conflict

CMakeLists.txt

@@ -29,7 +29,10 @@ if(DEBUGING)
     message(STATUS "debugging mode")
     add_definitions(-DPADDLE_MOBILE_DEBUG)
 else()
-    add_definitions(-fvisibility=hidden -fvisibility-inlines-hidden)
+    if(FPGA)
+    else()
+        add_definitions(-fvisibility=hidden -fvisibility-inlines-hidden)
+    endif()
 endif()
 
 if(USE_EXCEPTION)
@@ -93,8 +96,7 @@ else()
 endif()
 
 if(FPGA)
-    set(DEBUGING ON)
-    add_definitions(-DPADDLE_MOBILE_DEBUG)
+    message("FPGA mode enabled")
     add_definitions(-DPADDLE_MOBILE_FPGA)
 else()
     file(GLOB_RECURSE _tmp_list src/operators/kernel/fpga/*.cpp src/operators/kernel/fpga/*.cc)
@@ -177,6 +179,10 @@ if(DEBUGING)
     else()
         add_subdirectory(test)
     endif()
+elseif(FPGA)
+    add_subdirectory(test)
 endif()
 
 
+
+

README.md

@@ -83,7 +83,7 @@ Paddle-Mobile是PaddlePaddle组织下的项目，是一个致力于嵌入式平
 
 - **FPGA**
 
-    FPGA实现正在进行中，是基于Xilinx的ZU5目标开发板。
+    目前已经支持 ZCU102 开发板。
 
 - **灵活性**
 
@@ -112,6 +112,7 @@ Paddle-Mobile是PaddlePaddle组织下的项目，是一个致力于嵌入式平
 开发文档主要是关于编译、运行等问题。做为开发者，它可以和贡献文档共同结合使用。
 * [iOS](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_ios.md)
 * [Android](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_android.md)
+* [FPGA](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_fpga.md)
 
 ### 贡献文档
 - [贡献文档链接](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/CONTRIBUTING.md)

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";
@@ -34,6 +35,7 @@ const char *G_OP_TYPE_FUSION_CONV_ADD = "fusion_conv_add";
 const char *G_OP_TYPE_LRN = "lrn";
 const char *G_OP_TYPE_MUL = "mul";
 const char *G_OP_TYPE_MULTICLASS_NMS = "multiclass_nms";
+const char *G_OP_TYPE_POLYGON_BOX_TRANSFORM = "polygon_box_transform";
 const char *G_OP_TYPE_POOL2D = "pool2d";
 const char *G_OP_TYPE_PRIOR_BOX = "prior_box";
 const char *G_OP_TYPE_RELU = "relu";
@@ -94,9 +96,11 @@ std::unordered_map<
         {G_OP_TYPE_FUSION_CONV_BN_ADD_RELU, {{"Input"}, {"Out"}}},
         {G_OP_TYPE_PRIOR_BOX, {{"Image", "Input"}, {"Boxes", "Variances"}}},
         {G_OP_TYPE_MULTICLASS_NMS, {{"BBoxes", "Scores"}, {"Out"}}},
+        {G_OP_TYPE_POLYGON_BOX_TRANSFORM, {{"Input"}, {"Output"}}},
         {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. */
 
-#include <cstdlib>
+#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/api.cpp

@@ -22,7 +22,7 @@ limitations under the License. */
 #include "fpga/filter.h"
 #include "fpga/image.h"
 #define FPGA_TEST_MODE
-//#define PADDLE_MOBILE_OS_LINUX
+#define PADDLE_MOBILE_OS_LINUX
 
 namespace paddle_mobile {
 namespace fpga {
@@ -149,7 +149,7 @@ int ComputeBasicConv(const struct ConvArgs &args) {
   return do_ioctl(IOCTL_CONFIG_CONV, &args);
 }
 
-int ComputeFpgaConv(const struct WrapperConvArgs &args) {
+int ComputeFpgaConv(const struct SplitConvArgs &args) {
 #ifdef FPGA_TEST_MODE
   DLOG << "=============ComputeFPGAConv===========";
   DLOG << "   filter_num:" << args.filter_num
@@ -194,8 +194,8 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef FPGA_TEST_MODE
   DLOG << "=============ComputeFpgaEWAdd===========";
   DLOG << "   relu_enabled:" << args.relu_enabled
-       << "   const0:" << fp16_2_fp32(short(args.const0))
-       << "   const1:" << fp16_2_fp32(short(args.const1));
+       << "   const0:" << fp16_2_fp32(int16_t(args.const0))
+       << "   const1:" << fp16_2_fp32(int16_t(args.const1));
   DLOG << "   image0_address:" << args.image0.address
        << "   image0_scale_address:" << args.image0.scale_address
        << "   image0_channels:" << args.image0.channels
@@ -383,10 +383,10 @@ void format_concat_output(framework::Tensor *out, int height, int width,
   out->reset_data_ptr(data_ptr);
 }
 
-void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
-                   framework::Tensor *out, framework::Tensor *filter,
-                   bool relu_enabled, int group_num, int stride_h, int stride_w,
-                   int padding_h, int padding_w, float *bs_ptr) {
+void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
+                    framework::Tensor *out, framework::Tensor *filter,
+                    bool relu_enabled, int group_num, int stride_h,
+                    int stride_w, int padding_h, int padding_w, float *bs_ptr) {
   auto input_ptr = input->data<float>();
   auto filter_ptr = filter->data<float>();
   auto out_ptr = out->data<float>();

src/fpga/api.h

@@ -89,7 +89,7 @@ struct ConcatArgs {
   uint32_t width;
 };
 
-struct WrapperConvArgs {
+struct SplitConvArgs {
   uint32_t split_num;
   uint32_t group_num;
   uint32_t filter_num;
@@ -98,6 +98,14 @@ struct WrapperConvArgs {
   struct ConcatArgs concat_arg;
 };
 
+struct GroupConvArgs {
+  uint32_t group_num;
+  uint32_t filter_num;
+  struct ImageOutputArgs output;
+  struct SplitConvArgs* conv_args;
+  struct ConcatArgs concat_arg;
+};
+
 struct PoolingArgs {
   int16_t mode;  // mode: 0:max, 1:avg
   half kernel_reciprocal;
@@ -159,30 +167,6 @@ struct MemoryCacheArgs {
 #define IOCTL_FPGA_REG_READ _IOW(IOCTL_FPGA_MAGIC, 28, struct FpgaRegReadArgs)
 #define IOCTL_FPGA_REG_WRITE _IOW(IOCTL_FPGA_MAGIC, 29, struct FpgaRegWriteArgs)
 
-enum FPGA_ERR_TYPE {
-  ERR_IOCTL_CMD = -1,
-  ERR_TIMEOUT = -2,
-  ERR_COMPLETION_TIMEOUT = -3,
-  ERR_INVALID_FPGA_ADDR = -4,
-  ERR_NOMEM = -5,
-  ERR_NO_RESERVE_MEM = -6,
-  ERR_COPY_FROM_USER = -7,
-  ERR_COPY_TO_USER = -8,
-  ERR_DEL_TIMER = -9,
-  ERR_ENABLE_MSI = -10,
-  ERR_REGISTER_IRQ = -11,
-  ERR_PCIE_REGISTER = -12,
-  ERR_PCIE_PROBE = -13,
-  ERR_REGISTER_BLOCK = -14,
-  ERR_ALLOC_GENDISK = -15,
-  ERR_INIT_QUEUE = -16,
-  ERR_WAIT = -17,
-  ERR_ECC_ERROR = -31,
-  ERR_FPGA_FAIL_STOP = -64,
-  ERR_FPGA_DEBUG_STOP = -113,
-  DEV_TMP_UNAVAILABLE = -128
-};
-
 //============================== API =============================
 
 int open_device();
@@ -195,7 +179,7 @@ int fpga_flush(void* address, size_t size);
 int fpga_invalidate(void* address, size_t size);
 
 int PerformBypass(const struct BypassArgs& args);
-int ComputeFpgaConv(const struct WrapperConvArgs& args);
+int ComputeFpgaConv(const struct SplitConvArgs& args);
 int ComputeFpgaPool(const struct PoolingArgs& args);
 int ComputeFpgaEWAdd(const struct EWAddArgs& args);
 int ComputeFPGAConcat(const struct ConcatArgs& args);
@@ -220,10 +204,10 @@ void format_bias_scale_array(float** bias_scale_array,
 void format_concat_output(framework::Tensor* out, int height, int width,
                           int image_num, uint32_t* channel_num);
 
-void fill_conv_arg(struct WrapperConvArgs* arg, framework::Tensor* input,
-                   framework::Tensor* out, framework::Tensor* filter,
-                   bool relu_enabled, int group_num, int stride_h, int stride_w,
-                   int padding_h, int padding_w, float* bs_ptr);
+void fill_split_arg(struct SplitConvArgs* arg, framework::Tensor* input,
+                    framework::Tensor* out, framework::Tensor* filter,
+                    bool relu_enabled, int group_num, int stride_h,
+                    int stride_w, int padding_h, int padding_w, float* bs_ptr);
 
 half fp32_2_fp16(float fp32_num);
 float fp16_2_fp32(half fp16_num);

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/fpga/filter.cpp

@@ -21,7 +21,10 @@ namespace paddle_mobile {
 namespace fpga {
 namespace filter {
 
-int calc_division_capacity(int chw) { return 2048 / ((chw + 15) / 16) * 32; }
+int calc_division_capacity(int chw) {
+  int n = 2048 / ((chw + 15) / 16) * 32;
+  return n < 2048 ? n : 2048;
+}
 
 int calc_split_num(int num, int division_capacity) {
   return (num + division_capacity - 1) / division_capacity;

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
@@ -199,6 +202,9 @@ LOAD_OP3(pool2d, CPU, MALI_GPU, FPGA);
 #ifdef MULTICLASSNMS_OP
 LOAD_OP1(multiclass_nms, CPU);
 #endif
+#ifdef POLYGONBOXTRANSFORM_OP
+LOAD_OP1(polygon_box_transform, CPU);
+#endif
 #ifdef SUM_OP
 LOAD_OP1(sum, CPU);
 #endif

src/framework/operator.cpp

@@ -32,7 +32,7 @@ template <typename Dtype>
 vector<string> OperatorBase<Dtype>::GetInputKeys() const {
   auto it = op_input_output_key.find(type_);
   if (it == op_input_output_key.end()) {
-    DLOG << type_ << " has no outputs";
+    DLOG << type_ << " has no inputs";
     return {};
   }
   return it->second.first;

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

@@ -338,10 +338,14 @@ inline Print &operator<<(Print &printer, const Tensor &tensor) {
   for (int i = 0; i < tensor.numel(); i += stride) {
     if (tensor.type() == typeid(float)) {
       printer << tensor.data<float>()[i] << " ";
+    } else if (tensor.type() == typeid(int32_t)) {
+      printer << tensor.data<int32_t>()[i] << " ";
     } else if (tensor.type() == typeid(int64_t)) {
       printer << tensor.data<int64_t>()[i] << " ";
     } else if (tensor.type() == typeid(int8_t)) {
-      printer << 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/api_paddle_mobile.cc

@@ -29,7 +29,14 @@ PaddleMobilePredictor<Dtype, P>::PaddleMobilePredictor(
 template <typename Dtype, Precision P>
 bool PaddleMobilePredictor<Dtype, P>::Init(const PaddleMobileConfig &config) {
   paddle_mobile_.reset(new PaddleMobile<Dtype, P>());
-  if (!config.model_dir.empty()) {
+
+  if (config.memory_pack.from_memory) {
+    DLOG << "load from memory!";
+    paddle_mobile_->LoadCombinedMemory(config.memory_pack.model_size,
+                                       config.memory_pack.model_buf,
+                                       config.memory_pack.combined_params_size,
+                                       config.memory_pack.combined_params_buf);
+  } else if (!config.model_dir.empty()) {
     paddle_mobile_->Load(config.model_dir, config.optimize,
                          config.quantification, config.batch_size);
   } else if (!config.prog_file.empty() && !config.param_file.empty()) {

