Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 337
Star 4
Fork 1
P
Paddle-Lite
提交 1d76df93 编写于 作者: R Ray Liu 提交者： GitHub
浏览文件
操作

Merge pull request #1048 from hjchen2/dev-latest 

Add quantize and dequantize op and fix some code style
上级 a9c249d3 e789c61e
展开全部 隐藏空白更改
内联 并排
Showing with 1760 addition and 1566 deletion
CMakeLists.txt
浏览文件 @ 1d76df93
cmake_minimum_required(VERSION 3.0)
option(USE_OPENMP "openmp support" OFF)
project(paddle-mobile)
# select the platform to build
option(CPU "armv7 with neon support" ON)
option(MALI_GPU "mali gpu support" OFF)
option(FPGA "fpga support" OFF)
option(USE_OPENMP "openmp support" OFF)
option(DEBUGING "enable debug mode" ON)
option(USE_EXCEPTION "use std exception" OFF)
option(LOG_PROFILE "log profile" OFF)
# select the platform to build
option(CPU "armv7 with neon" ON)
option(MALI_GPU "mali gpu" OFF)
option(FPGA "fpga" OFF)
file(GLOB_RECURSE PADDLE_MOBILE_CC src/*.cc src/*.cpp src/*.c src/*.mm)
file(GLOB_RECURSE PADDLE_MOBILE_H src/*.h)
include_directories(src/)
set(CMAKE_BUILD_TYPE Release)
add_definitions(-O3 -s -DNDEBUG)
if(IS_IOS)
set(CMAKE_CXX_FLAGS "-mfpu=neon -marm -fobjc-abi-version=2 -fobjc-arc -std=gnu++11 -stdlib=libc++ -O3 -s -isysroot ${CMAKE_OSX_SYSROOT} ${CMAKE_CXX_FLAGS}")
set(CMAKE_CXX_FLAGS "-mfpu=neon -marm -fobjc-abi-version=2 -fobjc-arc \
-std=gnu++11 -stdlib=libc++ -isysroot ${CMAKE_OSX_SYSROOT} ${CMAKE_CXX_FLAGS}")
else()
set(CMAKE_CXX_FLAGS "-std=c++14 -O3 -s ${CMAKE_CXX_FLAGS}")
set(CMAKE_CXX_FLAGS "-std=c++11 ${CMAKE_CXX_FLAGS}")
endif()
if (DEBUGING)
message(STATUS "debug")
set(CMAKE_BUILD_TYPE Release)
set(CMAKE_CXX_FLAGS_RELEASE "-DNDEBUG")
if(DEBUGING)
message(STATUS "debugging mode")
add_definitions(-DPADDLE_MOBILE_DEBUG)
else ()
set(CMAKE_BUILD_TYPE Release)
set(CMAKE_CXX_FLAGS_RELEASE "-DNDEBUG")
else()
add_definitions(-fvisibility=hidden -fvisibility-inlines-hidden)
endif ()
endif()
if (USE_EXCEPTION)
if(USE_EXCEPTION)
message(STATUS "use exception")
add_definitions(-DENABLE_EXCEPTION)
add_definitions(-fexceptions)
add_definitions(-DENABLE_EXCEPTION -fexceptions)
else()
add_definitions(-fno-exceptions)
endif ()
endif()
if (LOG_PROFILE)
if(LOG_PROFILE)
add_definitions(-DPADDLE_MOBILE_PROFILE)
endif()
......@@ -50,12 +47,12 @@ if(USE_OPENMP)
endif()
# platform control
if (ARM_LINUX)
if(ARM_LINUX)
include("${CMAKE_CURRENT_LIST_DIR}/tools/arm-platform.cmake")
endif ()
endif()
if (CPU)
add_definitions(-DPADDLE_MOBILE_CPU)
if(CPU)
add_definitions(-DPADDLE_MOBILE_CPU)
else()
file(GLOB_RECURSE _tmp_list src/operators/kernel/arm/*.cpp src/operators/kernel/arm/*.cc)
foreach(f ${_tmp_list})
......@@ -68,7 +65,7 @@ else()
endforeach()
endif()
if (MALI_GPU)
if(MALI_GPU)
add_definitions(-DPADDLE_MOBILE_MALI_GPU)
add_definitions(-DUSE_ACL=1)
add_definitions(-DUSE_OPENCL)
......@@ -120,20 +117,20 @@ else()
endforeach()
endif()
if (ANDROID_NDK_TOOLCHAIN_INCLUDED)
if(ANDROID_NDK_TOOLCHAIN_INCLUDED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -llog")
else()
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/jni/paddle_mobile_jni.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/jni/paddle_mobile_jni.cpp)
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/operators/math/math_func_neon.h)
endif ()
endif()
if (IS_IOS)
if(IS_IOS)
else()
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobileCPU.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobileCPU.mm)
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/op_symbols.h)
endif ()
endif()
set(CMAKE_VERBOSE_MAKEFILE ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
......@@ -142,7 +139,7 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY build)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY build)
# NET default
if (FPGA)
if(FPGA)
set(NET "FPGAnets" CACHE STRING "select net type")
else()
set(NET "default" CACHE STRING "select net type")
......@@ -153,7 +150,7 @@ include("${CMAKE_CURRENT_LIST_DIR}/tools/op.cmake")
# build library
if (ANDROID_NDK_TOOLCHAIN_INCLUDED)
if(ANDROID_NDK_TOOLCHAIN_INCLUDED)
list(REMOVE_DUPLICATES CMAKE_CXX_FLAGS)
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
elseif(IS_IOS)
......@@ -168,9 +165,9 @@ elseif(IS_IOS)
else()
add_library(paddle-mobile STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
endif()
else ()
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
endif ()
else()
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
endif()
# unit test
if(DEBUGING)
......
src/common/dep_core.h
浏览文件 @ 1d76df93
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#ifdef PADDLE_EXECUTOR_MULTITHREAD
#include <string>
#include <unordered_map>
......@@ -60,6 +61,7 @@ class depCore {
std::vector<std::vector<int>> deps;
std::vector<std::vector<int>> next;
};
} // namespace paddle_mobile
#endif
src/common/types.cpp
浏览文件 @ 1d76df93
......@@ -63,6 +63,9 @@ const char *G_OP_TYPE_BILINEAR_INTERP = "bilinear_interp";
const char *G_OP_TYPE_FLATTEN = "flatten";
const char *G_OP_TYPE_SHAPE = "shape";
const char *G_OP_TYPE_QUANTIZE = "quantize";
const char *G_OP_TYPE_DEQUANTIZE = "dequantize";
std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
op_input_output_key = {
......@@ -111,6 +114,8 @@ std::unordered_map<
{G_OP_TYPE_BILINEAR_INTERP, {{"OutSize", "X"}, {"Out"}}},
{G_OP_TYPE_FLATTEN, {{"X"}, {"Out"}}},
{G_OP_TYPE_SHAPE, {{"Input"}, {"Out"}}},
{G_OP_TYPE_CONV_TRANSPOSE, {{"Input"}, {"Output"}}}};
{G_OP_TYPE_CONV_TRANSPOSE, {{"Input"}, {"Output"}}},
{G_OP_TYPE_QUANTIZE, {{"X"}, {"Out", "OutScale"}}},
{G_OP_TYPE_DEQUANTIZE, {{"X", "Scale"}, {"Out"}}}};
} // namespace paddle_mobile
src/common/types.h
浏览文件 @ 1d76df93
......@@ -79,6 +79,13 @@ enum PMStatus {
PMWrongDevice = 0x08 /*!< un-correct device. */
};
enum RoundType {
ROUND_UNK = 0,
ROUND_NEAREST_AWAY_ZERO = 1,
ROUND_NEAREST_TOWARDS_ZERO = 2,
ROUND_NEAREST_TO_EVEN = 3
};
extern const char *G_OP_TYPE_CONV;
extern const char *G_OP_TYPE_BATCHNORM;
extern const char *G_OP_TYPE_BOX_CODER;
......@@ -120,6 +127,9 @@ extern const char *G_OP_TYPE_FUSION_CONV_BN;
extern const char *G_OP_TYPE_CONV_TRANSPOSE;
extern const char *G_OP_TYPE_PRELU;
extern const char *G_OP_TYPE_QUANTIZE;
extern const char *G_OP_TYPE_DEQUANTIZE;
extern std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
op_input_output_key;
......
src/common/util.cpp 0 → 100644
浏览文件 @ 1d76df93
/* 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 "common/util.h"
namespace paddle_mobile {
char *ReadFileToBuff(std::string filename) {
FILE *file = fopen(filename.c_str(), "rb");
PADDLE_MOBILE_ENFORCE(file != nullptr, "can't open file: %s ",
filename.c_str());
fseek(file, 0, SEEK_END);
int64_t size = ftell(file);
PADDLE_MOBILE_ENFORCE(size > 0, "file should not be empty");
rewind(file);
char *data = new char[size];
size_t bytes_read = fread(data, 1, size, file);
PADDLE_MOBILE_ENFORCE(bytes_read == size,
"read binary file bytes do not match with fseek");
fclose(file);
return data;
}
} // namespace paddle_mobile
src/common/util.h 0 → 100644
浏览文件 @ 1d76df93
/* 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 "common/enforce.h"
namespace paddle_mobile {
char *ReadFileToBuff(std::string filename);
} // namespace paddle_mobile
src/fpga/api.cpp
浏览文件 @ 1d76df93
......@@ -12,15 +12,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "api.h"
#include "fpga/api.h"
#include <fcntl.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <algorithm>
#include <map>
#include "bias_scale.h"
#include "filter.h"
#include "image.h"
#include "fpga/bias_scale.h"
#include "fpga/filter.h"
#include "fpga/image.h"
#define FPGA_TEST_MODE
#define PADDLE_MOBILE_OS_LINUX
......@@ -263,7 +263,7 @@ void format_image(framework::Tensor *image_tensor) {
auto channel = dims[1], height = dims[2], width = dims[3];
auto data_ptr = image_tensor->data<float>();
size_t memory_size = channel * height * width * sizeof(float);
auto new_data = (float *)fpga_malloc(memory_size);
auto new_data = (float *)fpga_malloc(memory_size); // NOLINT
fpga_copy(new_data, data_ptr, memory_size);
image::format_image(&new_data, channel, height, width);
image_tensor->reset_data_ptr(new_data);
......@@ -334,13 +334,13 @@ int get_aligned_filter_num(int num) {
void format_filter(framework::Tensor *filter_tensor, float max_value,
int group_num) {
filter_tensor->scale[0] = float(max_value / 127.0);
filter_tensor->scale[1] = float(127.0 / max_value);
filter_tensor->scale[0] = float(max_value / 127.0); // NOLINT
filter_tensor->scale[1] = float(127.0 / max_value); // NOLINT
auto dims = filter_tensor->dims();
auto num = dims[0], channel = dims[1], height = dims[2], width = dims[3];
auto data_ptr = filter_tensor->data<float>();
size_t memory_size = num * channel * height * width * sizeof(float);
auto new_data = (float *)fpga_malloc(memory_size);
auto new_data = (float *)fpga_malloc(memory_size); // NOLINT
fpga_copy(new_data, data_ptr, memory_size);
filter::format_filter(&new_data, num, channel, height, width, group_num,
max_value);
......@@ -381,7 +381,8 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
arg->filter_num = (uint32_t)filter->dims()[0];
arg->output.address = out_ptr;
arg->output.scale_address = out->scale;
arg->conv_args = (ConvArgs *)fpga_malloc(arg->split_num * sizeof(ConvArgs));
arg->conv_args =
(ConvArgs *)fpga_malloc(arg->split_num * sizeof(ConvArgs)); // NOLINT
arg->concat_arg.image_num = arg->split_num;
arg->concat_arg.image_out = out_ptr;
......@@ -390,12 +391,15 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
arg->concat_arg.width = (uint32_t)filter->dims()[3];
int n = arg->split_num;
arg->concat_arg.images_in = (half **)fpga_malloc(n * sizeof(int *));
arg->concat_arg.scales_in = (float **)fpga_malloc(n * sizeof(float *));
arg->concat_arg.channel_num = (uint32_t *)fpga_malloc(n * sizeof(uint32_t));
arg->concat_arg.images_in =
(half **)fpga_malloc(n * sizeof(int *)); // NOLINT
arg->concat_arg.scales_in =
(float **)fpga_malloc(n * sizeof(float *)); // NOLINT
arg->concat_arg.channel_num =
(uint32_t *)fpga_malloc(n * sizeof(uint32_t)); // NOLINT
arg->concat_arg.image_out = out_ptr;
auto channel = (int)out->dims()[1];
auto channel = (int)out->dims()[1]; // NOLINT
int filter_num_per_div = get_filter_num_per_div(filter, group_num);
int element_num = get_aligned_filter_element_num(
filter->dims()[1] * filter->dims()[2] * filter->dims()[3]);
......@@ -415,29 +419,28 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
arg->conv_args[i].image.pad_height = (uint32_t)padding_h;
arg->conv_args[i].image.pad_width = (uint32_t)padding_w;
arg->conv_args[i].filter_scale_address = filter->scale;
arg->conv_args[i].filter_address =
&((int8_t *)filter_ptr)[i * element_num * filter_num_per_div];
arg->conv_args[i].filter_address = &(
(int8_t *)filter_ptr)[i * element_num * filter_num_per_div]; // NOLINT
arg->conv_args[i].sb_address = &bs_ptr[i * filter_num_per_div * 2];
arg->conv_args[i].filter_num =
(uint32_t)(i == n - 1 ? channel - (n - 1) * filter_num_per_div
: filter_num_per_div);
arg->conv_args[i].filter_num = (uint32_t)(
i == n - 1 ? channel - (n - 1) * filter_num_per_div // NOLINT
: filter_num_per_div);
if (n > 1) {
arg->conv_args[i].output.scale_address =
(float *)fpga_malloc(2 * sizeof(float));
(float *)fpga_malloc(2 * sizeof(float)); // NOLINT
arg->conv_args[i].output.address = fpga_malloc(
input->dims()[2] *
align_to_x(input->dims()[3] * arg->conv_args[i].filter_num,
IMAGE_ALIGNMENT) *
sizeof(half));
}
else {
} else {
arg->conv_args[i].output.scale_address = out->scale;
arg->conv_args[i].output.address = out_ptr;
}
arg->concat_arg.images_in[i] = (half *)arg->conv_args[i].output.address;
arg->concat_arg.images_in[i] =
(half *)arg->conv_args[i].output.address; // NOLINT
arg->concat_arg.scales_in[i] = arg->conv_args[i].output.scale_address;
arg->concat_arg.channel_num[i] = arg->conv_args[i].filter_num;
}
......
