Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_bug_for_lstmp

README.md

 # PaddlePaddle
 
+English | [简体中文](./README_cn.md)
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
 [![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
@@ -7,7 +8,6 @@
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
-
 Welcome to the PaddlePaddle GitHub.
 
 PaddlePaddle (PArallel Distributed Deep LEarning) is an easy-to-use,
@@ -18,16 +18,6 @@ learning to many products at Baidu.
 Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
-
-欢迎来到 PaddlePaddle GitHub
-
-PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
-
-我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
-
-跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
-
-
 ### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
 ### Install Latest Stable Release:
 ```
@@ -43,23 +33,6 @@ pip install paddlepaddle-gpu==1.2.0.post85
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```
 
-
-### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
-### 安装最新稳定版本:
-```
-# Linux CPU
-pip install paddlepaddle
-# Linux GPU cuda9cudnn7
-pip install paddlepaddle-gpu
-# Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.2.0.post87
-# Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.2.0.post85
-
-# 其他平台上的安装指引请参考 http://paddlepaddle.org/
-```
-
-
 ## Features
 
 - **Flexibility**
@@ -100,38 +73,10 @@ pip install paddlepaddle-gpu==1.2.0.post85
     Baidu and it has achieved a significant impact. We hope you can also explore
     the capability of PaddlePaddle to make an impact on your product.
 
-## 特点
-
-- **灵活性**
-
-    PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
-
--  **高效性**
-
-    为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
-    
-    - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
-    - 通过MKL-DNN库优化CNN网络
-    - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
-    - 针对高维稀疏数据模型，优化了局部和分布式训练。
-     
-
-- **稳定性**
-
-    有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
-
-- **连接产品**
-
-    另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
-
 ## Installation
 
 It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
 
-## 安装
-
-推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) 
-
 ## Documentation
 
 We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
@@ -153,37 +98,9 @@ We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarte
 
    We appreciate your contributions!
 
-## 文档
-
-我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
-[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
-
-- [深度学习101](https://github.com/PaddlePaddle/book)
-
-  或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
-
-- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
-
-  可以在MPI集群上运行分布式训练任务
-
-- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
-
-   新的API支持代码更少更简洁的程序
-
-- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
-
-   欢迎您的贡献!
-
 ## Ask Questions
 
 You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues).
 
-## 答疑
-
-欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
-
 ## Copyright and License
 PaddlePaddle is provided under the [Apache-2.0 license](LICENSE).
-
-## 版权和许可证
-PaddlePaddle由[Apache-2.0 license](LICENSE)提供

README_cn.md

+# PaddlePaddle
+
+[English](./README.md) | 简体中文
+
+[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
+[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
+[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
+
+欢迎来到 PaddlePaddle GitHub
+
+PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
+
+我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
+
+跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
+
+### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### 安装最新稳定版本:
+```
+# Linux CPU
+pip install paddlepaddle
+# Linux GPU cuda9cudnn7
+pip install paddlepaddle-gpu
+# Linux GPU cuda8cudnn7
+pip install paddlepaddle-gpu==1.2.0.post87
+# Linux GPU cuda8cudnn5
+pip install paddlepaddle-gpu==1.2.0.post85
+
+# 其他平台上的安装指引请参考 http://paddlepaddle.org/
+```
+
+## 特性
+
+- **灵活性**
+
+    PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
+
+-  **高效性**
+
+    为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
+
+    - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
+    - 通过MKL-DNN库优化CNN网络
+    - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
+    - 针对高维稀疏数据模型，优化了局部和分布式训练。
+
+
+- **稳定性**
+
+    有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
+
+- **与产品相连**
+
+    另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
+
+## 安装
+
+推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
+
+## 文档
+
+我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
+[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
+
+- [深度学习101](https://github.com/PaddlePaddle/book)
+
+  或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
+
+- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+
+  可以在MPI集群上运行分布式训练任务
+
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+
+   新的API支持代码更少更简洁的程序
+
+- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+
+   欢迎您的贡献!
+
+## 答疑
+
+欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
+
+## 版权和许可证
+PaddlePaddle由[Apache-2.0 license](LICENSE)提供

paddle/fluid/API.spec

@@ -324,7 +324,8 @@ paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes',
 paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0))
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None))
+paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.box_clip ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.multiclass_nms ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None))
 paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
 paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))
@@ -361,6 +362,9 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b
 paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None))
 paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.contrib.Calibrator.__init__ ArgSpec(args=['self'], varargs='args', keywords='kwargs', defaults=None)
+paddle.fluid.contrib.Calibrator.sample_data ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.contrib.Calibrator.save_int8_model ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.contrib.reader.ctr_reader.ctr_reader ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
 paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))

paddle/fluid/framework/ir/CMakeLists.txt

@@ -65,6 +65,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference)
 pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(conv_affine_channel_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
+pass_library(identity_scale_op_clean_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -117,11 +117,6 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
       // return false;
     }
   }
-  for (auto &item : pdnodes2nodes_) {
-    for (auto &n : item.second) {
-      GetMarkedNodes(const_cast<Graph *>(&graph)).insert(n);
-    }
-  }
   VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
 
   return !pdnodes2nodes_.empty();

paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc

+// Copyright (c) 2019 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/framework/ir/identity_scale_op_clean_pass.h"
+#include <string>
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  FusePassBase::Init("identity_scale_op_clean", graph.get());
+
+  // pre_op -> scale_in -> scale_op -> scale_out
+  // ->
+  // pre_op -> scale_out
+  GraphPatternDetector detector;
+  auto pre_op = detector.mutable_pattern()->NewNode("pre_op")->assert_is_op();
+  auto scale_in = detector.mutable_pattern()
+                      ->NewNode("scale_in")
+                      ->assert_is_op_input("scale")
+                      ->AsIntermediate();
+  auto scale_op = detector.mutable_pattern()
+                      ->NewNode("scale_fuse")
+                      ->assert_is_op("scale")
+                      ->assert_op_attr<float>("scale", 1.)
+                      ->assert_op_attr<float>("bias", 0.);
+  auto scale_out = detector.mutable_pattern()
+                       ->NewNode("scale_out")
+                       ->assert_is_op_output("scale");
+
+  pre_op->LinksTo({scale_in});
+  scale_op->LinksFrom({scale_in}).LinksTo({scale_out});
+
+  GraphPatternDetector::handle_t handler = [&](
+      const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
+    Node* scale_op_var = subgraph.at(scale_op);
+    Node* scale_in_var = subgraph.at(scale_in);
+    Node* scale_out_var = subgraph.at(scale_out);
+    Node* pre_op_var = subgraph.at(pre_op);
+    // Link pre_op directly to scale_out
+    const std::string scale_in_name = scale_in_var->Name();
+    const std::string scale_out_name = scale_out_var->Name();
+    // Remove links in graph
+    GraphSafeRemoveNodes(graph, {scale_in_var, scale_op_var});
+    // Modify proto message
+    auto* pre_op_desc = pre_op_var->Op();
+    for (auto& parameter : *pre_op_desc->Proto()->mutable_outputs()) {
+      auto* arguments = parameter.mutable_arguments();
+      auto it = std::find(arguments->begin(), arguments->end(), scale_in_name);
+      PADDLE_ENFORCE(it != arguments->end());
+      *it = scale_out_name;
+    }
+
+    IR_NODE_LINK_TO(pre_op_var, scale_out_var);
+  };
+
+  detector(graph.get(), handler);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(identity_scale_op_clean_pass,
+              paddle::framework::ir::IdentityScaleOpCleanPass);

paddle/fluid/framework/ir/identity_scale_op_clean_pass.h

+// Copyright (c) 2019 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 "paddle/fluid/framework/ir/fuse_pass_base.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class IdentityScaleOpCleanPass : public FusePassBase {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+
+ private:
+  virtual ~IdentityScaleOpCleanPass() = default;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/scope.cc

@@ -22,11 +22,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/string/printf.h"
 
-DEFINE_bool(benchmark, false,
-            "Doing memory benchmark. It will make deleting scope synchronized, "
-            "and add some memory usage logs."
-            "Default cuda is asynchronous device, set to True will"
-            "force op run in synchronous mode.");
+DECLARE_bool(benchmark);
 
 DEFINE_bool(
     eager_delete_scope, true,

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -83,7 +83,6 @@ void IRPassManager::CreatePasses(Argument *argument,
           new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
     }
 
-    // graph_ = pass->Apply(std::move(graph_));
     pre_pass = pass_name;
 
     passes_.emplace_back(std::move(pass));
@@ -97,8 +96,9 @@ std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
   PADDLE_ENFORCE(graph.get());
   // Apply all the passes
   for (const auto &pass : passes_) {
-    if (pass->Type() == "graph_viz_pass") continue;
-    PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
+    if (pass->Type() != "graph_viz_pass") {
+      PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
+    }
     graph = pass->Apply(std::move(graph));
   }
   return std::move(graph);

paddle/fluid/inference/api/analysis_config.cc

@@ -318,4 +318,9 @@ NativeConfig AnalysisConfig::ToNativeConfig() const {
   return config;
 }
 
+void AnalysisConfig::SwitchIrDebug(int x) {
+  ir_debug_ = x;
+  Update();
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor.cc

@@ -58,7 +58,8 @@ namespace {
 bool IsPersistable(const framework::VarDesc *var) {
   if (var->Persistable() &&
       var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
-      var->GetType() != framework::proto::VarType::FETCH_LIST) {
+      var->GetType() != framework::proto::VarType::FETCH_LIST &&
+      var->GetType() != framework::proto::VarType::RAW) {
     return true;
   }
   return false;

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -196,7 +196,7 @@ TEST(AnalysisPredictor, memory_optim) {
   AnalysisConfig config(FLAGS_dirname);
   config.DisableGpu();
   config.EnableMemoryOptim(true);
-  config.pass_builder()->TurnOnDebug();
+  config.SwitchIrDebug();
 
   auto native_predictor =
       CreatePaddlePredictor<NativeConfig>(config.ToNativeConfig());

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -140,9 +140,12 @@ struct AnalysisConfig {
    */
   bool tensorrt_engine_enabled() const { return use_tensorrt_; }
 
-  /** Control whther to debug IR graph analysis phase.
+  /** \brief Control whether to debug IR graph analysis phase.
+   *
+   * This will generate DOT files for visualizing the computation graph after
+   * each analysis pass applied.
    */
-  void SwitchIrDebug(int x = true) { ir_debug_ = x; }
+  void SwitchIrDebug(int x = true);
 
   /** Turn on MKLDNN.
    */

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -117,6 +117,7 @@ class CpuPassStrategy : public PassStrategy {
         "conv_bn_fuse_pass",             //
         "conv_eltwiseadd_bn_fuse_pass",  //
         "is_test_pass",                  //
+        "identity_scale_op_clean_pass",  //
     });
     use_gpu_ = false;
   }
@@ -155,6 +156,7 @@ class GpuPassStrategy : public PassStrategy {
   GpuPassStrategy() : PassStrategy({}) {
     passes_.assign({
       "infer_clean_graph_pass",                        //
+          "identity_scale_op_clean_pass",              //
           "conv_affine_channel_fuse_pass",             //
           "conv_eltwiseadd_affine_channel_fuse_pass",  //
           "conv_bn_fuse_pass",                         //

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -128,9 +128,9 @@ inference_analysis_api_test_with_fake_data(test_analyzer_resnet50
 inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet_depthwise_conv
   "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz" SERIAL)
 
-# bert, max_len=20
-set(BERT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/bert20")
-download_model_and_data(${BERT_INSTALL_DIR} "bert_model.tar.gz" "bert_data_len20.txt.tar.gz")
+# bert, max_len=20, embedding_dim=128
+set(BERT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/bert_emb128")
+download_model_and_data(${BERT_INSTALL_DIR} "bert_emb128_model.tar.gz" "bert_data_len20.txt.tar.gz")
 inference_analysis_api_test(test_analyzer_bert ${BERT_INSTALL_DIR} analyzer_bert_tester.cc SERIAL)
 
 # anakin

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -142,7 +142,7 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
   cfg->SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params");
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
-  cfg->pass_builder()->TurnOnDebug();
+  cfg->SwitchIrDebug();
   cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
   if (use_mkldnn) {
     cfg->EnableMKLDNN();

paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc

@@ -69,7 +69,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 TEST(Analyzer_Text_Classification, profile) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
-  cfg.pass_builder()->TurnOnDebug();
+  cfg.SwitchIrDebug();
   std::vector<PaddleTensor> outputs;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;

paddle/fluid/inference/utils/CMakeLists.txt

 cc_library(benchmark SRCS benchmark.cc DEPS enforce)
 cc_test(test_benchmark SRCS benchmark_tester.cc DEPS benchmark)
-#cc_binary(visualizer SRCS visualizer.cc DEPS analysis
-#    paddle_pass_builder ir_pass_manager pass graph_viz_pass analysis_passes)
+cc_binary(visualizer SRCS visualizer.cc DEPS analysis
+    paddle_pass_builder ir_pass_manager pass graph_viz_pass analysis_passes)

paddle/fluid/memory/allocation/legacy_allocator.cc

@@ -35,6 +35,7 @@ DEFINE_bool(init_allocated_mem, false,
             "To find this error in time, we use init_allocated_mem to indicate "
             "that initializing the allocated memory with a small value "
             "during unit testing.");
+DECLARE_bool(benchmark);
 DECLARE_double(fraction_of_gpu_memory_to_use);
 
 namespace paddle {
@@ -59,11 +60,6 @@ size_t memory_usage(const platform::Place &p);
 
 using BuddyAllocator = detail::BuddyAllocator;
 
-std::unordered_map</*device id*/ int,
-                   std::pair</*current memory usage*/ uint64_t,
-                             /*peak memory usage*/ uint64_t>>
-    gpu_mem_info;
-
 BuddyAllocator *GetCPUBuddyAllocator() {
   // We tried thread_local for inference::RNN1 model, but that not works much
   // for multi-thread test.
@@ -144,6 +140,8 @@ BuddyAllocator *GetGPUBuddyAllocator(int gpu_id) {
     devices = platform::GetSelectedDevices();
     int gpu_num = devices.size();
 
+    allocation::GPUMemMonitor.Initialize(devices.size());
+
     a_arr = new BuddyAllocator *[gpu_num];
     for (size_t i = 0; i < devices.size(); ++i) {
       int dev_id = devices[i];
@@ -204,12 +202,7 @@ void *Alloc<platform::CUDAPlace>(const platform::CUDAPlace &place,
                  << string::HumanReadableSize(Used<platform::CUDAPlace>(place));
     platform::SetDeviceId(cur_dev);
   } else {
-    gpu_mem_info[place.device].first += size;
-    if (gpu_mem_info[place.device].first > gpu_mem_info[place.device].second) {
-      gpu_mem_info[place.device].second = gpu_mem_info[place.device].first;
-      VLOG(3) << "device: " << place.device << " peak memory usage : "
-              << (gpu_mem_info[place.device].second >> 20) << " MiB";
-    }
+    if (FLAGS_benchmark) allocation::GPUMemMonitor.Add(place.device, size);
     if (FLAGS_init_allocated_mem) {
       cudaMemset(ptr, 0xEF, size);
     }
@@ -225,7 +218,7 @@ void Free<platform::CUDAPlace>(const platform::CUDAPlace &place, void *p,
                                size_t size) {
 #ifdef PADDLE_WITH_CUDA
   GetGPUBuddyAllocator(place.device)->Free(p);
-  gpu_mem_info[place.device].first -= size;
+  if (FLAGS_benchmark) allocation::GPUMemMonitor.Minus(place.device, size);
 #else
   PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
 #endif
@@ -335,6 +328,8 @@ size_t Usage::operator()(const platform::CUDAPinnedPlace &cuda_pinned) const {
 
 namespace allocation {
 
+LegacyMemMonitor GPUMemMonitor;
+
 Allocation *LegacyAllocator::AllocateImpl(size_t size, Allocator::Attr attr) {
   void *ptr = boost::apply_visitor(legacy::AllocVisitor(size), place_);
   return new Allocation(ptr, size, place_);
@@ -346,6 +341,63 @@ void LegacyAllocator::Free(Allocation *allocation) {
       allocation->place());
   delete allocation;
 }
+
+bool MemInfo::Add(const size_t &size) {
+  std::lock_guard<std::mutex> lock(mutex_);
+  usage_ += size;
+  bool peak_point = usage_ > peak_usage_;
+  if (peak_point) peak_usage_ = usage_;
+  return peak_point;
+}
+
+void MemInfo::Minus(const size_t &size) {
+  std::lock_guard<std::mutex> lock(mutex_);
+  usage_ -= size;
+}
+
+uint64_t MemInfo::GetPeakUsage() { return peak_usage_; }
+
+LegacyMemMonitor::~LegacyMemMonitor() {
+  for (auto &item : gpu_mem_info_) delete item.second;
+}
+
+void LegacyMemMonitor::Initialize(const int &device_num) {
+  for (auto i = 0; i < device_num; ++i) {
+    gpu_mem_info_[i] = new MemInfo();
+  }
+}
+
+void LegacyMemMonitor::Add(const int &device, const size_t &size) {
+  if (gpu_mem_info_[device]->Add(size)) {
+    VLOG(3) << "#LegacyMemMonitor# device: " << device
+            << " peak memory usage : "
+            << (gpu_mem_info_[device]->GetPeakUsage() >> 20) << " MiB";
+  }
+}
+
+void LegacyMemMonitor::Minus(const int &device, const size_t &size) {
+  gpu_mem_info_[device]->Minus(size);
+}
+
+uint64_t LegacyMemMonitor::GetMemUsage(const int &device) {
+  return gpu_mem_info_.find(device) == gpu_mem_info_.end()
+             ? 0
+             : gpu_mem_info_[device]->GetPeakUsage();
+}
+
+void LegacyMemMonitor::PrintMemUsage() {
+  std::vector<int> devices;
+  for (const auto &item : gpu_mem_info_) {
+    devices.emplace_back(item.first);
+  }
+  std::sort(devices.begin(), devices.end());
+  for (const auto &device : devices) {
+    std::cout << "Device : " << device << " Peak Memory Usage : "
+              << (gpu_mem_info_[device]->GetPeakUsage() >> 20) << " MiB"
+              << std::endl;
+  }
+}
+
 }  // namespace allocation
 }  // namespace memory
 }  // namespace paddle

paddle/fluid/memory/allocation/legacy_allocator.h

@@ -13,12 +13,59 @@
 // limitations under the License.
