OUTPUT_DIR)/ptxinfo.tab.c -o $(OUTPUT_DIR)/ptxinfo.tab.o
+ $(CPP) -c $(CXX_OPT) -DYYDEBUG $(OUTPUT_DIR)/ptxinfo.tab.c -o $(OUTPUT_DIR)/ptxinfo.tab.o
$(OUTPUT_DIR)/lex.ptxinfo_.o: $(OUTPUT_DIR)/lex.ptxinfo_.c $(OUTPUT_DIR)/ptxinfo.tab.c
- $(CPP) -c $(OPT) $(OUTPUT_DIR)/lex.ptxinfo_.c -o $(OUTPUT_DIR)/lex.ptxinfo_.o
+ $(CPP) -c $(CXX_OPT) $(OUTPUT_DIR)/lex.ptxinfo_.c -o $(OUTPUT_DIR)/lex.ptxinfo_.o
$(OUTPUT_DIR)/ptx.tab.c: ptx.y
bison --name-prefix=ptx_ -v -d ptx.y --file-prefix=$(OUTPUT_DIR)/ptx
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/cuda-sim/ptx.y b/src/cuda-sim/ptx.y
index 8447693..837fbe9 100644
--- a/src/cuda-sim/ptx.y
+++ b/src/cuda-sim/ptx.y
@@ -229,7 +229,6 @@ class ptx_recognizer;
#include <string.h>
#include <math.h>
void syntax_not_implemented(yyscan_t yyscanner, ptx_recognizer* recognizer);
- extern int g_func_decl;
int ptx_lex(YYSTYPE * yylval_param, yyscan_t yyscanner, ptx_recognizer* recognizer);
int ptx_error( yyscan_t yyscanner, ptx_recognizer* recognizer, const char *s );
%}
@@ -260,21 +259,21 @@ block_spec_list: block_spec
function_decl: function_decl_header LEFT_PAREN { recognizer->start_function($1); recognizer->func_header_info("(");} param_entry RIGHT_PAREN {recognizer->func_header_info(")");} function_ident_param { $$ = recognizer->reset_symtab(); }
| function_decl_header { recognizer->start_function($1); } function_ident_param { $$ = recognizer->reset_symtab(); }
- | function_decl_header { recognizer->start_function($1); recognizer->add_function_name(""); g_func_decl=0; $$ = recognizer->reset_symtab(); }
+ | function_decl_header { recognizer->start_function($1); recognizer->add_function_name(""); recognizer->g_func_decl=0; $$ = recognizer->reset_symtab(); }
;
-function_ident_param: IDENTIFIER { recognizer->add_function_name($1); } LEFT_PAREN {recognizer->func_header_info("(");} param_list RIGHT_PAREN { g_func_decl=0; recognizer->func_header_info(")"); }
- | IDENTIFIER { recognizer->add_function_name($1); g_func_decl=0; }
+function_ident_param: IDENTIFIER { recognizer->add_function_name($1); } LEFT_PAREN {recognizer->func_header_info("(");} param_list RIGHT_PAREN { recognizer->g_func_decl=0; recognizer->func_header_info(")"); }
+ | IDENTIFIER { recognizer->add_function_name($1); recognizer->g_func_decl=0; }
;
-function_decl_header: ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; recognizer->func_header(".entry"); }
- | VISIBLE_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; recognizer->func_header(".entry"); }
- | WEAK_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; recognizer->func_header(".entry"); }
- | FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; recognizer->func_header(".func"); }
- | VISIBLE_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; recognizer->func_header(".func"); }
- | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; recognizer->func_header(".func"); }
- | EXTERN_DIRECTIVE FUNC_DIRECTIVE { $$ = 2; g_func_decl=1; recognizer->func_header(".func"); }
- | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; recognizer->func_header(".func"); }
+function_decl_header: ENTRY_DIRECTIVE { $$ = 1; recognizer->g_func_decl=1; recognizer->func_header(".entry"); }
+ | VISIBLE_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; recognizer->g_func_decl=1; recognizer->func_header(".entry"); }
+ | WEAK_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; recognizer->g_func_decl=1; recognizer->func_header(".entry"); }
+ | FUNC_DIRECTIVE { $$ = 0; recognizer->g_func_decl=1; recognizer->func_header(".func"); }
+ | VISIBLE_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; recognizer->g_func_decl=1; recognizer->func_header(".func"); }
+ | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; recognizer->g_func_decl=1; recognizer->func_header(".func"); }
+ | EXTERN_DIRECTIVE FUNC_DIRECTIVE { $$ = 2; recognizer->g_func_decl=1; recognizer->func_header(".func"); }
+ | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; recognizer->g_func_decl=1; recognizer->func_header(".func"); }
;
param_list: /*empty*/
diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index a61dc16..897beaf 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -49,7 +49,6 @@ void set_ptx_warp_size(const struct core_config * warp_size)
static bool g_debug_ir_generation=false;
const char *g_filename;
-unsigned g_max_regs_per_thread = 0;
// the program intermediate representation...
