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]
这个提交包含在:
Wenkai Du
2025-07-01 14:38:49 -07:00
提交者 GitHub
父节点 fd0d9ac44c
当前提交 6db3b4cd4f
修改 4 个文件,包含 28 行新增1 行删除
+4
查看文件
@@ -122,7 +122,11 @@ ncclResult_t ncclCudaHostCallocDebug(T** ptr, size_t nelem, const char *filefunc
CUDACHECKGOTO(hipExtMallocWithFlags((void**)ptr, nelem*ncclSizeOfT<T>(), hipDeviceMallocFinegrained), result, finish);
#endif
} else
#if defined(HIP_HOST_UNCACHED_MEMORY)
CUDACHECKGOTO(hipHostMalloc(ptr, nelem*ncclSizeOfT<T>(), cudaHostAllocMapped | hipHostMallocUncached), result, finish);
#else
CUDACHECKGOTO(hipHostMalloc(ptr, nelem*ncclSizeOfT<T>(), cudaHostAllocMapped), result, finish);
#endif
memset(*ptr, 0, nelem*ncclSizeOfT<T>());
}
finish: