diff options
| author | Nick <[email protected]> | 2019-09-03 10:21:51 -0400 |
|---|---|---|
| committer | Nick <[email protected]> | 2019-09-03 10:21:51 -0400 |
| commit | 7bc0007c0dac14528409d33ca1a28491a2568d12 (patch) | |
| tree | 89a44f43be0e65ffd3f41fef50235400bf711457 /src | |
| parent | 13b2410ef3f178e7adb7d7b1d634648e6be37715 (diff) | |
get rid of more UB
Diffstat (limited to 'src')
| -rw-r--r-- | src/abstract_hardware_model.h | 32 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_loader.cc | 3 |
2 files changed, 21 insertions, 14 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/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"); |