src/fpga/bias_scale.cpp
浏览文件 @ 1d76df93
......@@ -12,9 +12,9 @@ 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 "bias_scale.h"
#include "fpga/bias_scale.h"
#include <memory.h>
#include "api.h"
#include "fpga/api.h"
namespace paddle_mobile {
namespace fpga {
......@@ -29,7 +29,8 @@ void align_element(float **data_in, int num_per_div_before_alignment, int num) {
align_to_x(num_per_div_before_alignment, BS_NUM_ALIGNMENT);
int num_element =
2 * div_num * num_per_div_after_alignment; // including bias & scale
float *ptr_aligned = (float *)fpga_malloc(num_element * sizeof(float));
float *ptr_aligned =
(float *)fpga_malloc(num_element * sizeof(float)); // NOLINT
memset(ptr_aligned, 0, num_element * sizeof(float));
......@@ -59,7 +60,7 @@ void interleave(float **data_in, int num_after_alignment) {
float *ptr_uninterleaved = *data_in;
float *ptr_interleaved =
(float *)fpga_malloc(2 * num_after_alignment * sizeof(float));
(float *)fpga_malloc(2 * num_after_alignment * sizeof(float)); // NOLINT
int num = num_after_alignment / 4;
for (int i = 0; i < num; i++) {
memcpy(ptr_interleaved + 8 * i, ptr_uninterleaved + 4 * i,
......
src/fpga/filter.cpp
浏览文件 @ 1d76df93
......@@ -11,9 +11,11 @@ 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 "filter.h"
#include "fpga/filter.h"
#include <memory.h>
#include "api.h"
#include <algorithm>
#include "fpga/api.h"
namespace paddle_mobile {
namespace fpga {
......@@ -55,7 +57,7 @@ void convert_to_hwc(char **data_in, int num, int channel, int height,
int width) {
char *tmp = *data_in;
int chw = channel * height * width;
char *data_tmp = (char *)fpga_malloc(chw * num * sizeof(char));
char *data_tmp = (char *)fpga_malloc(chw * num * sizeof(char)); // NOLINT
for (int n = 0; n < num; n++) {
int64_t amount_per_row = width * channel;
for (int c = 0; c < channel; c++) {
......@@ -102,7 +104,7 @@ void quantize(float **data_in, int data_size, float max) {
tmp_data[i] = float_to_int8(
(*data_in)[i] * scale); // (signed char)((*data_in)[i] * scale);
}
*data_in = (float *)tmp_data;
*data_in = (float *)tmp_data; // NOLINT
fpga_free(tmp);
}
......@@ -112,7 +114,8 @@ void align_element(char **data_in, int num, int chw) {
int align_chw = align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
if (align_chw != chw) {
char *tmp = *data_in;
char *data_tmp = (char *)fpga_malloc(num * align_chw * sizeof(char));
char *data_tmp =
(char *)fpga_malloc(num * align_chw * sizeof(char)); // NOLINT
memset(data_tmp, 0, num * align_chw);
for (j = 0; j < num; j++) {
......@@ -134,7 +137,7 @@ void align_num(char **data_in, int num_per_div_before_alignment, int num,
int div_num =
(num + num_per_div_before_alignment - 1) / num_per_div_before_alignment;
int num_element = div_num * num_per_div_after_alignment * align_chw;
char *data_tmp = (char *)fpga_malloc(num_element * sizeof(char));
char *data_tmp = (char *)fpga_malloc(num_element * sizeof(char)); // NOLINT
memset(data_tmp, 0, num_element * sizeof(char));
......@@ -156,7 +159,8 @@ void reorder(char **data_in, int num_after_alignment, int chw) {
int chw_align = align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
char *data_tmp =
(char *)fpga_malloc(chw_align * num_after_alignment * sizeof(char));
(char *)fpga_malloc(chw_align * num_after_alignment * // NOLINT
sizeof(char));
char *tmp = *data_in;
for (index = 0; index < num_after_alignment; index++) {
new_index = index / 32 * 32 + (index % 16 / 4 * 8) + (index % 16 % 4) +
......@@ -173,10 +177,11 @@ void interleave(char **data_in, int num_after_alignment, int chw) {
int j = 0;
int k = 0;
int interleave_per_num = 16;
;
int chw_align = align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
char *data_tmp =
(char *)fpga_malloc(chw_align * num_after_alignment * sizeof(char));
(char *)fpga_malloc(chw_align * num_after_alignment * // NOLINT
sizeof(char));
char *tmp = *data_in;
int interleave_num = chw_align * 2 / interleave_per_num;
for (i = 0; i < num_after_alignment; i += 2) {
......@@ -209,7 +214,7 @@ void format_filter(float **data_in, int num, int channel, int height, int width,
quantize(data_in, data_size, max);
char **quantize_data = (char **)data_in;
char **quantize_data = (char **)data_in; // NOLINT
convert_to_hwc(quantize_data, num, channel, height, width);
align_element(quantize_data, num, chw);
......
src/fpga/image.cpp
浏览文件 @ 1d76df93
......@@ -12,9 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "image.h"
#include "fpga/image.h"
#include <memory.h>
#include "api.h"
#include <algorithm>
#include "fpga/api.h"
namespace paddle_mobile {
namespace fpga {
......@@ -23,7 +24,7 @@ namespace image {
void convert_to_hwc(float **data_in, int channel, int height, int width) {
float *tmp = *data_in;
float *data_tmp =
(float *)fpga_malloc(channel * height * width * sizeof(float));
(float *)fpga_malloc(channel * height * width * sizeof(float)); // NOLINT
int64_t amount_per_row = width * channel;
for (int c = 0; c < channel; c++) {
for (int h = 0; h < height; h++) {
......@@ -42,12 +43,14 @@ void align_element_conv(float **data_in, int height, int cw) {
int align_cw = align_to_x(cw, IMAGE_ALIGNMENT);
if (align_cw != cw) {
float *tmp = *data_in;
float *data_tmp = (float *)fpga_malloc(height * align_cw * sizeof(float));
float *data_tmp =
(float *)fpga_malloc(height * align_cw * sizeof(float)); // NOLINT
memset(data_tmp, 0, height * align_cw * sizeof(float));
for (h = 0; h < height; h++) {
memcpy((void *)(data_tmp + h * align_cw), (void *)(*data_in + h * cw),
memcpy((void *)(data_tmp + h * align_cw), // NOLINT
(void *)(*data_in + h * cw), // NOLINT
cw * sizeof(float));
}
......@@ -95,7 +98,7 @@ void concat_images(int16_t **images_in, float **scales_in, void *image_out,
for (i = 0; i < image_num; i++) {
align_each_in_area_cw =
align_to_x(channel_num[i] * width, IMAGE_ALIGNMENT);
memcpy((int16_t *)image_out + tmp_channel +
memcpy((int16_t *)image_out + tmp_channel + // NOLINT
k * align_each_out_area_cw_differ,
images_in[i] + j * channel_num[i] + k * align_each_in_area_cw,
channel_num[i] * sizeof(int16_t));
......
