Merge remote-tracking branch 'upstream/develop' into develop

CMakeLists.txt

@@ -2,9 +2,9 @@ cmake_minimum_required(VERSION 3.0)
 project(paddle-mobile)
 
 # select the platform to build
-option(CPU "armv7 with neon support" OFF)
+option(CPU "armv7 with neon support" ON)
 option(MALI_GPU "mali gpu support" OFF)
-option(FPGA "fpga support" ON)
+option(FPGA "fpga support" OFF)
 
 option(USE_OPENMP "openmp support" OFF)
 option(DEBUGING "enable debug mode" ON)

doc/development_fpga.md

+# FPGA开发文档
+
+FPGA平台的代码在Xilinx ZCU102 revision 1.0开发板测试Resnet50成功，预测结果正确。
+
+## 准备硬件
+___
+
+1. 购买Xilinx ZCU102 revision1.0 开发板
+2. 另外下载Xilinx ZCU102 Ubuntu[镜像文件](https://www.xilinx.com/member/forms/download/xef.html?filename=Ubuntu_Desktop_Release_2018_1.zip)，并烧录进SD卡。
+ * Windowns系统可使用Win32DiskImager
+ * Linux系统使用dd命令：dd if=name.img of=/dev/sdb
+2. 将SD卡插入电脑，替换分区1中已有的BOOT.BIN、image.ub为[BOOT.BIN、image.ub](http://mms-graph.bj.bcebos.com/paddle-mobile/fpga/files.tar.gz)
+3. 将SD卡插入ZCU102开发板，设置板拨码开关为SD卡启动，上电启动Linux系统.
+3. 装载驱动：sudo insmod [fpgadrv.ko](http://mms-graph.bj.bcebos.com/paddle-mobile/fpga/files.tar.gz)
+
+
+## 编译工程
+___
+1. 将最新的paddle mobile 代码复制到ZCU102开发板中。
+2. 进入paddle-mobile根目录， CMakeLists.txt 设置平台为 option(FPGA "fpga support" ON)。CPU和MALI\_GPU选项设置为OFF。
+2. 执行以下命令，可在./test/build下生成test-resnet50可执行程序。
+    * mkdir build
+    * cd build
+    * cmake ..
+    * make
+
+## 准备模型和数据
+___
+1. 模型文件放在./test/models/resnet50中。将[\_\_model\_\_](http://mms-graph.bj.bcebos.com/paddle-mobile/fpga/files.tar.gz)文件复制到此文件夹下。
+2. 另外下载模型[权重文件](http://paddle-imagenet-models.bj.bcebos.com/resnet_50_model.tar),解压后也放在./test/models/resnet50 中。
+3. 将数据文件[image_src_float](http://mms-graph.bj.bcebos.com/paddle-mobile/fpga/files.tar.gz)复制到/test/images下。此数据文件对应着标准数据集中的ILSVRC2012_val_00000885.JPEG，分类标签为80， 对应着"black grouse".
+
+## 运行程序
+___
+1. 进入./test/build目录。
+2. sudo ./test-resnet50
+3. 如果于DEBUG选项是否打开，屏幕会输出很多中间打印信息。最终打印出预测分类结果为80。

src/common/types.cpp

@@ -22,6 +22,7 @@ const char *G_OP_TYPE_BATCHNORM = "batch_norm";
 const char *G_OP_TYPE_BOX_CODER = "box_coder";
 const char *G_OP_TYPE_CONCAT = "concat";
 const char *G_OP_TYPE_ELEMENTWISE_ADD = "elementwise_add";
+const char *G_OP_TYPE_FILL_CONSTANT = "fill_constant";
 const char *G_OP_TYPE_FUSION_CONV_ADD_RELU = "fusion_conv_add_relu";
 const char *G_OP_TYPE_FUSION_CONV_ADD_PRELU = "fusion_conv_add_prelu";
 const char *G_OP_TYPE_FUSION_CONV_ADD_ADD_PRELU = "fusion_conv_add_add_prelu";
@@ -99,6 +100,7 @@ std::unordered_map<
         {G_OP_TYPE_FC, {{"X", "Y", "Z"}, {"Out"}}},
         {G_OP_TYPE_RESHAPE, {{"X"}, {"Out"}}},
         {G_OP_TYPE_DEPTHWISE_CONV, {{"Input"}, {"Output"}}},
+        {G_OP_TYPE_FILL_CONSTANT, {{}, {"Out"}}},
         {G_OP_TYPE_FUSION_CONV_ADD_RELU, {{"Input"}, {"Out"}}},
         {G_OP_TYPE_FUSION_CONV_ADD_PRELU, {{"Input"}, {"Out"}}},
         {G_OP_TYPE_FUSION_CONV_ADD_ADD_PRELU, {{"Input"}, {"Out"}}},

src/common/variant.h

@@ -12,14 +12,16 @@ 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 <cstdlib>
 #include <cstring>
+#include <string>
 #include "common/enforce.h"
 #include "common/log.h"
 
-#pragma once
-
 namespace paddle_mobile {
+
 template <int ID, typename Type>
 struct IDToType {
   typedef Type type_t;

src/fpga/bias_scale.cpp

@@ -27,9 +27,6 @@ void align_element(float **data_in, int num_per_div_before_alignment, int num) {
       (num + num_per_div_before_alignment - 1) / num_per_div_before_alignment;
   int num_per_div_after_alignment =
       align_to_x(num_per_div_before_alignment, BS_NUM_ALIGNMENT);
-  if (num_per_div_before_alignment == num_per_div_after_alignment) {
-    return;
-  }
   int num_element =
       2 * div_num * num_per_div_after_alignment;  // including bias & scale
   float *ptr_aligned =

src/framework/attribute.h

@@ -156,7 +156,7 @@ class AttrReader {
   template <typename T>
   inline T Get(const string &name) const {
     PADDLE_MOBILE_ENFORCE(attrs_.count(name) != 0,
-                          "%s should  be in AttributeMap", name);
+                          "%s should  be in AttributeMap", name.c_str());
     return ((Attribute)attrs_.at(name)).Get<T>();
   }
 

src/framework/data_type.cpp

+/* 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 "framework/data_type.h"
+#include <stdint.h>
+#include <string>
+#include <unordered_map>
+
+namespace paddle_mobile {
+namespace framework {
+
+struct DataTypeMap {
+  std::unordered_map<std::type_index,
+                     _PaddleMobile__Framework__Proto__VarType__Type>
+      cpp_to_proto_;
+  std::unordered_map<int, std::type_index> proto_to_cpp_;
+  std::unordered_map<int, std::string> proto_to_str_;
+  std::unordered_map<std::type_index, size_t> cpp_to_size_;
+};
+
+static DataTypeMap* InitDataTypeMap();
+// C++11 removes the need for manual locking. Concurrent execution shall wait if
+// a static local variable is already being initialized.
+// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex
+static DataTypeMap& gDataTypeMap() {
+  static DataTypeMap* g_data_type_map_ = InitDataTypeMap();
+  return *g_data_type_map_;
+}
+
+template <typename T>
+static inline void RegisterType(
+    DataTypeMap* map, _PaddleMobile__Framework__Proto__VarType__Type proto_type,
+    const std::string& name) {
+  map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
+  map->cpp_to_proto_.emplace(typeid(T), proto_type);
+  map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
+  map->cpp_to_size_.emplace(typeid(T), sizeof(T));
+}
+
+static DataTypeMap* InitDataTypeMap() {
+  auto retv = new DataTypeMap();
+
+#define RegType(cc_type, proto_type) \
+  RegisterType<cc_type>(retv, proto_type, #cc_type)
+
+  // NOTE: Add your customize type here.
+  // RegType(float16, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP16);
+  RegType(float, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP32);
+  RegType(double, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP64);
+  RegType(int, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT32);
+  RegType(int64_t, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT64);
+  RegType(bool, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__BOOL);
+  RegType(size_t, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__SIZE_T);
+  RegType(int16_t, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT16);
+  RegType(uint8_t, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__UINT8);
+  RegType(int8_t, PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT8);
+
+#undef RegType
+  return retv;
+}
+
+_PaddleMobile__Framework__Proto__VarType__Type ToDataType(
+    std::type_index type) {
+  auto it = gDataTypeMap().cpp_to_proto_.find(type);
+  if (it != gDataTypeMap().cpp_to_proto_.end()) {
+    return it->second;
+  }
+  PADDLE_MOBILE_THROW_EXCEPTION("Not support %s as tensor type", type.name());
+}
+
+std::type_index ToTypeIndex(
+    _PaddleMobile__Framework__Proto__VarType__Type type) {
+  auto it = gDataTypeMap().proto_to_cpp_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_cpp_.end()) {
+    return it->second;
+  }
+  PADDLE_MOBILE_THROW_EXCEPTION(
+      "Not support _PaddleMobile__Framework__Proto__VarType__Type(%d) as "
+      "tensor type",
+      static_cast<int>(type));
+}
+
+std::string DataTypeToString(
+    const _PaddleMobile__Framework__Proto__VarType__Type type) {
+  auto it = gDataTypeMap().proto_to_str_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_str_.end()) {
+    return it->second;
+  }
+  PADDLE_MOBILE_THROW_EXCEPTION(
+      "Not support _PaddleMobile__Framework__Proto__VarType__Type(%d) as "
+      "tensor type",
+      static_cast<int>(type));
+}
+
+}  // namespace framework
+}  // namespace paddle_mobile