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/io/paddle_inference_api.h

@@ -111,6 +111,14 @@ class PaddlePredictor {
   PaddlePredictor() = default;
 };
 
+struct PaddleModelMemoryPack {
+  bool from_memory = false;
+  size_t model_size = 0;
+  uint8_t* model_buf = nullptr;
+  size_t combined_params_size = 0;
+  uint8_t* combined_params_buf = nullptr;
+};
+
 struct PaddleMobileConfig : public PaddlePredictor::Config {
   enum Precision { FP32 = 0 };
   enum Device { kCPU = 0, kFPGA = 1, kGPU_MALI = 2 };
@@ -124,6 +132,7 @@ struct PaddleMobileConfig : public PaddlePredictor::Config {
   int thread_num = 1;
   std::string prog_file;
   std::string param_file;
+  struct PaddleModelMemoryPack memory_pack;
 };
 
 // A factory to help create different predictors.

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/mul_kernel.cpp

@@ -31,6 +31,8 @@ void MulKernel<CPU, float>::Compute(const MulParam<CPU> &param) const {
   param.Out()->set_lod(param.InputX()->lod());
 }
 
+template class MulKernel<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/kernel/fpga/mul_kernel.cpp → src/operators/kernel/arm/polygon_box_transform_kernel.cpp

@@ -12,57 +12,24 @@ 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 MUL_OP
+#ifdef POLYGONBOXTRANSFORM_OP
 
-#include "operators/kernel/mul_kernel.h"
+#include "operators/kernel/polygon_box_transform_kernel.h"
+#include "operators/kernel/central-arm-func/polygon_box_transform_arm_func.h"
 
 namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool MulKernel<FPGA, float>::Init(MulParam<FPGA> *param) {
-  bool relu_enabled = false;
-  auto input_x = const_cast<LoDTensor *>(param->InputX());
-  auto filter = const_cast<LoDTensor *>(param->InputY());
-  auto out = param->Out();
-
-  PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == filter->dims()[0],
-                        "Image channel should be equal to weight number");
-  int channel = (uint32_t)out->dims()[1];
-  auto bs_ptr =
-      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // NOLINT
-  for (int i = 0; i < channel; i++) {
-    bs_ptr[i + channel] = 1;
-    bs_ptr[i] = 0;
-  }
-  int num = (uint32_t)filter->dims()[1];
-  int chw = (uint32_t)filter->dims()[0];
-  PADDLE_MOBILE_ENFORCE(
-      chw == input_x->numel(),
-      "Filter element num should be equal to IFM element num");
-  int height = (uint32_t)input_x->dims()[2];
-  int width = (uint32_t)input_x->dims()[3];
-  int filter_channel = chw / height / width;
-
-  out->Resize(framework::make_ddim({1, channel, 1, 1}));
-  filter->Resize(framework::make_ddim({num, filter_channel, height, width}));
-  float max_value = fpga::filter_find_max(filter);
-  fpga::format_fc_filter(filter, max_value);
-
-  int element_num_per_div = fpga::get_filter_num_per_div(filter, 1);
-  fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
-  fpga::format_fp16_ofm(out);
-
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,
-                      0, bs_ptr);
-  param->SetFpgaArgs(conv_arg);
+bool PolygonBoxTransformKernel<CPU, float>::Init(
+    PolygonBoxTransformParam<CPU> *param) {
   return true;
 }
 
 template <>
-void MulKernel<FPGA, float>::Compute(const MulParam<FPGA> &param) const {
-  fpga::ComputeFpgaConv(param.FpgaArgs());
+void PolygonBoxTransformKernel<CPU, float>::Compute(
+    const PolygonBoxTransformParam<CPU> &param) const {
+  PolygonBoxTransformCompute<float>(param);
 }
 
