Fix review issues.

Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
NVIDIA · Jun 8, 2022 · 30d4b9f · 30d4b9f
1 parent 1c45b0b
commit 30d4b9f
Show file tree

Hide file tree

Showing 3 changed files with 3 additions and 5 deletions.
diff --git a/dali/c_api/c_api.cc b/dali/c_api/c_api.cc
@@ -589,8 +589,6 @@ void daliOutputCopySamples(daliPipelineHandle *pipe_handle, void **dsts, int out
   if (ws->OutputIsType<CPUBackend>(output_idx)) {
     copy_order = is_pinned ? AccessOrder(stream) : AccessOrder::host();
     auto & src = ws->Output<CPUBackend>(output_idx);
-    if (!host_sync)
-      wait_order = src.order();  // if the copy order is host, then wait will be no-op
     CopyToExternal(dsts, dst_mem_kind, src, copy_order, use_copy_kernel);
     if (!host_sync)
       wait_order = src.order();  // if the copy order is host, then wait will be no-op

diff --git a/dali/c_api/c_api_test.cu b/dali/c_api/c_api_test.cu
@@ -101,8 +101,8 @@ void TestCopyOutput(Method method) {
 
   // This loop is tuned so that if the output buffer is recycled before the asynchronous copy
   // finishes, the buffer is clobbered and an error is detected.
-  // (michalz) Verified on my desktop. The changes in c_api that came with this test
-  // fix the synchronization problem.
+  // In order to trigger a failure, remove the `wait_order.wait` at the end of
+  // daliOutputCopy / daliOutputCopySamples
   for (int attempt = 0; attempt < 20; attempt++) {
     daliPipelineHandle handle;
 

diff --git a/dali/python/backend_impl.cc b/dali/python/backend_impl.cc
@@ -290,7 +290,7 @@ enum DALIDataTypePlaceholder {};
  * The copy will be scheduled on the provided `cuda_stream` or, if left out, on an internal DALI
  * stream.
  * If a non-blocking copy is requested, the function will synchronize the source buffer's
- * associated access order with the provided stream; otherwie, the function will wait until the
+ * associated access order with the provided stream; otherwise, the function will wait until the
  * copy completes.
  *
  * @tparam SourceObject  a data store on GPUBackend (Tensor, TensorList, TensorVector)