 
 #pragma once
+#include <algorithm>
+#include <mutex>  // NOLINT
+#include <unordered_map>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/memory/allocation/allocator.h"
 #include "paddle/fluid/platform/place.h"
 namespace paddle {
 namespace memory {
 namespace allocation {
 
+class MemInfo {
+ public:
+  MemInfo() : usage_(0), peak_usage_(0) {}
+  MemInfo(const MemInfo &) = delete;
+  MemInfo &operator=(const MemInfo &) = delete;
+
+  // return a flag to indicate current operation will create a peak point or not
+  bool Add(const size_t &);
+  void Minus(const size_t &);
+
+  uint64_t GetPeakUsage();
+
+ private:
+  /* current memory usage*/
+  uint64_t usage_;
+  uint64_t peak_usage_;
+  std::mutex mutex_;
+};
+
+class LegacyMemMonitor {
+ public:
+  // used to store the GPU memory usage of each devices
+  using MemUsage = std::unordered_map</*device id*/ int,
+                                      /*mem usage info node*/ MemInfo *>;
+
+  MemUsage GetMemUsageInfo() { return gpu_mem_info_; }
+  ~LegacyMemMonitor();
+
+  void Initialize(const int &);
+  void Add(const int &, const size_t &);
+  void Minus(const int &, const size_t &);
+
+  uint64_t GetMemUsage(const int &);
+
+  void PrintMemUsage();
+
+ protected:
+  MemUsage gpu_mem_info_;
+};
+
+extern LegacyMemMonitor GPUMemMonitor;
+
 class LegacyAllocatorPrivate;
 class LegacyAllocator : public Allocator {
  public:

paddle/fluid/operators/batch_norm_op.cc

@@ -589,8 +589,10 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker {
     op->SetInput("SavedVariance", Output("SavedVariance"));
 
     // used when setting use_global_stats True during training
-    op->SetInput("Mean", Output("MeanOut"));
-    op->SetInput("Variance", Output("VarianceOut"));
+    if (boost::get<bool>(GetAttr("use_global_stats"))) {
+      op->SetInput("Mean", Output("MeanOut"));
+      op->SetInput("Variance", Output("VarianceOut"));
+    }
 
     op->SetAttrMap(Attrs());
 

paddle/fluid/operators/detection/CMakeLists.txt

@@ -31,6 +31,8 @@ detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc
 polygon_box_transform_op.cu)
 detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc)
 detection_library(generate_proposal_labels_op SRCS generate_proposal_labels_op.cc)
+detection_library(box_clip_op SRCS box_clip_op.cc box_clip_op.cu)
+detection_library(yolov3_loss_op SRCS yolov3_loss_op.cc)
 
 if(WITH_GPU)
   detection_library(generate_proposals_op SRCS generate_proposals_op.cc generate_proposals_op.cu DEPS memory cub)

paddle/fluid/operators/detection/bbox_util.h

@@ -99,5 +99,29 @@ void BboxOverlaps(const framework::Tensor& r_boxes,
   }
 }
 
+template <class T>
+void ClipTiledBoxes(const platform::DeviceContext& ctx,
+                    const framework::Tensor& im_info,
+                    const framework::Tensor& input_boxes,
+                    framework::Tensor* out) {
+  T* out_data = out->mutable_data<T>(ctx.GetPlace());
+  const T* im_info_data = im_info.data<T>();
+  const T* input_boxes_data = input_boxes.data<T>();
+  T zero(0);
+  T im_w = round(im_info_data[1] / im_info_data[2]);
+  T im_h = round(im_info_data[0] / im_info_data[2]);
+  for (int64_t i = 0; i < input_boxes.numel(); ++i) {
+    if (i % 4 == 0) {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_w - 1), zero);
+    } else if (i % 4 == 1) {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_h - 1), zero);
+    } else if (i % 4 == 2) {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_w - 1), zero);
+    } else {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_h - 1), zero);
+    }
+  }
+}
+
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detection/box_clip_op.cc

+/* 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. */
+
+#include "paddle/fluid/operators/detection/box_clip_op.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+class BoxClipOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of BoxClipOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("ImInfo"),
+                   "Input(ImInfo) of BoxClipOp should not be null.");
+
+    auto input_box_dims = ctx->GetInputDim("Input");
+    auto im_info_dims = ctx->GetInputDim("ImInfo");
+
+    if (ctx->IsRuntime()) {
+      auto input_box_size = input_box_dims.size();
+      PADDLE_ENFORCE_EQ(input_box_dims[input_box_size - 1], 4,
+                        "The last dimension of Input must be 4");
+      PADDLE_ENFORCE_EQ(im_info_dims.size(), 2,
+                        "The rank of Input(Input) in BoxClipOp must be 2");
+      PADDLE_ENFORCE_EQ(im_info_dims[1], 3,
+                        "The last dimension of ImInfo must be 3");
+    }
+    ctx->ShareDim("Input", /*->*/ "Output");
+    ctx->ShareLoD("Input", /*->*/ "Output");
+  }
+};
+
+class BoxClipOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("Input",
+             "(LoDTensor) "
+             "Input is a LoDTensor with shape [..., 4] holds 4 points"
+             "in last dimension in format [xmin, ymin, xmax, ymax]");
+    AddInput("ImInfo",
+             "(Tensor) Information for image reshape is in shape (N, 3), "
+             "in format (height, width, im_scale)");
+    AddOutput("Output",
+              "(LoDTensor) "
+              "Output is a LoDTensor with the same shape as Input"
+              "and it is the result after clip");
+    AddComment(R"DOC(
+This operator clips input boxes to original input images.
+
+For each input box, The formula is given as follows:
+
+       $$xmin = \max(\min(xmin, im_w - 1), 0)$$
+       $$ymin = \max(\min(ymin, im_h - 1), 0)$$     
+       $$xmax = \max(\min(xmax, im_w - 1), 0)$$
+       $$ymax = \max(\min(ymax, im_h - 1), 0)$$
+
+where im_w and im_h are computed from ImInfo, the formula is given as follows:
+
+       $$im_w = \round(width / im_scale)$$
+       $$im_h = \round(height / im_scale)$$ 
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(box_clip, ops::BoxClipOp, ops::BoxClipOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    box_clip, ops::BoxClipKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BoxClipKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/detection/box_clip_op.cu

+/* 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. */
+
+#include <algorithm>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detection/box_clip_op.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+#include "paddle/fluid/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTenso = framework::LoDTensor;
+
+static constexpr int ImInfoSize = 3;
+
+template <typename T, int BlockSize>
+static __global__ void GPUBoxClip(const T *input, const size_t *lod,
+                                  const size_t width, const T *im_info,
+                                  T *output) {
+  T im_w = round(im_info[blockIdx.x * ImInfoSize + 1] /
+                 im_info[blockIdx.x * ImInfoSize + 2]);
+  T im_h = round(im_info[blockIdx.x * ImInfoSize] /
+                 im_info[blockIdx.x * ImInfoSize + 2]);
+  for (int i = threadIdx.x; i < (lod[blockIdx.x + 1] - lod[blockIdx.x]) * width;
+       i += BlockSize) {
+    int idx = lod[blockIdx.x] * width + i;
+    T im_size = (idx % 2 == 0) ? im_w : im_h;
+    output[idx] = max(min(input[idx], im_size - 1), T(0.));
+  }
+}
+
+template <typename DeviceContext, typename T>
+class GPUBoxClipKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto *input = context.Input<LoDTensor>("Input");
+    auto *im_info = context.Input<Tensor>("ImInfo");
+    auto *output = context.Output<LoDTensor>("Output");
+    const int64_t num = input->dims()[0];
+    const int64_t bbox_width = input->numel() / num;
+    auto lod = input->lod();
+    framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
+    auto &dev_ctx = context.template device_context<DeviceContext>();
+    auto stream = dev_ctx.stream();
+    const size_t batch_size = lod.back().size() - 1;
+    T *output_data = output->mutable_data<T>(dev_ctx.GetPlace());
+    GPUBoxClip<T, 512><<<batch_size, 512, 0, stream>>>(
+        input->data<T>(), abs_offset_lod[0].CUDAMutableData(dev_ctx.GetPlace()),
+        bbox_width, im_info->data<T>(), output_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    box_clip, ops::GPUBoxClipKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GPUBoxClipKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/detection/box_clip_op.h

+/* 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 <string>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detection/bbox_util.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+template <typename DeviceContext, typename T>
+class BoxClipKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input_box = context.Input<LoDTensor>("Input");
+    auto* im_info = context.Input<LoDTensor>("ImInfo");
+    auto* output_box = context.Output<LoDTensor>("Output");
+    auto& dev_ctx =
+        context.template device_context<platform::CPUDeviceContext>();
+    output_box->mutable_data<T>(context.GetPlace());
+    if (input_box->lod().size()) {
+      PADDLE_ENFORCE_EQ(input_box->lod().size(), 1UL,
+                        "Only support 1 level of LoD.");
+    }
+    auto box_lod = input_box->lod().back();
+    int64_t n = static_cast<int64_t>(box_lod.size() - 1);
+    for (int i = 0; i < n; ++i) {
+      Tensor im_info_slice = im_info->Slice(i, i + 1);
+      Tensor box_slice = input_box->Slice(box_lod[i], box_lod[i + 1]);
+      Tensor output_slice = output_box->Slice(box_lod[i], box_lod[i + 1]);
+      ClipTiledBoxes<T>(dev_ctx, im_info_slice, box_slice, &output_slice);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/detection/density_prior_box_op.h

@@ -52,6 +52,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
       step_height = step_h;
     }
     int num_priors = 0;
+
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for reduction(+ : num_priors)
+#endif
     for (size_t i = 0; i < densities.size(); ++i) {
       num_priors += (fixed_ratios.size()) * (pow(densities[i], 2));
     }
@@ -64,6 +68,17 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
     auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes).setConstant(0.0);
     int step_average = static_cast<int>((step_width + step_height) * 0.5);
 
+    std::vector<float> sqrt_fixed_ratios;
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for
+#endif
+    for (int i = 0; i < fixed_ratios.size(); i++) {
+      sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i]));
+    }
+
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for collapse(2)
+#endif
     for (int h = 0; h < feature_height; ++h) {
       for (int w = 0; w < feature_width; ++w) {
         T center_x = (w + offset) * step_width;
@@ -73,34 +88,25 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
         for (size_t s = 0; s < fixed_sizes.size(); ++s) {
           auto fixed_size = fixed_sizes[s];
           int density = densities[s];
+          int shift = step_average / density;
           // Generate density prior boxes with fixed ratios.
           for (size_t r = 0; r < fixed_ratios.size(); ++r) {
-            float ar = fixed_ratios[r];
-            int shift = step_average / density;
-            float box_width_ratio = fixed_size * sqrt(ar);
-            float box_height_ratio = fixed_size / sqrt(ar);
+            float box_width_ratio = fixed_size * sqrt_fixed_ratios[r];
+            float box_height_ratio = fixed_size / sqrt_fixed_ratios[r];
+            float density_center_x = center_x - step_average / 2. + shift / 2.;
+            float density_center_y = center_y - step_average / 2. + shift / 2.;
             for (int di = 0; di < density; ++di) {
               for (int dj = 0; dj < density; ++dj) {
-                float center_x_temp =
-                    center_x - step_average / 2. + shift / 2. + dj * shift;
-                float center_y_temp =
-                    center_y - step_average / 2. + shift / 2. + di * shift;
-                e_boxes(h, w, idx, 0) =
-                    (center_x_temp - box_width_ratio / 2.) / img_width >= 0
-                        ? (center_x_temp - box_width_ratio / 2.) / img_width
-                        : 0;
-                e_boxes(h, w, idx, 1) =
-                    (center_y_temp - box_height_ratio / 2.) / img_height >= 0
-                        ? (center_y_temp - box_height_ratio / 2.) / img_height
-                        : 0;
-                e_boxes(h, w, idx, 2) =
-                    (center_x_temp + box_width_ratio / 2.) / img_width <= 1
-                        ? (center_x_temp + box_width_ratio / 2.) / img_width
-                        : 1;
-                e_boxes(h, w, idx, 3) =
-                    (center_y_temp + box_height_ratio / 2.) / img_height <= 1
-                        ? (center_y_temp + box_height_ratio / 2.) / img_height
-                        : 1;
+                float center_x_temp = density_center_x + dj * shift;
+                float center_y_temp = density_center_y + di * shift;
+                e_boxes(h, w, idx, 0) = std::max(
+                    (center_x_temp - box_width_ratio / 2.) / img_width, 0.);
+                e_boxes(h, w, idx, 1) = std::max(
+                    (center_y_temp - box_height_ratio / 2.) / img_height, 0.);
+                e_boxes(h, w, idx, 2) = std::min(
+                    (center_x_temp + box_width_ratio / 2.) / img_width, 1.);
+                e_boxes(h, w, idx, 3) = std::min(
+                    (center_y_temp + box_height_ratio / 2.) / img_height, 1.);
                 idx++;
               }
             }
@@ -131,8 +137,14 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
     vars->Resize({box_num, static_cast<int>(variances.size())});
 
     auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
-
-    e_vars = var_et.broadcast(Eigen::DSizes<int, 2>(box_num, 1));
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for collapse(2)
+#endif
+    for (int i = 0; i < box_num; ++i) {
+      for (int j = 0; j < variances.size(); ++j) {
+        e_vars(i, j) = variances[j];
+      }
+    }
 
     vars->Resize(var_dim);
     boxes->Resize(box_dim);

paddle/fluid/operators/yolov3_loss_op.cc → paddle/fluid/operators/detection/yolov3_loss_op.cc

@@ -9,7 +9,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/fluid/operators/yolov3_loss_op.h"
+#include "paddle/fluid/operators/detection/yolov3_loss_op.h"
 #include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {
@@ -29,23 +29,33 @@ class Yolov3LossOp : public framework::OperatorWithKernel {
                    "Input(GTLabel) of Yolov3LossOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Loss"),
                    "Output(Loss) of Yolov3LossOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("ObjectnessMask"),
+        "Output(ObjectnessMask) of Yolov3LossOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("GTMatchMask"),
+                   "Output(GTMatchMask) of Yolov3LossOp should not be null.");
 
     auto dim_x = ctx->GetInputDim("X");
     auto dim_gtbox = ctx->GetInputDim("GTBox");
     auto dim_gtlabel = ctx->GetInputDim("GTLabel");
     auto anchors = ctx->Attrs().Get<std::vector<int>>("anchors");
+    int anchor_num = anchors.size() / 2;
+    auto anchor_mask = ctx->Attrs().Get<std::vector<int>>("anchor_mask");
+    int mask_num = anchor_mask.size();
     auto class_num = ctx->Attrs().Get<int>("class_num");
+
     PADDLE_ENFORCE_EQ(dim_x.size(), 4, "Input(X) should be a 4-D tensor.");
     PADDLE_ENFORCE_EQ(dim_x[2], dim_x[3],
                       "Input(X) dim[3] and dim[4] should be euqal.");
-    PADDLE_ENFORCE_EQ(dim_x[1], anchors.size() / 2 * (5 + class_num),
-                      "Input(X) dim[1] should be equal to (anchor_number * (5 "
-                      "+ class_num)).");
+    PADDLE_ENFORCE_EQ(
+        dim_x[1], mask_num * (5 + class_num),
+        "Input(X) dim[1] should be equal to (anchor_mask_number * (5 "
+        "+ class_num)).");
     PADDLE_ENFORCE_EQ(dim_gtbox.size(), 3,
                       "Input(GTBox) should be a 3-D tensor");
     PADDLE_ENFORCE_EQ(dim_gtbox[2], 4, "Input(GTBox) dim[2] should be 5");
     PADDLE_ENFORCE_EQ(dim_gtlabel.size(), 2,
-                      "Input(GTBox) should be a 2-D tensor");
+                      "Input(GTLabel) should be a 2-D tensor");
     PADDLE_ENFORCE_EQ(dim_gtlabel[0], dim_gtbox[0],
                       "Input(GTBox) and Input(GTLabel) dim[0] should be same");
     PADDLE_ENFORCE_EQ(dim_gtlabel[1], dim_gtbox[1],
@@ -54,11 +64,22 @@ class Yolov3LossOp : public framework::OperatorWithKernel {
                       "Attr(anchors) length should be greater then 0.");
     PADDLE_ENFORCE_EQ(anchors.size() % 2, 0,
                       "Attr(anchors) length should be even integer.");
+    for (size_t i = 0; i < anchor_mask.size(); i++) {
+      PADDLE_ENFORCE_LT(
+          anchor_mask[i], anchor_num,
+          "Attr(anchor_mask) should not crossover Attr(anchors).");
+    }
     PADDLE_ENFORCE_GT(class_num, 0,
                       "Attr(class_num) should be an integer greater then 0.");
 
-    std::vector<int64_t> dim_out({1});
+    std::vector<int64_t> dim_out({dim_x[0]});
     ctx->SetOutputDim("Loss", framework::make_ddim(dim_out));
+
+    std::vector<int64_t> dim_obj_mask({dim_x[0], mask_num, dim_x[2], dim_x[3]});
+    ctx->SetOutputDim("ObjectnessMask", framework::make_ddim(dim_obj_mask));
+
+    std::vector<int64_t> dim_gt_match_mask({dim_gtbox[0], dim_gtbox[1]});
+    ctx->SetOutputDim("GTMatchMask", framework::make_ddim(dim_gt_match_mask));
   }
 
  protected:
@@ -73,11 +94,11 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("X",
-             "The input tensor of YOLO v3 loss operator, "
+             "The input tensor of YOLOv3 loss operator, "
              "This is a 4-D tensor with shape of [N, C, H, W]."
