Detect HIP_UNCACHED_MEMORY support from HIP version (#842)

This commit is contained in:
Wenkai Du
2023-08-04 10:17:04 -07:00
committed by GitHub
parent 8e58b65873
commit d65c0830c6
3 changed files with 10 additions and 13 deletions
+3 -9
View File
@@ -167,9 +167,7 @@ ncclResult_t ncclCudaMallocDebug(const char *filefunc, int line, T** ptr, size_t
CUDACHECK(cudaThreadExchangeStreamCaptureMode(&mode));
if (isFineGrain) {
#if defined(HIP_UNCACHED_MEMORY)
hipDeviceProp_t prop;
CUDACHECK(hipGetDeviceProperties(&prop, 0));
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), prop.gcnArch/10 == 94 ? hipDeviceMallocUncached : hipDeviceMallocFinegrained), result, finish);
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocUncached), result, finish);
#else
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocFinegrained), result, finish);
#endif
@@ -195,9 +193,7 @@ ncclResult_t ncclCudaCallocDebug(const char *filefunc, int line, T** ptr, size_t
CUDACHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
if (isFineGrain) {
#if defined(HIP_UNCACHED_MEMORY)
hipDeviceProp_t prop;
CUDACHECK(hipGetDeviceProperties(&prop, 0));
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), prop.gcnArch/10 == 94 ? hipDeviceMallocUncached : hipDeviceMallocFinegrained), result, finish);
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocUncached), result, finish);
#else
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocFinegrained), result, finish);
#endif
@@ -229,9 +225,7 @@ ncclResult_t ncclCudaCallocAsyncDebug(const char *filefunc, int line, T** ptr, s
CUDACHECK(cudaThreadExchangeStreamCaptureMode(&mode));
if (isFineGrain) {
#if defined(HIP_UNCACHED_MEMORY)
hipDeviceProp_t prop;
CUDACHECK(hipGetDeviceProperties(&prop, 0));
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), prop.gcnArch/10 == 94 ? hipDeviceMallocUncached : hipDeviceMallocFinegrained), result, finish);
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocUncached), result, finish);
#else
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocFinegrained), result, finish);
#endif