src/framework/framework.pb-c.c
浏览文件 @ 1d76df93
......@@ -7,6 +7,35 @@
#endif
#include "framework.pb-c.h"
void paddle_mobile__framework__proto__version__init(
PaddleMobile__Framework__Proto__Version *message) {
static const PaddleMobile__Framework__Proto__Version init_value =
PADDLE_MOBILE__FRAMEWORK__PROTO__VERSION__INIT;
*message = init_value;
}
size_t paddle_mobile__framework__proto__version__get_packed_size(
const PaddleMobile__Framework__Proto__Version *message) {
assert(message->base.descriptor ==
&paddle_mobile__framework__proto__version__descriptor);
return protobuf_c_message_get_packed_size(
(const ProtobufCMessage *)(message));
}
PaddleMobile__Framework__Proto__Version *
paddle_mobile__framework__proto__version__unpack(ProtobufCAllocator *allocator,
size_t len,
const uint8_t *data) {
return (PaddleMobile__Framework__Proto__Version *)protobuf_c_message_unpack(
&paddle_mobile__framework__proto__version__descriptor, allocator, len,
data);
}
void paddle_mobile__framework__proto__version__free_unpacked(
PaddleMobile__Framework__Proto__Version *message,
ProtobufCAllocator *allocator) {
if (!message) return;
assert(message->base.descriptor ==
&paddle_mobile__framework__proto__version__descriptor);
protobuf_c_message_free_unpacked((ProtobufCMessage *)message, allocator);
}
void paddle_mobile__framework__proto__op_desc__attr__init(
PaddleMobile__Framework__Proto__OpDesc__Attr *message) {
static const PaddleMobile__Framework__Proto__OpDesc__Attr init_value =
......@@ -32,7 +61,6 @@ size_t paddle_mobile__framework__proto__op_desc__get_packed_size(
return protobuf_c_message_get_packed_size(
(const ProtobufCMessage *)(message));
}
PaddleMobile__Framework__Proto__OpDesc *
paddle_mobile__framework__proto__op_desc__unpack(ProtobufCAllocator *allocator,
size_t len,
......@@ -74,7 +102,6 @@ size_t paddle_mobile__framework__proto__op_proto__get_packed_size(
return protobuf_c_message_get_packed_size(
(const ProtobufCMessage *)(message));
}
PaddleMobile__Framework__Proto__OpProto *
paddle_mobile__framework__proto__op_proto__unpack(ProtobufCAllocator *allocator,
size_t len,
......@@ -171,7 +198,6 @@ size_t paddle_mobile__framework__proto__var_desc__get_packed_size(
return protobuf_c_message_get_packed_size(
(const ProtobufCMessage *)(message));
}
PaddleMobile__Framework__Proto__VarDesc *
paddle_mobile__framework__proto__var_desc__unpack(ProtobufCAllocator *allocator,
size_t len,
......@@ -201,7 +227,6 @@ size_t paddle_mobile__framework__proto__block_desc__get_packed_size(
return protobuf_c_message_get_packed_size(
(const ProtobufCMessage *)(message));
}
PaddleMobile__Framework__Proto__BlockDesc *
paddle_mobile__framework__proto__block_desc__unpack(
ProtobufCAllocator *allocator, size_t len, const uint8_t *data) {
......@@ -230,7 +255,6 @@ size_t paddle_mobile__framework__proto__program_desc__get_packed_size(
return protobuf_c_message_get_packed_size(
(const ProtobufCMessage *)(message));
}
PaddleMobile__Framework__Proto__ProgramDesc *
paddle_mobile__framework__proto__program_desc__unpack(
ProtobufCAllocator *allocator, size_t len, const uint8_t *data) {
......@@ -247,8 +271,46 @@ void paddle_mobile__framework__proto__program_desc__free_unpacked(
&paddle_mobile__framework__proto__program_desc__descriptor);
protobuf_c_message_free_unpacked((ProtobufCMessage *)message, allocator);
}
static const int64_t
paddle_mobile__framework__proto__version__version__default_value = 0ll;
static const ProtobufCFieldDescriptor
paddle_mobile__framework__proto__op_desc__attr__field_descriptors[12] = {
paddle_mobile__framework__proto__version__field_descriptors[1] = {
{
"version", 1, PROTOBUF_C_LABEL_OPTIONAL, PROTOBUF_C_TYPE_INT64,
offsetof(PaddleMobile__Framework__Proto__Version, has_version),
offsetof(PaddleMobile__Framework__Proto__Version, version), NULL,
&paddle_mobile__framework__proto__version__version__default_value,
0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
};
static const unsigned
paddle_mobile__framework__proto__version__field_indices_by_name[] = {
0, /* field[0] = version */
};
static const ProtobufCIntRange
paddle_mobile__framework__proto__version__number_ranges[1 + 1] = {{1, 0},
{0, 1}};
const ProtobufCMessageDescriptor
paddle_mobile__framework__proto__version__descriptor = {
PROTOBUF_C__MESSAGE_DESCRIPTOR_MAGIC,
"paddle_mobile.framework.proto.Version",
"Version",
"PaddleMobile__Framework__Proto__Version",
"paddle_mobile.framework.proto",
sizeof(PaddleMobile__Framework__Proto__Version),
1,
paddle_mobile__framework__proto__version__field_descriptors,
paddle_mobile__framework__proto__version__field_indices_by_name,
1,
paddle_mobile__framework__proto__version__number_ranges,
(ProtobufCMessageInit)paddle_mobile__framework__proto__version__init,
NULL,
NULL,
NULL /* reserved[123] */
};
static const ProtobufCFieldDescriptor
paddle_mobile__framework__proto__op_desc__attr__field_descriptors[13] = {
{
"name", 1, PROTOBUF_C_LABEL_REQUIRED, PROTOBUF_C_TYPE_STRING,
0, /* quantifier_offset */
......@@ -335,11 +397,20 @@ static const ProtobufCFieldDescriptor
NULL, 0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
{
"blocks_idx", 14, PROTOBUF_C_LABEL_REPEATED, PROTOBUF_C_TYPE_INT32,
offsetof(PaddleMobile__Framework__Proto__OpDesc__Attr,
n_blocks_idx),
offsetof(PaddleMobile__Framework__Proto__OpDesc__Attr, blocks_idx),
NULL, NULL, 0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
};
static const unsigned
paddle_mobile__framework__proto__op_desc__attr__field_indices_by_name[] = {
8, /* field[8] = b */
10, /* field[10] = block_idx */
12, /* field[12] = blocks_idx */
9, /* field[9] = bools */
3, /* field[3] = f */
6, /* field[6] = floats */
......@@ -353,7 +424,7 @@ static const unsigned
};
static const ProtobufCIntRange
paddle_mobile__framework__proto__op_desc__attr__number_ranges[2 + 1] = {
{1, 0}, {10, 8}, {0, 12}};
{1, 0}, {10, 8}, {0, 13}};
const ProtobufCMessageDescriptor
paddle_mobile__framework__proto__op_desc__attr__descriptor = {
PROTOBUF_C__MESSAGE_DESCRIPTOR_MAGIC,
......@@ -362,7 +433,7 @@ const ProtobufCMessageDescriptor
"PaddleMobile__Framework__Proto__OpDesc__Attr",
"paddle_mobile.framework.proto",
sizeof(PaddleMobile__Framework__Proto__OpDesc__Attr),
12,
13,
paddle_mobile__framework__proto__op_desc__attr__field_descriptors,
paddle_mobile__framework__proto__op_desc__attr__field_indices_by_name,
2,
......@@ -500,7 +571,7 @@ static const protobuf_c_boolean
paddle_mobile__framework__proto__op_proto__var__dispensable__default_value =
0;
static const ProtobufCFieldDescriptor
paddle_mobile__framework__proto__op_proto__var__field_descriptors[5] = {
paddle_mobile__framework__proto__op_proto__var__field_descriptors[6] = {
{
"name", 1, PROTOBUF_C_LABEL_REQUIRED, PROTOBUF_C_TYPE_STRING,
0, /* quantifier_offset */
......@@ -546,6 +617,13 @@ static const ProtobufCFieldDescriptor
0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
{
"reuse", 6, PROTOBUF_C_LABEL_OPTIONAL, PROTOBUF_C_TYPE_STRING,
0, /* quantifier_offset */
offsetof(PaddleMobile__Framework__Proto__OpProto__Var, reuse), NULL,
NULL, 0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
};
static const unsigned
paddle_mobile__framework__proto__op_proto__var__field_indices_by_name[] = {
......@@ -554,10 +632,11 @@ static const unsigned
2, /* field[2] = duplicable */
3, /* field[3] = intermediate */
0, /* field[0] = name */
5, /* field[5] = reuse */
};
static const ProtobufCIntRange
paddle_mobile__framework__proto__op_proto__var__number_ranges[1 + 1] = {
{1, 0}, {0, 5}};
{1, 0}, {0, 6}};
const ProtobufCMessageDescriptor
paddle_mobile__framework__proto__op_proto__var__descriptor = {
PROTOBUF_C__MESSAGE_DESCRIPTOR_MAGIC,
......@@ -566,7 +645,7 @@ const ProtobufCMessageDescriptor
"PaddleMobile__Framework__Proto__OpProto__Var",
"paddle_mobile.framework.proto",
sizeof(PaddleMobile__Framework__Proto__OpProto__Var),
5,
6,
paddle_mobile__framework__proto__op_proto__var__field_descriptors,
paddle_mobile__framework__proto__op_proto__var__field_indices_by_name,
1,
......@@ -1012,7 +1091,7 @@ const ProtobufCMessageDescriptor
NULL /* reserved[123] */
};
static const ProtobufCEnumValue
paddle_mobile__framework__proto__var_type__type__enum_values_by_number[19] =
paddle_mobile__framework__proto__var_type__type__enum_values_by_number[22] =
{
{"BOOL", "PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__BOOL",
0},
......@@ -1057,31 +1136,29 @@ static const ProtobufCEnumValue
{"RAW", "PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__RAW", 17},
{"TUPLE", "PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__TUPLE",
18},
{"SIZE_T",
"PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__SIZE_T", 19},
{"UINT8", "PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__UINT8",
20},
{"INT8", "PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT8",
21},
};
static const ProtobufCIntRange
paddle_mobile__framework__proto__var_type__type__value_ranges[] = {{0, 0},
{0, 19}};
{0, 22}};
static const ProtobufCEnumValueIndex
paddle_mobile__framework__proto__var_type__type__enum_values_by_name[19] = {
{"BOOL", 0},
{"CHANNEL", 16},
{"FEED_MINIBATCH", 9},
{"FETCH_LIST", 10},
{"FP16", 4},
{"FP32", 5},
{"FP64", 6},
{"INT16", 1},
{"INT32", 2},
{"INT64", 3},
{"LOD_RANK_TABLE", 12},
{"LOD_TENSOR", 7},
{"LOD_TENSOR_ARRAY", 13},
{"PLACE_LIST", 14},
{"RAW", 17},
{"READER", 15},
{"SELECTED_ROWS", 8},
{"STEP_SCOPES", 11},
{"TUPLE", 18},
paddle_mobile__framework__proto__var_type__type__enum_values_by_name[22] = {
{"BOOL", 0}, {"CHANNEL", 16},
{"FEED_MINIBATCH", 9}, {"FETCH_LIST", 10},
{"FP16", 4}, {"FP32", 5},
{"FP64", 6}, {"INT16", 1},
{"INT32", 2}, {"INT64", 3},
{"INT8", 21}, {"LOD_RANK_TABLE", 12},
{"LOD_TENSOR", 7}, {"LOD_TENSOR_ARRAY", 13},
{"PLACE_LIST", 14}, {"RAW", 17},
{"READER", 15}, {"SELECTED_ROWS", 8},
{"SIZE_T", 19}, {"STEP_SCOPES", 11},
{"TUPLE", 18}, {"UINT8", 20},
};
const ProtobufCEnumDescriptor
paddle_mobile__framework__proto__var_type__type__descriptor = {
......@@ -1090,9 +1167,9 @@ const ProtobufCEnumDescriptor
"Type",
"PaddleMobile__Framework__Proto__VarType__Type",
"paddle_mobile.framework.proto",
19,
22,
paddle_mobile__framework__proto__var_type__type__enum_values_by_number,
19,
22,
paddle_mobile__framework__proto__var_type__type__enum_values_by_name,
1,
paddle_mobile__framework__proto__var_type__type__value_ranges,
......@@ -1325,7 +1402,7 @@ const ProtobufCMessageDescriptor
NULL /* reserved[123] */
};
static const ProtobufCFieldDescriptor
paddle_mobile__framework__proto__program_desc__field_descriptors[1] = {
paddle_mobile__framework__proto__program_desc__field_descriptors[2] = {
{
"blocks", 1, PROTOBUF_C_LABEL_REPEATED, PROTOBUF_C_TYPE_MESSAGE,
offsetof(PaddleMobile__Framework__Proto__ProgramDesc, n_blocks),
......@@ -1334,14 +1411,23 @@ static const ProtobufCFieldDescriptor
0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
{
"version", 2, PROTOBUF_C_LABEL_OPTIONAL, PROTOBUF_C_TYPE_MESSAGE,
0, /* quantifier_offset */
offsetof(PaddleMobile__Framework__Proto__ProgramDesc, version),
&paddle_mobile__framework__proto__version__descriptor, NULL,
0, /* flags */
0, NULL, NULL /* reserved1,reserved2, etc */
},
};
static const unsigned
paddle_mobile__framework__proto__program_desc__field_indices_by_name[] = {
0, /* field[0] = blocks */
1, /* field[1] = version */
};
static const ProtobufCIntRange
paddle_mobile__framework__proto__program_desc__number_ranges[1 + 1] = {
{1, 0}, {0, 1}};
{1, 0}, {0, 2}};
const ProtobufCMessageDescriptor
paddle_mobile__framework__proto__program_desc__descriptor = {
PROTOBUF_C__MESSAGE_DESCRIPTOR_MAGIC,
......@@ -1350,7 +1436,7 @@ const ProtobufCMessageDescriptor
"PaddleMobile__Framework__Proto__ProgramDesc",
"paddle_mobile.framework.proto",
sizeof(PaddleMobile__Framework__Proto__ProgramDesc),
1,
2,
paddle_mobile__framework__proto__program_desc__field_descriptors,
paddle_mobile__framework__proto__program_desc__field_indices_by_name,
1,
......@@ -1362,7 +1448,7 @@ const ProtobufCMessageDescriptor
NULL /* reserved[123] */
};
static const ProtobufCEnumValue
paddle_mobile__framework__proto__attr_type__enum_values_by_number[10] = {
paddle_mobile__framework__proto__attr_type__enum_values_by_number[11] = {
{"INT", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__INT", 0},
{"FLOAT", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__FLOAT", 1},
{"STRING", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__STRING", 2},
......@@ -1373,15 +1459,16 @@ static const ProtobufCEnumValue
{"BOOLEANS", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BOOLEANS", 7},
{"BLOCK", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BLOCK", 8},
{"LONG", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__LONG", 9},
{"BLOCKS", "PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BLOCKS", 10},
};
static const ProtobufCIntRange
paddle_mobile__framework__proto__attr_type__value_ranges[] = {{0, 0},
{0, 10}};
{0, 11}};
static const ProtobufCEnumValueIndex
paddle_mobile__framework__proto__attr_type__enum_values_by_name[10] = {
{"BLOCK", 8}, {"BOOLEAN", 6}, {"BOOLEANS", 7}, {"FLOAT", 1},
{"FLOATS", 4}, {"INT", 0}, {"INTS", 3}, {"LONG", 9},
{"STRING", 2}, {"STRINGS", 5},
paddle_mobile__framework__proto__attr_type__enum_values_by_name[11] = {
{"BLOCK", 8}, {"BLOCKS", 10}, {"BOOLEAN", 6}, {"BOOLEANS", 7},
{"FLOAT", 1}, {"FLOATS", 4}, {"INT", 0}, {"INTS", 3},
{"LONG", 9}, {"STRING", 2}, {"STRINGS", 5},
};
const ProtobufCEnumDescriptor
paddle_mobile__framework__proto__attr_type__descriptor = {
......@@ -1390,9 +1477,9 @@ const ProtobufCEnumDescriptor
"AttrType",
"PaddleMobile__Framework__Proto__AttrType",
"paddle_mobile.framework.proto",
10,
11,
paddle_mobile__framework__proto__attr_type__enum_values_by_number,
10,
11,
paddle_mobile__framework__proto__attr_type__enum_values_by_name,
1,
paddle_mobile__framework__proto__attr_type__value_ranges,
......
src/framework/framework.pb-c.h
浏览文件 @ 1d76df93
......@@ -4,16 +4,18 @@
#ifndef PROTOBUF_C_framework_2eproto__INCLUDED
#define PROTOBUF_C_framework_2eproto__INCLUDED
#include "common/protobuf-c.h"
#include <protobuf-c/protobuf-c.h>
PROTOBUF_C__BEGIN_DECLS
#if PROTOBUF_C_VERSION_NUMBER < 1000000
# error This file was generated by a newer version of protoc-c which is incompatible with your libprotobuf-c headers. Please update your headers.