 }  // namespace operators

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/mul_arm_func.h

@@ -58,7 +58,7 @@ void MulCompute(const MulParam<CPU> &param) {
   const Tensor *input_x = param.InputX();
   const Tensor *input_y = param.InputY();
   Tensor *out = param.Out();
-  out->mutable_data<float>();
+
   const Tensor x_matrix =
       input_x->dims().size() > 2
           ? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
@@ -71,15 +71,21 @@ void MulCompute(const MulParam<CPU> &param) {
   if (out_dim.size() != 2) {
     out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
   }
-  math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
-                      out, static_cast<float>(0));
+  if (param.InputX()->type() == typeid(int8_t)) {
+    out->mutable_data<int32_t>();
+    math::matmul<int8_t>(x_matrix, false, y_matrix, false,
+                         static_cast<int8_t>(1), out, static_cast<int8_t>(0));
+
+  } else {
+    out->mutable_data<float>();
+    math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
+                        out, static_cast<float>(0));
+  }
   if (out_dim.size() != 2) {
     out->Resize(out_dim);
   }
 }
 
-template class MulKernel<CPU, float>;
-
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/kernel/central-arm-func/polygon_box_transform_arm_func.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 POLYGONBOXTRANSFORM_OP
+#pragma once
+
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename P>
+void PolygonBoxTransformCompute(const PolygonBoxTransformParam<CPU>& param) {
+  const auto* input = param.Input();
+  const auto& input_dims = input->dims();
+  const auto* input_data = input->data<float>();
+  auto* output = param.Output();
+  auto* output_data = output->mutable_data<float>();
+
+  int64_t batch_size = input_dims[0];
+  int64_t geo_channel = input_dims[1];
+  int64_t height = input_dims[2];
+  int64_t width = input_dims[3];
+  int64_t id = 0;
+  for (int64_t id_n = 0; id_n < batch_size * geo_channel; ++id_n) {
+    for (int64_t id_h = 0; id_h < height; ++id_h) {
+      for (int64_t id_w = 0; id_w < width; ++id_w) {
+        id = id_n * height * width + width * id_h + id_w;
+        if (id_n % 2 == 0) {
+          output_data[id] = id_w * 4 - input_data[id];
+        } else {
+          output_data[id] = id_h * 4 - input_data[id];
+        }
+      }
+    }
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

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/fpga/conv_add_bn_kernel.cpp

@@ -66,10 +66,11 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
-                      param->Groups(), param->Strides()[0], param->Strides()[1],
-                      param->Paddings()[0], param->Paddings()[1], bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
+                       param->Groups(), param->Strides()[0],
+                       param->Strides()[1], param->Paddings()[0],
+                       param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
 
   return true;

src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp

@@ -65,10 +65,11 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
 
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
-                      param->Groups(), param->Strides()[0], param->Strides()[1],
-                      param->Paddings()[0], param->Paddings()[1], bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
+                       param->Groups(), param->Strides()[0],
+                       param->Strides()[1], param->Paddings()[0],
+                       param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }

src/operators/kernel/fpga/conv_add_relu_kernel.cpp

@@ -47,10 +47,11 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
 
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
-                      param->Groups(), param->Strides()[0], param->Strides()[1],
-                      param->Paddings()[0], param->Paddings()[1], bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
+                       param->Groups(), param->Strides()[0],
+                       param->Strides()[1], param->Paddings()[0],
+                       param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }

src/operators/kernel/fpga/conv_bn_kernel.cpp

@@ -59,10 +59,11 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
 
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
-                      param->Groups(), param->Strides()[0], param->Strides()[1],
-                      param->Paddings()[0], param->Paddings()[1], bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
+                       param->Groups(), param->Strides()[0],
+                       param->Strides()[1], param->Paddings()[0],
+                       param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }

src/operators/kernel/fpga/conv_bn_relu_kernel.cpp

@@ -59,10 +59,11 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
 
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
-                      param->Groups(), param->Strides()[0], param->Strides()[1],
-                      param->Paddings()[0], param->Paddings()[1], bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
+                       param->Groups(), param->Strides()[0],
+                       param->Strides()[1], param->Paddings()[0],
+                       param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -53,9 +53,9 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,
-                      0, bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1,
+                       0, 0, bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -54,9 +54,9 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
   fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg = {0};
-  fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,
-                      0, bs_ptr);
+  fpga::SplitConvArgs conv_arg = {0};
+  fpga::fill_split_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1,
+                       0, 0, bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }

src/operators/kernel/polygon_box_transform_kernel.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 POLYGONBOXTRANSFORM_OP
+
+#pragma once
+
+#include "framework/operator.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename DeviceType, typename T>
+class PolygonBoxTransformKernel
+    : public framework::OpKernelBase<DeviceType,
+                                     PolygonBoxTransformParam<DeviceType>> {
+ public:
+  void Compute(const PolygonBoxTransformParam<DeviceType>& param) const;
+  bool Init(PolygonBoxTransformParam<DeviceType>* param);
+};
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

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;
@@ -3662,7 +3662,7 @@ void Gemm::AddDot6x8(int k, const float *a, const float *b, float *c, int ldc) {
   b_ptr = b;
   int kc1 = k / 8;
   int kc2 = k % 8;
-  int step = 4 * ldc;
+  int step = sizeof(float) * ldc;
   asm volatile(
       "pld        [%[a_ptr]]            \n\t"
       "pld        [%[a_ptr],  #64]      \n\t"
@@ -3866,11 +3866,10 @@ void Gemm::AddDot6x8(int k, const float *a, const float *b, float *c, int ldc) {
       :
       : [a_ptr] "r"(a_ptr), [b_ptr] "r"(b_ptr), [c] "r"(c), [kc1] "r"(kc1),
         [kc2] "r"(kc2), [step] "r"(step)
-      : "memory", "r5", "r6", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-        "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
+      : "cc", "memory", "r5", "r6", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
+        "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
 
 #endif  // __aarch64__
-#else
 
 #endif  // __ARM_NEON
 }

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
@@ -96,6 +98,7 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
   void InnerKernelWithPRelu(int mc, int nc, const float *a, const float *b,
                             float *c, float *C, int ldc, float *p,
                             std::string mode, float *bias, float *bias1);
+
   /*
   // 向量矩阵乘法 (M = 1)
   void VectorKernel(int m, int n, int k, float alpha, const float *A, int lda,
@@ -139,6 +142,7 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
                        float *new_scale, float *new_bias);
   void WriteWithBnAddRelu(int mc, int nc, float *c, float *C, int ldc,
                           float *new_scale, float *new_bias, float *bias1);
+
   /*
   // 向量矩阵乘法结果回写
   // C = A * B
@@ -185,15 +189,67 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
                           const float *B, int ldb, float *C, int ldc, float *p,
                           std::string mode, float *bias, float *bias1);
 
+  // 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);
+
+  // 8 bits int inner product
+  void InnerKernelWithBias(int32_t mc, int32_t nc, int8_t alpha,
+                           const int8_t *a, const int8_t *b, int8_t beta,
+                           int32_t *c, int32_t *C, int32_t ldc, bool relu,
+                           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,
+                      int32_t ldb, int8_t *buffer);
+
+  // 8 bits int matrix product
+  void Sgemm(int32_t m, int32_t n, int32_t k, int8_t alpha, const int8_t *A,
+             int32_t lda, const int8_t *B, int32_t ldb, int8_t beta, int32_t *C,
+             int32_t ldc, bool relu, int8_t *bias);
+
+  // 8 bits int write back
+  // C = alpha * A * B + beta * C
+  void WriteWithAlphaBeta(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
+                          int32_t ldc);
+  // C = A * B
+  void WriteBasic(int32_t mc, int32_t nc, int32_t *c, int32_t *C, int32_t ldc);
+  // C = A * B + C
+  void WriteWithAdd(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
+                    int32_t ldc);
+  // C = A * B + bias
+  void WriteWithAddV1(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
+                      int32_t ldc, int8_t *bias);
+  // C = A * B + C, relu(C)
+  void WriteWithAddRelu(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
+                        int32_t ldc);
+  // C = A * B + bias, relu(C)
+  void WriteWithAddReluV1(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
+                          int32_t ldc, int8_t *bias);
+
  private:
   int MC = 0;
   int KC = 0;
   int NC = 0;
 
