From d3f88584a77b119d965ac0b200ab4ddc47547331 Mon Sep 17 00:00:00 2001
From: xiebaiyuan <xiebaiyuan@baidu.com>
Date: Thu, 16 May 2019 17:56:05 +0800
Subject: [PATCH] fix crash when close optimise in gpu test close #1626

---
 src/framework/executor.cpp | 30 +++++++++++++++++---------
 src/framework/executor.h   |  5 +++++
 src/framework/type_trait.h | 44 ++++++++++++++++++++++++++++++++++++++
 src/operators/op_param.h   | 21 ++----------------
 4 files changed, 71 insertions(+), 29 deletions(-)
 create mode 100644 src/framework/type_trait.h

diff --git a/src/framework/executor.cpp b/src/framework/executor.cpp
index 47fd9c3c66..2ac7035a70 100644
--- a/src/framework/executor.cpp
+++ b/src/framework/executor.cpp
@@ -467,21 +467,32 @@ PMStatus Executor<Device, T>::Predict() {
     ++op_index;
 #endif
   }
+
+#ifdef PADDLE_MOBILE_PROFILE
+  PrintProfile(profile);
+#endif
+  return PMSuccess;
+}
+
 #ifdef PADDLE_MOBILE_PROFILE
+template <typename Device, typename T>
+void Executor<Device, T>::PrintProfile(
+    const vector<Executor<Device, T>::ProfInfo> &profile) const {
   std::unordered_map<std::string, uint64_t> _tp;
   for (int i = 0; i < profile.size(); i++) {
     const auto &pInfo = profile[i];
     uint64_t timeCost = pInfo.runEnd - pInfo.runBegin;
-    if (ops_of_block0_[i]->Type() == "conv2d" ||
-        ops_of_block0_[i]->Type() == "depthwise_conv2d") {
-      auto inputs = ops_of_block0_[i]->Inputs();
-      auto *filter =
-          GetVarValue<LoDTensor>("Filter", inputs, *(program_.scope));
+    if (this->ops_of_block0_[i]->Type() == "conv2d" ||
+        this->ops_of_block0_[i]->Type() == "depthwise_conv2d") {
+      auto inputs = this->ops_of_block0_[i]->Inputs();
+
+      auto *filter = GetVarValue<ProfileTensorType>("Filter", inputs,
+                                                    *(this->program_.scope));
       int kernel_size = filter->dims()[2];
-      _tp[ops_of_block0_[i]->Type() + "_" + std::to_string(kernel_size)] +=
-          timeCost;
+      _tp[this->ops_of_block0_[i]->Type() + "_" +
+          std::to_string(kernel_size)] += timeCost;
     } else {
-      _tp[ops_of_block0_[i]->Type()] += timeCost;
+      _tp[this->ops_of_block0_[i]->Type()] += timeCost;
     }
   }
   printf("====================[ profile ]======================\n");
@@ -502,9 +513,8 @@ PMStatus Executor<Device, T>::Predict() {
            static_cast<float>(p.second) / _ptotal * 100.0);
   }
   printf("====================[---------]======================\n");
-#endif
-  return PMSuccess;
 }
+#endif
 
 template <typename Device, typename T>
 void Executor<Device, T>::FeedTensorData(const vector<framework::Tensor> &v) {
diff --git a/src/framework/executor.h b/src/framework/executor.h
index 78a4bd61dd..57b05dbc6c 100644
--- a/src/framework/executor.h
+++ b/src/framework/executor.h
@@ -25,6 +25,7 @@ limitations under the License. */
 #include "framework/operator.h"
 #include "framework/program/program.h"
 #include "framework/tensor.h"
+#include "framework/type_trait.h"
 
 namespace paddle_mobile {
 namespace framework {
@@ -98,11 +99,15 @@ class Executor {
   DDim input_dim_last_;
 
 #ifdef PADDLE_MOBILE_PROFILE
+  typedef typename DtypeTensorTrait<Device>::gtype ProfileTensorType;
+
   struct ProfInfo {
     int tid = 0;
     uint64_t runBegin = 0UL;
     uint64_t runEnd = 0UL;
   };
+
+  void PrintProfile(const vector<Executor<Device, T>::ProfInfo> &profile) const;
 #endif
 };
 
diff --git a/src/framework/type_trait.h b/src/framework/type_trait.h
new file mode 100644
index 0000000000..d1a8e30522
--- /dev/null
+++ b/src/framework/type_trait.h
@@ -0,0 +1,44 @@
+/* Copyright (c) 2018 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. */
+
+#pragma once
+
+#include <common/types.h>
+#include <string>
+
+namespace paddle_mobile {
+namespace framework {
+
+template <typename Dtype>
+struct DtypeTensorTrait {
+  // This is the type we obtained in variable.
+  typedef framework::LoDTensor gtype;
+  // This type will be the parent class type
+  // or the same type.
+  typedef framework::Tensor rtype;
+};
+
+#ifdef PADDLE_MOBILE_CL
+template <>
+struct DtypeTensorTrait<GPU_CL> {
+  // This is the type we obtained in variable.
+  typedef framework::CLImage gtype;
+  // This type will be the parent class type
+  // or the same type.
+  typedef framework::CLImage rtype;
+};
+#endif
+
+}  // namespace framework
+}  // namespace paddle_mobile
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index fc34ea1e27..97dbd091a1 100644
--- a/src/operators/op_param.h
+++ b/src/operators/op_param.h
@@ -23,6 +23,7 @@ limitations under the License. */
 #include "framework/lod_tensor.h"
 #include "framework/scope.h"
 #include "framework/tensor.h"
+#include "framework/type_trait.h"
 #include "framework/variable.h"
 
 #ifdef PADDLE_MOBILE_FPGA_V1
@@ -53,25 +54,7 @@ using framework::Variable;
 using std::string;
 using std::vector;
 
-template <typename Dtype>
-struct DtypeTensorTrait {
-  // This is the type we obtained in variable.
-  typedef framework::LoDTensor gtype;
-  // This type will be the parent class type
-  // or the same type.
-  typedef framework::Tensor rtype;
-};
-
-#ifdef PADDLE_MOBILE_CL
-template <>
-struct DtypeTensorTrait<GPU_CL> {
-  // This is the type we obtained in variable.
-  typedef framework::CLImage gtype;
-  // This type will be the parent class type
-  // or the same type.
-  typedef framework::CLImage rtype;
-};
-#endif
+using framework::DtypeTensorTrait;
 
 class OpParam {
  public:
-- 
GitLab