#elif 1003000 < PROTOBUF_C_MIN_COMPILER_VERSION
#elif 1003001 < PROTOBUF_C_MIN_COMPILER_VERSION
# error This file was generated by an older version of protoc-c which is incompatible with your libprotobuf-c headers. Please regenerate this file with a newer version of protoc-c.
#endif
typedef struct _PaddleMobile__Framework__Proto__Version
PaddleMobile__Framework__Proto__Version;
typedef struct _PaddleMobile__Framework__Proto__OpDesc
PaddleMobile__Framework__Proto__OpDesc;
typedef struct _PaddleMobile__Framework__Proto__OpDesc__Attr
......@@ -60,6 +62,12 @@ typedef enum _PaddleMobile__Framework__Proto__VarType__Type {
PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP16 = 4,
PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP32 = 5,
PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__FP64 = 6,
/*
* Tensor<size_t> is used in C++.
*/
PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__SIZE_T = 19,
PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__UINT8 = 20,
PADDLE_MOBILE__FRAMEWORK__PROTO__VAR_TYPE__TYPE__INT8 = 21,
/*
* Other types that may need additional descriptions
*/
......@@ -93,13 +101,32 @@ typedef enum _PaddleMobile__Framework__Proto__AttrType {
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BOOLEAN = 6,
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BOOLEANS = 7,
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BLOCK = 8,
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__LONG =
9 PROTOBUF_C__FORCE_ENUM_TO_BE_INT_SIZE(
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__LONG = 9,
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__BLOCKS =
10 PROTOBUF_C__FORCE_ENUM_TO_BE_INT_SIZE(
PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE)
} PaddleMobile__Framework__Proto__AttrType;
/* --- messages --- */
/*
* Any incompatible changes to ProgramDesc and its dependencies should
* raise the version defined version.h.
* Serailization and Deserialization codes should be modified in a way
* that supports old versions following the version and compatibility policy.
*/
struct _PaddleMobile__Framework__Proto__Version {
ProtobufCMessage base;
protobuf_c_boolean has_version;
int64_t version;
};
#define PADDLE_MOBILE__FRAMEWORK__PROTO__VERSION__INIT \
{ \
PROTOBUF_C_MESSAGE_INIT( \
&paddle_mobile__framework__proto__version__descriptor) \
, 0, 0ll \
}
struct _PaddleMobile__Framework__Proto__OpDesc__Attr {
ProtobufCMessage base;
char *name;
......@@ -123,13 +150,15 @@ struct _PaddleMobile__Framework__Proto__OpDesc__Attr {
int32_t block_idx;
protobuf_c_boolean has_l;
int64_t l;
size_t n_blocks_idx;
int32_t *blocks_idx;
};
#define PADDLE_MOBILE__FRAMEWORK__PROTO__OP_DESC__ATTR__INIT \
{ \
PROTOBUF_C_MESSAGE_INIT( \
&paddle_mobile__framework__proto__op_desc__attr__descriptor) \
, NULL, PADDLE_MOBILE__FRAMEWORK__PROTO__ATTR_TYPE__INT, 0, 0, 0, 0, NULL, \
0, NULL, 0, NULL, 0, NULL, 0, 0, 0, NULL, 0, 0, 0, 0 \
0, NULL, 0, NULL, 0, NULL, 0, 0, 0, NULL, 0, 0, 0, 0, 0, NULL \
}
struct _PaddleMobile__Framework__Proto__OpDesc__Var {
......@@ -181,12 +210,13 @@ struct _PaddleMobile__Framework__Proto__OpProto__Var {
protobuf_c_boolean intermediate;
protobuf_c_boolean has_dispensable;
protobuf_c_boolean dispensable;
char *reuse;
};
#define PADDLE_MOBILE__FRAMEWORK__PROTO__OP_PROTO__VAR__INIT \
{ \
PROTOBUF_C_MESSAGE_INIT( \
&paddle_mobile__framework__proto__op_proto__var__descriptor) \
, NULL, NULL, 0, 0, 0, 0, 0, 0 \
, NULL, NULL, 0, 0, 0, 0, 0, 0, NULL \
}
/*
......@@ -375,14 +405,27 @@ struct _PaddleMobile__Framework__Proto__ProgramDesc {
ProtobufCMessage base;
size_t n_blocks;
PaddleMobile__Framework__Proto__BlockDesc **blocks;
PaddleMobile__Framework__Proto__Version *version;
};
#define PADDLE_MOBILE__FRAMEWORK__PROTO__PROGRAM_DESC__INIT \
{ \
PROTOBUF_C_MESSAGE_INIT( \
&paddle_mobile__framework__proto__program_desc__descriptor) \
, 0, NULL \
, 0, NULL, NULL \
}
/* PaddleMobile__Framework__Proto__Version methods */
void paddle_mobile__framework__proto__version__init(
PaddleMobile__Framework__Proto__Version *message);
size_t paddle_mobile__framework__proto__version__get_packed_size(
const PaddleMobile__Framework__Proto__Version *message);
PaddleMobile__Framework__Proto__Version *
paddle_mobile__framework__proto__version__unpack(ProtobufCAllocator *allocator,
size_t len,
const uint8_t *data);
void paddle_mobile__framework__proto__version__free_unpacked(
PaddleMobile__Framework__Proto__Version *message,
ProtobufCAllocator *allocator);
/* PaddleMobile__Framework__Proto__OpDesc__Attr methods */
void paddle_mobile__framework__proto__op_desc__attr__init(
PaddleMobile__Framework__Proto__OpDesc__Attr *message);
......@@ -392,10 +435,8 @@ void paddle_mobile__framework__proto__op_desc__var__init(
/* PaddleMobile__Framework__Proto__OpDesc methods */
void paddle_mobile__framework__proto__op_desc__init(
PaddleMobile__Framework__Proto__OpDesc *message);
size_t paddle_mobile__framework__proto__op_desc__get_packed_size(
const PaddleMobile__Framework__Proto__OpDesc *message);
PaddleMobile__Framework__Proto__OpDesc *
paddle_mobile__framework__proto__op_desc__unpack(ProtobufCAllocator *allocator,
size_t len,
......@@ -487,6 +528,8 @@ void paddle_mobile__framework__proto__program_desc__free_unpacked(
ProtobufCAllocator *allocator);
/* --- per-message closures --- */
typedef void (*PaddleMobile__Framework__Proto__Version_Closure)(
const PaddleMobile__Framework__Proto__Version *message, void *closure_data);
typedef void (*PaddleMobile__Framework__Proto__OpDesc__Attr_Closure)(
const PaddleMobile__Framework__Proto__OpDesc__Attr *message,
void *closure_data);
......@@ -539,6 +582,8 @@ typedef void (*PaddleMobile__Framework__Proto__ProgramDesc_Closure)(
extern const ProtobufCEnumDescriptor
paddle_mobile__framework__proto__attr_type__descriptor;
extern const ProtobufCMessageDescriptor
paddle_mobile__framework__proto__version__descriptor;
extern const ProtobufCMessageDescriptor
paddle_mobile__framework__proto__op_desc__descriptor;
extern const ProtobufCMessageDescriptor
......
src/framework/framework.proto
浏览文件 @ 1d76df93
......@@ -16,6 +16,13 @@ syntax = "proto2";
option optimize_for = LITE_RUNTIME;
package paddle_mobile.framework.proto;
// Any incompatible changes to ProgramDesc and its dependencies should
// raise the version defined version.h.
//
// Serailization and Deserialization codes should be modified in a way
// that supports old versions following the version and compatibility policy.
message Version { optional int64 version = 1 [ default = 0 ]; }
enum AttrType {
INT = 0;
FLOAT = 1;
......@@ -27,6 +34,7 @@ enum AttrType {
BOOLEANS = 7;
BLOCK = 8;
LONG = 9;
BLOCKS = 10;
}
// OpDesc describes an instance of a C++ framework::OperatorBase
......@@ -46,6 +54,7 @@ message OpDesc {
repeated bool bools = 11;
optional int32 block_idx = 12;
optional int64 l = 13;
repeated int32 blocks_idx = 14;
};
message Var {
......@@ -71,6 +80,7 @@ message OpProto {
optional bool duplicable = 3 [ default = false ];
optional bool intermediate = 4 [ default = false ];
optional bool dispensable = 5 [ default = false ];
optional string reuse = 6;
}
// AttrProto describes the C++ type Attribute.
......@@ -101,6 +111,10 @@ message VarType {
FP16 = 4;
FP32 = 5;
FP64 = 6;
// Tensor<size_t> is used in C++.
SIZE_T = 19;
UINT8 = 20;
INT8 = 21;
// Other types that may need additional descriptions
LOD_TENSOR = 7;
......@@ -173,4 +187,8 @@ message BlockDesc {
// for more details.
// TODO(panyx0718): A model can have multiple programs. Need a
// way to distinguish them. Maybe ID or name?
