Fix FLAGS_alloc_fill_value (#63428)

sneaxiy · web-flow · commit 9d5f31687cce · 2024-04-12T21:56:12.000+08:00
diff --git a/paddle/fluid/memory/allocation/allocator.h b/paddle/fluid/memory/allocation/allocator.h
@@ -147,7 +147,10 @@ static T&& FillValue(T&& allocation) {
 #if defined(PADDLE_WITH_CUDA)
   if (allocation != nullptr) {
     if (FLAGS_sync_after_alloc || FLAGS_alloc_fill_value >= 0) {
-      PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());
+      bool need_sync = !platform::is_cpu_place(allocation->place());
+      if (need_sync) {
+        PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());
+      }
       if (FLAGS_alloc_fill_value >= 0) {
         VLOG(10) << "Set " << FLAGS_alloc_fill_value << " on "
                  << allocation->ptr() << " " << allocation->place() << " "
@@ -159,7 +162,9 @@ static T&& FillValue(T&& allocation) {
           std::memset(
               allocation->ptr(), FLAGS_alloc_fill_value, allocation->size());
         }
-        PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());
+        if (need_sync) {
+          PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());
+        }
       }
     }
   }

Original file line number	Diff line number	Diff line change
`@@ -147,7 +147,10 @@ static T&& FillValue(T&& allocation) {`
`147`	`147`	`#if defined(PADDLE_WITH_CUDA)`
`148`	`148`	`if (allocation != nullptr) {`
`149`	`149`	`if (FLAGS_sync_after_alloc \|\| FLAGS_alloc_fill_value >= 0) {`
`150`		`- PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());`
	`150`	`+ bool need_sync = !platform::is_cpu_place(allocation->place());`
	`151`	`+ if (need_sync) {`
	`152`	`+ PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());`
	`153`	`+ }`
`151`	`154`	`if (FLAGS_alloc_fill_value >= 0) {`
`152`	`155`	`VLOG(10) << "Set " << FLAGS_alloc_fill_value << " on "`
`153`	`156`	`<< allocation->ptr() << " " << allocation->place() << " "`
`@@ -159,7 +162,9 @@ static T&& FillValue(T&& allocation) {`
`159`	`162`	`std::memset(`
`160`	`163`	`allocation->ptr(), FLAGS_alloc_fill_value, allocation->size());`
`161`	`164`	`}`
`162`		`- PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());`
	`165`	`+ if (need_sync) {`
	`166`	`+ PADDLE_ENFORCE_GPU_SUCCESS(cudaDeviceSynchronize());`
	`167`	`+ }`
`163`	`168`	`}`
`164`	`169`	`}`
`165`	`170`	`}`