Franscobec/gpufetch
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[v0.22] Fix for previous commit (dont show tensor cores in TU116)

Browse Source
This commit is contained in:
Dr-Noob
2022-01-20 22:57:19 +01:00
parent d3aaf7cfe5
commit 23586a18e9
1 changed files with 14 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
src/cuda/cuda.cpp
View File

@@ -26,19 +26,27 @@ struct cache* get_cache_info(cudaDeviceProp prop) {
return cach; return cach;
} }
int get_tensor_cores(int sm, int major) { int get_tensor_cores(struct uarch* arch, int sm, int major) {
if(major == 7) return sm * 8; if(major == 7) {
// TU116 does not have tensor cores!
// https://www.anandtech.com/show/13973/nvidia-gtx-1660-ti-review-feat-evga-xc-gaming/2
if(arch->chip == CHIP_TU116 || arch->chip == CHIP_TU116BM ||
arch->chip == CHIP_TU116GL || arch->chip == CHIP_TU116M) {
return 0;
}
return sm * 8;
}
else if(major == 8) return sm * 4; else if(major == 8) return sm * 4;
else return 0; else return 0;
} }
struct topology_c* get_topology_info(cudaDeviceProp prop) { struct topology_c* get_topology_info(struct uarch* arch, cudaDeviceProp prop) {
struct topology_c* topo = (struct topology_c*) emalloc(sizeof(struct topology_c)); struct topology_c* topo = (struct topology_c*) emalloc(sizeof(struct topology_c));
topo->streaming_mp = prop.multiProcessorCount; topo->streaming_mp = prop.multiProcessorCount;
topo->cores_per_mp = _ConvertSMVer2Cores(prop.major, prop.minor); topo->cores_per_mp = _ConvertSMVer2Cores(prop.major, prop.minor);
topo->cuda_cores = topo->streaming_mp * topo->cores_per_mp; topo->cuda_cores = topo->streaming_mp * topo->cores_per_mp;
topo->tensor_cores = get_tensor_cores(topo->streaming_mp, prop.major); topo->tensor_cores = get_tensor_cores(arch, topo->streaming_mp, prop.major);
return topo; return topo;
} }
@@ -85,15 +93,7 @@ int64_t get_peak_performance_cuda(struct gpu_info* gpu) {
int64_t get_peak_performance_tcu(cudaDeviceProp prop, struct gpu_info* gpu) { int64_t get_peak_performance_tcu(cudaDeviceProp prop, struct gpu_info* gpu) {
// Volta / Turing tensor cores performs 4x4x4 FP16 matrix multiplication // Volta / Turing tensor cores performs 4x4x4 FP16 matrix multiplication
// Ampere tensor cores performs 8x4x8 FP16 matrix multiplicacion // Ampere tensor cores performs 8x4x8 FP16 matrix multiplicacion
if(prop.major == 7) { if(prop.major == 7) return gpu->freq * 1000000 * 4 * 4 * 4 * 2 * gpu->topo_c->tensor_cores;
// TU116 does not have tensor cores!
// https://www.anandtech.com/show/13973/nvidia-gtx-1660-ti-review-feat-evga-xc-gaming/2
if(gpu->arch->chip == CHIP_TU116 || gpu->arch->chip == CHIP_TU116BM ||
gpu->arch->chip == CHIP_TU116GL || gpu->arch->chip == CHIP_TU116M) {
return 0;
}
return gpu->freq * 1000000 * 4 * 4 * 4 * 2 * gpu->topo_c->tensor_cores;
}
else if(prop.major == 8) return gpu->freq * 1000000 * 8 * 4 * 8 * 2 * gpu->topo_c->tensor_cores; else if(prop.major == 8) return gpu->freq * 1000000 * 8 * 4 * 8 * 2 * gpu->topo_c->tensor_cores;
else return 0; else return 0;
} }
@@ -151,7 +151,7 @@ struct gpu_info* get_gpu_info_cuda(int gpu_idx) {
gpu->arch = get_uarch_from_cuda(gpu); gpu->arch = get_uarch_from_cuda(gpu);
gpu->cach = get_cache_info(deviceProp); gpu->cach = get_cache_info(deviceProp);
gpu->mem = get_memory_info(gpu, deviceProp); gpu->mem = get_memory_info(gpu, deviceProp);
gpu->topo_c = get_topology_info(deviceProp); gpu->topo_c = get_topology_info(gpu->arch, deviceProp);
gpu->peak_performance = get_peak_performance_cuda(gpu); gpu->peak_performance = get_peak_performance_cuda(gpu);
gpu->peak_performance_tcu = get_peak_performance_tcu(deviceProp, gpu); gpu->peak_performance_tcu = get_peak_performance_tcu(deviceProp, gpu);