message ProgramDesc { repeated BlockDesc blocks = 1; }
message ProgramDesc {
repeated BlockDesc blocks = 1;
optional Version version = 2;
}
src/framework/load_ops.h 0 → 100644
浏览文件 @ 1d76df93
/* 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
#ifdef PADDLE_MOBILE_CPU
#define LOAD_CPU_OP(op_type) \
extern int TouchOpRegistrar_##op_type##_##cpu(); \
static int use_op_itself_##op_type##_##cpu __attribute__((unused)) = \
TouchOpRegistrar_##op_type##_##cpu()
#else
#define LOAD_CPU_OP(op_type)
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#define LOAD_MALI_GPU_OP(op_type) \
extern int TouchOpRegistrar_##op_type##_##mali_gpu(); \
static int use_op_itself_##op_type##_##mali_gpu __attribute__((unused)) = \
TouchOpRegistrar_##op_type##_##mali_gpu()
#else
#define LOAD_MALI_GPU_OP(op_type)
#endif
#ifdef PADDLE_MOBILE_FPGA
#define LOAD_FPGA_OP(op_type) \
extern int TouchOpRegistrar_##op_type##_##fpga(); \
static int use_op_itself_##op_type##_##fpga __attribute__((unused)) = \
TouchOpRegistrar_##op_type##_##mali_gpu()
#else
#define LOAD_FPGA_OP(op_type)
#endif
#define LOAD_FUSION_MATCHER(op_type) \
extern int TouchFusionMatcherRegistrar_##op_type(); \
static int use_fusion_matcher_itself_##op_type __attribute__((unused)) = \
TouchFusionMatcherRegistrar_##op_type();
#define LOAD_OP(op_type) \
LOAD_CPU_OP(op_type); \
LOAD_MALI_GPU_OP(op_type); \
LOAD_FPGA_OP(op_type);
#define LOAD_OP1(op_type, device_type) LOAD_##device_type##_OP(op_type);
#define LOAD_OP2(op_type, device_type1, device_type2) \
LOAD_OP1(op_type, device_type1) \
LOAD_OP1(op_type, device_type2)
#define LOAD_OP3(op_type, device_type1, device_type2, device_type3) \
LOAD_OP2(op_type, device_type1, device_type2) \
LOAD_OP1(op_type, device_type3)
// load requared ops
LOAD_OP(feed)
LOAD_OP(fetch)
#ifdef BATCHNORM_OP
LOAD_OP2(batch_norm, CPU, MALI_GPU);
#endif
#ifdef BILINEAR_INTERP_OP
LOAD_OP1(bilinear_interp, CPU);
#endif
#ifdef BOXCODER_OP
LOAD_OP1(box_coder, CPU);
#endif
#ifdef CONCAT_OP
LOAD_OP3(concat, CPU, MALI_GPU, FPGA);
#endif
#ifdef CONV_OP
LOAD_OP3(conv2d, CPU, MALI_GPU, FPGA);
#endif
#ifdef LRN_OP
LOAD_OP2(lrn, CPU, MALI_GPU);
#endif
#ifdef SIGMOID_OP
LOAD_OP1(sigmoid, CPU);
#endif
#ifdef FUSION_FC_RELU_OP
LOAD_OP3(fusion_fc_relu, CPU, MALI_GPU, FPGA);
LOAD_FUSION_MATCHER(fusion_fc_relu);
#endif
#ifdef FUSION_ELEMENTWISEADDRELU_OP
LOAD_OP3(fusion_elementwise_add_relu, CPU, MALI_GPU, FPGA);
LOAD_FUSION_MATCHER(fusion_elementwise_add_relu);
#endif
#ifdef SPLIT_OP
LOAD_OP1(split, CPU);
#endif
#ifdef RESIZE_OP
LOAD_OP2(resize, CPU, MALI_GPU);
#endif
#ifdef FUSION_CONVADDBNRELU_OP
LOAD_OP2(fusion_conv_add_bn_relu, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_add_bn_relu);
#endif
#ifdef RESHAPE_OP
LOAD_OP2(reshape, CPU, MALI_GPU);
#endif
#ifdef TRANSPOSE_OP
LOAD_OP1(transpose, CPU);
#endif
#ifdef PRIORBOX_OP
LOAD_OP1(prior_box, CPU);
#endif
#ifdef FUSION_CONVADDRELU_OP
LOAD_OP2(fusion_conv_add_relu, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_add_relu);
#endif
#ifdef FUSION_CONVADDADDPRELU_OP
LOAD_OP2(fusion_conv_add_add_prelu, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_add_add_prelu);
#endif
#ifdef FUSION_CONVADD_OP
LOAD_OP2(fusion_conv_add, CPU, MALI_GPU);
LOAD_FUSION_MATCHER(fusion_conv_add);
#endif
#ifdef SOFTMAX_OP
LOAD_OP2(softmax, CPU, MALI_GPU);
#endif
#ifdef SHAPE_OP
LOAD_OP1(shape, CPU);
#endif
#ifdef DEPTHWISECONV_OP
LOAD_OP1(depthwise_conv2d, CPU);
#endif
#ifdef CONV_TRANSPOSE_OP
LOAD_OP1(conv2d_transpose, CPU);
#endif
#ifdef SCALE_OP
LOAD_OP2(scale, CPU, MALI_GPU);
#endif
#ifdef ELEMENTWISEADD_OP
LOAD_OP2(elementwise_add, CPU, MALI_GPU);
#endif
#ifdef PRELU_OP
LOAD_OP2(prelu, CPU, MALI_GPU);
#endif
#ifdef FLATTEN_OP
LOAD_OP1(flatten, CPU);
#endif
#ifdef FUSION_CONVBNADDRELU_OP
LOAD_OP2(fusion_conv_bn_add_relu, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_bn_add_relu);
#endif
#ifdef FUSION_CONVBNRELU_OP
LOAD_OP2(fusion_conv_bn_relu, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_bn_relu);
#endif
#ifdef GRU_OP
LOAD_OP1(gru, CPU);
#endif
#ifdef FUSION_CONVADDBN_OP
LOAD_OP2(fusion_conv_add_bn, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_add_bn);
#endif
#ifdef DROPOUT_OP
LOAD_OP2(dropout, CPU, FPGA);
#endif
#ifdef FUSION_CONVADDPRELU_OP
LOAD_OP2(fusion_conv_add_prelu, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_add_prelu);
#endif
#ifdef FUSION_DWCONVBNRELU_OP
LOAD_OP1(fusion_dwconv_bn_relu, CPU);
LOAD_FUSION_MATCHER(fusion_dwconv_bn_relu);
#endif
#ifdef CRF_OP
LOAD_OP1(crf_decoding, CPU);
#endif
#ifdef MUL_OP
LOAD_OP2(mul, CPU, MALI_GPU);
#endif
#ifdef RELU_OP
LOAD_OP2(relu, CPU, MALI_GPU);
#endif
#ifdef IM2SEQUENCE_OP
LOAD_OP1(im2sequence, CPU);
#endif
#ifdef LOOKUP_OP
LOAD_OP1(lookup_table, CPU);
#endif
#ifdef FUSION_FC_OP
LOAD_OP3(fusion_fc, CPU, MALI_GPU, FPGA);
LOAD_FUSION_MATCHER(fusion_fc);
#endif
#ifdef POOL_OP
LOAD_OP3(pool2d, CPU, MALI_GPU, FPGA);
#endif
#ifdef MULTICLASSNMS_OP
LOAD_OP1(multiclass_nms, CPU);
#endif
#ifdef SLICE_OP
LOAD_OP2(slice, CPU, MALI_GPU);
#endif
#ifdef FUSION_CONVBN_OP
LOAD_OP2(fusion_conv_bn, CPU, FPGA);
LOAD_FUSION_MATCHER(fusion_conv_bn);
#endif
LOAD_OP1(quantize, CPU);
LOAD_OP1(dequantize, CPU);
src/framework/op_registry.h
浏览文件 @ 1d76df93
......@@ -97,6 +97,7 @@ class OpRegistry {
};
#define REGISTER_OPERATOR(op_type, op_class, device_name, device_type) \
template class op_class<device_type, float>; \
template <typename Dtype, typename T> \
class _OpClass_##op_type##_##device_name : public op_class<Dtype, T> { \
public: \
......@@ -119,16 +120,5 @@ class OpRegistry {
#define REGISTER_OPERATOR_FPGA(op_type, op_class) \
REGISTER_OPERATOR(op_type, op_class, fpga, paddle_mobile::FPGA);
#define USE_OP(op_type, device_name) \
extern int TouchOpRegistrar_##op_type##_##device_name(); \
static int use_op_itself_##op_type##_##device_name __attribute__((unused)) = \
TouchOpRegistrar_##op_type##_##device_name()
#define USE_OP_CPU(op_type) USE_OP(op_type, cpu);
#define USE_OP_MALI_GPU(op_type) USE_OP(op_type, mali_gpu);
#define USE_OP_FPGA(op_type) USE_OP(op_type, fpga);
} // namespace framework
} // namespace paddle_mobile
src/framework/program/program-optimize/fusion_op_register.h
浏览文件 @ 1d76df93
......@@ -67,7 +67,16 @@ class FusionOpRegistrar {
explicit FusionOpRegistrar(FusionOpMatcher* matcher) {
FusionOpRegister::Instance()->regist(matcher);
}
void Touch() {}
};
} // namespace framework
} // namespace paddle_mobile
#define REGISTER_FUSION_MATCHER(op_type, matcher) \
static paddle_mobile::framework::FusionOpRegistrar \
__fusion_matcher_registrar_##op_type(new matcher()); \
int TouchFusionMatcherRegistrar_##op_type() { \
__fusion_matcher_registrar_##op_type.Touch(); \
return 0; \
}
src/framework/program/program.h
浏览文件 @ 1d76df93
......@@ -33,8 +33,6 @@ class Program {
bool quantification = false;
size_t combined_params_len;
const uint8_t *combined_params_buf;
private:
};
} // namespace framework
......
src/framework/program/tensor_desc.h
浏览文件 @ 1d76df93
......@@ -40,7 +40,10 @@ enum VarType_Type {
VARTYPE_TYPE_READER = 15,
VARTYPE_TYPE_CHANNEL = 16,
VARTYPE_TYPE_RAW = 17,
VARTYPE_TYPE_TUPLE = 18
VARTYPE_TYPE_TUPLE = 18,
VARTYPE_TYPE_SIZE_T = 19,
VARTYPE_TYPE_UINT8 = 20,
VARTYPE_TYPE_INT8 = 21,
};
class TensorDesc {
......@@ -58,8 +61,9 @@ class TensorDesc {
}
data_type_ = (VarType_Type)desc->data_type;
}
// return tensor dim as a vector
std::vector<int64_t> Dims() const { return dims_; };
// return tensor data type
VarType_Type DataType() const { return data_type_; }
private:
......
src/framework/program/var_desc.h
浏览文件 @ 1d76df93
......@@ -31,6 +31,7 @@ class VarDesc {
this->tensor_desc_ = var_desc.tensor_desc_;
this->type_ = var_desc.type_;
}
VarDesc(PaddleMobile__Framework__Proto__VarDesc *desc) {
type_ = (VarType_Type)desc->type->type;
name_ = std::string(desc->name);
......@@ -44,9 +45,7 @@ class VarDesc {
tensor_desc_ = TensorDesc(desc->type->lod_tensor->tensor);
break;
case VARTYPE_TYPE_STEP_LOD_TENSOR_ARRAY:
desc->type->tensor_array->tensor->data_type;
tensor_desc_ = TensorDesc(desc->type->tensor_array->tensor);
break;
default:
break;
......@@ -60,6 +59,7 @@ class VarDesc {
break;
}
}
std::string Name() const { return name_; }
VarType_Type Type() const { return type_; }
......
src/framework/tensor.h
浏览文件 @ 1d76df93
......@@ -319,10 +319,11 @@ class Tensor {
* begins.
*/
size_t offset_;
#ifdef PADDLE_MOBILE_FPGA
public:
public: // NOLINT
inline void reset_data_ptr(void *p) {
((PlaceholderImpl *)(holder_.get()))->ptr_.reset((uint8_t *)p);
((PlaceholderImpl *)(holder_.get()))->ptr_.reset((uint8_t *)p); // NOLINT
}
float scale[2]; // scale[0]= MAX/127.0, scale[1]= 127.0/MAX
#endif
......@@ -335,11 +336,12 @@ inline Print &operator<<(Print &printer, const Tensor &tensor) {
stride = stride > 0 ? stride : 1;
#ifndef PADDLE_MOBILE_FPGA
for (int i = 0; i < tensor.numel(); i += stride) {
// 这不一定是float的
if (tensor.type() == typeid(float)) {
printer << tensor.data<float>()[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] << " ";
}
}
#endif
......
