From 454f4f21406483845d6206863a5b3edb50fe3e0b Mon Sep 17 00:00:00 2001
From: Yiqun Liu <liuyiqun01@baidu.com>
Date: Wed, 27 Feb 2019 10:51:34 +0800
Subject: [PATCH] Rewrite is_empty op to avoid unnecessary data transform.
 (#15509)

* Rewrite is_empty op to avoid unnecessary data transform.
test=develop

* Add the implementation of InferShape and InferVarType for is_empty op.
test=develop

* Rewrite is_empty op to avoid directly inherit OperatorBase.
test=develop
---
 paddle/fluid/operators/is_empty_op.cc    |  6 ++----
 paddle/fluid/operators/is_empty_op.cu.cc | 23 +++++++++++++++++++++++
 paddle/fluid/operators/is_empty_op.h     |  3 +++
 3 files changed, 28 insertions(+), 4 deletions(-)
 create mode 100644 paddle/fluid/operators/is_empty_op.cu.cc

diff --git a/paddle/fluid/operators/is_empty_op.cc b/paddle/fluid/operators/is_empty_op.cc
index ba50bdf34..092a6eae6 100644
--- a/paddle/fluid/operators/is_empty_op.cc
+++ b/paddle/fluid/operators/is_empty_op.cc
@@ -34,9 +34,8 @@ class IsEmptyOp : public framework::OperatorWithKernel {
 
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = framework::OpKernelType(
-        ctx.Input<framework::LoDTensor>("X")->type(), platform::CPUPlace());
-    return kt;
+    auto *x = ctx.Input<framework::LoDTensor>("X");
+    return framework::OpKernelType(x->type(), x->place());
   }
 };
 
@@ -58,7 +57,6 @@ It will just return product(tensor.ddims()) > 0;
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-
 REGISTER_OPERATOR(is_empty, ops::IsEmptyOp, ops::IsEmptyOpMaker,
                   paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(
diff --git a/paddle/fluid/operators/is_empty_op.cu.cc b/paddle/fluid/operators/is_empty_op.cu.cc
new file mode 100644
index 000000000..3c256503b
--- /dev/null
+++ b/paddle/fluid/operators/is_empty_op.cu.cc
@@ -0,0 +1,23 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/is_empty_op.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    is_empty, ops::IsEmptyOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::IsEmptyOpKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::IsEmptyOpKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::IsEmptyOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
diff --git a/paddle/fluid/operators/is_empty_op.h b/paddle/fluid/operators/is_empty_op.h
index 3e3af22fa..4f6419eb5 100644
--- a/paddle/fluid/operators/is_empty_op.h
+++ b/paddle/fluid/operators/is_empty_op.h
@@ -28,6 +28,9 @@ class IsEmptyOpKernel : public framework::OpKernel<T> {
     // get output
     auto* output_tensor = context.Output<framework::LoDTensor>("Out");
 
+    // Note: is_empty is always executed on CPU and the output data should
+    // always be allocated for CPUPlace. We reigister CUDA kernel for this op to
+    // avoid the unnecessary data transform.
     output_tensor->mutable_data<bool>(platform::CPUPlace())[0] =
         framework::product(input_tensor->dims()) == 0;
   }
-- 
GitLab