From a19acd20f07eee82622701a3015e6e9c073a5e0b Mon Sep 17 00:00:00 2001
From: Zhen Wang <wangzhen31@baidu.com>
Date: Mon, 6 May 2019 12:59:10 +0800
Subject: [PATCH] change the output into inplace. test=develop

---
 paddle/fluid/operators/fake_quantize_op.cc | 16 +++++++++++++++-
 paddle/fluid/operators/fake_quantize_op.h  |  5 ++---
 2 files changed, 17 insertions(+), 4 deletions(-)

diff --git a/paddle/fluid/operators/fake_quantize_op.cc b/paddle/fluid/operators/fake_quantize_op.cc
index 25ca1f7e0a0..beaa4392f20 100644
--- a/paddle/fluid/operators/fake_quantize_op.cc
+++ b/paddle/fluid/operators/fake_quantize_op.cc
@@ -458,6 +458,18 @@ $$Out = X$$
   }
 };
 
+class MovingAverageAbsMaxScaleOpInplaceInToOut
+    : public framework::InplaceOpInference {
+ public:
+  std::unordered_map<std::string, std::string> operator()(
+      const framework::OpDesc& op_desc) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        {"X", "Out"},
+    };
+    return inplace_in_to_out;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -482,6 +494,7 @@ REGISTER_OPERATOR(fake_quantize_moving_average_abs_max,
                   paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(fake_quantize_moving_average_abs_max,
                        ops::FakeQuantizeMovingAverageAbsMaxKernel<CPU, float>);
+
 REGISTER_OPERATOR(fake_channel_wise_quantize_abs_max,
                   ops::FakeChannelWiseQuantizeAbsMaxOp,
                   ops::FakeChannelWiseQuantizeAbsMaxOpMaker,
@@ -491,6 +504,7 @@ REGISTER_OP_CPU_KERNEL(fake_channel_wise_quantize_abs_max,
 
 REGISTER_OPERATOR(moving_average_abs_max_scale, ops::MovingAverageAbsMaxScaleOp,
                   ops::MovingAverageAbsMaxScaleOpMaker,
-                  paddle::framework::EmptyGradOpMaker);
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::MovingAverageAbsMaxScaleOpInplaceInToOut);
 REGISTER_OP_CPU_KERNEL(moving_average_abs_max_scale,
                        ops::MovingAverageAbsMaxScaleKernel<CPU, float>);
diff --git a/paddle/fluid/operators/fake_quantize_op.h b/paddle/fluid/operators/fake_quantize_op.h
index 87bcece5824..b0fd4590951 100644
--- a/paddle/fluid/operators/fake_quantize_op.h
+++ b/paddle/fluid/operators/fake_quantize_op.h
@@ -15,9 +15,9 @@ limitations under the License. */
 #pragma once
 
 #include <string>
+#include <unordered_map>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/operators/math/blas.h"
 
 namespace paddle {
@@ -204,9 +204,8 @@ class MovingAverageAbsMaxScaleKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* in = context.Input<framework::Tensor>("X");
     auto* out = context.Output<framework::Tensor>("Out");
-    out->mutable_data<T>(context.GetPlace());
     auto& dev_ctx = context.template device_context<DeviceContext>();
-    framework::TensorCopy(*in, context.GetPlace(), dev_ctx, out);
+    out->ShareDataWith(*in);
 
     bool is_test = context.Attr<bool>("is_test");
     // testing
-- 
GitLab