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

src/include/alloc.h

@@ -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: