summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/abstract_hardware_model.h32
-rw-r--r--src/cuda-sim/cuda-sim.cc6
-rw-r--r--src/cuda-sim/half.h6
-rw-r--r--src/cuda-sim/instructions.cc8
-rw-r--r--src/cuda-sim/ptx_ir.cc2
-rw-r--r--src/cuda-sim/ptx_loader.cc3
6 files changed, 29 insertions, 28 deletions
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index fdc4dc3..29e4a30 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -442,19 +442,25 @@ protected:
class gpgpu_sim * m_gpu;
};
-#define GLOBAL_HEAP_START 0xC0000000
- // start allocating from this address (lower values used for allocating globals in .ptx file)
-#define SHARED_MEM_SIZE_MAX (96*1024)
-#define LOCAL_MEM_SIZE_MAX (16*1024)
-#define MAX_STREAMING_MULTIPROCESSORS 80 //scale it to Volta
-#define MAX_THREAD_PER_SM 2048
-#define MAX_WARP_PER_SM 64
-#define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX)
-#define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX)
-#define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX)
-#define SHARED_GENERIC_START (GLOBAL_HEAP_START-TOTAL_SHARED_MEM)
-#define LOCAL_GENERIC_START (SHARED_GENERIC_START-TOTAL_LOCAL_MEM)
-#define STATIC_ALLOC_LIMIT (GLOBAL_HEAP_START - (TOTAL_LOCAL_MEM+TOTAL_SHARED_MEM))
+// Let's just upgrade to C++11 so we can use constexpr here...
+// start allocating from this address (lower values used for allocating globals in .ptx file)
+const unsigned long long GLOBAL_HEAP_START = 0xC0000000;
+// Volta max shmem size is 96kB
+const unsigned long long SHARED_MEM_SIZE_MAX = 96 * (1 << 10);
+// Volta max local mem is 16kB
+const unsigned long long LOCAL_MEM_SIZE_MAX = 1 << 14;
+// Volta Titan V has 80 SMs
+const unsigned MAX_STREAMING_MULTIPROCESSORS = 80;
+// Max 2048 threads / SM
+const unsigned MAX_THREAD_PER_SM = 1 << 11;
+// MAX 64 warps / SM
+const unsigned MAX_WARP_PER_SM = 1 << 6;
+const unsigned long long TOTAL_LOCAL_MEM_PER_SM = MAX_THREAD_PER_SM * LOCAL_MEM_SIZE_MAX;
+const unsigned long long TOTAL_SHARED_MEM = MAX_STREAMING_MULTIPROCESSORS * SHARED_MEM_SIZE_MAX;
+const unsigned long long TOTAL_LOCAL_MEM = MAX_STREAMING_MULTIPROCESSORS * MAX_THREAD_PER_SM * LOCAL_MEM_SIZE_MAX;
+const unsigned long long SHARED_GENERIC_START = GLOBAL_HEAP_START - TOTAL_SHARED_MEM;
+const unsigned long long LOCAL_GENERIC_START = SHARED_GENERIC_START - TOTAL_LOCAL_MEM;
+const unsigned long long STATIC_ALLOC_LIMIT = GLOBAL_HEAP_START - (TOTAL_LOCAL_MEM + TOTAL_SHARED_MEM);
#if !defined(__CUDA_RUNTIME_API_H__)
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index fbce75b..28b4bf9 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -1309,7 +1309,7 @@ void function_info::ptx_jit_config(std::map<unsigned long long, size_t> mallocPt
char buff[1024];
std::string filename_c(filename+"_c");
snprintf(buff,1024,"c++filt %s > %s", get_name().c_str(), filename_c.c_str());
- system(buff);
+ assert(system(buff) != NULL);
FILE *fp = fopen(filename_c.c_str(), "r");
fgets(buff, 1024, fp);
fclose(fp);
@@ -1432,13 +1432,13 @@ void function_info::ptx_jit_config(std::map<unsigned long long, size_t> mallocPt
fout = fopen(ptx_config_fn.c_str(), "a");
assert(fout!=NULL);
for (unsigned i = 0; i<line_number; i++){
- fgets(buff, 1024, fin);
+ assert(fgets(buff, 1024, fin) != NULL);
assert(!feof(fin));
}
fprintf(fout, "\n\n");
do{
fprintf(fout, "%s", buff);
- fgets(buff, 1024, fin);
+ assert(fgets(buff, 1024, fin) != NULL);
if(feof(fin)){
break;
}
diff --git a/src/cuda-sim/half.h b/src/cuda-sim/half.h
index 8f1a8eb..9f74bb7 100644
--- a/src/cuda-sim/half.h
+++ b/src/cuda-sim/half.h
@@ -642,10 +642,10 @@ namespace half_float
if(exp > 16)
{
if(R == std::round_toward_infinity)
- return hbits | 0x7C00 - (hbits>>15);
+ return hbits | (0x7C00 - (hbits>>15));
else if(R == std::round_toward_neg_infinity)
- return hbits | 0x7BFF + (hbits>>15);
- return hbits | 0x7BFF + (R!=std::round_toward_zero);
+ return hbits | (0x7BFF + (hbits>>15));
+ return hbits | (0x7BFF + (R!=std::round_toward_zero));
}
if(exp < -13)
value = std::ldexp(value, 24);
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index a44b03f..014e588 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -224,7 +224,6 @@ void ptx_thread_info::resume_reg_thread(char * fname, symbol_table * symtab)
{
symbol *reg;
char * pch;
- unsigned size;
pch = strtok (line," ");
char * name =pch;
reg= symtab->lookup(name);
@@ -233,9 +232,6 @@ void ptx_thread_info::resume_reg_thread(char * fname, symbol_table * symtab)
data = atoi(pch);
pch = strtok (NULL," ");
pch = strtok (NULL," ");
- size = atoi(pch);
-
-
m_regs.back()[reg] = data;
}
fclose ( fp2 );
@@ -3080,7 +3076,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst )
size_t size;
unsigned smid;
int t;
- int thrd, odd, inx, k;
+ int thrd, k;
ptx_thread_info *thread;
const operand_info &src = pI->operand_lookup(1);
@@ -3100,8 +3096,6 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst )
_memory_op_t insn_memory_op = pI->has_memory_read() ? memory_load : memory_store;
for (thrd=0; thrd < core->get_warp_size(); thrd++) {
thread = core->get_thread_info()[tid+thrd];
- odd= thrd % 2;
- inx= thrd / 2;
ptx_reg_t addr_reg = thread->get_operand_value(src1, src, type, thread, 1);
ptx_reg_t src2_data = thread->get_operand_value(src2, src, type, thread, 1);
const operand_info &src_a= pI->operand_lookup(1);
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index 6978cc1..d8943d2 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -1415,7 +1415,7 @@ unsigned function_info::print_insn( unsigned pc, FILE * fp ) const
snprintf(command,1024,"c++filt -p %s",m_name.c_str());
FILE *p = popen(command,"r");
buffer[0]=0;
- fgets(buffer, 1023, p);
+ assert(fgets(buffer, 1023, p) != NULL);
// Remove trailing "\n" in buffer
char *c;
if ((c=strchr(buffer, '\n')) != NULL) *c = '\0';
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index dca3cec..33bcf45 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -214,7 +214,8 @@ void fix_duplicate_errors(char fname2[1024]) {
long filesize = ftell(ptxsource);
rewind(ptxsource);
char *ptxdata = (char*)malloc((filesize+1)*sizeof(char));
- fread(ptxdata, filesize, 1, ptxsource);
+ // Fail if we do not read the file
+ assert(fread(ptxdata, filesize, 1, ptxsource) == 1);
fclose(ptxsource);
FILE *ptxdest = fopen(fname2,"w");
generated by cgit v1.3 (git 2.53.0) at 2026-07-11 02:20:37 +0000