src/io/executor.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/io/executor.h
浏览文件 @ 1d76df93
......@@ -18,19 +18,12 @@ limitations under the License. */
#include <memory>
#include <string>
#include <vector>
#include "common/types.h"
#include "common/util.h"
#include "framework/lod_tensor.h"
#include "framework/operator.h"
#include "framework/program/program.h"
#include "framework/tensor.h"
#ifdef PADDLE_EXECUTOR_MULTITHREAD
#include <condition_variable>
#include <mutex>
#include <thread>
#include "common/dep_core.h"
#endif
using std::string;
namespace paddle_mobile {
......@@ -38,50 +31,61 @@ template <typename Dtype = CPU, Precision P = Precision::FP32>
class Executor {
public:
typedef typename PrecisionTrait<P>::ptype Ptype;
// exector constructor
// @param program program converted from proto program in PaddlePaddle
// @param use_optimize bool whether use operator fusion to speed up or not
// @param loddable bool
Executor(const framework::Program<Dtype> program, int batch_size = 1,
const bool use_optimize = true, const bool loddable = false);
/*
* @b init executor with program load by Loader class
* @b 用 loader load 的 program 实例化 executor
* */
Executor(const framework::Program<Dtype> p, int batch_size = 1,
bool use_optimize = true, bool loddable = false);
/*
* @b to predict
* */
// predict with tensor input
// @param t input tensor to do prediction
// @return predicted tensor
std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t);
/*
* @b to predict
* */
// predict with lod tensor input
// @param t input lod tensor to do prediction
// @return predicted lod tensor
std::shared_ptr<framework::LoDTensor> PredictLod(
const framework::LoDTensor &t);
/*
* @b to predict with vector and dim
*
* @b 使用 输入 和 输入的维度信息 进行预测
* */
// predict with vector input and dims
// @param input vector whose elements will be formed
// @param input lod tensor to do prediction
// @param dims vector whose elements will be formed
// @param input tensor shape
// @return vector which is flatted from predicted tensor
std::vector<Ptype> Predict(const std::vector<Ptype> &input,
const std::vector<int64_t> &dims);
#ifdef PADDLE_MOBILE_FPGA
void InjectVariable(const framework::Tensor &t, string var_name);
void FeedData(const framework::Tensor &t);
std::shared_ptr<framework::Tensor> FetchResult(int id = -1);
void Predict_From_To(int start = 0, int end = -1);
void Predict_From(int start);
void Predict_To(int end);
#endif
protected:
Executor() = default;
std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t,
int block_id);
bool varInputMemory(const std::shared_ptr<framework::VarDesc> &var_desc,
framework::Variable *var,
framework::LoDTensor *tensor) const;
void InitMemory();
void LoadMemory(const framework::VarDesc var_desc,
framework::LoDTensor *tensor, char **data);
void InitCombineMemory();
void LoadMemory(void **data,
const std::shared_ptr<framework::VarDesc> var_desc,
framework::LoDTensor *tensor);
framework::Program<Dtype> program_;
int batch_size_ = 1;
std::shared_ptr<framework::ProgramDesc> to_predict_program_;
std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t,
int block_id);
std::map<framework::BlockDesc,
std::vector<std::shared_ptr<framework::OperatorBase<Dtype>>>>
ops_of_block_;
bool use_optimize_ = false;
bool loddable_ = false;
#ifdef PADDLE_EXECUTOR_MULTITHREAD
std::vector<depCore> depManager;
#endif
#ifdef PADDLE_MOBILE_PROFILE
struct ProfInfo {
int tid = 0;
......@@ -89,21 +93,8 @@ class Executor {
uint64_t runEnd = 0UL;
};
#endif
bool varInputMemory(const std::shared_ptr<framework::VarDesc> &var_desc,
framework::Variable *var,
framework::LoDTensor *tensor) const;
#ifdef PADDLE_MOBILE_FPGA
public:
void InjectVariable(const framework::Tensor &t, string var_name);
void FeedData(const framework::Tensor &t);
std::shared_ptr<framework::Tensor> FetchResult(int id = -1);
void Predict_From_To(int start = 0, int end = -1);
void Predict_From(int start);
void Predict_To(int end);
#endif
bool use_optimize_ = false;
bool loddable_ = false;
};
} // namespace paddle_mobile
src/io/loader.cpp
浏览文件 @ 1d76df93
......@@ -27,8 +27,8 @@ using framework::Variable;
* @param scope
*/
void InitMemoryFromProgram(
std::shared_ptr<framework::ProgramDesc> &originProgramDesc,
std::shared_ptr<framework::Scope> &scope) {
std::shared_ptr<framework::ProgramDesc> &originProgramDesc, // NOLINT
std::shared_ptr<framework::Scope> &scope) { // NOLINT
for (const auto &block : originProgramDesc.get()->Blocks()) {
for (const auto &var_desc : block->Vars()) {
auto var = scope.get()->Var(var_desc->Name());
......@@ -61,12 +61,16 @@ void InitMemoryFromProgram(
*/
template <typename Dtype, Precision P>
void FusionAndPrintInfos(
bool &optimize, bool &can_add_split, framework::Program<Dtype, P> &program,
bool optimize, bool can_add_split,
framework::Program<Dtype, P> &program, // NOLINT
const std::shared_ptr<framework::ProgramDesc> &originProgramDesc) {
if (optimize) {
framework::ProgramOptimize program_optimize;
program.optimizeProgram =
program_optimize.FusionOptimize(originProgramDesc, can_add_split);
if (!program.optimizeProgram) {
program.optimizeProgram = originProgramDesc;
}
}
if (optimize) {
program.optimizeProgram->Description("optimize: ");
......@@ -74,6 +78,7 @@ void FusionAndPrintInfos(
originProgramDesc->Description("program: ");
}
}
static size_t ReadBuffer(const char *file_name, uint8_t **out) {
FILE *fp;
fp = fopen(file_name, "rb");
......
src/io/loader.h
浏览文件 @ 1d76df93
......@@ -24,19 +24,11 @@ namespace paddle_mobile {
template <typename Dtype = CPU, Precision P = Precision::FP32>
class Loader {
public:
/*
* @b load separate format fluid model
* @b 加载分开形式的 fluid 模型
* */
const framework::Program<Dtype, P> Load(const std::string &dirname,
bool optimize = false,
bool quantification = false,
bool can_add_split = false);
/*
* @b load combine format fluid mode
* @b 加载结合在一起格式的模型
* */
const framework::Program<Dtype, P> Load(const std::string &model_path,
const std::string &para_path,
bool optimize = false,
......
src/io/paddle_mobile.cpp
浏览文件 @ 1d76df93
......@@ -19,10 +19,9 @@ namespace paddle_mobile {
template <typename Dtype, Precision P>
void PaddleMobile<Dtype, P>::SetThreadNum(int num) {
#ifdef _OPENMP
// omp_set_dynamic(0);
omp_set_num_threads(num);
#endif
};
}
template <typename Dtype, Precision P>
bool PaddleMobile<Dtype, P>::Load(const std::string &dirname, bool optimize,
......@@ -135,33 +134,31 @@ void PaddleMobile<Dtype, P>::InjectVariable(const framework::Tensor &t,
template <typename Dtype, Precision P>
void PaddleMobile<Dtype, P>::FeedData(const framework::Tensor &t) {
executor_->FeedData(t);
};
}
template <typename Dtype, Precision P>
std::shared_ptr<framework::Tensor> PaddleMobile<Dtype, P>::FetchResult(int id) {
return executor_->FetchResult(id);
};
}
template <typename Dtype, Precision P>
void PaddleMobile<Dtype, P>::Predict_From_To(int start, int end) {
executor_->Predict_From_To(start, end);
};
}
template <typename Dtype, Precision P>
void PaddleMobile<Dtype, P>::Predict_From(int start) {
executor_->Predict_From(start);
};
}
template <typename Dtype, Precision P>
void PaddleMobile<Dtype, P>::Predict_To(int end) {
executor_->Predict_To(end);
};
}
#endif
template class PaddleMobile<CPU, Precision::FP32>;
template class PaddleMobile<FPGA, Precision::FP32>;
template class PaddleMobile<GPU_MALI, Precision::FP32>;
} // namespace paddle_mobile
src/io/paddle_mobile.h
浏览文件 @ 1d76df93
......@@ -22,6 +22,7 @@ limitations under the License. */
#endif // _OPENMP
#include "common/types.h"
#include "framework/load_ops.h"
#include "framework/tensor.h"
#include "io/executor.h"
#include "io/loader.h"
......@@ -34,67 +35,31 @@ class PaddleMobile {
public:
PaddleMobile() {}
/*
* @b load separate format fluid model
* @b 加载分开形式的 fluid 模型
* */
bool Load(const std::string &dirname, bool optimize = false,
bool quantification = false, int batch_size = 1,
bool loddable = false);
/*
* @b load combine format fluid mode
* @b 加载结合在一起格式的模型
* */
bool Load(const std::string &model_path, const std::string &para_path,
bool optimize = false, bool quantification = false,
int batch_size = 1, bool loddable = false);
/*
* @b 设置线程数, 当 cmake 中开启 openmp 时生效
* */
void SetThreadNum(int num);
/*
* @b to predict
* */
std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t);
/*
* @b to predict
* */
std::shared_ptr<framework::Tensor> PredictLod(const framework::LoDTensor &t);
/*
* @b to predict with vector and dim
*
* @b 使用 输入 和 输入的维度信息 进行预测
* */
std::vector<Ptype> Predict(const std::vector<Ptype> &input,
const std::vector<int64_t> &dims);
/**
* 从内存加载model 以及 combinedparams的接口
*
* @param model_len model 文件的内存大小
* @param model_buf model文件的内存
* @param combined_params_len params文件的内存大小
* @param combined_params_buf params文件的内存
* @return
*/
bool LoadCombinedMemory(size_t model_len, const uint8_t *model_buf,
size_t combined_params_len,
const uint8_t *combined_params_buf);
void SetThreadNum(int num);
void Clear();
~PaddleMobile();
private:
std::shared_ptr<Loader<Dtype, P>> loader_;
std::shared_ptr<Executor<Dtype, P>> executor_;
#ifdef PADDLE_MOBILE_FPGA
public:
void InjectVariable(const framework::Tensor &t, string var_name);
void FeedData(const framework::Tensor &t);
std::shared_ptr<framework::Tensor> FetchResult(int id = -1);
......@@ -102,6 +67,10 @@ class PaddleMobile {
void Predict_From(int start);
void Predict_To(int end);
#endif
private:
std::shared_ptr<Loader<Dtype, P>> loader_;
std::shared_ptr<Executor<Dtype, P>> executor_;
};
} // namespace paddle_mobile
src/ios_io/PaddleMobileCPU.mm
浏览文件 @ 1d76df93
......@@ -13,15 +13,12 @@
limitations under the License. */
#import "PaddleMobileCPU.h"
#import "op_symbols.h"
#include "framework/tensor.h"
#import "framework/load_ops.h"
#import "framework/tensor.h"
#import "io/paddle_mobile.h"
#import <memory>
#import <vector>
@interface PaddleMobileCPUResult()
-(void)toSetOutput:(float *)output;
......
src/ios_io/op_symbols.h 已删除 100644 → 0
浏览文件 @ a9c249d3
/* 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 "operators/batchnorm_op.h"
#include "operators/bilinear_interp_op.h"
#include "operators/box_coder_op.h"
#include "operators/concat_op.h"
#include "operators/conv_op.h"
#include "operators/conv_transpose_op.h"
#include "operators/crf_op.h"
#include "operators/depthwise_conv_op.h"
#include "operators/dropout_op.h"
#include "operators/elementwise_add_op.h"
#include "operators/feed_op.h"
#include "operators/fetch_op.h"
#include "operators/flatten_op.h"
#include "operators/fusion_conv_add.h"
#include "operators/fusion_conv_add_add_prelu_op.h"
#include "operators/fusion_conv_add_bn_op.h"
#include "operators/fusion_conv_add_bn_relu_op.h"
#include "operators/fusion_conv_add_prelu_op.h"
#include "operators/fusion_conv_add_relu_op.h"
#include "operators/fusion_conv_bn_add_relu_op.h"
#include "operators/fusion_conv_bn_relu_op.h"
#include "operators/fusion_dwconv_bn_relu_op.h"
#include "operators/fusion_elementwise_add_relu_op.h"
#include "operators/fusion_fc_op.h"
#include "operators/fusion_fc_relu_op.h"
#include "operators/gru_op.h"
#include "operators/im2sequence_op.h"
#include "operators/lookup_op.h"
#include "operators/lrn_op.h"
#include "operators/mul_op.h"
#include "operators/multiclass_nms_op.h"
#include "operators/pool_op.h"
#include "operators/prelu_op.h"
#include "operators/prior_box_op.h"
#include "operators/relu_op.h"
#include "operators/reshape_op.h"
#include "operators/resize_op.h"
#include "operators/scale_op.h"
#include "operators/shape_op.h"
#include "operators/sigmoid_op.h"
#include "operators/slice_op.h"
#include "operators/softmax_op.h"
#include "operators/split_op.h"
#include "operators/transpose_op.h"
src/jni/paddle_mobile_jni.cpp
浏览文件 @ 1d76df93
......@@ -84,14 +84,14 @@ Java_com_baidu_paddle_PML_loadnlp(JNIEnv *env, jclass thiz, jstring modelPath) {
#ifdef ENABLE_EXCEPTION
try {
isLoadOk = getPaddleMobileInstance()->Load(
jstring2cppstring(env, modelPath), optimize, false, 1, true);
jstring2cppstring(env, modelPath), optimize, false, true);
} catch (paddle_mobile::PaddleMobileException &e) {
ANDROIDLOGE("jni got an PaddleMobileException! ", e.what());
isLoadOk = false;
}
#else
isLoadOk = getPaddleMobileInstance()->Load(jstring2cppstring(env, modelPath),
optimize, false, 1, true);
optimize, false, true);
#endif
return static_cast<jboolean>(isLoadOk);
}
......
src/operators/batchnorm_op.h
浏览文件 @ 1d76df93
......@@ -46,13 +46,4 @@ class BatchNormOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(batch_norm);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(batch_norm);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/bilinear_interp_op.h
浏览文件 @ 1d76df93
......@@ -50,12 +50,4 @@ class BilinearOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(bilinear_interp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/box_coder_op.h
浏览文件 @ 1d76df93
......@@ -51,12 +51,4 @@ class BoxCoderOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(box_coder);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/concat_op.h
浏览文件 @ 1d76df93
......@@ -46,14 +46,4 @@ class ConcatOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(concat);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(concat);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(concat);
#endif
#endif
src/operators/conv_op.h
浏览文件 @ 1d76df93
......@@ -46,14 +46,4 @@ class ConvOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv2d);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv2d);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(conv2d);
#endif
#endif
src/operators/conv_transpose_op.cpp
浏览文件 @ 1d76df93
......@@ -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 CONV_TRANSPOSE
#ifdef CONV_TRANSPOSE_OP
#include "operators/conv_transpose_op.h"
......
src/operators/conv_transpose_op.h
浏览文件 @ 1d76df93
......@@ -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 CONV_TRANSPOSE
#ifdef CONV_TRANSPOSE_OP
#pragma once
......@@ -88,14 +88,4 @@ class ConvOpTranspose : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv2d_transpose);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv2d_transpose);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(conv2d_transpose);
#endif
#endif
src/operators/crf_op.h
浏览文件 @ 1d76df93
......@@ -47,12 +47,4 @@ class CrfOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(crf_decoding);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/depthwise_conv_op.cpp
浏览文件 @ 1d76df93
......@@ -56,9 +56,5 @@ namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(depthwise_conv2d, ops::DepthwiseConvOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/depthwise_conv_op.h
浏览文件 @ 1d76df93
......@@ -48,12 +48,4 @@ class DepthwiseConvOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(depthwise_conv2d);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/dequantize_op.cpp 0 → 100644
浏览文件 @ 1d76df93
/* 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/dequantize_op.h"
namespace paddle_mobile {
namespace operators {
template <typename DeviceType, typename T>
void DequantizeOp<DeviceType, T>::InferShape() const {
const auto& input_dims = this->param_.input_->dims();
this->param_.out_->Resize(input_dims);
}
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(dequantize, ops::DequantizeOp);
#endif
src/operators/dequantize_op.h 0 → 100644
浏览文件 @ 1d76df93
/* 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 "framework/operator.h"
#include "operators/kernel/dequantize_kernel.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
template <typename DeviceType, typename T>
class DequantizeOp
: public framework::OperatorWithKernel<DeviceType,
DequantizeParam<DeviceType>,
DequantizeKernel<DeviceType, T>> {
public:
DequantizeOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
std::shared_ptr<framework::Scope> scope)
: framework::OperatorWithKernel<DeviceType, DequantizeParam<DeviceType>,
DequantizeKernel<DeviceType, T>>(
type, inputs, outputs, attrs, scope) {}
// inference output shape
void InferShape() const override;
};
} // namespace operators
} // namespace paddle_mobile
src/operators/dropout_op.cpp
浏览文件 @ 1d76df93
......@@ -30,8 +30,6 @@ namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(dropout, ops::DropoutOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(dropout, ops::DropoutOp);
#endif
......
src/operators/dropout_op.h
浏览文件 @ 1d76df93
......@@ -50,13 +50,4 @@ class DropoutOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(dropout);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(dropout);
#endif
#endif
src/operators/elementwise_add_op.cpp
浏览文件 @ 1d76df93
......@@ -35,7 +35,5 @@ REGISTER_OPERATOR_CPU(elementwise_add, ops::ElementwiseAddOp);
#ifdef PADDLE_MOBILE_MALI_GPU
REGISTER_OPERATOR_MALI_GPU(elementwise_add, ops::ElementwiseAddOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/elementwise_add_op.h
浏览文件 @ 1d76df93
......@@ -48,13 +48,4 @@ class ElementwiseAddOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(elementwise_add);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(elementwise_add);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/feed_op.cpp
浏览文件 @ 1d76df93
......@@ -12,12 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "feed_op.h"
namespace paddle_mobile {
namespace operators {}
} // namespace paddle_mobile
#include "operators/feed_op.h"
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(feed, ops::FeedOp);
#endif
......
src/operators/feed_op.h
浏览文件 @ 1d76df93
......@@ -20,11 +20,11 @@ limitations under the License. */
namespace paddle_mobile {
namespace operators {
using std::string;
template <typename DeviceType, typename T>
class FeedOp : public framework::OperatorBase<DeviceType> {
public:
FeedOp(const string &type, const VariableNameMap &inputs,
FeedOp(const std::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,
......@@ -35,10 +35,6 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
auto out_dims = param_.Out()->dims();
out_dims[0] = param_.BatchSize();
param_.Out()->Resize(out_dims);
// note : mobile infershape iscalled when executer is created. so do not
// pass lod here .
// it is empty
}
#ifdef PADDLE_MOBILE_FPGA
......@@ -49,7 +45,7 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
}
void RunImpl() const {
auto input = (Tensor *)const_cast<LoDTensor *>(param_.InputX());
auto input = (Tensor *)const_cast<LoDTensor *>(param_.InputX()); // NOLINT
fpga::format_image(input);
auto input_ptr = input->data<float>();
Tensor *output = param_.Out();
......@@ -61,7 +57,7 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
args.output_data_type = fpga::DATA_TYPE_FP16;
args.input_layout_type = fpga::LAYOUT_CHW;
args.output_layout_type = fpga::LAYOUT_HWC;
args.image.address = (void *)input_ptr;
args.image.address = (void *)input_ptr; // NOLINT
args.image.channels = (uint32_t)input->dims()[1];
args.image.height = (uint32_t)input->dims()[2];
args.image.width = (uint32_t)input->dims()[3];
......@@ -86,13 +82,3 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(feed);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(feed);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(feed);
#endif
src/operators/fetch_op.cpp
浏览文件 @ 1d76df93
......@@ -12,10 +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. */
#include "fetch_op.h"
namespace paddle_mobile {
namespace operators {}
} // namespace paddle_mobile
#include "operators/fetch_op.h"
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
......
src/operators/fetch_op.h
浏览文件 @ 1d76df93
......@@ -46,13 +46,3 @@ class FetchOp : public framework::OperatorBase<DeviceType> {
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fetch);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fetch);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fetch);
#endif
src/operators/flatten_op.cpp
浏览文件 @ 1d76df93
......@@ -53,8 +53,6 @@ namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(flatten, ops::FlattenOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
......
src/operators/flatten_op.h
浏览文件 @ 1d76df93
......@@ -63,12 +63,4 @@ class FlattenOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(flatten);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add_add_prelu.cpp → src/operators/fusion_conv_add_add_prelu_op.cpp
浏览文件 @ 1d76df93
......@@ -14,7 +14,7 @@ limitations under the License. */
#ifdef FUSION_CONVADDADDPRELU_OP
#include "fusion_conv_add_add_prelu_op.h"
#include "operators/fusion_conv_add_add_prelu_op.h"
#include "operators/math/conv_func.h"
namespace paddle_mobile {
......@@ -48,13 +48,14 @@ void FusionConvAddAddPReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_add_add_prelu,
ops::FusionConvAddAddPReluOpMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add_add_prelu, ops::FusionConvAddAddPReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_add_add_prelu, ops::FusionConvAddAddPReluOp);
#endif
#endif
#endif // FUSION_CONVADDADDPRELU_OP
src/operators/fusion_conv_add_add_prelu_op.h
浏览文件 @ 1d76df93
......@@ -76,37 +76,7 @@ class FusionConvAddAddPReluOp
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_ADD_PRELU_REGISTER
#define CONV_ADD_ADD_PRELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_add_prelu_registrar(
new FusionConvAddAddPReluOpMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef CONV_ADD_ADD_PRELU_REGISTER
#define CONV_ADD_ADD_PRELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_add_prelu_registrar(
new FusionConvAddAddPReluOpMatcher());
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_add_prelu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_add_add_prelu);
#endif
#endif
src/operators/fusion_conv_add_bn_op.cpp
浏览文件 @ 1d76df93
......@@ -49,11 +49,11 @@ void FusionConvAddBNOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_add_bn, ops::FusionConvAddBNMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add_bn, ops::FusionConvAddBNOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_add_bn, ops::FusionConvAddBNOp);
#endif
......
src/operators/fusion_conv_add_bn_op.h
浏览文件 @ 1d76df93
......@@ -70,46 +70,7 @@ class FusionConvAddBNOp : public framework::OperatorWithKernel<
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_CONV_ADD_BN_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_registrar(
new FusionConvAddBNMatcher());
#define FUSION_CONV_ADD_BN_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_CONV_ADD_BN_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_registrar(
new FusionConvAddBNMatcher());
#define FUSION_CONV_ADD_BN_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_CONV_ADD_BN_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_registrar(
new FusionConvAddBNMatcher());
#define FUSION_CONV_ADD_BN_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_bn);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_add_bn);
#endif
#endif
src/operators/fusion_conv_add_bn_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -49,11 +49,12 @@ void FusionConvAddBNReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_add_bn_relu,
ops::FusionConvAddBNReluMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add_bn_relu, ops::FusionConvAddBNReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_add_bn_relu, ops::FusionConvAddBNReluOp);
#endif
......
src/operators/fusion_conv_add_bn_relu_op.h
浏览文件 @ 1d76df93
......@@ -75,46 +75,7 @@ class FusionConvAddBNReluOp
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
new FusionConvAddBNReluMatcher());
#define FUSION_CONV_ADD_BN_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
new FusionConvAddBNReluMatcher());
#define FUSION_CONV_ADD_BN_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
new FusionConvAddBNReluMatcher());
#define FUSION_CONV_ADD_BN_RELU_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_bn_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_add_bn_relu);
#endif
#endif
src/operators/fusion_conv_add.cpp → src/operators/fusion_conv_add_op.cpp
浏览文件 @ 1d76df93
......@@ -14,7 +14,7 @@ limitations under the License. */
#ifdef FUSION_CONVADD_OP
#include "operators/fusion_conv_add.h"
#include "operators/fusion_conv_add_op.h"
#include "operators/math/conv_func.h"
namespace paddle_mobile {
......@@ -49,13 +49,13 @@ void FusionConvAddOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_add, ops::FusionConvAddMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add, ops::FusionConvAddOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
REGISTER_OPERATOR_MALI_GPU(fusion_conv_add, ops::FusionConvAddOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add.h → src/operators/fusion_conv_add_op.h
浏览文件 @ 1d76df93
......@@ -65,40 +65,7 @@ class FusionConvAddOp : public framework::OperatorWithKernel<
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_REGISTER
static framework::FusionOpRegistrar convadd_registrar(
new FusionConvAddMatcher());
#define CONV_ADD_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef CONV_ADD_REGISTER
static framework::FusionOpRegistrar convadd_registrar(
new FusionConvAddMatcher());
#define CONV_ADD_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fusion_conv_add);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add_prelu_op.cpp
浏览文件 @ 1d76df93
......@@ -14,7 +14,7 @@ limitations under the License. */
#ifdef FUSION_CONVADDPRELU_OP
#include "fusion_conv_add_prelu_op.h"
#include "operators/fusion_conv_add_prelu_op.h"
#include "operators/math/conv_func.h"
namespace paddle_mobile {
......@@ -48,11 +48,12 @@ void FusionConvAddPReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_add_prelu,
ops::FusionConvAddPReluOpMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add_prelu, ops::FusionConvAddPReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_add_prelu, ops::FusionConvAddPReluOp);
#endif
......
src/operators/fusion_conv_add_prelu_op.h
浏览文件 @ 1d76df93
......@@ -71,37 +71,7 @@ class FusionConvAddPReluOp
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_PRELU_REGISTER
#define CONV_ADD_PRELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_prelu_registrar(
new FusionConvAddPReluOpMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef CONV_ADD_PRELU_REGISTER
#define CONV_ADD_PRELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_prelu_registrar(
new FusionConvAddPReluOpMatcher());
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_prelu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_add_prelu);
#endif
#endif
src/operators/fusion_conv_add_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -14,7 +14,7 @@ limitations under the License. */
#ifdef FUSION_CONVADDRELU_OP
#include "fusion_conv_add_relu_op.h"
#include "operators/fusion_conv_add_relu_op.h"
#include "operators/math/conv_func.h"
namespace paddle_mobile {
......@@ -48,11 +48,11 @@ void FusionConvAddReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_add_relu, ops::FusionConvAddReluOpMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add_relu, ops::FusionConvAddReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_add_relu, ops::FusionConvAddReluOp);
#endif
......