static symbol_table *g_global_allfiles_symbol_table = NULL;
@@ -61,27 +60,6 @@ static symbol *g_last_symbol = NULL;
int g_error_detected = 0;
-// type specifier stuff:
-memory_space_t g_space_spec = undefined_space;
-memory_space_t g_ptr_spec = undefined_space;
-int g_scalar_type_spec = -1;
-int g_vector_spec = -1;
-int g_alignment_spec = -1;
-int g_extern_spec = 0;
-
-// variable declaration stuff:
-type_info *g_var_type = NULL;
-
-// instruction definition stuff:
-const symbol *g_pred;
-int g_neg_pred;
-int g_pred_mod;
-symbol *g_label;
-int g_opcode = -1;
-std::list<operand_info> g_operands;
-std::list<int> g_options;
-std::list<int> g_wmma_options;
-std::list<int> g_scalar_type;
#define PTX_PARSE_DPRINTF(...) \
if( g_debug_ir_generation ) { \
@@ -362,10 +340,6 @@ bool check_for_duplicates( const char *identifier )
extern std::set<std::string> g_globals;
extern std::set<std::string> g_constants;
-int g_func_decl = 0;
-int g_ident_add_uid = 0;
-unsigned g_const_alloc = 1;
-
// Returns padding that needs to be inserted ahead of address to make it aligned to min(size, maxalign)
/*
* @param address the address in bytes
diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h
index fb1b7f2..c52fe20 100644
--- a/src/cuda-sim/ptx_parser.h
+++ b/src/cuda-sim/ptx_parser.h
@@ -29,6 +29,7 @@
#define ptx_parser_INCLUDED
#include "../abstract_hardware_model.h"
+#include "ptx_ir.h"
extern const char *g_filename;
extern int g_error_detected;
@@ -44,7 +45,20 @@ class ptx_recognizer {
scanner = NULL;
g_size = -1;
g_add_identifier_cached__identifier = NULL;
+ g_alignment_spec = -1;
+ g_var_type = NULL;
+ g_opcode = -1;
+ g_space_spec = undefined_space;
+ g_ptr_spec = undefined_space;
+ g_scalar_type_spec = -1;
+ g_vector_spec = -1;
+ g_extern_spec = 0;
+ g_func_decl = 0;
+ g_ident_add_uid = 0;
+ g_const_alloc = 1;
+ g_max_regs_per_thread = 0;
}
+ // global list
yyscan_t scanner;
#define LINEBUF_SIZE (4*1024)
char linebuf[LINEBUF_SIZE];
@@ -53,7 +67,31 @@ class ptx_recognizer {
char *g_add_identifier_cached__identifier;
int g_add_identifier_cached__array_dim;
int g_add_identifier_cached__array_ident;
+ int g_alignment_spec;
+ // variable declaration stuff:
+ type_info *g_var_type;
+ // instruction definition stuff:
+ const symbol *g_pred;
+ int g_neg_pred;
+ int g_pred_mod;
+ symbol *g_label;
+ int g_opcode;
+ std::list<operand_info> g_operands;
+ std::list<int> g_options;
+ std::list<int> g_wmma_options;
+ std::list<int> g_scalar_type;
+ // type specifier stuff:
+ memory_space_t g_space_spec;
+ memory_space_t g_ptr_spec;
+ int g_scalar_type_spec;
+ int g_vector_spec;
+ int g_extern_spec;
+ int g_func_decl;
+ int g_ident_add_uid;
+ unsigned g_const_alloc;
+ unsigned g_max_regs_per_thread;
+ // member function list
void init_directive_state();
void init_instruction_state();
void start_function( int entry_point );
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 +=
generated by cgit v1.3 (git 2.53.0) at 2026-07-20 05:11:10 +0000