src/framework/data_type.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 <typeindex>
+#include "common/enforce.h"
+#include "framework/framework.pb-c.h"
+
+namespace paddle_mobile {
+
+namespace framework {
+
+extern _PaddleMobile__Framework__Proto__VarType__Type ToDataType(
+    std::type_index type);
+extern std::type_index ToTypeIndex(
+    _PaddleMobile__Framework__Proto__VarType__Type type);
+
+template <typename Visitor>
+inline void VisitDataType(_PaddleMobile__Framework__Proto__VarType__Type type,
+                          Visitor visitor) {
+  switch (type) {
+    // case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP16:
+    //   visitor.template apply<float16>();
+    //   break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP32:
+      visitor.template apply<float>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP64:
+      visitor.template apply<double>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT32:
+      visitor.template apply<int>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT64:
+      visitor.template apply<int64_t>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__BOOL:
+      visitor.template apply<bool>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__UINT8:
+      visitor.template apply<uint8_t>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT16:
+      visitor.template apply<int16_t>();
+      break;
+    case PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT8:
+      visitor.template apply<int8_t>();
+      break;
+    default:
+      PADDLE_MOBILE_THROW_EXCEPTION("Not supported %d", type);
+  }
+}
+
+extern std::string DataTypeToString(
+    const _PaddleMobile__Framework__Proto__VarType__Type type);
+inline std::ostream& operator<<(
+    std::ostream& out,
+    const _PaddleMobile__Framework__Proto__VarType__Type& type) {
+  out << DataTypeToString(type);
+  return out;
+}
+
+}  // namespace framework
+}  // namespace paddle_mobile

src/framework/load_ops.h

@@ -64,6 +64,9 @@ limitations under the License. */
 // load requared ops
 LOAD_OP(feed)
 LOAD_OP(fetch)
+#ifdef FILL_CONSTANT_OP
+LOAD_OP(fill_constant)
+#endif
 #ifdef BATCHNORM_OP
 LOAD_OP2(batch_norm, CPU, MALI_GPU);
 #endif

src/framework/selected_rows.h

@@ -18,9 +18,9 @@ limitations under the License. */
 #include <vector>
 
 #include "framework/lod_tensor.h"
+#include "framework/mixed_vector.h"
 #include "framework/tensor.h"
 #include "memory/t_malloc.h"
-#include "mixed_vector.h"
 
 namespace paddle_mobile {
 namespace framework {

src/framework/tensor.h

@@ -343,7 +343,9 @@ inline Print &operator<<(Print &printer, const Tensor &tensor) {
     } else if (tensor.type() == typeid(int64_t)) {
       printer << tensor.data<int64_t>()[i] << " ";
     } else if (tensor.type() == typeid(int8_t)) {
-      printer << static_cast<int32_t>(tensor.data<int8_t>()[i]) << " ";
+      printer << static_cast<int>(tensor.data<int8_t>()[i]) << " ";
+    } else if (tensor.type() == typeid(int32_t)) {
+      printer << tensor.data<int32_t>()[i] << " ";
     }
   }
 #endif

src/io/executor.cpp

@@ -80,12 +80,13 @@ Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
 }
 
 template <typename Dtype>
-void LoadMemInternal(void **data, framework::LoDTensor *tensor) {
+static void LoadMemInternal(void **data, framework::LoDTensor *tensor,
+                            bool quant_uint8 = false) {
   char **data_buf = reinterpret_cast<char **>(data);
   int64_t size = tensor->numel();
   Dtype *tensor_data = tensor->mutable_data<Dtype>();
-  if (0) {
-    // TODO(hjchen2) should be moved into operator init function
+  if (quant_uint8) {
+    // should be moved into operator init function
     float min_value;
     float max_value;
     memcpy(&min_value, data_buf, sizeof(float));
@@ -141,7 +142,8 @@ void Executor<Dtype, P>::LoadMemory(
   // parse tensor from stream
   switch (tensor_desc.DataType()) {
     case framework::VARTYPE_TYPE_FP32:
-      LoadMemInternal<float>(reinterpret_cast<void **>(data_buf), tensor);
+      LoadMemInternal<float>(reinterpret_cast<void **>(data_buf), tensor,
+                             program_.quantification);
       break;
     case framework::VARTYPE_TYPE_INT8:
       LoadMemInternal<int8_t>(reinterpret_cast<void **>(data_buf), tensor);

src/operators/dequantize_op.cpp

@@ -12,6 +12,8 @@ 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. */
 
+#ifdef DEQUANT_OP
+
 #include "operators/dequantize_op.h"
 
 namespace paddle_mobile {
@@ -30,3 +32,5 @@ namespace ops = paddle_mobile::operators;
 #ifdef PADDLE_MOBILE_CPU
 REGISTER_OPERATOR_CPU(dequantize, ops::DequantizeOp);
 #endif
+
+#endif

src/operators/dequantize_op.h

@@ -12,6 +12,8 @@ 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. */
 
+#ifdef DEQUANT_OP
+
 #pragma once
 
 #include <string>
@@ -41,3 +43,5 @@ class DequantizeOp
 
 }  // namespace operators
 }  // namespace paddle_mobile
+
+#endif

src/operators/elementwise_mul_op.cpp

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #ifdef ELEMENTWISEMUL_OP
 
-#include "elementwise_mul_op.h"
+#include "operators/elementwise_mul_op.h"
 
 namespace paddle_mobile {
 namespace operators {

src/operators/fill_constant_op.cpp

+/* 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. */
+
+#ifdef FILL_CONSTANT_OP
+
+#include "operators/fill_constant_op.h"
+
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_CPU
+REGISTER_OPERATOR_CPU(fill_constant, ops::FillConstantOp);
+#endif
+#ifdef PADDLE_MOBILE_MALI_GPU
+REGISTER_OPERATOR_MALI_GPU(fill_constant, ops::FillConstantOp);
+#endif
+#ifdef PADDLE_MOBILE_FPGA
+REGISTER_OPERATOR_FPGA(fill_constant, ops::FillConstantOp);
+#endif
+
+#endif

src/operators/fill_constant_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. */
+
+#ifdef FILL_CONSTANT_OP
+
+#pragma once
+
+#include <string>
+#include "framework/data_type.h"
+#include "framework/operator.h"
+#include "framework/selected_rows.h"
+#include "operators/math/math_function.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+using std::string;
+
+template <typename DeviceType, typename T>
+class FillConstantOp : public framework::OperatorBase<DeviceType> {
+ public:
+  FillConstantOp(const string &type, const VariableNameMap &inputs,
+                 const VariableNameMap &outputs,
+                 const framework::AttributeMap attrs,
+                 std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorBase<DeviceType>(type, inputs, outputs, attrs,
+                                            scope),
+        param_(inputs, outputs, attrs, *scope) {}
+  void RunImpl() const {
+    auto data_type =
+        static_cast<_PaddleMobile__Framework__Proto__VarType__Type>(
+            param_.DataDtype());
+    framework::Tensor *tensor = nullptr;
+    auto value = param_.Value();
+    auto *outvar = param_.OutVar();
+
+    if (outvar->template IsType<framework::LoDTensor>()) {
+      tensor = outvar->template GetMutable<framework::LoDTensor>();
+    } else if (outvar->template IsType<framework::SelectedRows>()) {
+      tensor = outvar->template GetMutable<framework::SelectedRows>()
+                   ->mutable_value();
+    } else {
+      PADDLE_MOBILE_THROW_EXCEPTION(
+          "fill constant op's output only"
+          "supports SelectedRows and LoDTensor");
+    }
+    tensor->Resize(framework::make_ddim(param_.Shape()));
+    tensor->mutable_data(framework::ToTypeIndex(data_type));
+
+    math::set_constant(tensor, value);
+  }
+
+  void Init() {}
+
+  void InferShape() const {
+    PADDLE_MOBILE_ENFORCE(
+        param_.Out() != nullptr,
+        "Output (Out) of fill_constant op should not be null.");
+    framework::DDim ddim = framework::make_ddim(param_.Shape());
+    param_.Out()->Resize(ddim);
+  }
+
+ protected:
+  FillConstantParam<DeviceType> param_;
+};
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/kernel/arm/dequantize_kernel.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_MOBILE_CPU
+#ifdef DEQUANT_OP
 
 #include "operators/kernel/dequantize_kernel.h"
 
@@ -38,7 +38,8 @@ void DequantizeKernel<CPU, float>::Compute(
   const int32_t *x = input->data<const int32_t>();
   float *y = output->mutable_data<float>();
   size_t size = output->numel();
-  float scale = 1.f / (activation_scale * weight_scale);
+  // float scale = 1.f / (activation_scale * weight_scale);
+  float scale = activation_scale / weight_scale;
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;

src/operators/kernel/arm/quantize_kernel.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_MOBILE_CPU
+#ifdef QUANT_OP
 
 #include "operators/kernel/quantize_kernel.h"
 #include <cmath>
@@ -225,7 +225,7 @@ static void quantize_round_to_nearest(const Tensor *input, const float scale,
   const float *x = input->data<const float>();
   int8_t *y = output->mutable_data<int8_t>();
   size_t size = input->numel();
-#ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
   for (size_t i = 0; i < loop; ++i) {
@@ -280,17 +280,18 @@ void QuantizeKernel<CPU, float>::Compute(
   }
   max_abs = std::max(max_abs, 1e-6f);
   // only support int8 currently
-  float online_scale = 127 / max_abs;
-  param.online_scale_->mutable_data<float>()[0] = online_scale;
+  float scale = 127 / max_abs;
+  param.online_scale_->mutable_data<float>()[0] = max_abs;
   switch (param.round_type_) {
     case ROUND_NEAREST_TO_EVEN:
-      quantize_round_to_even(input, online_scale, output);
+      quantize_round_to_even(input, scale, output);
       break;
     case ROUND_NEAREST_TOWARDS_ZERO:
-      quantize_round_to_zero(input, online_scale, output);
+      quantize_round_to_zero(input, scale, output);
       break;
     case ROUND_NEAREST_AWAY_ZERO:
-      quantize_round_to_nearest(input, online_scale, output);
+      quantize_round_to_nearest(input, scale, output);
+      break;
     default:
       LOG(kLOG_ERROR) << "round type is not supported.";
       break;

src/operators/kernel/central-arm-func/conv_arm_func.h

@@ -16,24 +16,27 @@ limitations under the License. */
 
 #pragma once
 #include <vector>
+#include "operators/math/conv_arm_int8.h"
 #include "operators/math/conv_func.h"
 #include "operators/math/depthwise_conv_3x3.h"
 #include "operators/math/im2col.h"
 #include "operators/math/math_function.h"
+#include "operators/math/pad.h"
 #include "operators/math/vol2col.h"
 #include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {
+
+template <typename Dtype>
 inline void ConvBasic(const ConvParam<CPU> &param) {
   const Tensor *input = param.Input();
   Tensor filter = *param.Filter();
   Tensor *output = param.Output();
-  output->mutable_data<float>();
   int groups = param.Groups();
-  std::vector<int> strides = param.Strides();
-  std::vector<int> paddings = param.Paddings();
-  std::vector<int> dilations = param.Dilations();
+  const std::vector<int> strides = param.Strides();
+  const std::vector<int> paddings = param.Paddings();
+  const std::vector<int> dilations = param.Dilations();
 
   const int batch_size = static_cast<int>(input->dims()[0]);
 
@@ -57,7 +60,7 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
   Tensor col;
   Tensor col_matrix;
   if (is_expand) {
-    col.mutable_data<float>(col_shape);
+    col.mutable_data<Dtype>(col_shape);
     col_matrix.ShareDataWith(col);
     col_matrix.Resize(col_matrix_shape);
   }
@@ -76,8 +79,8 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
   int in_step = static_cast<int>(input->dims()[1]) / groups;
   int out_step = static_cast<int>(output->dims()[1]) / groups;
 
-  math::Vol2ColFunctor<CPU, float> vol2col;
-  math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
+  math::Vol2ColFunctor<CPU, Dtype> vol2col;
+  math::Im2ColFunctor<math::ColFormat::kCFO, CPU, Dtype> im2col;
 
   for (int i = 0; i < batch_size; i++) {
     Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
@@ -96,6 +99,7 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
                std::vector<int>{paddings[0], paddings[1], paddings[0],
                                 paddings[1]},
                &col);
+
       } else if (data_dim == 3U) {
         // vol2col
         vol2col(in_slice, dilations, strides, paddings, &col);
@@ -104,29 +108,85 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
       // gemm
       Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
       Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
-      math::matmul<float>(filter_slice, false, col_matrix, false,
+
+      math::matmul<Dtype>(filter_slice, false, col_matrix, false,
                           static_cast<float>(1), &out_slice,
                           static_cast<float>(0));
     }
   }
 }
 
+inline void ConvCompute_int8(const ConvParam<CPU> &param) {
+  typedef void (*ConvFunc)(const Tensor &input, const Tensor &kernel,
+                           Tensor *output);
+  static ConvFunc conv_funcs_table[7][5] = {
+      {0, 0, 0, 0, 0},                                // k = 1
+      {0, 0, 0, 0, 0}, {conv3x3s1_int8, 0, 0, 0, 0},  // k = 3
+      {0, 0, 0, 0, 0}, {conv5x5s1_int8, 0, 0, 0, 0},  // k = 5
+      {0, 0, 0, 0, 0}, {0, 0, 0, 0, 0},               // k = 7
+  };
+  const Tensor *input = param.Input();
+  Tensor *filter = param.Filter();
+  Tensor *output = param.Output();
+  int groups = param.Groups();
+  const std::vector<int> &strides = param.Strides();
+  const std::vector<int> &paddings = param.Paddings();
+  const std::vector<int> &dilations = param.Dilations();
+  int kernel_h = filter->dims()[2];
+  int kernel_w = filter->dims()[3];
+  output->mutable_data<int32_t>();
+
+  ConvFunc conv_func = 0;
+  if (strides[1] == strides[0] && strides[1] < 6 && kernel_h == kernel_w &&
+      kernel_h < 8 && groups == 1 && dilations[0] == dilations[1] &&
+      dilations[1] == 1) {
+    conv_func = conv_funcs_table[kernel_h - 1][strides[0] - 1];
+  }
+  if (conv_func) {
+    int batch_size = input->dims()[0];
+    math::PadFunctor<CPU, int8_t> pad;
+
+    Tensor input_pad;
+    for (int i = 0; i < batch_size; ++i) {
+      Tensor in_batch = input->Slice(i, i + 1);
+      Tensor out_batch = output->Slice(i, i + 1);
+      if (paddings[0] == 0 && paddings[1] == 0) {
+        input_pad = in_batch;
+      } else {
+        framework::DDim pad_shape = in_batch.dims();
+        pad_shape[2] += 2 * paddings[0];
+        pad_shape[3] += 2 * paddings[1];
+        input_pad.mutable_data<int8_t>(pad_shape);
+        pad(in_batch, paddings[0], paddings[1], &input_pad);
+      }
+      conv_func(input_pad, *filter, &out_batch);
+    }
+  } else {
+    ConvBasic<int8_t>(param);
+  }
+}
+
 template <typename P>
 void ConvCompute(const ConvParam<CPU> &param) {
-  if (param.Groups() == param.Input()->dims()[1] &&
-      param.Input()->dims()[1] == param.Output()->dims()[1] &&
-      param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
-      param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
-    math::DepthwiseConv3x3s1p1(param.Input(), param.Filter(), param.Output(),
-                               nullptr, false);
-  } else if (param.Groups() == param.Input()->dims()[1] &&
-             param.Input()->dims()[1] == param.Output()->dims()[1] &&
-             param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
-             param.Filter()->dims()[2] == 3) {
-    math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
-                           param.Filter(), nullptr, param.Output(), false);
+  if (param.Input()->type() == typeid(int8_t)) {
+    ConvCompute_int8(param);
   } else {
-    ConvBasic(param);
+    param.Output()->mutable_data<float>();
+    if (param.Groups() == param.Input()->dims()[1] &&
+        param.Input()->dims()[1] == param.Output()->dims()[1] &&
+        param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
+        param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
+      math::DepthwiseConv3x3s1p1(param.Input(), param.Filter(), param.Output(),
+                                 nullptr, false);
+    } else if (param.Groups() == param.Input()->dims()[1] &&
+               param.Input()->dims()[1] == param.Output()->dims()[1] &&
+               param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
+               param.Filter()->dims()[2] == 3) {
+      math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
+                             param.Filter(), nullptr, param.Output(), false);
+    } else {
+      ConvBasic<float>(param);
+    }
   }
 }
 

src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h

@@ -44,7 +44,7 @@ void DepthwiseConvCompute(const ConvParam<CPU> &param) {
                                  Bias, false);
 
   } else {
-    ConvBasic(param);
+    ConvBasic<float>(param);
   }
 }
 

src/operators/kernel/central-arm-func/elementwise_add_arm_func.h

@@ -15,8 +15,12 @@ limitations under the License. */
 #ifdef ELEMENTWISEADD_OP
 
 #pragma once
+
 #include "operators/math/elementwise_op_function.h"
 #include "operators/op_param.h"
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#include <arm_neon.h>
+#endif
 
 namespace paddle_mobile {
 namespace operators {
@@ -33,8 +37,61 @@ void ElementwiseAddCompute(const ElementwiseAddParam<CPU> &param) {
   Tensor *Out = param.Out();
   Out->mutable_data<float>();
   int axis = param.Axis();
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+  const auto &x_dims = input_x->dims();
+  const auto &y_dims = input_y->dims();
+  /// axis = -1 represent the last dimensions.
+  axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+  size_t batch = 1;
+  size_t channels = 1;
+  size_t elementwise_num = 1;
+  for (int i = 0; i < axis; ++i) {
+    batch *= x_dims[i];
+  }
+  for (int i = 0; i < y_dims.size(); ++i) {
+    channels *= y_dims[i];
+  }
+  for (int i = y_dims.size() + axis; i < x_dims.size(); ++i) {
+    elementwise_num *= x_dims[i];
+  }
+  const float *bias_data = input_y->data<float>();
+  const float *input_data = input_x->data<float>();
+  float *output_data = Out->mutable_data<float>();
+  for (int i = 0; i < batch; ++i) {
+    for (int j = 0; j < channels; ++j) {
+      size_t offset = (i * channels + j) * elementwise_num;
+      const float *input = input_data + offset;
+      const float *bias = bias_data + j;
+      float *output = output_data + offset;
+
+      int loop = elementwise_num >> 0x4;
+      int remain = elementwise_num & 0xF;
+      for (int k = 0; k < loop; ++k) {
+        float32x4_t rb = vdupq_n_f32(*bias);
+        float32x4_t r0 = vld1q_f32(input);
+        float32x4_t r1 = vld1q_f32(input + 4);
+        float32x4_t r2 = vld1q_f32(input + 8);
+        float32x4_t r3 = vld1q_f32(input + 12);
+        r0 = vaddq_f32(r0, rb);
+        r1 = vaddq_f32(r1, rb);
+        r2 = vaddq_f32(r2, rb);
+        r3 = vaddq_f32(r3, rb);
+        vst1q_f32(output, r0);
+        vst1q_f32(output + 4, r1);
+        vst1q_f32(output + 8, r2);
+        vst1q_f32(output + 12, r3);
+        input += 16;
+        output += 16;
+      }
+      for (int k = 0; k < remain; ++k) {
+        output[k] = input[k] + *bias;
+      }
+    }
+  }
+#else
   ElementwiseComputeEx<AddFunctor<float>, float>(input_x, input_y, axis,
                                                  AddFunctor<float>(), Out);
+#endif
 }
 
 template class ElementwiseAddKernel<CPU, float>;

src/operators/kernel/central-arm-func/relu_arm_func.h

@@ -17,6 +17,9 @@ limitations under the License. */
 
 #include <operators/math/transform.h>
 #include "operators/op_param.h"
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#include <arm_neon.h>
+#endif
 
 namespace paddle_mobile {
 namespace operators {
@@ -37,71 +40,100 @@ void ReluCompute(const ReluParam<CPU> &param) {
   auto *out_ptr = out->mutable_data<float>();
 
   int numel = input_x->numel();
-  //  if (numel > 64) {
-  //    asm volatile(
-  //        "pld        [%[input_x_ptr], #0]        \n\t"
-  //        "vmov.f32   q8,    #0.0                 \n\t"
-  //        "subs %[num], %[num], #32                \n\t"
-  //        "blt        end_num_%=                  \n\t"
-  //        "loop_num_%=:                           \n\t"
-  //        "pld        [%[input_x_ptr], #1024]      \n\t"
-  //
-  //        "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
-  //        "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
-  //        "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
-  //        "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
-  //
-  //        "vmax.f32 q0, q0, q8                   \n\t"
-  //        "vmax.f32 q1, q1, q8                    \n\t"
-  //        "vmax.f32 q2, q2, q8                   \n\t"
-  //        "vmax.f32 q3, q3, q8                   \n\t"
-  //        "vmax.f32 q4, q4, q8                   \n\t"
-  //        "vmax.f32 q5, q5, q8                   \n\t"
-  //        "vmax.f32 q6, q6, q8                   \n\t"
-  //        "vmax.f32 q7, q7, q8                   \n\t"
-  //
-  //        "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
-  //        "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
-  //        "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
-  //        "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
-  //
-  //        "subs %[num], %[num], #32              \n\t"
-  //        "bge        loop_num_%=                \n\t"
-  //        "end_num_%=:                           \n\t"
-  //        "cmp %[num], #0                         \n\t"
-  //        "bge   end_%=                          \n\t"
-  //        "mov r6, #4                             \n\t"
-  //        "mul r5, %[num], r6                     \n\t"
-  //        "add %[input_x_ptr], %[input_x_ptr], r5     \n\t"
-  //        "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
-  //        "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
-  //        "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
-  //        "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
-  //        "vmax.f32 q0, q0, q8                   \n\t"
-  //        "vmax.f32 q1, q1, q8                    \n\t"
-  //        "vmax.f32 q2, q2, q8                   \n\t"
-  //        "vmax.f32 q3, q3, q8                   \n\t"
-  //        "vmax.f32 q4, q4, q8                   \n\t"
-  //        "vmax.f32 q5, q5, q8                   \n\t"
-  //        "vmax.f32 q6, q6, q8                   \n\t"
-  //        "vmax.f32 q7, q7, q8                   \n\t"
-  //        "add %[out_ptr], %[out_ptr], r5       \n\t"
-  //        "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
-  //        "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
-  //        "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
-  //        "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
-  //        "end_%=:                                \n\t"
-  //        :
-  //        :
-  //        [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num]
-  //        "r"(numel) : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
-  //        "q7", "q8", "r5",
-  //          "r6");
-  //  } else {
-  ReluFunctor<float> func_;
-  math::Transform trans;
-  trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
-  //  }
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#if __aarch64__
+  if (numel > 0) {
+    int loop = numel >> 0x4;
+    int remain = numel & 0xF;
+    float32x4_t zero = vdupq_n_f32(0.f);
+    for (int i = 0; i < loop; ++i) {
+      float32x4_t r0 = vld1q_f32(input_x_ptr);
+      float32x4_t r1 = vld1q_f32(input_x_ptr + 4);
+      float32x4_t r2 = vld1q_f32(input_x_ptr + 8);
+      float32x4_t r3 = vld1q_f32(input_x_ptr + 12);
+      r0 = vmaxq_f32(r0, zero);
+      r1 = vmaxq_f32(r1, zero);
+      r2 = vmaxq_f32(r2, zero);
+      r3 = vmaxq_f32(r3, zero);
+      vst1q_f32(out_ptr, r0);
+      vst1q_f32(out_ptr + 4, r1);
+      vst1q_f32(out_ptr + 8, r2);
+      vst1q_f32(out_ptr + 12, r3);
+      input_x_ptr += 16;
+      out_ptr += 16;
+    }
+    for (int i = 0; i < remain; ++i) {
+      out_ptr[i] = (input_x_ptr[i] > 0) * input_x_ptr[i];
+    }
+#else
+  if (numel > 64) {
+    asm volatile(
+        "pld        [%[input_x_ptr], #0]        \n\t"
+        "vmov.f32   q8,    #0.0                 \n\t"
+        "subs %[num], %[num], #32                \n\t"
+        "blt        end_num_%=                  \n\t"
+        "loop_num_%=:                           \n\t"
+        "pld        [%[input_x_ptr], #1024]      \n\t"
+
+        "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
+
+        "vmax.f32 q0, q0, q8                   \n\t"
+        "vmax.f32 q1, q1, q8                    \n\t"
+        "vmax.f32 q2, q2, q8                   \n\t"
+        "vmax.f32 q3, q3, q8                   \n\t"
+        "vmax.f32 q4, q4, q8                   \n\t"
+        "vmax.f32 q5, q5, q8                   \n\t"
+        "vmax.f32 q6, q6, q8                   \n\t"
+        "vmax.f32 q7, q7, q8                   \n\t"
+
+        "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
+        "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
+
+        "subs %[num], %[num], #32              \n\t"
+        "bge        loop_num_%=                \n\t"
+        "end_num_%=:                           \n\t"
+        "cmp %[num], #0                         \n\t"
+        "bge   end_%=                          \n\t"
+        "mov r6, #4                             \n\t"
+        "mul r5, %[num], r6                     \n\t"
+        "add %[input_x_ptr], %[input_x_ptr], r5     \n\t"
+        "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
+        "vmax.f32 q0, q0, q8                   \n\t"
+        "vmax.f32 q1, q1, q8                    \n\t"
+        "vmax.f32 q2, q2, q8                   \n\t"
+        "vmax.f32 q3, q3, q8                   \n\t"
+        "vmax.f32 q4, q4, q8                   \n\t"
+        "vmax.f32 q5, q5, q8                   \n\t"
+        "vmax.f32 q6, q6, q8                   \n\t"
+        "vmax.f32 q7, q7, q8                   \n\t"
+        "add %[out_ptr], %[out_ptr], r5       \n\t"
+        "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
+        "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
+        "end_%=:                                \n\t"
+        :
+        :
+        [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num] "r"(numel)
+        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "r5",
+          "r6");
+#endif
+  } else {
+#endif
+    ReluFunctor<float> func_;
+    math::Transform trans;
+    trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+  }
+#endif
 }
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/central-arm-func/sum_arm_func.h

@@ -15,11 +15,14 @@ limitations under the License. */
 #ifdef SUM_OP
 #pragma once
 
+#include <vector>
 #include "operators/math/selected_rows_functor.h"
 
 namespace paddle_mobile {
 namespace operators {
+
 using LoDTensorArray = std::vector<LoDTensor>;
+
 template <typename P>
 void SumCompute(const SumParam<CPU> &param) {
   auto inputsvars = param.InputsVars();
@@ -63,31 +66,21 @@ void SumCompute(const SumParam<CPU> &param) {
     std::unique_ptr<framework::SelectedRows> in0;
     if (in_place) {
       // If is in_place, we store the input[0] to in0
-      auto *in_sel0 = inputsvars[0]->Get<SelectedRows>();
+      auto *in_sel0 = inputsvars[0]->Get<framework::SelectedRows>();
       auto &rows = in_sel0->rows();
-      //#ifdef PADDLE_WITH_CUDA
-      //                    std::vector<int64_t> rows_in_cpu;
-      //        rows_in_cpu.reserve(rows.size());
-      //        for (auto item : rows) {
-      //          rows_in_cpu.push_back(item);
-      //        }
-      //        in0.reset(new framework::SelectedRows(rows_in_cpu,
-      //        in_sel0.height()));
-      //#else
       in0.reset(new framework::SelectedRows(rows, in_sel0->height()));
-      //#endif
       in0->mutable_value()->ShareDataWith(in_sel0->value());
     }
 
-    auto get_selected_row = [&](size_t i) -> const SelectedRows & {
+    auto get_selected_row = [&](size_t i) -> const framework::SelectedRows & {
       if (i == 0 && in0) {
         return *in0.get();
       } else {
-        return *(inputsvars[i]->Get<SelectedRows>());
+        return *(inputsvars[i]->Get<framework::SelectedRows>());
       }
     };
 
-    auto *out = outvar->GetMutable<SelectedRows>();
+    auto *out = outvar->GetMutable<framework::SelectedRows>();
     out->mutable_rows()->clear();
     auto *out_value = out->mutable_value();
 
@@ -150,8 +143,6 @@ void SumCompute(const SumParam<CPU> &param) {
       }
     }
   } else {
-    if (outvar->IsType<framework::Tensor>()) {
-    }
     PADDLE_MOBILE_THROW_EXCEPTION(
         "Unexpected branch, output variable type is %s", outvar->Type().name());
   }

src/operators/kernel/dequantize_kernel.h

@@ -12,6 +12,8 @@ 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. */
 
+#ifdef DEQUANT_OP
+
 #pragma once
 
 #include "framework/operator.h"
@@ -30,3 +32,5 @@ class DequantizeKernel
 
 }  // namespace operators
 }  // namespace paddle_mobile
+
+#endif

src/operators/kernel/elementwise_mul_kernel.h

@@ -23,8 +23,6 @@ limitations under the License. */
 namespace paddle_mobile {
 namespace operators {
 
-using namespace framework;
-
 template <typename DeviceType, typename T>
 class ElementwiseMulKernel
     : public framework::OpKernelBase<DeviceType,

src/operators/kernel/quantize_kernel.h

@@ -12,6 +12,8 @@ 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. */
 
+#ifdef QUANT_OP
+
 #pragma once
 
 #include "framework/operator.h"
@@ -30,3 +32,5 @@ class QuantizeKernel
 
 }  // namespace operators
 }  // namespace paddle_mobile
+
+#endif

src/operators/kernel/sum_kernel.h

@@ -21,8 +21,6 @@ limitations under the License. */
 namespace paddle_mobile {
 namespace operators {
 
-using namespace framework;
-
 template <typename DeviceType, typename T>
 class SumKernel
     : public framework::OpKernelBase<DeviceType, SumParam<DeviceType>> {

src/operators/math/conv_arm_int8.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. */
+
+#ifdef CONV_OP
+
+#pragma once
+
+#include "framework/tensor.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+void conv3x3s1_int8(const framework::Tensor& input,
+                    const framework::Tensor& weight, framework::Tensor* output);
+
+void conv3x3s1_int8_4c(const framework::Tensor& input,
+                       const framework::Tensor& weight,
+                       framework::Tensor* output);
+
+void conv5x5s1_int8(const framework::Tensor& input,
+                    const framework::Tensor& weight, framework::Tensor* output);
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/math/gemm.cpp

@@ -3379,7 +3379,7 @@ void Gemm::SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A,
     // 对 B 分块
     NC = L1 / (KC * sizeof(float));
     if (NC == 0) {
-      NC == NR;
+      NC = NR;
     } else {
       int nblock_num = (n + NC - 1) / NC;
       NC = (n + nblock_num - 1) / nblock_num;

src/operators/math/gemm.h

@@ -22,9 +22,11 @@ limitations under the License. */
 #define C(i, j) C[(i)*ldc + (j)]
 
 #if __aarch64__
+#define MR_INT8 4
 #define MR 6
 #define NR 16
 #else
+#define MR_INT8 4
 #define MR 6
 #define NR 8
 #endif
@@ -189,6 +191,8 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
 
   // 8 bits function cluster begins
   // 8 bits int small block inner product
+  void AddDot4x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
+                 int32_t ldc);
   void AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
                  int32_t ldc);
 
@@ -199,6 +203,8 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
                            int8_t *bias);
 
   // 8 bits int pack function
+  void PackMatrixA_4r(int32_t m, int32_t k, int32_t m_tail, const int8_t *A,
+                      int32_t lda, int8_t *buffer);
   void PackMatrixA_6r(int32_t m, int32_t k, int32_t m_tail, const int8_t *A,
                       int32_t lda, int8_t *buffer);
   void PackMatrixB_8c(int32_t k, int32_t n, int32_t n_tail, const int8_t *B,

src/operators/math/math_function.cpp

@@ -15,12 +15,31 @@ limitations under the License. */
 #include "operators/math/math_function.h"
 #include <cstring>
 #include <string>
+#include "framework/data_type.h"
+#include "framework/tensor.h"
 #include "operators/math/gemm.h"
 
 namespace paddle_mobile {
 namespace operators {
 namespace math {
 
+struct TensorSetConstant {
+  TensorSetConstant(framework::Tensor *tensor, float value)
+      : tensor_(tensor), value_(value) {}
+  template <typename T>
+  void apply() const {
+    auto *begin = tensor_->mutable_data<T>();
+    std::fill(begin, begin + tensor_->numel(), static_cast<T>(value_));
+  }
+  framework::Tensor *tensor_;
+  float value_;
+};
+
+void set_constant(framework::Tensor *tensor, float value) {
+  framework::VisitDataType(framework::ToDataType(tensor->type()),
+                           TensorSetConstant(tensor, value));
+}
+
 template <>
 void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
                    const framework::Tensor &matrix_b, bool trans_b, float alpha,

src/operators/math/math_function.h

@@ -15,12 +15,15 @@ limitations under the License. */
 #pragma once
 
 #include <cmath>
+#include <string>
 #include "framework/tensor.h"
 
 namespace paddle_mobile {
 namespace operators {
 namespace math {
 
+void set_constant(framework::Tensor *tensor, float value);
+
 template <typename T>
 void matmul(const framework::Tensor &matrix_a, bool trans_a,
             const framework::Tensor &matrix_b, bool trans_b, T alpha,

src/operators/math/pad.cpp

+/* 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 "operators/math/pad.h"
+
+namespace paddle_mobile {
+namespace operators {
+namespace math {
+
+template <typename T>
+class PadFunctor<CPU, T> {
+ public:
+  void operator()(const framework::Tensor &input, const int pad_h,
+                  const int pad_w, framework::Tensor *output) {
+    const T *in_data = input.data<T>();
+    T *out_data = output->mutable_data<T>();
+    const framework::DDim &input_shape = input.dims();
+    const framework::DDim &output_shape = output->dims();
+    // fill output with 0
+    memset(out_data, 0, sizeof(T) * output->numel());
+    // should make sure the shape of output is match with input
+    for (int i = 0; i < input_shape[0]; ++i) {
+      for (int c = 0; c < input_shape[1]; ++c) {
+        out_data += pad_h * output_shape[3];
+        for (int h = 0; h < input_shape[2]; ++h) {
+          memcpy(out_data + pad_w, in_data, sizeof(T) * input_shape[3]);
+          out_data += output_shape[3];
+          in_data += input_shape[3];
+        }
+        out_data += pad_h * output_shape[3];
+      }
+    }
+  }
+};
+
+template class PadFunctor<CPU, float>;
+template class PadFunctor<CPU, int8_t>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/math/pad.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 "framework/tensor.h"
+
+namespace paddle_mobile {
+namespace operators {
+namespace math {
+
+template <typename DeviceType, typename T>
+class PadFunctor {
+ public:
+  void operator()(const framework::Tensor &input, const int pad_h,
+                  const int pad_w, framework::Tensor *output);
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/math/vol2col.cpp

@@ -32,9 +32,6 @@ class Vol2ColFunctor<CPU, T> {
   void operator()(const Tensor &vol, const std::vector<int> &dilations,
                   const std::vector<int> &strides,
                   const std::vector<int> &paddings, Tensor *col) const {
-    //    PADDLE_ENFORCE(vol.dims().size() == 4);
-    //    PADDLE_ENFORCE(col->dims().size() == 7);
-
     int input_channels = vol.dims()[0];
     int input_depth = vol.dims()[1];
     int input_height = vol.dims()[2];
@@ -48,32 +45,6 @@ class Vol2ColFunctor<CPU, T> {
     int channels_col =
         input_channels * filter_depth * filter_height * filter_width;
 
-    //    PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] -
-    //                       ((dilations[0] * (filter_depth - 1)
-    //                       + 1))) /
-    //                              strides[0] +
-    //                          1,
-    //                      output_depth,
-    //                      "input_depth and output_depth are "
-    //                      "mismatching.");
-    //    PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] -
-    //                       ((dilations[1] * (filter_height -
-    //                       1) + 1))) /
-    //                              strides[1] +
-    //                          1,
-    //                      output_height,
-    //                      "input_height and output_height are
-    //                      "
-    //                      "mismatching.");
-    //    PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] -
-    //                       ((dilations[2] * (filter_width - 1)
-    //                       + 1))) /
-    //                              strides[2] +
-    //                          1,
-    //                      output_width,
-    //                      "input_width and output_width are "
-    //                      "mismatching.");
-
     const T *vol_data = vol.data<T>();
     T *col_data = col->data<T>();
 
@@ -119,9 +90,6 @@ class Col2VolFunctor<CPU, T> {
   void operator()(const Tensor &col, const std::vector<int> &dilations,
                   const std::vector<int> &strides,
                   const std::vector<int> &paddings, Tensor *vol) const {
-    //    PADDLE_ENFORCE(vol->dims().size() == 4);
-    //    PADDLE_ENFORCE(col.dims().size() == 7);
-
     int input_channels = vol->dims()[0];
     int input_depth = vol->dims()[1];
     int input_height = vol->dims()[2];
@@ -135,31 +103,6 @@ class Col2VolFunctor<CPU, T> {
     int channels_col =
         input_channels * filter_depth * filter_height * filter_width;
 
-    //    PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] -
-    //                       ((dilations[0] * (filter_depth - 1)
-    //                       + 1))) /
-    //                              strides[0] +
-    //                          1,
-    //                      output_depth,
-    //                      "input_depth and output_depth are "
-    //                      "mismatching.");
-    //    PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] -
-    //                       ((dilations[1] * (filter_height -
-    //                       1) + 1))) /
-    //                              strides[1] +
-    //                          1,
-    //                      output_height,
-    //                      "input_height and output_height are
-    //                      "
-    //                      "mismatching.");
-    //    PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] -
-    //                       ((dilations[2] * (filter_width - 1)
-    //                       + 1))) /
-    //                              strides[2] +
-    //                          1,
-    //                      output_width,
-    //                      "input_width and output_width are "
-    //                      "mismatching.");
     T *vol_data = vol->data<T>();
     const T *col_data = col.data<T>();
 
@@ -195,9 +138,9 @@ class Col2VolFunctor<CPU, T> {
 };
 
 template class Vol2ColFunctor<CPU, float>;
-template class Vol2ColFunctor<CPU, double>;
+template class Vol2ColFunctor<CPU, int8_t>;
 template class Col2VolFunctor<CPU, float>;
-template class Col2VolFunctor<CPU, double>;
+template class Col2VolFunctor<CPU, int8_t>;
 
 }  // namespace math
 }  // namespace operators

src/operators/op_param.h

@@ -1063,6 +1063,42 @@ class FetchParam : public OpParam {
   RType *out_;
 };
 
+#ifdef FILL_CONSTANT_OP
+template <typename Dtype>
+class FillConstantParam : public OpParam {
+  typedef typename DtypeTensorTrait<Dtype>::gtype GType;
+  typedef typename DtypeTensorTrait<Dtype>::rtype RType;
+
+ public:
+  FillConstantParam(const VariableNameMap &inputs,
+                    const VariableNameMap &outputs, const AttributeMap &attrs,
+                    const Scope &scope) {
+    out_var_ = OutVarFrom(outputs, scope);
+    out_ = OutFrom<GType>(outputs, scope);
+    dtype_ = GetAttr<int>("dtype", attrs);
+    shape_ = GetAttr<vector<int>>("shape", attrs);
+    value_ = GetAttr<float>("value", attrs);
+  }
+
+  Variable *OutVar() const { return out_var_; }
+
+  RType *Out() const { return out_; }
+
+  const int &DataDtype() const { return dtype_; }
+
+  const vector<int> &Shape() const { return shape_; }
+
+  const float &Value() const { return value_; }
+
+ private:
+  Variable *out_var_;
+  RType *out_;
+  int dtype_;
+  vector<int> shape_;
+  float value_;
+};
+#endif
+
 #ifdef TRANSPOSE_OP
 template <typename Dtype>
 class TransposeParam : public OpParam {
@@ -2294,6 +2330,7 @@ class ShapeParam : public OpParam {
 };
 #endif
 
+#ifdef QUANT_OP
 template <typename Dtype>
 class QuantizeParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
@@ -2304,14 +2341,12 @@ class QuantizeParam : public OpParam {
                 const AttributeMap &attrs, const Scope &scope) {
     input_ = InputXFrom<GType>(inputs, scope);
     out_ = OutFrom<GType>(outputs, scope);
-    if (HasAttr("is_static", attrs)) {
-      is_static_ = GetAttr<bool>("is_static", attrs);
-    }
     // online
     // scale = max(abs(x))
     online_scale_ = GetVarValue<GType>("OutScale", outputs, scope);
     // offline
     if (HasAttr("static_scale", attrs)) {
+      is_static_ = true;
       static_scale_ = GetAttr<float>("static_scale", attrs);
     }
     // x = round(scale * x)
@@ -2333,9 +2368,11 @@ class QuantizeParam : public OpParam {
   float static_scale_ = 1.0f;
   // round method type
   // nearest_zero and nearest_even is valid currently
-  RoundType round_type_ = ROUND_NEAREST_TO_EVEN;
+  RoundType round_type_ = ROUND_NEAREST_AWAY_ZERO;
 };
+#endif
 
+#ifdef DEQUANT_OP
 template <typename Dtype>
 class DequantizeParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
@@ -2363,6 +2400,7 @@ class DequantizeParam : public OpParam {
   RType *activation_scale_;
   float weight_scale_;
 };
+#endif
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/quantize_op.cpp

@@ -12,6 +12,8 @@ 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. */
 
+#ifdef QUANT_OP
+
 #include "operators/quantize_op.h"
 #include <vector>
 
@@ -33,3 +35,5 @@ namespace ops = paddle_mobile::operators;
 #ifdef PADDLE_MOBILE_CPU
 REGISTER_OPERATOR_CPU(quantize, ops::QuantizeOp);
 #endif
+
+#endif

src/operators/quantize_op.h

@@ -12,6 +12,8 @@ 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. */
 
+#ifdef QUANT_OP
+
 #pragma once
 
 #include <string>
@@ -40,3 +42,5 @@ class QuantizeOp : public framework::OperatorWithKernel<
 
 }  // namespace operators
 }  // namespace paddle_mobile
+
+#endif

src/operators/sum_op.cpp

@@ -26,7 +26,7 @@ void SumOp<Dtype, T>::InferShape() const {
   auto inputs = this->param_.Inputs();
   const size_t n = inputs.size();
 
-  std::vector<DDim> inputs_dims;
+  std::vector<framework::DDim> inputs_dims;
   inputs_dims.reserve(n);
   for (int i = 0; i < n; i++) {
     inputs_dims.push_back(inputs[i]->dims());

test/CMakeLists.txt

@@ -192,6 +192,10 @@ if (NOT FOUND_MATCH)
     ADD_EXECUTABLE(test-polygon-box-transform-op operators/test_polygon_box_transform_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-polygon-box-transform-op paddle-mobile)
 
+    # gen test
+    ADD_EXECUTABLE(test-fill-constant-op operators/test_fill_constant_op.cpp test_helper.h test_include.h)
+    target_link_libraries(test-fill-constant-op paddle-mobile)
+
     # gen test
     ADD_EXECUTABLE(test-reshape-op operators/test_reshape_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-reshape-op paddle-mobile)
@@ -216,6 +220,10 @@ if (NOT FOUND_MATCH)
     ADD_EXECUTABLE(test-dequantize-op operators/test_dequantize_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-dequantize-op paddle-mobile)
 
+    # test int8 conv op
+    ADD_EXECUTABLE(test-int8-conv-op operators/test_int8_conv_op.cpp test_helper.h test_include.h)
+    target_link_libraries(test-int8-conv-op paddle-mobile)
+
     # gen test log
     ADD_EXECUTABLE(test-log common/test_log.cpp)
     target_link_libraries(test-log paddle-mobile)

test/net/test_googlenet.cpp

@@ -25,27 +25,31 @@ int main() {
   paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
 #endif
 
-  paddle_mobile.SetThreadNum(4);
-  bool optimize = true;
+  paddle_mobile.SetThreadNum(1);
+  bool optimize = false;
   auto time1 = time();
   if (paddle_mobile.Load(g_googlenet, optimize)) {
     auto time2 = time();
     std::cout << "load cost :" << time_diff(time1, time2) << "ms" << std::endl;
     std::vector<float> input;
+    std::vector<float> output;
     std::vector<int64_t> dims{1, 3, 224, 224};
     GetInput<float>(g_test_image_1x3x224x224, &input, dims);
-    // 预热十次
-    for (int i = 0; i < 10; ++i) {
-      auto vec_result = paddle_mobile.Predict(input, dims);
-    }
+    //    // 预热十次
+    //    for (int i = 0; i < 10; ++i) {
+    //      output = paddle_mobile.Predict(input, dims);
+    //    }
     auto time3 = time();
     for (int i = 0; i < 10; ++i) {
-      auto vec_result = paddle_mobile.Predict(input, dims);
+      output = paddle_mobile.Predict(input, dims);
     }
     auto time4 = time();
 
     std::cout << "predict cost :" << time_diff(time3, time4) / 10 << "ms"
               << std::endl;
+    for (int i = 0; i < output.size(); ++i) {
+      DLOG << "result[" << i << "] = " << output[i];
+    }
   }
   return 0;
 }

test/operators/test_dequantize_op.cpp

@@ -59,7 +59,7 @@ int TestDequqntizeOp() {
 
   framework::Tensor output_cmp;
   output_cmp.Resize(dim);
-  float dequant_scale = 1.f / (1.27 * 1.74);
+  float dequant_scale = 1.27 / 1.74;
   dequantize(input, dequant_scale, &output_cmp);
   const float* output_cmp_data = output_cmp.data<float>();
   for (int i = 0; i < output->numel(); ++i) {

test/operators/test_fill_constant_op.cpp

+/* 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 "../test_include.h"
+#include "operators/fill_constant_op.h"
+
+namespace paddle_mobile {
+namespace framework {
+
+template <typename Dtype>
+class TestFillConstantOp {
+ public:
+  explicit TestFillConstantOp(const Program<Dtype> p) : program_(p) {
+    if (use_optimize_) {
+      to_predict_program_ = program_.optimizeProgram;
+    } else {
+      to_predict_program_ = program_.originProgram;
+    }
+    const std::vector<std::shared_ptr<BlockDesc>> blocks =
+        to_predict_program_->Blocks();
+    for (auto block_desc : blocks) {
+      std::vector<std::shared_ptr<OpDesc>> ops = block_desc->Ops();
+      for (auto op : ops) {
+        if (op->Type() == "fill_constant") {
+          DLOG << " attr size: " << op->GetAttrMap().size();
+          std::unordered_map<std::string, Attribute> attrs = op->GetAttrMap();
+          for (std::unordered_map<std::string, Attribute>::iterator it =
+                   attrs.begin();
+               it != attrs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+          DLOG << " inputs size: " << op->GetInputs().size();
+          DLOG << " outputs size: " << op->GetOutputs().size();
+          DLOG << " output is : " << op->Output("Out")[0];
+          output_var_name = op->Output("Out")[0];
+          std::shared_ptr<operators::FillConstantOp<Dtype, float>> op_ptr =
+              std::make_shared<operators::FillConstantOp<Dtype, float>>(
+                  op->Type(), op->GetInputs(), op->GetOutputs(),
+                  op->GetAttrMap(), program_.scope);
+          ops_of_block_[*block_desc.get()].push_back(op_ptr);
+        }
+      }
+    }
+  }
+
+  std::shared_ptr<Tensor> predict() {
+    auto scope = program_.scope;
+
+    Variable *output = scope->Var(output_var_name);
+    auto *output_tensor = output->GetMutable<LoDTensor>();
+
+    std::shared_ptr<Tensor> out_tensor = std::make_shared<LoDTensor>();
+    out_tensor.reset(output_tensor);
+
+    predict(0);
+
+    return out_tensor;
+  }
+
+ private:
+  const framework::Program<Dtype> program_;
+  std::shared_ptr<ProgramDesc> to_predict_program_;
+  std::map<framework::BlockDesc,
+           std::vector<std::shared_ptr<OperatorBase<Dtype>>>>
+      ops_of_block_;
+  bool use_optimize_ = false;
+  string output_var_name;
+
+  void predict(int block_id) {
+    std::shared_ptr<BlockDesc> to_predict_block =
+        to_predict_program_->Block(block_id);
+    for (int j = 0; j < ops_of_block_[*to_predict_block.get()].size(); ++j) {
+      auto op = ops_of_block_[*to_predict_block.get()][j];
+      op->Run();
+    }
+  }
+};
+
+template class TestFillConstantOp<CPU>;
+}  // namespace framework
+}  // namespace paddle_mobile
+
+int main() {
+  DLOG << "----------**********----------";
+  DLOG << "begin to run FillConstant Test";
+  paddle_mobile::Loader<paddle_mobile::CPU> loader;
+  auto program = loader.Load(std::string(g_ocr) + "/model",
+                             std::string(g_ocr) + "/params");
+
+  paddle_mobile::framework::TestFillConstantOp<paddle_mobile::CPU>
+      testFillConstantOp(program);
+
+  auto output = testFillConstantOp.predict();
+  auto *output_ptr = output->data<float>();
+
+  DLOG << "output : ";
+  for (int i = 0; i < output->numel(); ++i) {
+    DLOG << " index " << i << " : " << output_ptr[i];
+  }
+  return 0;
+}

test/operators/test_mul_op.cpp

@@ -12,7 +12,6 @@ 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 <stdint-gcc.h>
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/mul_op.h"

test/operators/test_quantize_op.cpp

@@ -18,14 +18,6 @@ limitations under the License. */
 
 namespace paddle_mobile {
 
-// static float g_test_data[50] = {
-//   -5.55, -5.5, -5.45, -5.0, -4.55, -4.5, -4.45, -4.0, -3.55, -3.5,
-//   -3.45, -3.01, -2.75, -2.5, -2.501, -2.49, -2.01, -1.75, -1.5, -1.25,
-//   -1.0, -0.75, -0.5, -0.25, 0.0, 0.25, 0.5, 0.75, 1.0, 1.25,
-//   1.5, 1.75, 2.01, 2.49, 2.501, 2.5, 2.75, 3.01, 3.45, 3.5,
-//   3.55, 4.0, 4.45, 4.5, 4.55, 5.0, 5.45, 5.5, 5.55, 6.0,
-// };
-
 static float find_abs_max(const Tensor *input) {
   float max_abs = 0.f;
   const float *x = input->data<const float>();
@@ -60,6 +52,16 @@ static void quantize_round_to_even(const Tensor *input, const float scale,
   }
 }
 
+static void quantize_round_to_nearest(const Tensor *input, const float scale,
+                                      Tensor *output) {
+  const float *x = input->data<const float>();
+  int8_t *y = output->mutable_data<int8_t>();
+  size_t size = input->numel();
+  for (size_t i = 0; i < size; ++i) {
+    y[i] = round(x[i] * scale);
+  }
+}
+
 int TestQuqntizeOp() {
   framework::DDim dim = framework::make_ddim({1, 3, 224, 224});
 
@@ -88,15 +90,16 @@ int TestQuqntizeOp() {
   auto output_scale = output_scale_var->template Get<framework::LoDTensor>();
   const float *output_scale_data = output_scale->data<float>();
 
-  float max_abs = find_abs_max(input);
-  float output_scale_cmp = 127 / max_abs;
+  float output_scale_cmp = find_abs_max(input);
   PADDLE_MOBILE_ENFORCE(output_scale_cmp == output_scale_data[0],
                         "output_scale = %.6f, output_scale_cmp = %.6f",
                         output_scale_cmp, output_scale_data[0]);
 
   framework::Tensor output_cmp;
   output_cmp.Resize(dim);
-  quantize_round_to_even(input, output_scale_cmp, &output_cmp);
+  float scale = 127 / output_scale_cmp;
+  // quantize_round_to_even(input, scale, &output_cmp);
+  quantize_round_to_nearest(input, scale, &output_cmp);
   int8_t *output_cmp_data = output_cmp.data<int8_t>();
   for (int i = 0; i < output->numel(); ++i) {
     PADDLE_MOBILE_ENFORCE(output_data[i] == output_cmp_data[i],

tools/op.cmake

@@ -188,6 +188,7 @@ if(NOT FOUND_MATCH)
   set(ELEMENTWISEADD_OP ON)
   set(ELEMENTWISESUB_OP ON)
   set(IM2SEQUENCE_OP ON)
+  set(FILL_CONSTANT_OP ON)
   set(FUSION_CONVADD_OP ON)
   set(FUSION_CONVADDPRELU_OP ON)
   set(FUSION_CONVADDRELU_OP ON)
@@ -223,6 +224,8 @@ if(NOT FOUND_MATCH)
   set(SHAPE_OP ON)
   set(ELEMENTWISEMUL_OP ON)
   set(SUM_OP ON)
+  set(QUANT_OP ON)
+  set(DEQUANT_OP ON)
 endif()
 
   # option(BATCHNORM_OP "" ON)
@@ -231,6 +234,7 @@ endif()
   # option(CONV_OP "" ON)
   # option(DEPTHWISECONV_OP "" ON)
   # option(ELEMENTWISEADD_OP "" ON)
+  # option(FILL_CONSTANT_OP "" ON)
   # option(FUSION_CONVADD_OP "" ON)
   # option(FUSION_CONVADDRELU_OP "" ON)
   # option(FUSION_FC_OP "" ON)
@@ -268,6 +272,9 @@ endif()
 if (ELEMENTWISESUB_OP)
   add_definitions(-DELEMENTWISESUB_OP)
 endif()
+if (FILL_CONSTANT_OP)
+  add_definitions(-DFILL_CONSTANT_OP)
+endif()
 if (FUSION_CONVADD_OP)
   add_definitions(-DFUSION_CONVADD_OP)
 endif()
@@ -406,3 +413,10 @@ if (SUM_OP)
   add_definitions(-DSUM_OP)
 endif()
 
+if (QUANT_OP)
+  add_definitions(-DQUANT_OP)
+endif()
+if (DEQUANT_OP)
+  add_definitions(-DDEQUANT_OP)
+endif()
+