              "H and W should be same, and the second dimention(C) stores"
              "box locations, confidence score and classification one-hot"
-             "key of each anchor box");
+             "keys of each anchor box");
     AddInput("GTBox",
              "The input tensor of ground truth boxes, "
              "This is a 3-D tensor with shape of [N, max_box_num, 5], "
@@ -89,32 +110,39 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("GTLabel",
              "The input tensor of ground truth label, "
              "This is a 2-D tensor with shape of [N, max_box_num], "
-             "and each element shoudl be an integer to indicate the "
+             "and each element should be an integer to indicate the "
              "box class id.");
     AddOutput("Loss",
               "The output yolov3 loss tensor, "
-              "This is a 1-D tensor with shape of [1]");
+              "This is a 1-D tensor with shape of [N]");
+    AddOutput("ObjectnessMask",
+              "This is an intermediate tensor with shape of [N, M, H, W], "
+              "M is the number of anchor masks. This parameter caches the "
+              "mask for calculate objectness loss in gradient kernel.")
+        .AsIntermediate();
+    AddOutput("GTMatchMask",
+              "This is an intermediate tensor with shape of [N, B], "
+              "B is the max box number of GT boxes. This parameter caches "
+              "matched mask index of each GT boxes for gradient calculate.")
+        .AsIntermediate();
 
     AddAttr<int>("class_num", "The number of classes to predict.");
     AddAttr<std::vector<int>>("anchors",
                               "The anchor width and height, "
-                              "it will be parsed pair by pair.");
+                              "it will be parsed pair by pair.")
+        .SetDefault(std::vector<int>{});
+    AddAttr<std::vector<int>>("anchor_mask",
+                              "The mask index of anchors used in "
+                              "current YOLOv3 loss calculation.")
+        .SetDefault(std::vector<int>{});
+    AddAttr<int>("downsample_ratio",
+                 "The downsample ratio from network input to YOLOv3 loss "
+                 "input, so 32, 16, 8 should be set for the first, second, "
+                 "and thrid YOLOv3 loss operators.")
+        .SetDefault(32);
     AddAttr<float>("ignore_thresh",
-                   "The ignore threshold to ignore confidence loss.");
-    AddAttr<float>("loss_weight_xy", "The weight of x, y location loss.")
-        .SetDefault(1.0);
-    AddAttr<float>("loss_weight_wh", "The weight of w, h location loss.")
-        .SetDefault(1.0);
-    AddAttr<float>(
-        "loss_weight_conf_target",
-        "The weight of confidence score loss in locations with target object.")
-        .SetDefault(1.0);
-    AddAttr<float>("loss_weight_conf_notarget",
-                   "The weight of confidence score loss in locations without "
-                   "target object.")
-        .SetDefault(1.0);
-    AddAttr<float>("loss_weight_class", "The weight of classification loss.")
-        .SetDefault(1.0);
+                   "The ignore threshold to ignore confidence loss.")
+        .SetDefault(0.7);
     AddComment(R"DOC(
          This operator generate yolov3 loss by given predict result and ground
          truth boxes.
@@ -147,17 +175,28 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
          thresh, the confidence score loss of this anchor box will be ignored.
 
          Therefore, the yolov3 loss consist of three major parts, box location loss,
-         confidence score loss, and classification loss. The MSE loss is used for 
-         box location, and binary cross entropy loss is used for confidence score 
-         loss and classification loss.
+         confidence score loss, and classification loss. The L2 loss is used for 
+         box coordinates (w, h), and sigmoid cross entropy loss is used for box 
+         coordinates (x, y), confidence score loss and classification loss.
+
+         Each groud truth box find a best matching anchor box in all anchors, 
+         prediction of this anchor box will incur all three parts of losses, and
+         prediction of anchor boxes with no GT box matched will only incur objectness
+         loss.
+
+         In order to trade off box coordinate losses between big boxes and small 
+         boxes, box coordinate losses will be mutiplied by scale weight, which is
+         calculated as follow.
+
+         $$
+         weight_{box} = 2.0 - t_w * t_h
+         $$
 
          Final loss will be represented as follow.
 
          $$
-         loss = \loss_weight_{xy} * loss_{xy} + \loss_weight_{wh} * loss_{wh}
-              + \loss_weight_{conf_target} * loss_{conf_target}
-              + \loss_weight_{conf_notarget} * loss_{conf_notarget}
-              + \loss_weight_{class} * loss_{class}
+         loss = (loss_{xy} + loss_{wh}) * weight_{box}
+              + loss_{conf} + loss_{class}
          $$
          )DOC");
   }
@@ -196,6 +235,8 @@ class Yolov3LossGradMaker : public framework::SingleGradOpDescMaker {
     op->SetInput("GTBox", Input("GTBox"));
     op->SetInput("GTLabel", Input("GTLabel"));
     op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss"));
+    op->SetInput("ObjectnessMask", Output("ObjectnessMask"));
+    op->SetInput("GTMatchMask", Output("GTMatchMask"));
 
     op->SetAttrMap(Attrs());
 

paddle/fluid/operators/detection/yolov3_loss_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+   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 <algorithm>
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename T>
+static inline bool LessEqualZero(T x) {
+  return x < 1e-6;
+}
+
+template <typename T>
+static T SigmoidCrossEntropy(T x, T label) {
+  return (x > 0 ? x : 0.0) - x * label + std::log(1.0 + std::exp(-std::abs(x)));
+}
+
+template <typename T>
+static T L2Loss(T x, T y) {
+  return 0.5 * (y - x) * (y - x);
+}
+
+template <typename T>
+static T SigmoidCrossEntropyGrad(T x, T label) {
+  return 1.0 / (1.0 + std::exp(-x)) - label;
+}
+
+template <typename T>
+static T L2LossGrad(T x, T y) {
+  return x - y;
+}
+
+static int GetMaskIndex(std::vector<int> mask, int val) {
+  for (size_t i = 0; i < mask.size(); i++) {
+    if (mask[i] == val) {
+      return i;
+    }
+  }
+  return -1;
+}
+
+template <typename T>
+struct Box {
+  T x, y, w, h;
+};
+
+template <typename T>
+static inline T sigmoid(T x) {
+  return 1.0 / (1.0 + std::exp(-x));
+}
+
+template <typename T>
+static inline Box<T> GetYoloBox(const T* x, std::vector<int> anchors, int i,
+                                int j, int an_idx, int grid_size,
+                                int input_size, int index, int stride) {
+  Box<T> b;
+  b.x = (i + sigmoid<T>(x[index])) / grid_size;
+  b.y = (j + sigmoid<T>(x[index + stride])) / grid_size;
+  b.w = std::exp(x[index + 2 * stride]) * anchors[2 * an_idx] / input_size;
+  b.h = std::exp(x[index + 3 * stride]) * anchors[2 * an_idx + 1] / input_size;
+  return b;
+}
+
+template <typename T>
+static inline Box<T> GetGtBox(const T* gt, int batch, int max_boxes, int idx) {
+  Box<T> b;
+  b.x = gt[(batch * max_boxes + idx) * 4];
+  b.y = gt[(batch * max_boxes + idx) * 4 + 1];
+  b.w = gt[(batch * max_boxes + idx) * 4 + 2];
+  b.h = gt[(batch * max_boxes + idx) * 4 + 3];
+  return b;
+}
+
+template <typename T>
+static inline T BoxOverlap(T c1, T w1, T c2, T w2) {
+  T l1 = c1 - w1 / 2.0;
+  T l2 = c2 - w2 / 2.0;
+  T left = l1 > l2 ? l1 : l2;
+  T r1 = c1 + w1 / 2.0;
+  T r2 = c2 + w2 / 2.0;
+  T right = r1 < r2 ? r1 : r2;
+  return right - left;
+}
+
+template <typename T>
+static inline T CalcBoxIoU(Box<T> b1, Box<T> b2) {
+  T w = BoxOverlap(b1.x, b1.w, b2.x, b2.w);
+  T h = BoxOverlap(b1.y, b1.h, b2.y, b2.h);
+  T inter_area = (w < 0 || h < 0) ? 0.0 : w * h;
+  T union_area = b1.w * b1.h + b2.w * b2.h - inter_area;
+  return inter_area / union_area;
+}
+
+static inline int GetEntryIndex(int batch, int an_idx, int hw_idx, int an_num,
+                                int an_stride, int stride, int entry) {
+  return (batch * an_num + an_idx) * an_stride + entry * stride + hw_idx;
+}
+
+template <typename T>
+static void CalcBoxLocationLoss(T* loss, const T* input, Box<T> gt,
+                                std::vector<int> anchors, int an_idx,
+                                int box_idx, int gi, int gj, int grid_size,
+                                int input_size, int stride) {
+  T tx = gt.x * grid_size - gi;
+  T ty = gt.y * grid_size - gj;
+  T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
+  T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);
+
+  T scale = (2.0 - gt.w * gt.h);
+  loss[0] += SigmoidCrossEntropy<T>(input[box_idx], tx) * scale;
+  loss[0] += SigmoidCrossEntropy<T>(input[box_idx + stride], ty) * scale;
+  loss[0] += L2Loss<T>(input[box_idx + 2 * stride], tw) * scale;
+  loss[0] += L2Loss<T>(input[box_idx + 3 * stride], th) * scale;
+}
+
+template <typename T>
+static void CalcBoxLocationLossGrad(T* input_grad, const T loss, const T* input,
+                                    Box<T> gt, std::vector<int> anchors,
+                                    int an_idx, int box_idx, int gi, int gj,
+                                    int grid_size, int input_size, int stride) {
+  T tx = gt.x * grid_size - gi;
+  T ty = gt.y * grid_size - gj;
+  T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
+  T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);
+
+  T scale = (2.0 - gt.w * gt.h);
+  input_grad[box_idx] =
+      SigmoidCrossEntropyGrad<T>(input[box_idx], tx) * scale * loss;
+  input_grad[box_idx + stride] =
+      SigmoidCrossEntropyGrad<T>(input[box_idx + stride], ty) * scale * loss;
+  input_grad[box_idx + 2 * stride] =
+      L2LossGrad<T>(input[box_idx + 2 * stride], tw) * scale * loss;
+  input_grad[box_idx + 3 * stride] =
+      L2LossGrad<T>(input[box_idx + 3 * stride], th) * scale * loss;
+}
+
+template <typename T>
+static inline void CalcLabelLoss(T* loss, const T* input, const int index,
+                                 const int label, const int class_num,
+                                 const int stride) {
+  for (int i = 0; i < class_num; i++) {
+    T pred = input[index + i * stride];
+    loss[0] += SigmoidCrossEntropy<T>(pred, (i == label) ? 1.0 : 0.0);
+  }
+}
+
+template <typename T>
+static inline void CalcLabelLossGrad(T* input_grad, const T loss,
+                                     const T* input, const int index,
+                                     const int label, const int class_num,
+                                     const int stride) {
+  for (int i = 0; i < class_num; i++) {
+    T pred = input[index + i * stride];
+    input_grad[index + i * stride] =
+        SigmoidCrossEntropyGrad<T>(pred, (i == label) ? 1.0 : 0.0) * loss;
+  }
+}
+
+template <typename T>
+static inline void CalcObjnessLoss(T* loss, const T* input, const T* objness,
+                                   const int n, const int an_num, const int h,
+                                   const int w, const int stride,
+                                   const int an_stride) {
+  for (int i = 0; i < n; i++) {
+    for (int j = 0; j < an_num; j++) {
+      for (int k = 0; k < h; k++) {
+        for (int l = 0; l < w; l++) {
+          T obj = objness[k * w + l];
+          if (obj > 1e-5) {
+            // positive sample: obj = 1
+            loss[i] += SigmoidCrossEntropy<T>(input[k * w + l], 1.0);
+          } else if (obj > -0.5) {
+            // negetive sample: obj = 0
+            loss[i] += SigmoidCrossEntropy<T>(input[k * w + l], 0.0);
+          }
+        }
+      }
+      objness += stride;
+      input += an_stride;
+    }
+  }
+}
+
+template <typename T>
+static inline void CalcObjnessLossGrad(T* input_grad, const T* loss,
+                                       const T* input, const T* objness,
+                                       const int n, const int an_num,
+                                       const int h, const int w,
+                                       const int stride, const int an_stride) {
+  for (int i = 0; i < n; i++) {
+    for (int j = 0; j < an_num; j++) {
+      for (int k = 0; k < h; k++) {
+        for (int l = 0; l < w; l++) {
+          T obj = objness[k * w + l];
+          if (obj > 1e-5) {
+            input_grad[k * w + l] =
+                SigmoidCrossEntropyGrad<T>(input[k * w + l], 1.0) * loss[i];
+          } else if (obj > -0.5) {
+            input_grad[k * w + l] =
+                SigmoidCrossEntropyGrad<T>(input[k * w + l], 0.0) * loss[i];
+          }
+        }
+      }
+      objness += stride;
+      input += an_stride;
+      input_grad += an_stride;
+    }
+  }
+}
+
+template <typename T>
+static void inline GtValid(bool* valid, const T* gtbox, const int n,
+                           const int b) {
+  for (int i = 0; i < n; i++) {
+    for (int j = 0; j < b; j++) {
+      if (LessEqualZero(gtbox[j * 4 + 2]) || LessEqualZero(gtbox[j * 4 + 3])) {
+        valid[j] = false;
+      } else {
+        valid[j] = true;
+      }
+    }
+    valid += b;
+    gtbox += b * 4;
+  }
+}
+
+template <typename T>
+class Yolov3LossKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* input = ctx.Input<Tensor>("X");
+    auto* gt_box = ctx.Input<Tensor>("GTBox");
+    auto* gt_label = ctx.Input<Tensor>("GTLabel");
+    auto* loss = ctx.Output<Tensor>("Loss");
+    auto* objness_mask = ctx.Output<Tensor>("ObjectnessMask");
+    auto* gt_match_mask = ctx.Output<Tensor>("GTMatchMask");
+    auto anchors = ctx.Attr<std::vector<int>>("anchors");
+    auto anchor_mask = ctx.Attr<std::vector<int>>("anchor_mask");
+    int class_num = ctx.Attr<int>("class_num");
+    float ignore_thresh = ctx.Attr<float>("ignore_thresh");
+    int downsample_ratio = ctx.Attr<int>("downsample_ratio");
+
+    const int n = input->dims()[0];
+    const int h = input->dims()[2];
+    const int w = input->dims()[3];
+    const int an_num = anchors.size() / 2;
+    const int mask_num = anchor_mask.size();
+    const int b = gt_box->dims()[1];
+    int input_size = downsample_ratio * h;
+
+    const int stride = h * w;
+    const int an_stride = (class_num + 5) * stride;
+
+    const T* input_data = input->data<T>();
+    const T* gt_box_data = gt_box->data<T>();
+    const int* gt_label_data = gt_label->data<int>();
+    T* loss_data = loss->mutable_data<T>({n}, ctx.GetPlace());
+    memset(loss_data, 0, loss->numel() * sizeof(T));
+    T* obj_mask_data =
+        objness_mask->mutable_data<T>({n, mask_num, h, w}, ctx.GetPlace());
+    memset(obj_mask_data, 0, objness_mask->numel() * sizeof(T));
+    int* gt_match_mask_data =
+        gt_match_mask->mutable_data<int>({n, b}, ctx.GetPlace());
+
+    // calc valid gt box mask, avoid calc duplicately in following code
+    Tensor gt_valid_mask;
+    bool* gt_valid_mask_data =
+        gt_valid_mask.mutable_data<bool>({n, b}, ctx.GetPlace());
+    GtValid<T>(gt_valid_mask_data, gt_box_data, n, b);
+
+    for (int i = 0; i < n; i++) {
+      for (int j = 0; j < mask_num; j++) {
+        for (int k = 0; k < h; k++) {
+          for (int l = 0; l < w; l++) {
+            // each predict box find a best match gt box, if overlap is bigger
+            // then ignore_thresh, ignore the objectness loss.
+            int box_idx =
+                GetEntryIndex(i, j, k * w + l, mask_num, an_stride, stride, 0);
+            Box<T> pred = GetYoloBox(input_data, anchors, l, k, anchor_mask[j],
+                                     h, input_size, box_idx, stride);
+            T best_iou = 0;
+            for (int t = 0; t < b; t++) {
+              if (!gt_valid_mask_data[i * b + t]) {
+                continue;
+              }
+              Box<T> gt = GetGtBox(gt_box_data, i, b, t);
+              T iou = CalcBoxIoU(pred, gt);
+              if (iou > best_iou) {
+                best_iou = iou;
+              }
+            }
+
+            // If best IoU is bigger then ignore_thresh,
+            // ignore the objectness loss.
+            if (best_iou > ignore_thresh) {
+              int obj_idx = (i * mask_num + j) * stride + k * w + l;
+              obj_mask_data[obj_idx] = static_cast<T>(-1);
+            }
+            // all losses should be calculated if best IoU
+            // is bigger then truth thresh, but currently,
+            // truth thresh is an unreachable value as 1.0.
+          }
+        }
+      }
+      for (int t = 0; t < b; t++) {
+        if (!gt_valid_mask_data[i * b + t]) {
+          gt_match_mask_data[i * b + t] = -1;
+          continue;
+        }
+        Box<T> gt = GetGtBox(gt_box_data, i, b, t);
+        int gi = static_cast<int>(gt.x * w);
+        int gj = static_cast<int>(gt.y * h);
+        Box<T> gt_shift = gt;
+        gt_shift.x = 0.0;
+        gt_shift.y = 0.0;
+        T best_iou = 0.0;
+        int best_n = 0;
+        // each gt box find a best match anchor box as positive sample,
+        // for positive sample, all losses should be calculated, and for
+        // other samples, only objectness loss is required.
+        for (int an_idx = 0; an_idx < an_num; an_idx++) {
+          Box<T> an_box;
+          an_box.x = 0.0;
+          an_box.y = 0.0;
+          an_box.w = anchors[2 * an_idx] / static_cast<T>(input_size);
+          an_box.h = anchors[2 * an_idx + 1] / static_cast<T>(input_size);
+          float iou = CalcBoxIoU<T>(an_box, gt_shift);
+          if (iou > best_iou) {
+            best_iou = iou;
+            best_n = an_idx;
+          }
+        }
+
+        int mask_idx = GetMaskIndex(anchor_mask, best_n);
+        gt_match_mask_data[i * b + t] = mask_idx;
+        if (mask_idx >= 0) {
+          int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
+                                      an_stride, stride, 0);
+          CalcBoxLocationLoss<T>(loss_data + i, input_data, gt, anchors, best_n,
+                                 box_idx, gi, gj, h, input_size, stride);
+
+          int obj_idx = (i * mask_num + mask_idx) * stride + gj * w + gi;
+          obj_mask_data[obj_idx] = 1.0;
+
+          int label = gt_label_data[i * b + t];
+          int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
+                                        an_stride, stride, 5);
+          CalcLabelLoss<T>(loss_data + i, input_data, label_idx, label,
+                           class_num, stride);
+        }
+      }
+    }
+
+    CalcObjnessLoss<T>(loss_data, input_data + 4 * stride, obj_mask_data, n,
+                       mask_num, h, w, stride, an_stride);
+  }
+};
+
+template <typename T>
+class Yolov3LossGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* input = ctx.Input<Tensor>("X");
+    auto* gt_box = ctx.Input<Tensor>("GTBox");
+    auto* gt_label = ctx.Input<Tensor>("GTLabel");
+    auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
+    auto* objness_mask = ctx.Input<Tensor>("ObjectnessMask");
+    auto* gt_match_mask = ctx.Input<Tensor>("GTMatchMask");
+    auto anchors = ctx.Attr<std::vector<int>>("anchors");
+    auto anchor_mask = ctx.Attr<std::vector<int>>("anchor_mask");
+    int class_num = ctx.Attr<int>("class_num");
+    int downsample_ratio = ctx.Attr<int>("downsample_ratio");
+
+    const int n = input_grad->dims()[0];
+    const int c = input_grad->dims()[1];
+    const int h = input_grad->dims()[2];
+    const int w = input_grad->dims()[3];
+    const int mask_num = anchor_mask.size();
+    const int b = gt_match_mask->dims()[1];
+    int input_size = downsample_ratio * h;
+
+    const int stride = h * w;
+    const int an_stride = (class_num + 5) * stride;
+
+    const T* input_data = input->data<T>();
+    const T* gt_box_data = gt_box->data<T>();
+    const int* gt_label_data = gt_label->data<int>();
+    const T* loss_grad_data = loss_grad->data<T>();
+    const T* obj_mask_data = objness_mask->data<T>();
+    const int* gt_match_mask_data = gt_match_mask->data<int>();
+    T* input_grad_data =
+        input_grad->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
+    memset(input_grad_data, 0, input_grad->numel() * sizeof(T));
+
+    for (int i = 0; i < n; i++) {
+      for (int t = 0; t < b; t++) {
+        int mask_idx = gt_match_mask_data[i * b + t];
+        if (mask_idx >= 0) {
+          Box<T> gt = GetGtBox(gt_box_data, i, b, t);
+          int gi = static_cast<int>(gt.x * w);
+          int gj = static_cast<int>(gt.y * h);
+
+          int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
+                                      an_stride, stride, 0);
+          CalcBoxLocationLossGrad<T>(
+              input_grad_data, loss_grad_data[i], input_data, gt, anchors,
+              anchor_mask[mask_idx], box_idx, gi, gj, h, input_size, stride);
+
+          int label = gt_label_data[i * b + t];
+          int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
+                                        an_stride, stride, 5);
+          CalcLabelLossGrad<T>(input_grad_data, loss_grad_data[i], input_data,
+                               label_idx, label, class_num, stride);
+        }
+      }
+    }
+
+    CalcObjnessLossGrad<T>(input_grad_data + 4 * stride, loss_grad_data,
+                           input_data + 4 * stride, obj_mask_data, n, mask_num,
+                           h, w, stride, an_stride);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc

@@ -79,17 +79,17 @@ void FusionRepeatedFCReluOpMaker::Make() {
 }
 
 template <typename T>
-static void fc_relu(const T* x, const T* w, const T* b, T* y, int m, int n,
-                    int k) {
+static void fc_relu(const T* x, const T* w, const T* b, T* y,
+                    const jit::matmul_attr_t& attr) {
   auto matmul =
-      jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(k);
+      jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(attr);
   auto addbias_relu =
-      jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(n);
-  matmul(x, w, y, m, n, k);
+      jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(attr.n);
+  matmul(x, w, y, &attr);
   T* dst = y;
-  for (int i = 0; i < m; ++i) {
-    addbias_relu(b, dst, dst, n);
-    dst += n;
+  for (int i = 0; i < attr.m; ++i) {
+    addbias_relu(b, dst, dst, attr.n);
+    dst += attr.n;
   }
 }
 
@@ -107,32 +107,33 @@ class FusionRepeatedFCReluKernel : public framework::OpKernel<T> {
 
     auto i_dims = in->dims();
     auto w_dims = weights[0]->dims();
-    int m = i_dims[0];
-    int n = w_dims[1];
-    int k = w_dims[0];
-    relus[0]->Resize({m, n});
+    jit::matmul_attr_t attr;
+    attr.m = i_dims[0];
+    attr.n = w_dims[1];
+    attr.k = w_dims[0];
+    relus[0]->Resize({attr.m, attr.n});
     fc_relu(in->data<T>(), weights[0]->data<T>(), biases[0]->data<T>(),
-            relus[0]->mutable_data<T>(place), m, n, k);
+            relus[0]->mutable_data<T>(place), attr);
 
     for (int i = 1; i < weight_sz - 1; ++i) {
       auto i_dims = relus[i - 1]->dims();
       auto w_dims = weights[i]->dims();
-      int m = i_dims[0];
-      int n = w_dims[1];
-      int k = w_dims[0];
-      relus[i]->Resize({m, n});
+      attr.m = i_dims[0];
+      attr.n = w_dims[1];
+      attr.k = w_dims[0];
+      relus[i]->Resize({attr.m, attr.n});
       fc_relu(relus[i - 1]->data<T>(), weights[i]->data<T>(),
-              biases[i]->data<T>(), relus[i]->mutable_data<T>(place), m, n, k);
+              biases[i]->data<T>(), relus[i]->mutable_data<T>(place), attr);
     }
 
     auto i_dims_last = relus[weight_sz - 2]->dims();
     auto w_dims_last = weights[weight_sz - 1]->dims();
-    m = i_dims_last[0];
-    n = w_dims_last[1];
-    k = w_dims_last[0];
+    attr.m = i_dims_last[0];
+    attr.n = w_dims_last[1];
+    attr.k = w_dims_last[0];
     fc_relu(relus[weight_sz - 2]->data<T>(), weights[weight_sz - 1]->data<T>(),
-            biases[weight_sz - 1]->data<T>(), out->mutable_data<T>(place), m, n,
-            k);
+            biases[weight_sz - 1]->data<T>(), out->mutable_data<T>(place),
+            attr);
   }
 };
 

paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc

@@ -87,15 +87,18 @@ class FusionSquaredMatSubKernel : public framework::OpKernel<T> {
 
     auto x_dims = x->dims();
     auto y_dims = y->dims();
-    int m = x_dims[0];
-    int k = x_dims[1];
-    int n = y_dims[1];
-    int o_numel = m * n;
+    jit::matmul_attr_t attr;
+    attr.m = x_dims[0];
+    attr.k = x_dims[1];
+    attr.n = y_dims[1];
+    int o_numel = attr.m * attr.n;
 
     auto vsquare_x =
-        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(m * k);
+        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(attr.m *
+                                                                       attr.k);
     auto vsquare_y =
-        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(k * n);
+        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(attr.k *
+                                                                       attr.n);
     auto vsquare_xy =
         jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(o_numel);
     auto vsub =
@@ -103,7 +106,7 @@ class FusionSquaredMatSubKernel : public framework::OpKernel<T> {
     auto vscal =
         jit::Get<jit::kVScal, jit::AXYNTuples<T>, platform::CPUPlace>(o_numel);
     auto matmul =
-        jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(k);
+        jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(attr);
 
     const T* x_data = x->data<T>();
     const T* y_data = y->data<T>();
@@ -112,12 +115,12 @@ class FusionSquaredMatSubKernel : public framework::OpKernel<T> {
     T* squared_xy_data = squared_xy->mutable_data<T>(place);
     T* o_data = out->mutable_data<T>(place);
 
-    matmul(x_data, y_data, squared_xy_data, m, n, k);
+    matmul(x_data, y_data, squared_xy_data, &attr);
     vsquare_xy(squared_xy_data, squared_xy_data, o_numel);
 
-    vsquare_x(x_data, squared_x_data, m * k);
-    vsquare_y(y_data, squared_y_data, k * n);
-    matmul(squared_x_data, squared_y_data, o_data, m, n, k);
+    vsquare_x(x_data, squared_x_data, attr.m * attr.k);
+    vsquare_y(y_data, squared_y_data, attr.k * attr.n);
+    matmul(squared_x_data, squared_y_data, o_data, &attr);
 
     vsub(squared_xy_data, o_data, o_data, o_numel);
     vscal(&scalar, o_data, o_data, o_numel);

paddle/fluid/operators/jit/benchmark.cc

@@ -93,6 +93,7 @@ std::vector<int> TestSizes() {
 template <typename KernelTuples, typename... Args>
 struct BenchFunc {
   // return this function avg time
+  // TODO(TJ): clear cache every time
   double operator()(const typename KernelTuples::func_type tgt, Args... args) {
     for (int i = 0; i < FLAGS_burning; ++i) {
       tgt(args...);
@@ -172,6 +173,9 @@ void BenchXYZNKernel() {
     RandomVec<T>(d, y_data);
     BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data<T>(),
                                                      y.data<T>(), z_data, d);
+    // test inplace
+    BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data<T>(), z_data,
+                                                     z_data, d);
   }
 }
 
@@ -311,8 +315,9 @@ void BenchMatMulKernel() {
         const T* a_data = a.data<T>();
         const T* b_data = b.data<T>();
         T* c_data = c.mutable_data<T>(PlaceType());
-        BenchAllImpls<KT, jit::MatMulTuples<T>, PlaceType>(k, a_data, b_data,
-                                                           c_data, m, n, k);
+        const jit::matmul_attr_t attr{m, n, k};
+        BenchAllImpls<KT, jit::MatMulTuples<T>, PlaceType>(attr, a_data, b_data,
+                                                           c_data, &attr);
       }
     }
   }

paddle/fluid/operators/jit/gen/CMakeLists.txt

@@ -9,6 +9,7 @@ function(USE_JITKERNEL_GEN TARGET)
 endfunction()
 
 # use gen jitcode kernel by name
+USE_JITKERNEL_GEN(kMatMul)
 USE_JITKERNEL_GEN(kVMul)
 USE_JITKERNEL_GEN(kVAdd)
 USE_JITKERNEL_GEN(kVSub)

paddle/fluid/operators/jit/gen/blas.cc

@@ -155,7 +155,7 @@ class NCHW16CMulNCCreator : public JitCodeCreator<int> {
   class name##Creator : public JitCodeCreator<int> {                         \
    public:                                                                   \
     bool UseMe(const int& attr) const override {                             \
-      return platform::MayIUse(platform::avx);                               \
+      return platform::MayIUse(platform::avx) && attr <= 1024;               \
     }                                                                        \
     size_t CodeSize(const int& d) const override {                           \
       return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \

paddle/fluid/operators/jit/gen/blas.h

@@ -61,6 +61,7 @@ class VXXJitCode : public JitCode {
       base += "_Vec";
     }
     base += (with_relu_ ? "_Relu" : "");
+    base += "_D" + std::to_string(num_);
     return base.c_str();
   }
   void genCode() override;

paddle/fluid/operators/jit/gen/matmul.cc

+/* 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. */
+
+#include "paddle/fluid/operators/jit/gen/matmul.h"
+#include <stddef.h>  // offsetof
+#include <vector>
+
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+void MatMulJitCode::genCode() {
+  preCode();
+  int block, rest;
+  const auto groups = packed_groups(n_, k_, &block, &rest);
+  PADDLE_ENFORCE_GT(groups.front(), 0);
+
+  const int block_len = sizeof(float) * block;
+  const int x_reg_idx = (block == ZMM_FLOAT_BLOCK ? 32 : 16) - 1;
+  const int w_reg_idx = x_reg_idx - 1;
+  // from packed mov(reg_ptr_wgt, ptr[param_attr + offsetof(matmul_attr_t,
+  // packed_weight)]);
+  mov(reg_ptr_wgt, param_y);
+  size_t z_offset = 0;
+  size_t wgt_offset = 0;
+  for (size_t g = 0; g < groups.size(); ++g) {
+    size_t x_offset = 0;
+    for (int k = 0; k < k_; ++k) {
+      vbroadcastss(zmm_t(x_reg_idx), ptr[param_x + x_offset]);
+      // clean
+      if (k == 0) {
+        for (int i = 0; i < groups[g]; ++i) {
