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

[v0.04] Fix bug in which NVML was using the wrong GPU id

Browse Source
This commit is contained in:
Dr-Noob
2021-08-16 15:18:55 +02:00
parent fd15008ab2
commit dea5211791
3 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/cuda/cuda.cpp
View File

@@ -67,8 +67,9 @@ int64_t get_peak_performance(struct gpu_info* gpu) {
struct gpu_info* get_gpu_info(int gpu_idx) {
struct gpu_info* gpu = (struct gpu_info*) emalloc(sizeof(struct gpu_info));
gpu->pci = NULL;
gpu->idx = gpu_idx;
if(gpu_idx < 0) {
if(gpu->idx < 0) {
printErr("GPU index must be equal or greater than zero");
return NULL;
}
@@ -89,13 +90,13 @@ struct gpu_info* get_gpu_info(int gpu_idx) {
return NULL;
}
if(gpu_idx+1 > num_gpus) {
printErr("Requested GPU index %d in a system with %d GPUs", gpu_idx, num_gpus);
if(gpu->idx+1 > num_gpus) {
printErr("Requested GPU index %d in a system with %d GPUs", gpu->idx, num_gpus);
return NULL;
}
cudaDeviceProp deviceProp;
if ((err = cudaGetDeviceProperties(&deviceProp, gpu_idx)) != cudaSuccess) {
if ((err = cudaGetDeviceProperties(&deviceProp, gpu->idx)) != cudaSuccess) {
printErr("%s: %s", cudaGetErrorName(err), cudaGetErrorString(err));
return NULL;
}
@@ -106,7 +107,7 @@ struct gpu_info* get_gpu_info(int gpu_idx) {
strcpy(gpu->name, deviceProp.name);
gpu->nvmld = nvml_init();
if(nvml_get_pci_info(gpu_idx, gpu->nvmld)) {
if(nvml_get_pci_info(gpu->idx, gpu->nvmld)) {
gpu->pci = get_pci_from_nvml(gpu->nvmld);
}

7
src/cuda/uarch.cpp
View File

@@ -252,9 +252,12 @@ void map_chip_to_uarch(struct uarch* arch) {
struct uarch* get_uarch_from_cuda(struct gpu_info* gpu) {
struct uarch* arch = (struct uarch*) emalloc(sizeof(struct uarch));
int dev = 0;
cudaError_t err = cudaSuccess;
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, dev);
if ((err = cudaGetDeviceProperties(&deviceProp, gpu->idx)) != cudaSuccess) {
printErr("%s: %s", cudaGetErrorName(err), cudaGetErrorString(err));
return NULL;
}
arch->chip_str = NULL;
arch->cc_major = deviceProp.major;