RCCL 2.4 update

src/include/alloc.h

@@ -12,14 +12,14 @@
 #include <sys/mman.h>
 
 static inline ncclResult_t ncclCudaHostAlloc(void** ptr, void** devPtr, size_t size) {
-  CUDACHECK(cudaHostAlloc(ptr, size, cudaHostAllocMapped));
+  CUDACHECK(hipHostMalloc(ptr, size, hipHostMallocMapped));
   memset(*ptr, 0, size);
   *devPtr = *ptr;
   return ncclSuccess;
 }
 
 static inline ncclResult_t ncclCudaHostFree(void* ptr) {
-  CUDACHECK(cudaFreeHost(ptr));
+  CUDACHECK(hipHostFree(ptr));
   return ncclSuccess;
 }
 
@@ -36,15 +36,18 @@ static ncclResult_t ncclCalloc(T** ptr, size_t nelem) {
 }
 
 template <typename T>
-static ncclResult_t ncclCudaCalloc(T** ptr, size_t nelem) {
-  CUDACHECK(cudaMalloc(ptr, nelem*sizeof(T)));
-  CUDACHECK(cudaMemset(*ptr, 0, nelem*sizeof(T)));
+static ncclResult_t ncclCudaCalloc(T** ptr, size_t nelem, bool isFineGrain = false) {
+  if (isFineGrain)
+    CUDACHECK(hipExtMallocWithFlags((void**)ptr, nelem*sizeof(T), hipDeviceMallocFinegrained));
+  else
+    CUDACHECK(hipMalloc(ptr, nelem*sizeof(T)));
+  CUDACHECK(hipMemset(*ptr, 0, nelem*sizeof(T)));
   return ncclSuccess;
 }
 
 template <typename T>
 static ncclResult_t ncclCudaMemcpy(T* dst, T* src, size_t nelem) {
-  CUDACHECK(cudaMemcpy(dst, src, nelem*sizeof(T), cudaMemcpyDefault));
+  CUDACHECK(hipMemcpy(dst, src, nelem*sizeof(T), hipMemcpyDefault));
   return ncclSuccess;
 }