+          vxorps(zmm_t(i), zmm_t(i), zmm_t(i));
+        }
+      }
+      for (int i = 0; i < groups[g]; ++i) {
+        vmovups(zmm_t(w_reg_idx), ptr[reg_ptr_wgt + wgt_offset]);
+        vfmadd231ps(zmm_t(i), zmm_t(w_reg_idx), zmm_t(x_reg_idx));
+        wgt_offset += block_len;
+      }
+      // last one, save
+      if (k == k_ - 1) {
+        for (int i = 0; i < groups[g]; ++i) {
+          // only rest save should be careful
+          if (rest != 0 && g == groups.size() - 1 && i == groups[g] - 1) {
+            break;
+          }
+          vmovups(ptr[param_z + z_offset + i * block_len], zmm_t(i));
+        }
+      }
+      x_offset += sizeof(float);
+    }
+    z_offset += block_len * groups[g];
+  }
+
+  if (rest != 0) {
+    // below should refine with mask
+    int reg_idx = groups.back() - 1;
+    z_offset = (n_ - rest) * sizeof(float);
+    int inner_block = 8;
+    while (rest > 0) {
+      if (rest >= 8) {
+        inner_block = 8;
+        vmovups(ptr[param_z + z_offset], ymm_t(reg_idx));
+        // shift zmm of inner_block, change reg_idx if update
+      } else if (rest >= 4) {
+        inner_block = 4;
+        vmovups(ptr[param_z + z_offset], xmm_t(reg_idx));
+      } else if (rest >= 2) {
+        inner_block = 2;
+        vmovq(ptr[param_z + z_offset], xmm_t(reg_idx));
+      } else {
+        inner_block = 1;
+        vmovss(ptr[param_z + z_offset], xmm_t(reg_idx));
+      }
+      z_offset += inner_block * sizeof(float);
+      rest -= inner_block;
+    }
+  }
+
+  postCode();
+}
+
+class MatMulCreator : public JitCodeCreator<matmul_attr_t> {
+ public:
+  bool UseMe(const matmul_attr_t& attr) const override {
+    return attr.m == 1 && platform::MayIUse(platform::avx512f) &&
+           attr.n % ZMM_FLOAT_BLOCK == 0 && attr.k < 512;
+  }
+  size_t CodeSize(const matmul_attr_t& attr) const override {
+    int block = YMM_FLOAT_BLOCK;
+    if (platform::MayIUse(platform::avx512f)) {
+      block = ZMM_FLOAT_BLOCK;
+    }
+    return 96 + 4 * attr.k * (attr.n / block + 1) * 8;
+  }
+  std::unique_ptr<GenBase> CreateJitCode(
+      const matmul_attr_t& attr) const override {
+    PADDLE_ENFORCE_GT(attr.m, 0);
+    PADDLE_ENFORCE_GT(attr.n, 0);
+    PADDLE_ENFORCE_GT(attr.k, 0);
+    return make_unique<MatMulJitCode>(attr, CodeSize(attr));
+  }
+};
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kMatMul, gen::MatMulCreator);

paddle/fluid/operators/jit/gen/matmul.h

+/* 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 <stdlib.h>  // for malloc and free
+#include <string>
+#include <vector>
+#include "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+class MatMulJitCode : public JitCode {
+ public:
+  explicit MatMulJitCode(const matmul_attr_t& attr,
+                         size_t code_size = 256 * 1024,
+                         void* code_ptr = nullptr)
+      : JitCode(code_size, code_ptr), m_(attr.m), n_(attr.n), k_(attr.k) {
+    PADDLE_ENFORCE_EQ(m_, 1, "Only support m==1 yet");
+    this->genCode();
+  }
+
+  virtual const char* name() const {
+    std::string base = "MatMulJitCode";
+    base = base + "_M" + std::to_string(m_) + "_N" + std::to_string(n_) + "_K" +
+           std::to_string(k_);
+    return base.c_str();
+  }
+  void genCode() override;
+
+ private:
+  int m_, n_, k_;
+
+  reg64_t param_x{abi_param1};
+  reg64_t param_y{abi_param2};
+  reg64_t param_z{abi_param3};
+  reg64_t param_attr{abi_param4};
+  reg64_t reg_tmp{rax};
+
+  reg64_t reg_ptr_wgt{r10};
+};
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen_base.cc

@@ -16,6 +16,8 @@
 #include <fstream>
 #include <iostream>
 #include <sstream>
+#include <vector>
+#include "paddle/fluid/platform/cpu_info.h"
 
 DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file");
 
@@ -38,6 +40,35 @@ void GenBase::dumpCode(const unsigned char* code) const {
   }
 }
 
+std::vector<int> packed_groups(int n, int k, int* block_out, int* rest_out) {
+  int block;
+  int max_num_regs;
+  if (platform::MayIUse(platform::avx512f)) {
+    block = ZMM_FLOAT_BLOCK;
+    max_num_regs = 32;
+  } else {
+    block = YMM_FLOAT_BLOCK;
+    max_num_regs = 16;
+  }
+  // one for x, one for y, others for z
+  const int max_used_regs_for_n = max_num_regs - 2;
+  const int aligned_n = n % block == 0 ? n : (n / block + 1) * block;
+  const int num_block = aligned_n / block;
+  const int num_groups = num_block / max_used_regs_for_n;
+  std::vector<int> groups(num_groups, max_used_regs_for_n);
+  int rest_num_regs = num_block % max_used_regs_for_n;
+  if (rest_num_regs != 0) {
+    groups.push_back(rest_num_regs);
+  }
+  if (block_out) {
+    *block_out = block;
+  }
+  if (rest_out) {
+    *rest_out = n % block;
+  }
+  return groups;
+}
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/gen_base.h

@@ -16,6 +16,7 @@
 
 #include <gflags/gflags.h>
 #include <memory>  // for unique_ptr
+#include <vector>
 #include "paddle/fluid/operators/jit/kernel_base.h"
 
 DECLARE_bool(dump_jitcode);
@@ -67,6 +68,11 @@ class JitCodeCreator : public GenCreator {
   virtual std::unique_ptr<GenBase> CreateJitCode(const Attr& attr) const = 0;
 };
 
+// unify the method of packed groups
+// output the packed groups which used in weights, the block size and rest size
+std::vector<int> packed_groups(int n, int k, int* block = nullptr,
+                               int* rest = nullptr);
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/helper.cc

@@ -14,6 +14,8 @@
 
 #include "paddle/fluid/operators/jit/helper.h"
 #include <algorithm>  // tolower
+#include <numeric>
+#include <string>
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
@@ -91,6 +93,41 @@ KernelType to_kerneltype(const std::string& act) {
   return kNone;
 }
 
+template <>
+void pack_weights<float>(const float* src, float* dst, int n, int k) {
+  int block, rest;
+  const auto groups = packed_groups(n, k, &block, &rest);
+  std::for_each(groups.begin(), groups.end(), [&](int i) {
+    PADDLE_ENFORCE_GT(i, 0, "each element of groups should be larger than 0.");
+  });
+  int sum = std::accumulate(groups.begin(), groups.end(), 0);
+  std::memset(dst, 0, k * sum * block * sizeof(float));
+  PADDLE_ENFORCE_GE(sum * block, n,
+                    "The packed n should be equal to or larger than n");
+
+  const int block_len = sizeof(float) * block;
+  int n_offset = 0;
+
+  for (size_t g = 0; g < groups.size(); ++g) {
+    const float* from = src + n_offset;
+    for (int j = 0; j < k; ++j) {
+      size_t copy_sz = groups[g] * block_len;
+      if (g == groups.size() - 1 && rest != 0) {
+        copy_sz = (groups[g] - 1) * block_len + rest * sizeof(float);
+      }
+      std::memcpy(dst, from + j * n, copy_sz);
+      dst += groups[g] * block;
+    }
+    n_offset += groups[g] * block;
+  }
+}
+
+template <typename T>
+typename std::enable_if<!std::is_same<T, float>::value>::type pack_weights(
+    const T* src, T* dst, int n, int k) {
+  PADDLE_THROW("Only support pack with float type.");
+}
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/helper.h

@@ -118,26 +118,33 @@ typename KernelTuples::func_type Get(
   return GetRefer<KT, KernelTuples>();
 }
 
-template <KernelType KT, typename KernelTuples>
-class KernelFuncsCache {
+template <KernelType KT, typename KernelTuples, typename PlaceType>
+class KernelFuncs {
  public:
-  KernelFuncsCache() = default;
-  static KernelFuncsCache& Instance() {
-    static thread_local KernelFuncsCache<KT, KernelTuples> g_func_cache;
+  KernelFuncs() = default;
+  static KernelFuncs& Cache() {
+    static thread_local KernelFuncs<KT, KernelTuples, PlaceType> g_func_cache;
     return g_func_cache;
   }
 
   bool Has(int key) const { return funcs_.find(key) != funcs_.end(); }
 
-  typename KernelTuples::func_type At(int key) { return funcs_.at(key); }
-
   void Insert(int key, typename KernelTuples::func_type func) {
     funcs_.emplace(key, func);
   }
 
+  typename KernelTuples::func_type At(int key) {
+    if (Has(key)) {
+      return funcs_.at(key);
+    }
+    auto func = Get<KT, KernelTuples, PlaceType>(key);
+    Insert(key, func);
+    return func;
+  }
+
  private:
   std::unordered_map<int, typename KernelTuples::func_type> funcs_;
-  DISABLE_COPY_AND_ASSIGN(KernelFuncsCache);
+  DISABLE_COPY_AND_ASSIGN(KernelFuncs);
 };
 
 const char* to_string(KernelType kt);
@@ -152,17 +159,28 @@ inline std::ostream& operator<<(std::ostream& os, const lstm_attr_t& attr) {
      << (attr.use_peephole ? "True" : "False") << "]";
   return os;
 }
+
 inline std::ostream& operator<<(std::ostream& os, const gru_attr_t& attr) {
   os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate)
      << "],act_cand[" << to_string(attr.act_cand) << "]";
   return os;
 }
+
 inline std::ostream& operator<<(std::ostream& os, const seq_pool_attr_t& attr) {
   os << "height_size[" << attr.h << "],width_size[" << attr.w << "],pool_type["
      << to_string(attr.type) << "]";
   return os;
 }
 
+inline std::ostream& operator<<(std::ostream& os, const matmul_attr_t& attr) {
+  os << "M[" << attr.m << "],N[" << attr.n << "],K[" << attr.k << "]";
+  return os;
+}
+
+// expose the method to pack matmul weight
+template <typename T>
+void pack_weights(const T* src, T* dst, int n, int k);
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/kernel_base.h

@@ -145,11 +145,19 @@ struct SeqPoolTuples {
   typedef void (*func_type)(const T*, T*, const seq_pool_attr_t*);
 };
 
+typedef struct matmul_attr_s {
+  int m, n, k;
+  void* packed_weight{nullptr};
+  matmul_attr_s() = default;
+  explicit matmul_attr_s(int m_, int n_, int k_, void* packed_weight_ = nullptr)
+      : m(m_), n(n_), k(k_), packed_weight(packed_weight_) {}
+} matmul_attr_t;
+
 template <typename T>
 struct MatMulTuples {
   typedef T data_type;
-  typedef int attr_type;
-  typedef void (*func_type)(const T*, const T*, T*, int, int, int);
+  typedef matmul_attr_t attr_type;
+  typedef void (*func_type)(const T*, const T*, T*, const matmul_attr_t*);
 };
 
 template <typename T>

paddle/fluid/operators/jit/kernel_key.cc

@@ -49,6 +49,13 @@ size_t JitCodeKey<seq_pool_attr_t>(const seq_pool_attr_t& attr) {
   return (key << pool_type_shift) + static_cast<int>(attr.type);
 }
 
+template <>
+size_t JitCodeKey<matmul_attr_t>(const matmul_attr_t& attr) {
+  size_t key = attr.m;
+  constexpr int shift = 21;
+  return (key << shift * 2) + ((static_cast<size_t>(attr.n)) << shift) + attr.k;
+}
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/more/mix/mix.cc

@@ -49,49 +49,16 @@ void VTanh(const T* x, T* y, int n) {
 }
 
 void Softmax(const T* x, T* y, int n, int bs) {
-  typename XRNTuples<T>::func_type compute_hmax{nullptr};
-  typename XRNTuples<T>::func_type compute_hsum{nullptr};
-  typename AXYNTuples<T>::func_type compute_vscal{nullptr};
-  typename AXYNTuples<T>::func_type compute_vaddbias{nullptr};
-  typename XYNTuples<T>::func_type compute_vexp{nullptr};
-
-  if (!KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Has(n)) {
-    compute_hmax = Get<kHMax, XRNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Insert(n, compute_hmax);
-  } else {
-    compute_hmax = KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Has(n)) {
-    compute_hsum = Get<kHSum, XRNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Insert(n, compute_hsum);
-  } else {
-    compute_hsum = KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Has(n)) {
-    compute_vscal = Get<kVScal, AXYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Insert(n,
-                                                               compute_vscal);
-  } else {
-    compute_vscal = KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Has(n)) {
-    compute_vaddbias = Get<kVAddBias, AXYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Insert(
-        n, compute_vaddbias);
-  } else {
-    compute_vaddbias =
-        KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Has(n)) {
-    compute_vexp = Get<KernelType::kVExp, XYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Insert(n, compute_vexp);
-  } else {
-    compute_vexp = KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().At(n);
-  }
+  auto compute_hmax =
+      KernelFuncs<kHMax, XRNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_hsum =
+      KernelFuncs<kHSum, XRNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vscal =
+      KernelFuncs<kVScal, AXYNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vaddbias =
+      KernelFuncs<kVAddBias, AXYNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vexp =
+      KernelFuncs<kVExp, XYNTuples<T>, platform::CPUPlace>::Cache().At(n);
 
   for (int i = 0; i < bs; ++i) {
     T scalar;

paddle/fluid/operators/jit/more/mkl/mkl.cc

@@ -25,17 +25,19 @@ namespace more {
 namespace mkl {
 
 template <>
-void MatMul<float>(const float* a, const float* b, float* c, int m, int n,
-                   int k) {
-  platform::dynload::cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m,
-                                 n, k, 1.f, a, k, b, n, 0.f, c, n);
+void MatMul<float>(const float* a, const float* b, float* c,
+                   const matmul_attr_t* attr) {
+  platform::dynload::cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
+                                 attr->m, attr->n, attr->k, 1.f, a, attr->k, b,
+                                 attr->n, 0.f, c, attr->n);
 }
 
 template <>
-void MatMul<double>(const double* a, const double* b, double* c, int m, int n,
-                    int k) {
-  platform::dynload::cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m,
-                                 n, k, 1.0, a, k, b, n, 0.0, c, n);
+void MatMul<double>(const double* a, const double* b, double* c,
+                    const matmul_attr_t* attr) {
+  platform::dynload::cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
+                                 attr->m, attr->n, attr->k, 1.0, a, attr->k, b,
+                                 attr->n, 0.0, c, attr->n);
 }
 
 template <>
@@ -127,11 +129,6 @@ void ASum<double>(const double* x, double* res, int n) {
 }
 
 // TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
-template <>
-bool MatMulKernel<float>::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx);
-}
-
 template <>
 bool VMulKernel<float>::UseMe(const int& d) const {
   return platform::MayIUse(platform::avx512f) && d > 512;
@@ -139,7 +136,7 @@ bool VMulKernel<float>::UseMe(const int& d) const {
 
 template <>
 bool VAddKernel<float>::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx512f) && d > 512;
+  return platform::MayIUse(platform::avx) && d > 512;
 }
 
 template <>
@@ -177,6 +174,16 @@ bool SeqPoolKernel<double>::UseMe(const seq_pool_attr_t& attr) const {
   return true;
 }
 
+template <>
+bool MatMulKernel<float>::UseMe(const matmul_attr_t& attr) const {
+  return platform::MayIUse(platform::avx);
+}
+
+template <>
+bool MatMulKernel<double>::UseMe(const matmul_attr_t& attr) const {
+  return true;
+}
+
 template <>
 bool SoftmaxKernel<float>::UseMe(const int& d) const {
   // tuned on avx2
@@ -189,7 +196,6 @@ bool SoftmaxKernel<float>::UseMe(const int& d) const {
     return true;                                         \
   }
 
-AWALYS_USE_ME_WITH_DOUBLE(MatMul);
 AWALYS_USE_ME_WITH_DOUBLE(VMul);
 AWALYS_USE_ME_WITH_DOUBLE(VAdd);
 AWALYS_USE_ME_WITH_DOUBLE(VScal);

paddle/fluid/operators/jit/more/mkl/mkl.h

@@ -26,7 +26,7 @@ namespace more {
 namespace mkl {
 
 template <typename T>
-void MatMul(const T* a, const T* b, T* c, int m, int n, int k);
+void MatMul(const T* a, const T* b, T* c, const matmul_attr_t* attr);
 
 template <typename T>
 void VMul(const T* x, const T* y, T* z, int n);

paddle/fluid/operators/jit/refer/refer.h

@@ -363,17 +363,19 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
 
 // A(M,K) * B(K,N) = C(M,N)
 template <typename T>
-void MatMul(const T* A, const T* B, T* C, int M, int N, int K) {
+void MatMul(const T* A, const T* B, T* C, const matmul_attr_t* attr) {
+  int M = attr->m;
+  int N = attr->n;
+  int K = attr->k;
   for (int m = 0; m < M; ++m) {
     const T* pa = A + m * K;
     T* pc = C + m * N;
     for (int n = 0; n < N; ++n) {
       const T* pb = B + n;
-      T sum = static_cast<T>(0);
-      for (int k = 0; k < K; ++k) {
-        sum += (pa[k] * pb[k * N]);
+      pc[n] = pa[0] * pb[0];
+      for (int k = 1; k < K; ++k) {
+        pc[n] += pa[k] * pb[k * N];
       }
-      *(pc + n) = sum;
     }
   }
 }

paddle/fluid/operators/jit/test.cc

@@ -22,7 +22,7 @@
 #include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/place.h"
 
-static double acc = 1e-5;
+DEFINE_double(acc, 1e-5, "Test accuracy threshold.");
 
 template <typename T>
 void RandomVec(const int n, T* a, const T lower = static_cast<T>(-20.f),
@@ -39,7 +39,7 @@ template <typename T>
 void ExpectEQ(const T* target, const T* refer, int n) {
   if (std::is_floating_point<T>::value) {
     for (int i = 0; i < n; ++i) {
-      EXPECT_NEAR(target[i], refer[i], acc);
+      EXPECT_NEAR(target[i], refer[i], FLAGS_acc);
     }
   } else {
     for (int i = 0; i < n; ++i) {
@@ -272,21 +272,23 @@ struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>, std::vector<T>,
 
 template <typename T>
 struct TestFuncWithRefer<jit::MatMulTuples<T>, std::vector<T>, std::vector<T>,
-                         std::vector<T>, int, int, int> {
+                         std::vector<T>,
+                         typename jit::MatMulTuples<T>::attr_type> {
   void operator()(const typename jit::MatMulTuples<T>::func_type tgt,
                   const std::vector<T>& a, const std::vector<T>& b,
-                  const std::vector<T>& cref, int m, int n, int k) {
+                  const std::vector<T>& cref,
+                  const typename jit::MatMulTuples<T>::attr_type& attr) {
     EXPECT_TRUE(tgt != nullptr);
-    EXPECT_EQ(a.size(), static_cast<size_t>(m * k));
-    EXPECT_EQ(b.size(), static_cast<size_t>(k * n));
-    EXPECT_EQ(cref.size(), static_cast<size_t>(m * n));
+    EXPECT_EQ(a.size(), static_cast<size_t>(attr.m * attr.k));
+    EXPECT_EQ(b.size(), static_cast<size_t>(attr.k * attr.n));
+    EXPECT_EQ(cref.size(), static_cast<size_t>(attr.m * attr.n));
     std::vector<T> c(cref.size());
     const T* a_data = a.data();
     const T* b_data = b.data();
     const T* cref_data = cref.data();
     T* c_data = c.data();
-    tgt(a_data, b_data, c_data, m, n, k);
-    ExpectEQ<T>(c_data, cref_data, m * n);
+    tgt(a_data, b_data, c_data, &attr);
+    ExpectEQ<T>(c_data, cref_data, attr.m * attr.n);
   }
 };
 
@@ -383,8 +385,8 @@ void TestAXYNKernel() {
 template <jit::KernelType KT, typename T, typename PlaceType>
 void TestXRNKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
-  auto last_acc = acc;
-  acc = 1e-4;
+  auto last_acc = FLAGS_acc;
+  FLAGS_acc = 1e-4;
   for (int d : TestSizes()) {
     auto ref = jit::GetRefer<KT, jit::XRNTuples<T>>();
     EXPECT_TRUE(ref != nullptr);
@@ -395,7 +397,7 @@ void TestXRNKernel() {
     TestAllImpls<KT, jit::XRNTuples<T>, PlaceType, std::vector<T>, T>(d, x,
                                                                       ref_res);
   }
-  acc = last_acc;
+  FLAGS_acc = last_acc;
 }
 
 template <jit::KernelType KT, typename T, typename PlaceType>
@@ -535,9 +537,10 @@ void TestSeqPoolKernel() {
 template <jit::KernelType KT, typename T, typename PlaceType>
 void TestMatMulKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
-  auto last_acc = acc;
-  // TODO(intel): this should be acc issue of MKL
-  acc = 1e-3;
+  auto last_acc = FLAGS_acc;
+  // TODO(intel): fix MKL acc issue
+  // https://github.com/PaddlePaddle/Paddle/issues/15447
+  FLAGS_acc = 1e-3;
   for (int m : {1, 2, 3, 4}) {
     for (int n : {1, 2, 3, 4}) {
       for (int k : TestSizes()) {
@@ -549,13 +552,14 @@ void TestMatMulKernel() {
         const T* a_data = a.data();
         const T* b_data = b.data();
         T* c_data = c.data();
-        ref(a_data, b_data, c_data, m, n, k);
+        const jit::matmul_attr_t attr{m, n, k};
+        ref(a_data, b_data, c_data, &attr);
         TestAllImpls<KT, jit::MatMulTuples<T>, PlaceType, std::vector<T>,
-                     std::vector<T>, std::vector<T>>(k, a, b, c, m, n, k);
+                     std::vector<T>, std::vector<T>>(attr, a, b, c, attr);
       }
     }
   }
-  acc = last_acc;
+  FLAGS_acc = last_acc;
 }
 
 template <jit::KernelType KT, typename T, typename PlaceType>

paddle/fluid/operators/math/fc_compute.h

@@ -30,15 +30,17 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
     return;
   }
   if (relu) {
-    auto compute =
-        jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(N);
+    auto compute = jit::KernelFuncs<jit::kVAddRelu, jit::XYZNTuples<T>,
+                                    platform::CPUPlace>::Cache()
+                       .At(N);
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
       compute(B, dst, dst, N);
     }
   } else {
-    auto compute =
-        jit::Get<jit::kVAdd, jit::XYZNTuples<T>, platform::CPUPlace>(N);
+    auto compute = jit::KernelFuncs<jit::kVAdd, jit::XYZNTuples<T>,
+                                    platform::CPUPlace>::Cache()
+                       .At(N);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif

paddle/fluid/operators/math/softmax_impl.h

@@ -82,8 +82,9 @@ class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
     const int kClassDim = 1;
     // 2D data. Batch x C
     auto compute_softmax =
-        jit::Get<jit::kSoftmax, jit::SoftmaxTuples<float>, platform::CPUPlace>(
-            in_dims[kClassDim]);
+        jit::KernelFuncs<jit::kSoftmax, jit::SoftmaxTuples<float>,
+                         platform::CPUPlace>::Cache()
+            .At(in_dims[kClassDim]);
     compute_softmax(in_data, out_data, in_dims[kClassDim], in_dims[kBatchDim]);
   }
 };

paddle/fluid/operators/ngraph/ngraph_bridge.cc

@@ -31,6 +31,8 @@ std::map<std::string,
                             std::shared_ptr<std::unordered_map<
                                 std::string, std::shared_ptr<ngraph::Node>>>)>>
     NgraphBridge::NG_NODE_MAP = {
+        {"conv2d", NG_OPS::BuildConv2dNode},
+        {"conv2d_grad", NG_OPS::BuildConv2dGradNode},
         {"elementwise_add", NG_OPS::BuildElementwiseAddNode},
         {"elementwise_add_grad", NG_OPS::BuildElementwiseAddGradNode},
         {"fill_constant", NG_OPS::BuildFillConstantNode},
@@ -38,6 +40,8 @@ std::map<std::string,
         {"mean_grad", NG_OPS::BuildMeanGradNode},
         {"mul", NG_OPS::BuildMulNode},
         {"mul_grad", NG_OPS::BuildMulGradNode},
+        {"pool2d", NG_OPS::BuildPool2dNode},
+        {"pool2d_grad", NG_OPS::BuildPool2dGradNode},
         {"softmax", NG_OPS::BuildSoftmaxNode},
         {"softmax_grad", NG_OPS::BuildSoftmaxGradNode},
         {"scale", NG_OPS::BuildScaleNode},

paddle/fluid/operators/ngraph/ngraph_ops.h

@@ -22,10 +22,12 @@ limitations under the License. */
 #pragma once
 
 #include "ops/binary_unnary_op.h"
+#include "ops/conv2d_op.h"
 #include "ops/elementwise_add_op.h"
 #include "ops/fill_constant_op.h"
 #include "ops/mean_op.h"
 #include "ops/mul_op.h"
+#include "ops/pool2d_op.h"
 #include "ops/scale_op.h"
 #include "ops/softmax_op.h"
 #include "ops/top_k_op.h"

paddle/fluid/operators/ngraph/ops/conv2d_op.h

+/* 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 <string>
+#include <vector>
+#include "ngraph/ngraph.hpp"
+#include "paddle/fluid/platform/ngraph_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace ngraphs {
+
+std::shared_ptr<ngraph::Node> GroupedConvolution(
+    const std::shared_ptr<ngraph::Node>& data_batch,
+    const std::shared_ptr<ngraph::Node>& filters, const ngraph::Strides strides,
+    const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings,
+    size_t groups) {
+  auto& data_shape = data_batch->get_shape();
+  auto& filter_shape = filters->get_shape();
+  ngraph::NodeVector ng_slices;
+
+  for (size_t i = 0; i < groups; ++i) {
+    size_t channel_step = filter_shape.at(1);
+    const std::vector<size_t> lower_bound{0, i * channel_step, 0, 0};
+    const std::vector<size_t> upper_bound{data_shape.at(0),
+                                          (i + 1) * channel_step,
+                                          data_shape.at(2), data_shape.at(3)};
+    auto data_slice = std::make_shared<ngraph::op::Slice>(
+        data_batch, lower_bound, upper_bound);
+
+    size_t filter_step = filter_shape.at(0) / groups;
+    const std::vector<size_t> filter_lower_bound{i * filter_step, 0, 0, 0};
+    const std::vector<size_t> filter_upper_bound{
+        (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2),
+        filter_shape.at(3)};
+    auto filter_slice = std::make_shared<ngraph::op::Slice>(
+        filters, filter_lower_bound, filter_upper_bound);
+    auto ng_conv = std::make_shared<ngraph::op::Convolution>(
+        data_slice, filter_slice, strides, dilations, paddings, paddings);
+    ng_slices.push_back(ng_conv);
+  }
+
+  size_t concat_axis = 1;
+  return std::make_shared<ngraph::op::Concat>(ng_slices, concat_axis);
+}
+
+std::shared_ptr<ngraph::Node> GroupedGradConvolutionFilter(
+    const std::shared_ptr<ngraph::Node>& data_batch,
+    const std::shared_ptr<ngraph::Node>& filters,
+    const std::shared_ptr<ngraph::Node>& doutput, const ngraph::Strides strides,
+    const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings,
+    size_t groups) {
+  auto& data_shape = data_batch->get_shape();
+  auto& filter_shape = filters->get_shape();
+  auto& out_shape = doutput->get_shape();
+  ngraph::NodeVector ng_slices;
+
+  for (size_t i = 0; i < groups; ++i) {
+    size_t channel_step = filter_shape.at(1);
+    const std::vector<size_t> lower_bound{0, i * channel_step, 0, 0};
+    const std::vector<size_t> upper_bound{data_shape.at(0),
+                                          (i + 1) * channel_step,
+                                          data_shape.at(2), data_shape.at(3)};
+    auto data_slice = std::make_shared<ngraph::op::Slice>(
+        data_batch, lower_bound, upper_bound);
+
+    size_t filter_step = data_shape.at(0);
+
+    const std::vector<size_t> filter_lower_bound{i * filter_step, 0, 0, 0};
+    const std::vector<size_t> filter_upper_bound{
+        (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2),
+        filter_shape.at(3)};
+    auto filter_slice = std::make_shared<ngraph::op::Slice>(
+        filters, filter_lower_bound, filter_upper_bound);
+
+    const std::vector<size_t> olower_bound{0, i * filter_step, 0, 0};
+    const std::vector<size_t> oupper_bound{out_shape.at(0),
+                                           (i + 1) * filter_step,
+                                           out_shape.at(2), out_shape.at(3)};
+    auto out_slice = std::make_shared<ngraph::op::Slice>(doutput, olower_bound,
+                                                         oupper_bound);
+
+    auto ng_conv = std::make_shared<ngraph::op::ConvolutionBackpropFilters>(
+        data_slice, filter_slice->get_shape(), out_slice, strides, dilations,
+        paddings, paddings, ngraph::Strides{1, 1});
+
+    ng_slices.push_back(ng_conv);
+  }
+
+  size_t concat_axis = 0;
+  return std::make_shared<ngraph::op::Concat>(ng_slices, concat_axis);
+}
+
+std::shared_ptr<ngraph::Node> GroupedGradConvolutionData(
+    const std::shared_ptr<ngraph::Node>& data_batch,
+    const std::shared_ptr<ngraph::Node>& filters,
+    const std::shared_ptr<ngraph::Node>& doutput, const ngraph::Strides strides,
+    const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings,
+    size_t groups) {
+  auto& data_shape = data_batch->get_shape();
+  auto& filter_shape = filters->get_shape();
+  auto& out_shape = doutput->get_shape();
+  ngraph::NodeVector ng_slices;
+
+  for (size_t i = 0; i < groups; ++i) {
+    size_t channel_step = filter_shape.at(1);
+    const std::vector<size_t> lower_bound{0, i * channel_step, 0, 0};
+    const std::vector<size_t> upper_bound{data_shape.at(0),
+                                          (i + 1) * channel_step,
+                                          data_shape.at(2), data_shape.at(3)};
+    auto data_slice = std::make_shared<ngraph::op::Slice>(
+        data_batch, lower_bound, upper_bound);
+
+    size_t filter_step = data_shape.at(0);
+
+    const std::vector<size_t> filter_lower_bound{i * filter_step, 0, 0, 0};
+    const std::vector<size_t> filter_upper_bound{
+        (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2),
+        filter_shape.at(3)};
+    auto filter_slice = std::make_shared<ngraph::op::Slice>(
+        filters, filter_lower_bound, filter_upper_bound);
+
+    const std::vector<size_t> olower_bound{0, i * filter_step, 0, 0};
+    const std::vector<size_t> oupper_bound{out_shape.at(0),
+                                           (i + 1) * filter_step,
+                                           out_shape.at(2), out_shape.at(3)};
+    auto out_slice = std::make_shared<ngraph::op::Slice>(doutput, olower_bound,
+                                                         oupper_bound);
+
+    auto ng_conv = std::make_shared<ngraph::op::ConvolutionBackpropData>(
+        data_slice->get_shape(), filter_slice, out_slice, strides, dilations,
+        paddings, paddings, ngraph::Strides{1, 1});
+    ng_slices.push_back(ng_conv);
+  }
+
+  size_t concat_axis = 1;
+  return std::make_shared<ngraph::op::Concat>(ng_slices, concat_axis);
+}
+
+void BuildConv2dNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto filters = paddle::platform::GetInputNode(op, "Filter", ngb_node_map);
+  auto input = paddle::platform::GetInputNode(op, "Input", ngb_node_map);
+
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+  std::vector<int> dilations = op_attrs.Get<std::vector<int>>("dilations");
+
+  const ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                                   static_cast<size_t>(strides.at(1))};
+  const ngraph::Strides ng_dilations{static_cast<size_t>(dilations.at(0)),
+                                     static_cast<size_t>(dilations.at(1))};
+  const ngraph::CoordinateDiff ng_paddings{
+      static_cast<std::ptrdiff_t>(paddings.at(0)),
+      static_cast<std::ptrdiff_t>(paddings.at(1))};
+
+  int groups = static_cast<size_t>(op_attrs.Get<int>("groups"));
+  PADDLE_ENFORCE_GE(groups, 1, "conv groups needs be no less than 1");
+
+  std::shared_ptr<ngraph::Node> result;
+  if (groups == 1) {
+    result = std::make_shared<ngraph::op::Convolution>(
+        input, filters, ng_strides, ng_dilations, ng_paddings, ng_paddings);
+  } else {
+    result = GroupedConvolution(input, filters, ng_strides, ng_dilations,
+                                ng_paddings, groups);
+  }
+  paddle::platform::SetOutputNode(op, "Output", result, ngb_node_map);
+}
+
+void BuildConv2dGradNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto filter = paddle::platform::GetInputNode(op, "Filter", ngb_node_map);
+  auto input = paddle::platform::GetInputNode(op, "Input", ngb_node_map);
+  auto doutput =
+      paddle::platform::GetInputNode(op, "Output@GRAD", ngb_node_map);
+
+  int groups = op_attrs.Get<int>("groups");
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+  std::vector<int> dilations = op_attrs.Get<std::vector<int>>("dilations");
+
+  const ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                                   static_cast<size_t>(strides.at(1))};
+  const ngraph::Strides ng_dilations{static_cast<size_t>(dilations.at(0)),
+                                     static_cast<size_t>(dilations.at(1))};
+  const ngraph::CoordinateDiff ng_paddings{
+      static_cast<std::ptrdiff_t>(paddings.at(0)),
+      static_cast<std::ptrdiff_t>(paddings.at(1))};
+
+  std::shared_ptr<ngraph::Node> dfilter;
+  std::shared_ptr<ngraph::Node> dinput;
+  if (groups == 1) {
+    dfilter = std::make_shared<ngraph::op::ConvolutionBackpropFilters>(
+        input, filter->get_shape(), doutput, ng_strides, ng_dilations,
+        ng_paddings, ng_paddings, ngraph::Strides{1, 1});
+
+    dinput = std::make_shared<ngraph::op::ConvolutionBackpropData>(
+        input->get_shape(), filter, doutput, ng_strides, ng_dilations,
+        ng_paddings, ng_paddings, ngraph::Strides{1, 1});
+
+  } else {
+    dfilter = GroupedGradConvolutionFilter(input, filter, doutput, ng_strides,
+                                           ng_dilations, ng_paddings, groups);
+    dinput = GroupedGradConvolutionData(input, filter, doutput, ng_strides,
+                                        ng_dilations, ng_paddings, groups);
+  }
+
+  paddle::platform::SetOutputNode(op, "Filter@GRAD", dfilter, ngb_node_map);
+  paddle::platform::SetOutputNode(op, "Input@GRAD", dinput, ngb_node_map);
+}
+}  // namespace ngraphs
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/ngraph/ops/pool2d_op.h

+/*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 <string>
+#include <vector>
+
+#include "ngraph/ngraph.hpp"
+#include "paddle/fluid/platform/ngraph_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace ngraphs {
+
+void BuildPool2dNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map);
+  auto x_shape = x->get_shape();
+
+  std::string pooling_type = op_attrs.Get<std::string>("pooling_type");
+  std::vector<int> ksize = op_attrs.Get<std::vector<int>>("ksize");
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+
+  PADDLE_ENFORCE_EQ(x_shape.size() - 2, ksize.size(),
+                    "Handling 2d pooling only");
+
+  if (op_attrs.Get<bool>("global_pooling")) {
+    for (size_t i = 0; i < ksize.size(); ++i) {
+      paddings[i] = 0;
+      ksize[i] = static_cast<int>(x_shape.at(i + 2));
+    }
+  }
+
+  ngraph::Shape ng_padding_below{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_padding_above{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_ksize_shape{static_cast<size_t>(ksize.at(0)),
+                               static_cast<size_t>(ksize.at(1))};
+  ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                             static_cast<size_t>(strides.at(1))};
+
+  auto ComputeCeiledOutput = [](size_t in, size_t k, size_t p, size_t s) {
+    return (in - k + 2 * p) / s + 1;
+  };
+
+  if (op_attrs.Get<bool>("ceil_mode")) {
+    auto dummy_out = paddle::platform::GetOutputNode(op, "Out", ngb_node_map);
+    auto dummpy_shape = dummy_out->get_shape();
+    for (size_t i = 0; i < ng_padding_above.size(); ++i) {
+      auto desired_size = ComputeCeiledOutput(x_shape[i + 2], ksize[i],
+                                              paddings[i], strides[i]);
+      if (desired_size != dummpy_shape[i + 2]) {
+        ng_padding_above[i] += strides[i];
+      }
+    }
+  }
+
+  bool padding_exclusive = op_attrs.Get<bool>("exclusive");
+  if (pooling_type == "max") {
+    auto pool2d = std::make_shared<ngraph::op::MaxPool>(
+        x, ng_ksize_shape, ng_strides, ng_padding_below, ng_padding_above);
+    paddle::platform::SetOutputNode(op, "Out", pool2d, ngb_node_map);
+  } else if (pooling_type == "avg") {
+    std::shared_ptr<ngraph::Node> pool2d;
+    if (op_attrs.Get<bool>("adaptive")) {
+      auto ComputeAdaptive = [](size_t in, size_t k) {
+        return std::floor(in / k);
+      };
+      ng_strides[0] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[0];
+      ng_strides[1] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[1];
+      pool2d =
+          std::make_shared<ngraph::op::AvgPool>(x, ng_ksize_shape, ng_strides);
+    } else {
+      pool2d = std::make_shared<ngraph::op::AvgPool>(
+          x, ng_ksize_shape, ng_strides, ng_padding_below, ng_padding_above,
+          !padding_exclusive);
+    }
+    paddle::platform::SetOutputNode(op, "Out", pool2d, ngb_node_map);
+  } else {
+    PADDLE_THROW("Support max and avg pooling only");
+  }
+}
+
+void BuildPool2dGradNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto out = paddle::platform::GetInputNode(op, "Out", ngb_node_map);
+  auto dout = paddle::platform::GetInputNode(op, "Out@GRAD", ngb_node_map);
+  auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map);
+  auto x_shape = x->get_shape();
+
+  std::string pooling_type = op_attrs.Get<std::string>("pooling_type");
+  std::vector<int> ksize = op_attrs.Get<std::vector<int>>("ksize");
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+
+  PADDLE_ENFORCE_EQ(x_shape.size() - 2, ksize.size(),
+                    "Handling 2d pooling only");
+
+  if (op_attrs.Get<bool>("global_pooling")) {
+    for (size_t i = 0; i < ksize.size(); ++i) {
+      paddings[i] = 0;
+      ksize[i] = static_cast<int>(x_shape.at(i + 2));
+    }
+  }
+
+  ngraph::Shape ng_padding_below{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_padding_above{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_ksize_shape{static_cast<size_t>(ksize.at(0)),
+                               static_cast<size_t>(ksize.at(1))};
+  ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                             static_cast<size_t>(strides.at(1))};
+
+  bool padding_exclusive = op_attrs.Get<bool>("exclusive");
+  if (pooling_type == "max") {
+    auto pool2d_grad = std::make_shared<ngraph::op::MaxPoolBackprop>(
+        x, dout, out, ng_ksize_shape, ng_strides, ng_padding_below,
+        ng_padding_above);
+    paddle::platform::SetOutputNode(op, "X@GRAD", pool2d_grad, ngb_node_map);
+  } else if (pooling_type == "avg") {
+    std::shared_ptr<ngraph::Node> pool2d_grad;
+    if (op_attrs.Get<bool>("adaptive")) {
+      auto ComputeAdaptive = [](size_t in, size_t k) {
+        return std::floor(in / k);
+      };
+      ng_strides[0] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[0];
+      ng_strides[1] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[1];
+      pool2d_grad = std::make_shared<ngraph::op::AvgPoolBackprop>(
+          x->get_shape(), dout, ng_ksize_shape, ng_strides, ng_padding_below,
+          ng_padding_above, !padding_exclusive);
+    } else {
+      pool2d_grad = std::make_shared<ngraph::op::AvgPoolBackprop>(
+          x->get_shape(), dout, ng_ksize_shape, ng_strides, ng_padding_below,
+          ng_padding_above, !padding_exclusive);
+    }
+    paddle::platform::SetOutputNode(op, "X@GRAD", pool2d_grad, ngb_node_map);
+  } else {
+    PADDLE_THROW("Support max and avg pooling only");
+  }
+}
+}  // namespace ngraphs
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/platform/cuda_device_function.h

@@ -53,10 +53,12 @@ inline static int RoundToPowerOfTwo(int dim) {
     __VA_ARGS__;                           \
   } break
 
-#define CUDA_LAUNCH_KERNEL_HELPER(...)         \
-  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__);  \
+#define CUDA_LAUNCH_KERNEL_HELPER(...)          \
+  CUDA_LAUNCH_KERNEL_BASE(1024, ##__VA_ARGS__); \
+  CUDA_LAUNCH_KERNEL_BASE(512, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__);   \
   CUDA_LAUNCH_KERNEL_BASE(32, ##__VA_ARGS__);
 
 template <typename T>

paddle/fluid/platform/place.cc

@@ -14,6 +14,12 @@ limitations under the License. */
 
 #include "paddle/fluid/platform/place.h"
 
+DEFINE_bool(benchmark, false,
+            "Doing memory benchmark. It will make deleting scope synchronized, "
+            "and add some memory usage logs."
+            "Default cuda is asynchronous device, set to True will"
+            "force op run in synchronous mode.");
+
 namespace paddle {
 namespace platform {
 

paddle/fluid/pybind/pybind.cc

@@ -37,6 +37,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/version.h"
 #include "paddle/fluid/imperative/layer.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
+#include "paddle/fluid/memory/allocation/legacy_allocator.h"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/operators/py_func_op.h"
 #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
@@ -127,6 +128,13 @@ PYBIND11_MODULE(core, m) {
   m.add_object("_cleanup",
                py::capsule([]() { ScopePool::Instance().Clear(); }));
 
+  m.def("get_mem_usage", [](int device) {
+    return memory::allocation::GPUMemMonitor.GetMemUsage(device);
+  });
+
+  m.def("print_mem_usage",
+        []() { return memory::allocation::GPUMemMonitor.PrintMemUsage(); });
+
   py::class_<imperative::VarBase>(m, "VarBase", R"DOC()DOC")
       // .def(py::init<>())
       .def(py::init<bool>(), py::arg("stop_gradient") = false)

python/paddle/fluid/contrib/__init__.py

@@ -22,6 +22,8 @@ from . import op_frequence
 from .op_frequence import *
 from . import quantize
 from .quantize import *
+from . import int8_inference
+from .int8_inference import *
 from . import reader
 from .reader import *
 from . import slim
@@ -34,6 +36,7 @@ __all__ += decoder.__all__
 __all__ += memory_usage_calc.__all__
 __all__ += op_frequence.__all__
 __all__ += quantize.__all__
+__all__ += int8_inference.__all__
 __all__ += reader.__all__
 __all__ += slim.__all__
 __all__ += utils.__all__

python/paddle/fluid/contrib/int8_inference/README.md

+# Offline INT8 Calibration Tool
+
+PaddlePaddle supports offline INT8 calibration to accelerate the inference speed. In this document, we provide the instructions on how to enable INT8 calibration and show the ResNet-50 and MobileNet-V1 results in accuracy.
+
+## 0. Prerequisite
+You need to install at least PaddlePaddle-1.3 python package `pip install paddlepaddle==1.3`.
+
+## 1. How to generate INT8 model
+You can refer to the unit test in [test_calibration.py](../tests/test_calibration.py). Basically, there are three steps:
+* Construct calibration object.
+
+```python
+calibrator = int8_utility.Calibrator( # Step 1
+    program=infer_program, # required, FP32 program
+    pretrained_model=model_path, # required, FP32 pretrained model
+    algo=algo, # required, calibration algorithm; default is max, the alternative is KL (Kullback–Leibler divergence)
+    exe=exe, # required, executor
+    output=int8_model, # required, INT8 model
+    feed_var_names=feed_dict, # required, feed dict
+    fetch_list=fetch_targets) # required, fetch targets
+```
+
+* Call the calibrator.sample_data() after executor run.
+```python
+_, acc1, _ = exe.run(
+    program,
+    feed={feed_dict[0]: image,
+          feed_dict[1]: label},
+    fetch_list=fetch_targets)
+
+calibrator.sample_data() # Step 2
+```
+
+* Call the calibrator.save_int8_model() after sampling over specified iterations (e.g., iterations = 50)
+```python
+calibrator.save_int8_model() # Step 3
+```
+
+## 2. How to run INT8 model
+You can load INT8 model by load_inference_model [API](https://github.com/PaddlePaddle/Paddle/blob/8b50ad80ff6934512d3959947ac1e71ea3fb9ea3/python/paddle/fluid/io.py#L991) and run INT8 inference similar as [FP32](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/object_detection/eval.py "FP32").
+
+```python
+[infer_program, feed_dict,
+    fetch_targets] = fluid.io.load_inference_model(model_path, exe)
+```
+
+## 3. Result
+We provide the results of accuracy measurd on [Intel® Xeon® Platinum Gold Processor](https://ark.intel.com/products/120489/Intel-Xeon-Gold-6148-Processor-27-5M-Cache-2-40-GHz- "Intel® Xeon® Gold 6148 Processor") (also known as Intel® Xeon® Skylake6148).
+
+| Model  | Dataset  | FP32 Accuracy  | INT8 Accuracy  | Accuracy Diff  |
+| ------------ | ------------ | ------------ | ------------ | ------------ |
+| ResNet-50  | Small  | 72.00%  | 72.00%  |  0.00% |
+| MobileNet-V1  | Small  | 62.00%  | 62.00%  | 0.00%  |
+| ResNet-50  | Full ImageNet Val  |  76.63%  | 76.17%  | 0.46% |
+| MobileNet-V1 | Full ImageNet Val  | 70.78%  | 70.49%  | 0.29%  |
+
+Please note that [Small](http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz "Small") is a subset of [full ImageNet validation dataset](http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar "full ImageNet validation dataset"). 
+
+Notes:
+* The accuracy measurement requires the model with `label`.
+* The INT8 theoretical speedup is ~1.33X on Intel® Xeon® Skylake Server (please refer to `This allows for 4x more input at the cost of 3x more instructions or 33.33% more compute` in  [Reference](https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training "Reference")).
+
+## 4. How to reproduce the results
+* Small dataset
+```bash
+python python/paddle/fluid/contrib/tests/test_calibration.py
+```
+
+* Full dataset
+```bash
+DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py
+```

python/paddle/fluid/contrib/int8_inference/__init__.py

@@ -11,3 +11,10 @@
 # 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.
+
+from __future__ import print_function
+
+from . import utility
+from .utility import *
+
+__all__ = utility.__all__

python/paddle/fluid/contrib/int8_inference/utility.py

@@ -11,11 +11,15 @@
 # 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.
-import paddle.fluid.core as core
+
+from paddle.fluid import core
 import numpy as np
 import math
 import os
-import paddle.fluid as fluid
+from paddle.fluid.executor import global_scope
+from paddle.fluid import io
+
+__all__ = ['Calibrator']
 
 
 class Calibrator(object):
@@ -76,8 +80,7 @@ class Calibrator(object):
         '''
         for i in self.sampling_program.list_vars():
             if i.name in self.sampling_vars:
-                np_data = np.array(fluid.global_scope().find_var(i.name)
-                                   .get_tensor())
+                np_data = np.array(global_scope().find_var(i.name).get_tensor())
                 if i.name not in self._sampling_data:
                     self._sampling_data[i.name] = []
                 self._sampling_data[i.name].append(np_data)
@@ -86,9 +89,9 @@ class Calibrator(object):
         '''
         Save the quantized model to the disk.
         '''
-        fluid.io.save_inference_model(self.output, self.feed_var_names,
-                                      self.fetch_list, self.exe,
-                                      self.sampling_program)
+        io.save_inference_model(self.output, self.feed_var_names,
+                                self.fetch_list, self.exe,
+                                self.sampling_program)
 
     def __display_debug(self):
         if self.debug:

python/paddle/fluid/contrib/tests/test_calibration.py

@@ -24,8 +24,7 @@ import contextlib
 from paddle.dataset.common import download
 from PIL import Image, ImageEnhance
 import math
-sys.path.append('..')
-import int8_inference.utility as int8_utility
+import paddle.fluid.contrib.int8_inference.utility as int8_utility
 
 random.seed(0)
 np.random.seed(0)

python/paddle/fluid/io.py

@@ -21,9 +21,10 @@ import shutil
 import six
 from functools import reduce
 
+from paddle.fluid import layers
 from paddle.fluid.executor import Executor
 from paddle.fluid.evaluator import Evaluator
-from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable
+from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable, program_guard
 from . import core
 
 __all__ = [
@@ -931,6 +932,17 @@ def save_inference_model(dirname,
     if main_program is None:
         main_program = default_main_program()
 
+    # fix the bug that the activation op's output as target will be pruned.
+    # will affect the inference performance.
+    # TODO(Superjomn) add an IR pass to remove 1-scale op.
+    with program_guard(main_program):
+        uniq_target_vars = []
+        for var in target_vars:
+            if isinstance(var, Variable):
+                var1 = layers.scale(var, 1.)
+            uniq_target_vars.append(var1)
+        target_vars = uniq_target_vars
+
     # when a pserver and a trainer running on the same machine, mkdir may conflict
     try:
         os.makedirs(dirname)

python/paddle/fluid/layers/detection.py

@@ -49,6 +49,7 @@ __all__ = [
     'box_coder',
     'polygon_box_transform',
     'yolov3_loss',
+    'box_clip',
     'multiclass_nms',
 ]
 
@@ -508,13 +509,10 @@ def yolov3_loss(x,
                 gtbox,
                 gtlabel,
                 anchors,
+                anchor_mask,
                 class_num,
                 ignore_thresh,
-                loss_weight_xy=None,
-                loss_weight_wh=None,
-                loss_weight_conf_target=None,
-                loss_weight_conf_notarget=None,
-                loss_weight_class=None,
+                downsample_ratio,
                 name=None):
     """
     ${comment}
@@ -526,16 +524,13 @@ def yolov3_loss(x,
                           and x, y, w, h should be relative value of input image.
                           N is the batch number and B is the max box number in 
                           an image.
-        gtlabel (Variable): class id of ground truth boxes, shoud be ins shape
+        gtlabel (Variable): class id of ground truth boxes, shoud be in shape
                             of [N, B].
         anchors (list|tuple): ${anchors_comment}
+        anchor_mask (list|tuple): ${anchor_mask_comment}
         class_num (int): ${class_num_comment}
         ignore_thresh (float): ${ignore_thresh_comment}
-        loss_weight_xy (float|None): ${loss_weight_xy_comment}
-        loss_weight_wh (float|None): ${loss_weight_wh_comment}
-        loss_weight_conf_target (float|None): ${loss_weight_conf_target_comment}
-        loss_weight_conf_notarget (float|None): ${loss_weight_conf_notarget_comment}
-        loss_weight_class (float|None): ${loss_weight_class_comment}
+        downsample_ratio (int): ${downsample_ratio_comment}
         name (string): the name of yolov3 loss
 
     Returns:
@@ -555,9 +550,10 @@ def yolov3_loss(x,
         x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32')
         gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32')
         gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32')
-        anchors = [10, 13, 16, 30, 33, 23]
-        loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80
-                                        anchors=anchors, ignore_thresh=0.5)
+        anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]
+        anchors = [0, 1, 2]
+        loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80, anchors=anchors, 
+                                        ignore_thresh=0.5, downsample_ratio=32)
     """
     helper = LayerHelper('yolov3_loss', **locals())
 
@@ -569,6 +565,8 @@ def yolov3_loss(x,
         raise TypeError("Input gtlabel of yolov3_loss must be Variable")
     if not isinstance(anchors, list) and not isinstance(anchors, tuple):
         raise TypeError("Attr anchors of yolov3_loss must be list or tuple")
+    if not isinstance(anchor_mask, list) and not isinstance(anchor_mask, tuple):
+        raise TypeError("Attr anchor_mask of yolov3_loss must be list or tuple")
     if not isinstance(class_num, int):
         raise TypeError("Attr class_num of yolov3_loss must be an integer")
     if not isinstance(ignore_thresh, float):
@@ -581,31 +579,29 @@ def yolov3_loss(x,
         loss = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
 
+    objectness_mask = helper.create_variable_for_type_inference(dtype='int32')
+    gt_match_mask = helper.create_variable_for_type_inference(dtype='int32')
+
     attrs = {
         "anchors": anchors,
+        "anchor_mask": anchor_mask,
         "class_num": class_num,
         "ignore_thresh": ignore_thresh,
+        "downsample_ratio": downsample_ratio,
     }
 
-    if loss_weight_xy is not None and isinstance(loss_weight_xy, float):
-        self.attrs['loss_weight_xy'] = loss_weight_xy
-    if loss_weight_wh is not None and isinstance(loss_weight_wh, float):
-        self.attrs['loss_weight_wh'] = loss_weight_wh
-    if loss_weight_conf_target is not None and isinstance(
-            loss_weight_conf_target, float):
-        self.attrs['loss_weight_conf_target'] = loss_weight_conf_target
-    if loss_weight_conf_notarget is not None and isinstance(
-            loss_weight_conf_notarget, float):
-        self.attrs['loss_weight_conf_notarget'] = loss_weight_conf_notarget
-    if loss_weight_class is not None and isinstance(loss_weight_class, float):
-        self.attrs['loss_weight_class'] = loss_weight_class
-
     helper.append_op(
         type='yolov3_loss',
-        inputs={"X": x,
-                "GTBox": gtbox,
-                "GTLabel": gtlabel},
-        outputs={'Loss': loss},
+        inputs={
+            "X": x,
+            "GTBox": gtbox,
+            "GTLabel": gtlabel,
+        },
+        outputs={
+            'Loss': loss,
+            'ObjectnessMask': objectness_mask,
+            'GTMatchMask': gt_match_mask
+        },
         attrs=attrs)
     return loss
 
@@ -2060,6 +2056,54 @@ def generate_proposals(scores,
     return rpn_rois, rpn_roi_probs
 
 
+def box_clip(input, im_info, name=None):
+    """
+    Clip the box into the size given by im_info
+    For each input box, The formula is given as follows:
+        
+    .. code-block:: text
+
+        xmin = max(min(xmin, im_w - 1), 0)
+        ymin = max(min(ymin, im_h - 1), 0) 
+        xmax = max(min(xmax, im_w - 1), 0)
+        ymax = max(min(ymax, im_h - 1), 0)
+    
+    where im_w and im_h are computed from im_info:
+ 
+    .. code-block:: text
+
+        im_h = round(height / scale)
+        im_w = round(weight / scale)
+
+    Args:
+        input(variable): The input box, the last dimension is 4.
+        im_info(variable): The information of image with shape [N, 3] with 
+                            layout (height, width, scale). height and width
+                            is the input size and scale is the ratio of input
+                            size and original size.
+        name (str): The name of this layer. It is optional.
+    
+    Returns:
+        Variable: The cliped tensor variable.
+        
+    Examples:
+        .. code-block:: python
+        
+            boxes = fluid.layers.data(
+                name='data', shape=[8, 4], dtype='float32', lod_level=1)
+            im_info = fluid.layers.data(name='im_info', shape=[3])
+            out = fluid.layers.box_clip(
+                input=boxes, im_info=im_info, inplace=True)
+    """
+
+    helper = LayerHelper("box_clip", **locals())
+    output = helper.create_variable_for_type_inference(dtype=input.dtype)
+    inputs = {"Input": input, "ImInfo": im_info}
+    helper.append_op(type="box_clip", inputs=inputs, outputs={"Output": output})
+
+    return output
+
+
 def multiclass_nms(bboxes,
                    scores,
                    score_threshold,
@@ -2137,9 +2181,11 @@ def multiclass_nms(bboxes,
              (After version 1.3, when no boxes detected, the lod is changed 
              from {0} to {1}) 
 
+
     Examples:
         .. code-block:: python
 
+
             boxes = fluid.layers.data(name='bboxes', shape=[81, 4],
                                       dtype='float32', lod_level=1)
             scores = fluid.layers.data(name='scores', shape=[81],

python/paddle/fluid/tests/test_detection.py

@@ -476,12 +476,23 @@ class TestYoloDetection(unittest.TestCase):
             x = layers.data(name='x', shape=[30, 7, 7], dtype='float32')
             gtbox = layers.data(name='gtbox', shape=[10, 4], dtype='float32')
             gtlabel = layers.data(name='gtlabel', shape=[10], dtype='int32')
-            loss = layers.yolov3_loss(x, gtbox, gtlabel, [10, 13, 30, 13], 10,
-                                      0.5)
+            loss = layers.yolov3_loss(x, gtbox, gtlabel, [10, 13, 30, 13],
+                                      [0, 1], 10, 0.7, 32)
 
             self.assertIsNotNone(loss)
 
 
+class TestBoxClip(unittest.TestCase):
+    def test_box_clip(self):
+        program = Program()
+        with program_guard(program):
+            input_box = layers.data(
+                name='input_box', shape=[7, 4], dtype='float32', lod_level=1)
+            im_info = layers.data(name='im_info', shape=[3], dtype='float32')
+            out = layers.box_clip(input_box, im_info)
+            self.assertIsNotNone(out)
+
+
 class TestMulticlassNMS(unittest.TestCase):
     def test_multiclass_nms(self):
         program = Program()

python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py

+# 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.
+
+from __future__ import print_function
+
+import unittest
+from paddle.fluid.tests.unittests.test_conv2d_op import *
+
+
+class TestNGRAPH(TestConv2dOp):
+    def init_kernel_type(self):
+        super(TestNGRAPH, self).init_kernel_type()
+
+
+class TestNGRAPHWithPad(TestWithPad):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithPad, self).init_kernel_type()
+
+
+class TestNGRAPHWithStride(TestWithStride):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithStride, self).init_kernel_type()
+
+
+class TestNGRAPHWithGroup(TestWithGroup):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithGroup, self).init_kernel_type()
+
+
+class TestNGRAPHWith1x1(TestWith1x1):
+    def init_kernel_type(self):
+        super(TestNGRAPHWith1x1, self).init_kernel_type()
+
+
+class TestNGRAPHWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithInput1x1Filter1x1, self).init_kernel_type()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py

+#   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.
+
+from __future__ import print_function
+
+from paddle.fluid.tests.unittests.test_pool2d_op import *
+
+
+class TestNGRAPHPool2D_Op(TestPool2D_Op):
+    def init_test_case(self):
+        super(TestNGRAPHPool2D_Op, self).init_test_case()
+
+
+class TestNGRAPHCase1(TestCase1):
+    def init_test_case(self):
+        super(TestNGRAPHCase1, self).init_test_case()
+
+
+class TestNGRAPHCase2(TestCase2):
+    def init_test_case(self):
+        super(TestNGRAPHCase2, self).init_test_case()
+
+
+class TestNGRAPHCase3(TestCase3):
+    def init_pool_type(self):
+        super(TestNGRAPHCase3, self).init_pool_type()
+
+
+class TestNGRAPHCase4(TestCase4):
+    def init_pool_type(self):
+        super(TestNGRAPHCase4, self).init_pool_type()
+
+
+class TestNGRAPHCase5(TestCase5):
+    def init_pool_type(self):
+        super(TestNGRAPHCase5, self).init_pool_type()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_box_clip_op.py

+#   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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+import copy
+
+
+def box_clip(input_box, im_info, output_box):
+    im_w = round(im_info[1] / im_info[2])
+    im_h = round(im_info[0] / im_info[2])
+    output_box[:, :, 0] = np.maximum(
+        np.minimum(input_box[:, :, 0], im_w - 1), 0)
+    output_box[:, :, 1] = np.maximum(
+        np.minimum(input_box[:, :, 1], im_h - 1), 0)
+    output_box[:, :, 2] = np.maximum(
+        np.minimum(input_box[:, :, 2], im_w - 1), 0)
+    output_box[:, :, 3] = np.maximum(
+        np.minimum(input_box[:, :, 3], im_h - 1), 0)
+
+
+def batch_box_clip(input_boxes, im_info, lod):
+    n = input_boxes.shape[0]
+    m = input_boxes.shape[1]
+    output_boxes = np.zeros((n, m, 4), dtype=np.float32)
+    cur_offset = 0
+    for i in range(len(lod)):
+        box_clip(input_boxes[cur_offset:(cur_offset + lod[i]), :, :],
+                 im_info[i, :],
+                 output_boxes[cur_offset:(cur_offset + lod[i]), :, :])
+        cur_offset += lod[i]
+    return output_boxes
+
+
+class TestBoxClipOp(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_clip"
+        lod = [[1, 2, 3]]
+        input_boxes = np.random.random((6, 10, 4)) * 5
+        im_info = np.array([[5, 8, 1.], [6, 6, 1.], [7, 5, 1.]])
+        output_boxes = batch_box_clip(input_boxes, im_info, lod[0])
+
+        self.inputs = {
+            'Input': (input_boxes.astype('float32'), lod),
+            'ImInfo': im_info.astype('float32'),
+        }
+        self.outputs = {'Output': output_boxes}
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py

@@ -268,9 +268,6 @@ class TestImperativePtbRnn(unittest.TestCase):
                 sgd.minimize(dy_loss)
                 for param in ptb_model.parameters():
                     dy_param_updated[param.name] = param._numpy()
-                # print("dy_loss is {}".format(dy_loss._numpy()))
-                # print("last_hidden is {}".format(last_hidden._numpy()))
-                # print("last_cell is {}".format(last_cell._numpy()))
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed

python/paddle/fluid/tests/unittests/test_inference_model_io.py

@@ -82,7 +82,8 @@ class TestBook(unittest.TestCase):
 
         self.assertEqual(feed_var_names, ["x", "y"])
         self.assertEqual(len(fetch_vars), 1)
-        self.assertEqual(str(fetch_vars[0]), str(avg_cost))
+        print("fetch %s" % str(fetch_vars[0]))
+        self.assertTrue("scale" in str(fetch_vars[0]))
         self.assertEqual(expected, actual)
 
 

python/paddle/fluid/transpiler/details/__init__.py

@@ -17,3 +17,4 @@ from __future__ import print_function
 from .program_utils import *
 from .ufind import *
 from .checkport import *
+from .vars_distributed import *

python/setup.py.in

@@ -109,6 +109,7 @@ packages=['paddle',
           'paddle.fluid.contrib',
           'paddle.fluid.contrib.decoder',
           'paddle.fluid.contrib.quantize',
+          'paddle.fluid.contrib.int8_inference',
           'paddle.fluid.contrib.reader',
           'paddle.fluid.contrib.slim',
           'paddle.fluid.contrib.slim.core',