+  // 32位 float
   float *packedA;
   float *packedB;
   float *packedC;
   float *zero;
+
+  // 8 bits int
+  int8_t *packedA_int8;
+  int8_t *packedB_int8;
+  int32_t *packedC_int8;
+  int8_t *zero_int8;
 };
 
 }  // namespace math

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,
@@ -135,7 +154,7 @@ template <typename T>
 struct ClearTensor<CPU, T> {
   void operator()(framework::Tensor *tensor) {
     auto size = tensor->numel();
-    auto *tensor_data = tensor->data<float>();
+    auto *tensor_data = tensor->data<T>();
     memset((void *)tensor_data, 0, sizeof(T) * size);  // NOLINT
   }
 };
@@ -151,9 +170,9 @@ struct RowwiseAdd<CPU, T> {
     PADDLE_MOBILE_ENFORCE((output->dims() == in_dims),
                           "output->dims() must be equal to in_dims.");
 
-    auto *input_data = input.data<float>();
-    auto *out_data = output->data<float>();
-    auto *vec_data = vector.data<float>();
+    auto *input_data = input.data<T>();
+    auto *out_data = output->data<T>();
+    auto *vec_data = vector.data<T>();
     for (int64_t i = 0; i < in_dims[0]; ++i) {
       for (int64_t j = 0; j < size; ++j) {
         out_data[i * size + j] = input_data[i * size + j] + vec_data[j];

src/operators/math/math_function.h

@@ -15,17 +15,20 @@ 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,
             framework::Tensor *matrix_out, T beta, bool relu = false,
-            float *bias = nullptr);
+            T *bias = nullptr);
 
 template <typename T>
 void matmulWithBn(const framework::Tensor &matrix_a, bool trans_a,

src/operators/math/math_function_int8.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 <cstring>
+#include <string>
+#include "operators/math/gemm.h"
+#include "operators/math/math_function.h"
+
+namespace paddle_mobile {
+namespace operators {
+namespace math {
+template <>
+void matmul<int8_t>(const framework::Tensor &matrix_a, bool trans_a,
+                    const framework::Tensor &matrix_b, bool trans_b,
+                    int8_t alpha, framework::Tensor *matrix_out, int8_t beta,
+                    bool relu, int8_t *bias) {
+  auto dim_a = matrix_a.dims();
+  auto dim_b = matrix_b.dims();
+  auto dim_out = matrix_out->dims();
+  PADDLE_MOBILE_ENFORCE(
+      dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+      "The input and output of matmul be matrix");
+
+  int32_t M = dim_out[0];
+  int32_t N = dim_out[1];
+  int32_t K = (!trans_a) ? dim_a[1] : dim_a[0];
+  Gemm gemm;
+
+  if (trans_a) {
+    int32_t numel = matrix_a.numel();
+    int32_t m = matrix_a.dims()[0];
+    int32_t n = matrix_a.dims()[1];
+    int8_t *tmp = (int8_t *)(matrix_a.data<int8_t>());  // NOLINT
+    int8_t *a = static_cast<int8_t *>(
+        paddle_mobile::memory::Alloc(sizeof(int8_t) * numel));
+    int32_t index = 0;
+    for (int32_t j = 0; j < n; j++) {
+      for (int32_t i = 0; i < m; i++) {
+        a[index++] = tmp[i * n + j];
+      }
+    }
+
+    gemm.Sgemm(M, N, K, alpha, a, K, matrix_b.data<int8_t>(), N, beta,
+               matrix_out->data<int32_t>(), N, relu, bias);
+  } else {
+    gemm.Sgemm(M, N, K, alpha, matrix_a.data<int8_t>(), K,
+               matrix_b.data<int8_t>(), N, beta, matrix_out->data<int32_t>(), N,
+               relu, bias);
+  }
+}
+}  // namespace math
+}  // 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