diff --git a/src/cuda/cuda.cpp b/src/cuda/cuda.cpp index 20f1d93..027ce2b 100644 --- a/src/cuda/cuda.cpp +++ b/src/cuda/cuda.cpp @@ -1,3 +1,6 @@ + +// patched cuda.cpp for cuda13 by cloudy + #include #include #include @@ -14,25 +17,20 @@ bool print_gpu_cuda(struct gpu_info* gpu) { char* cc = get_str_cc(gpu->arch); printf("%s (Compute Capability %s)\n", gpu->name, cc); free(cc); - return true; } struct cache* get_cache_info(cudaDeviceProp prop) { struct cache* cach = (struct cache*) emalloc(sizeof(struct cache)); - cach->L2 = (struct cach*) emalloc(sizeof(struct cach)); cach->L2->size = prop.l2CacheSize; cach->L2->num_caches = 1; cach->L2->exists = true; - return cach; } int get_tensor_cores(struct uarch* arch, int sm, int major) { 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 (is_chip_TU116(arch)) return 0; return sm * 8; @@ -43,57 +41,57 @@ int get_tensor_cores(struct uarch* arch, int sm, int major) { struct topology_c* get_topology_info(struct uarch* arch, cudaDeviceProp prop) { struct topology_c* topo = (struct topology_c*) emalloc(sizeof(struct topology_c)); - topo->streaming_mp = prop.multiProcessorCount; topo->cores_per_mp = _ConvertSMVer2Cores(prop.major, prop.minor); topo->cuda_cores = topo->streaming_mp * topo->cores_per_mp; topo->tensor_cores = get_tensor_cores(arch, topo->streaming_mp, prop.major); - return topo; } int32_t guess_clock_multipilier(struct gpu_info* gpu, struct memory* mem) { - // Guess clock multiplier int32_t clk_mul = 1; - int32_t clk8 = abs((mem->freq/8) - gpu->freq); int32_t clk4 = abs((mem->freq/4) - gpu->freq); int32_t clk2 = abs((mem->freq/2) - gpu->freq); int32_t clk1 = abs((mem->freq/1) - gpu->freq); - int32_t min = mem->freq; if(clkm_possible_for_uarch(8, gpu->arch) && min > clk8) { clk_mul = 8; min = clk8; } if(clkm_possible_for_uarch(4, gpu->arch) && min > clk4) { clk_mul = 4; min = clk4; } if(clkm_possible_for_uarch(2, gpu->arch) && min > clk2) { clk_mul = 2; min = clk2; } if(clkm_possible_for_uarch(1, gpu->arch) && min > clk1) { clk_mul = 1; min = clk1; } - return clk_mul; } struct memory* get_memory_info(struct gpu_info* gpu, cudaDeviceProp prop) { struct memory* mem = (struct memory*) emalloc(sizeof(struct memory)); + int val = 0; mem->size_bytes = (unsigned long long) prop.totalGlobalMem; - mem->freq = prop.memoryClockRate * 0.001f; + + if (cudaDeviceGetAttribute(&val, cudaDevAttrMemoryClockRate, gpu->idx) == cudaSuccess) { + if (val > 1000000) + mem->freq = (float)val / 1000000.0f; + else + mem->freq = (float)val * 0.001f; + } else { + mem->freq = 0.0f; + } + mem->bus_width = prop.memoryBusWidth; mem->clk_mul = guess_clock_multipilier(gpu, mem); mem->type = guess_memtype_from_cmul_and_uarch(mem->clk_mul, gpu->arch); - // Fix frequency returned from CUDA to show real frequency - mem->freq = mem->freq / mem->clk_mul; + if (mem->clk_mul > 0) + mem->freq = mem->freq / mem->clk_mul; return mem; } -// Compute peak performance when using CUDA cores int64_t get_peak_performance_cuda(struct gpu_info* gpu) { return gpu->freq * 1000000 * gpu->topo_c->cuda_cores * 2; } -// Compute peak performance when using tensor cores int64_t get_peak_performance_tcu(cudaDeviceProp prop, struct gpu_info* gpu) { - // Volta / Turing tensor cores performs 4x4x4 FP16 matrix multiplication - // Ampere tensor cores performs 8x4x8 FP16 matrix multiplicacion if(prop.major == 7) 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 return 0; @@ -115,8 +113,7 @@ struct gpu_info* get_gpu_info_cuda(struct pci_dev *devices, int gpu_idx) { } int num_gpus = -1; - cudaError_t err = cudaSuccess; - err = cudaGetDeviceCount(&num_gpus); + cudaError_t err = cudaGetDeviceCount(&num_gpus); if(gpu_idx == 0) { printf("\r%*c\r", (int) strlen(CUDA_DRIVER_START_WARNING), ' '); @@ -134,7 +131,6 @@ struct gpu_info* get_gpu_info_cuda(struct pci_dev *devices, int gpu_idx) { } if(gpu->idx+1 > num_gpus) { - // Master is trying to query an invalid GPU return NULL; } @@ -144,15 +140,25 @@ struct gpu_info* get_gpu_info_cuda(struct pci_dev *devices, int gpu_idx) { return NULL; } - gpu->freq = deviceProp.clockRate * 1e-3f; + int core_clk = 0; + if (cudaDeviceGetAttribute(&core_clk, cudaDevAttrClockRate, gpu->idx) == cudaSuccess) { + if (core_clk > 1000000) + gpu->freq = core_clk / 1000000.0f; + else + gpu->freq = core_clk * 0.001f; + } else { + gpu->freq = 0.0f; + } + gpu->vendor = GPU_VENDOR_NVIDIA; - gpu->name = (char *) emalloc(sizeof(char) * (strlen(deviceProp.name) + 1)); + gpu->name = (char *) emalloc(strlen(deviceProp.name) + 1); strcpy(gpu->name, deviceProp.name); if((gpu->pci = get_pci_from_pciutils(devices, PCI_VENDOR_ID_NVIDIA, gpu_idx)) == NULL) { printErr("Unable to find a valid device for vendor id 0x%.4X using pciutils", PCI_VENDOR_ID_NVIDIA); return NULL; } + gpu->arch = get_uarch_from_cuda(gpu); gpu->cach = get_cache_info(deviceProp); gpu->mem = get_memory_info(gpu, deviceProp); @@ -163,19 +169,7 @@ struct gpu_info* get_gpu_info_cuda(struct pci_dev *devices, int gpu_idx) { return gpu; } -char* get_str_sm(struct gpu_info* gpu) { - return get_str_generic(gpu->topo_c->streaming_mp); -} - -char* get_str_cores_sm(struct gpu_info* gpu) { - return get_str_generic(gpu->topo_c->cores_per_mp); -} - -char* get_str_cuda_cores(struct gpu_info* gpu) { - return get_str_generic(gpu->topo_c->cuda_cores); -} - -char* get_str_tensor_cores(struct gpu_info* gpu) { - return get_str_generic(gpu->topo_c->tensor_cores); -} - +char* get_str_sm(struct gpu_info* gpu) { return get_str_generic(gpu->topo_c->streaming_mp); } +char* get_str_cores_sm(struct gpu_info* gpu) { return get_str_generic(gpu->topo_c->cores_per_mp); } +char* get_str_cuda_cores(struct gpu_info* gpu) { return get_str_generic(gpu->topo_c->cuda_cores); } +char* get_str_tensor_cores(struct gpu_info* gpu) { return get_str_generic(gpu->topo_c->tensor_cores); }