Add support for extended fine grained system memory pool (#1770)

* Add support for extended fine-grained system memory pool
* Use hipHostRegisterUncached
* Add "sc0 sc1" flags for LL store on gfx950
* Update after HIP flag is changed to hipExtHostRegisterUncached

[ROCm/rccl commit: 4640ab19b3]
Этот коммит содержится в:
Wenkai Du authored and GitHub committed 2025-07-01 16:38:49 -05:00
1 родитель fd0d9ac44c
Коммит 6db3b4cd4f
4 files changed
+28 -1

No files matched your search

+4
Просмотреть файл
@@ -114,7 +114,11 @@ ncclResult_t ncclShmOpen(char* shmPath, size_t shmPathSize, size_t shmSize, void
}
if (devShmPtr) {
#if defined(HIP_HOST_UNCACHED_MEMORY)
CUDACHECKGOTO(cudaHostRegister((void*)hptr, realShmSize, cudaHostRegisterPortable | cudaHostRegisterMapped | hipExtHostRegisterUncached), ret, fail);
#else
CUDACHECKGOTO(cudaHostRegister((void*)hptr, realShmSize, cudaHostRegisterPortable | cudaHostRegisterMapped), ret, fail);
#endif
CUDACHECKGOTO(cudaHostGetDevicePointer(&dptr, (void*)hptr, 0), ret, fail);
}