AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Use hipExtMallocWithFlags to allocate host memory on APU (#1149)

Browse Source 
Also use SM60 as CUDA compatibility level.
This commit is contained in:
Wenkai Du
2024-04-17 16:56:38 -07:00
committed by GitHub
parent a14137c062
commit 220066197a
2 changed files with 15 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/topo.cc
+5
View File
@@ -371,7 +371,12 @@ ncclResult_t ncclTopoAddNic(struct ncclXmlNode* xmlNic, struct ncclTopoSystem* s
}
ncclResult_t ncclTopoAddGpu(struct ncclXmlNode* xmlGpu, struct ncclTopoSystem* system, struct ncclTopoNode* gpu) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
// There is no direct mapping between CUDA SM to HIP GFX. Use SM60 as compatibility level.
gpu->gpu.cudaCompCap = 60;
#else
NCCLCHECK(xmlGetAttrInt(xmlGpu, "sm", &gpu->gpu.cudaCompCap));
#endif
const char* gcnArch;
const char* gcnArchName;
NCCLCHECK(xmlGetAttr(xmlGpu, "gcn", &gcnArch));
src/include/alloc.h
+10 -1
View File
@@ -27,7 +27,16 @@ ncclResult_t ncclCudaHostCallocDebug(T** ptr, size_t nelem, const char *filefunc
cudaStreamCaptureMode mode = cudaStreamCaptureModeRelaxed;
*ptr = nullptr;
CUDACHECK(cudaThreadExchangeStreamCaptureMode(&mode));
CUDACHECKGOTO(hipHostMalloc(ptr, nelem*sizeof(T), cudaHostAllocMapped), result, finish);
int managed = 0;
CUDACHECK(hipDeviceGetAttribute(&managed, hipDeviceAttributeDirectManagedMemAccessFromHost, 0));
if (managed) {
#if defined(HIP_UNCACHED_MEMORY)
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocUncached), result, finish);
#else
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocFinegrained), result, finish);
#endif
} else
CUDACHECKGOTO(hipHostMalloc(ptr, nelem*sizeof(T), cudaHostAllocMapped), result, finish);
memset(*ptr, 0, nelem*sizeof(T));
finish:
CUDACHECK(cudaThreadExchangeStreamCaptureMode(&mode));