use default stream and make 1D copy async

oneapi-src · Mar 1, 2024 · aa04d5d · aa04d5d
1 parent 549d56d
commit aa04d5d
Showing 1 changed file with 13 additions and 4 deletions.
diff --git a/source/adapters/cuda/image.cpp b/source/adapters/cuda/image.cpp
@@ -642,7 +642,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
 
   try {
     ScopedContext Active(hQueue->getContext());
-    CUstream Stream = hQueue->getNextTransferStream();
+    CUstream Stream = 0;
     enqueueEventsWait(hQueue, Stream, numEventsInWaitList, phEventWaitList);
 
     // We have to use a different copy function for each image dimensionality.
@@ -817,9 +817,18 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
       }
     } else {
       if (pImageDesc->type == UR_MEM_TYPE_IMAGE1D) {
-        UR_CHECK_ERROR(cuMemcpyAtoA((CUarray)pDst, dstOffset.x * PixelSizeBytes,
-                                    (CUarray)pSrc, 0,
-                                    PixelSizeBytes * copyExtent.width));
+        CUDA_MEMCPY2D cpy_desc = {};
+        cpy_desc.srcXInBytes = srcOffset.x;
+        cpy_desc.srcY = 0;
+        cpy_desc.dstXInBytes = dstOffset.x;
+        cpy_desc.dstY = 0;
+        cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
+        cpy_desc.srcArray = (CUarray)pSrc;
+        cpy_desc.dstMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
+        cpy_desc.dstArray = (CUarray)pDst;
+        cpy_desc.WidthInBytes = PixelSizeBytes * copyExtent.width;
+        cpy_desc.Height = 1;
+        UR_CHECK_ERROR(cuMemcpy2DAsync(&cpy_desc, Stream));
       } else if (pImageDesc->type == UR_MEM_TYPE_IMAGE2D) {
         CUDA_MEMCPY2D cpy_desc = {};
         cpy_desc.srcXInBytes = srcOffset.x;