modify API.spec

Author: sneaxiy

paddle/fluid/API.spec

@@ -162,6 +162,7 @@ paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.prelu ArgSpec(args=['x', 'mode', 'param_attr', 'name'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
+paddle.fluid.layers.stack ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))

paddle/fluid/operators/stack_op.h

@@ -154,17 +154,22 @@ class StackKernel : public framework::OpKernel<T> {
     if (std::is_same<DeviceContext, platform::CPUDeviceContext>::value ||
         n > kMaxThreshold) {
 #ifdef __NVCC__
+      VLOG(10) << "Stack more than " << kMaxThreshold
+               << " tensors on GPU may be slow.";
       thrust::device_vector<const T *> device_x_vec(x_datas);
       auto x_data_arr = device_x_vec.data().get();
 #else
       auto x_data_arr = x_datas.data();
 #endif
       StackFunctorForRange(dev_ctx, x_data_arr, y_data, total_num, n, post);
+#ifdef __NVCC__
+      // Wait() must be called because device_x_vec may be destructed before
+      // kernel ends
+      dev_ctx.Wait();
+#endif
     }
 #ifdef __NVCC__
     else {  // NOLINT
-      VLOG(10) << "Stack more than " << kMaxThreshold
-               << " tensors on GPU may be slow.";
       framework::Array<const T *, kMaxThreshold> x_data_arr;
       for (int i = 0; i < n; ++i) x_data_arr[i] = x_datas[i];
       StackFunctorForRange(dev_ctx, x_data_arr, y_data, total_num, n, post);
@@ -243,18 +248,23 @@ class StackGradKernel : public framework::OpKernel<T> {
     if (std::is_same<DeviceContext, platform::CPUDeviceContext>::value ||
         n > kMaxThreshold) {
 #ifdef __NVCC__
+      VLOG(10) << "Stack more than " << kMaxThreshold
+               << " tensors on GPU may be slow.";
       thrust::device_vector<T *> device_dx_vec(dx_datas);
       auto dx_data_arr = device_dx_vec.data().get();
 #else
       auto dx_data_arr = dx_datas.data();
 #endif
       StackGradFunctorForRange(dev_ctx, dx_data_arr, dy_data, total_num, n,
                                post);
+#ifdef __NVCC__
+      // Wait() must be called because device_dx_vec may be destructed before
+      // kernel ends
+      dev_ctx.Wait();
+#endif
     }
 #ifdef __NVCC__
     else {  // NOLINT
-      VLOG(10) << "Stack more than " << kMaxThreshold
-               << " tensors on GPU may be slow.";
       framework::Array<T *, kMaxThreshold> dx_data_arr;
       for (int i = 0; i < n; ++i) dx_data_arr[i] = dx_datas[i];
       StackGradFunctorForRange(dev_ctx, dx_data_arr, dy_data, total_num, n,