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
This commit is contained in:
Wenkai Du
2025-07-01 14:38:49 -07:00
committed by GitHub
parent 3e51c41dcb
commit 4640ab19b3
4 changed files with 28 additions and 1 deletions
+4
View File
@@ -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);
}