src/operators/fusion_conv_add_relu_op.h
浏览文件 @ 1d76df93
......@@ -65,37 +65,7 @@ class FusionConvAddReluOp : public framework::OperatorWithKernel<
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_RELU_REGISTER
#define CONV_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_relu_registrar(
new FusionConvAddReluOpMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef CONV_ADD_RELU_REGISTER
#define CONV_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_relu_registrar(
new FusionConvAddReluOpMatcher());
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_add_relu);
#endif
#endif
src/operators/fusion_conv_bn_add_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -49,11 +49,12 @@ void FusionConvBNAddReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_bn_add_relu,
ops::FusionConvBNAddReluMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_bn_add_relu, ops::FusionConvBNAddReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_bn_add_relu, ops::FusionConvBNAddReluOp);
#endif
......
src/operators/fusion_conv_bn_add_relu_op.h
浏览文件 @ 1d76df93
......@@ -80,46 +80,7 @@ class FusionConvBNAddReluOp
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_CONV_BN_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_add_relu_registrar(
new FusionConvBNAddReluMatcher());
#define FUSION_CONV_BN_ADD_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_CONV_BN_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_add_relu_registrar(
new FusionConvBNAddReluMatcher());
#define FUSION_CONV_BN_ADD_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_CONV_BN_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_add_relu_registrar(
new FusionConvBNAddReluMatcher());
#define FUSION_CONV_BN_ADD_RELU_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_bn_add_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_bn_add_relu);
#endif
#endif
src/operators/fusion_conv_bn_op.cpp
浏览文件 @ 1d76df93
......@@ -48,11 +48,11 @@ void FusionConvBNOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_bn, ops::FusionConvBNMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_bn, ops::FusionConvBNOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_bn, ops::FusionConvBNOp);
#endif
......
src/operators/fusion_conv_bn_op.h
浏览文件 @ 1d76df93
......@@ -67,39 +67,7 @@ class FusionConvBNOp : public framework::OperatorWithKernel<
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_CONV_BN_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_registrar(
new FusionConvBNMatcher());
#define FUSION_CONV_BN_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_CONV_BN_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_registrar(
new FusionConvBNMatcher());
#define FUSION_CONV_BN_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_bn);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_bn);
#endif
#endif
src/operators/fusion_conv_bn_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -49,11 +49,11 @@ void FusionConvBNReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_conv_bn_relu, ops::FusionConvBNReluMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_bn_relu, ops::FusionConvBNReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
REGISTER_OPERATOR_FPGA(fusion_conv_bn_relu, ops::FusionConvBNReluOp);
#endif
......
src/operators/fusion_conv_bn_relu_op.h
浏览文件 @ 1d76df93
......@@ -72,39 +72,7 @@ class FusionConvBNReluOp : public framework::OperatorWithKernel<
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_CONV_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_relu_registrar(
new FusionConvBNReluMatcher());
#define FUSION_CONV_BN_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_CONV_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_bn_relu_registrar(
new FusionConvBNReluMatcher());
#define FUSION_CONV_BN_RELU_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_bn_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_conv_bn_relu);
#endif
#endif
src/operators/fusion_dwconv_bn_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -49,11 +49,11 @@ void FusionDWConvBNReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_dwconv_bn_relu, ops::FusionDWConvBNReluMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_dwconv_bn_relu, ops::FusionDWConvBNReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
......
src/operators/fusion_dwconv_bn_relu_op.h
浏览文件 @ 1d76df93
......@@ -73,38 +73,7 @@ class FusionDWConvBNReluOp
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_DWCONV_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_dwconv_bn_relu_registrar(
new FusionDWConvBNReluMatcher());
#define FUSION_DWCONV_BN_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_DWCONV_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_dwconv_bn_relu_registrar(
new FusionDWConvBNReluMatcher());
#define FUSION_DWCONV_BN_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_dwconv_bn_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_elementwise_add_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -14,7 +14,7 @@ limitations under the License. */
#ifdef FUSION_ELEMENTWISEADDRELU_OP
#include "fusion_elementwise_add_relu_op.h"
#include "operators/fusion_elementwise_add_relu_op.h"
namespace paddle_mobile {
namespace operators {
......@@ -29,6 +29,9 @@ void FusionElementwiseAddReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_elementwise_add_relu,
ops::FusioneElementwiseAddReluMatcher);
#ifdef PADDLE_MOBILE_CPU
// REGISTER_OPERATOR_CPU(fusion_elementwise_add_relu,
// ops::FusionElementwiseAddReluOp);
......
src/operators/fusion_elementwise_add_relu_op.h
浏览文件 @ 1d76df93
......@@ -61,39 +61,7 @@ class FusionElementwiseAddReluOp
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_ELEMENTWISE_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_elementwise_relu_registrar(
new FusioneElementwiseAddReluMatcher());
#define FUSION_ELEMENTWISE_ADD_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_ELEMENTWISE_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_elementwise_relu_registrar(
new FusioneElementwiseAddReluMatcher());
#define FUSION_ELEMENTWISE_ADD_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_ELEMENTWISE_ADD_RELU_REGISTER
static framework::FusionOpRegistrar fusion_elementwise_relu_registrar(
new FusioneElementwiseAddReluMatcher());
#define FUSION_ELEMENTWISE_ADD_RELU_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_elementwise_add_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_elementwise_add_relu);
#endif
#endif
src/operators/fusion_fc_op.cpp
浏览文件 @ 1d76df93
......@@ -15,6 +15,7 @@ limitations under the License. */
#ifdef FUSION_FC_OP
#include "operators/fusion_fc_op.h"
namespace paddle_mobile {
namespace operators {
......@@ -54,6 +55,8 @@ void FusionFcOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_fc, ops::FusionFcMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_fc, ops::FusionFcOp);
#endif
......@@ -64,4 +67,4 @@ REGISTER_OPERATOR_MALI_GPU(fusion_fc, ops::FusionFcOp);
REGISTER_OPERATOR_FPGA(fusion_fc, ops::FusionFcOp);
#endif
#endif
#endif // FUSION_FC_OP
src/operators/fusion_fc_op.h
浏览文件 @ 1d76df93
......@@ -25,8 +25,7 @@ limitations under the License. */
namespace paddle_mobile {
namespace operators {
using std::string;
using std::vector;
class FusionFcMatcher : public framework::FusionOpMatcher {
public:
FusionFcMatcher() {
......@@ -49,7 +48,7 @@ class FusionFcOp : public framework::OperatorWithKernel<
DeviceType, FusionFcParam<DeviceType>,
operators::FusionFcKernel<DeviceType, T>> {
public:
FusionFcOp(const string &type, const VariableNameMap &inputs,
FusionFcOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
std::shared_ptr<framework::Scope> scope)
......@@ -60,42 +59,11 @@ class FusionFcOp : public framework::OperatorWithKernel<
using framework::OperatorWithKernel<
DeviceType, FusionFcParam<DeviceType>,
operators::FusionFcKernel<DeviceType, T>>::OperatorWithKernel;
void InferShape() const override;
protected:
void InferShape() const override;
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_FC_REGISTER
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#define FUSION_FC_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_FC_REGISTER
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#define FUSION_FC_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_FC_REGISTER
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#define FUSION_FC_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_fc);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fusion_fc);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_fc);
#endif
#endif
#endif // FUSION_FC_OP
src/operators/fusion_fc_relu_op.cpp
浏览文件 @ 1d76df93
......@@ -54,6 +54,9 @@ void FusionFcReluOp<Dtype, T>::InferShape() const {
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
REGISTER_FUSION_MATCHER(fusion_fc_relu, ops::FusionFcReluMatcher);
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_fc_relu, ops::FusionFcReluOp);
#endif
......
src/operators/fusion_fc_relu_op.h
浏览文件 @ 1d76df93
......@@ -64,39 +64,7 @@ class FusionFcReluOp : public framework::OperatorWithKernel<
protected:
};
#ifdef PADDLE_MOBILE_CPU
#ifndef FUSION_FC_RELU_REGISTER
static framework::FusionOpRegistrar fc_relu_registrar(
new FusionFcReluMatcher());
#define FUSION_FC_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_FC_RELU_REGISTER
static framework::FusionOpRegistrar fc_relu_registrar(
new FusionFcReluMatcher());
#define FUSION_FC_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#ifndef FUSION_FC_RELU_REGISTER
static framework::FusionOpRegistrar fc_relu_registrar(
new FusionFcReluMatcher());
#define FUSION_FC_RELU_REGISTER
#endif
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_fc_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fusion_fc_relu);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(fusion_fc_relu);
#endif
#endif // FUSION_FC_RELU_OP
src/operators/gru_op.cpp
浏览文件 @ 1d76df93
......@@ -64,8 +64,6 @@ namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(gru, ops::GruOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
......
src/operators/gru_op.h
浏览文件 @ 1d76df93
......@@ -47,12 +47,4 @@ class GruOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(gru);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/im2sequence_op.h
浏览文件 @ 1d76df93
......@@ -50,12 +50,4 @@ class Im2SequenceOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(im2sequence);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/kernel/arm/bilinear_interp_kernel.cpp
浏览文件 @ 1d76df93
......@@ -11,7 +11,6 @@ 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
#ifdef BILINEAR_INTERP_OP
......
src/operators/kernel/arm/conv_transpose_kernel.cpp
浏览文件 @ 1d76df93
......@@ -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 CONV_TRANSPOSE
#ifdef CONV_TRANSPOSE_OP
#include "operators/kernel/conv_transpose_kernel.h"
#include "operators/kernel/central-arm-func/conv_transpose_arm_func.h"
......
src/operators/kernel/arm/dequantize_kernel.cpp 0 → 100644
浏览文件 @ 1d76df93
/* 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 PADDLE_MOBILE_CPU
#include "operators/kernel/dequantize_kernel.h"
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#include <arm_neon.h>
#endif
namespace paddle_mobile {
namespace operators {
template <>
bool DequantizeKernel<CPU, float>::Init(DequantizeParam<CPU> *param) {
return true;
}
template <>
void DequantizeKernel<CPU, float>::Compute(
const DequantizeParam<CPU> &param) const {
const Tensor *input = param.input_;
Tensor *output = param.out_;
float activation_scale = param.activation_scale_->data<float>()[0];
float weight_scale = param.weight_scale_;
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);
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
size_t loop = size >> 4;
size_t remain = size & 0xF;
float32x4_t s = vdupq_n_f32(scale);
for (size_t i = 0; i < loop; ++i) {
int32x4_t r0 = vld1q_s32(x);
int32x4_t r1 = vld1q_s32(x + 4);
int32x4_t r2 = vld1q_s32(x + 8);
int32x4_t r3 = vld1q_s32(x + 12);
float32x4_t f0 = vcvtq_f32_s32(r0);
float32x4_t f1 = vcvtq_f32_s32(r1);
float32x4_t f2 = vcvtq_f32_s32(r2);
float32x4_t f3 = vcvtq_f32_s32(r3);
f0 = vmulq_f32(f0, s);
f1 = vmulq_f32(f1, s);
f2 = vmulq_f32(f2, s);
f3 = vmulq_f32(f3, s);
vst1q_f32(y, f0);
vst1q_f32(y + 4, f1);
vst1q_f32(y + 8, f2);
vst1q_f32(y + 12, f3);
x += 16;
y += 16;
}
size = remain;
#endif
for (size_t i = 0; i < size; ++i) {
y[i] = x[i] * scale;
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/arm/flatten_kernel.cpp
浏览文件 @ 1d76df93
......@@ -11,7 +11,6 @@ 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
#ifdef FLATTEN_OP
......
src/operators/kernel/arm/quantize_kernel.cpp 0 → 100644
浏览文件 @ 1d76df93
/* 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 PADDLE_MOBILE_CPU
#include "operators/kernel/quantize_kernel.h"
#include <cmath>
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#include <arm_neon.h>
#ifndef __aarch64__
float32_t vmaxvq_f32(float32x4_t r) {
float32x2_t v = vmax_f32(vget_high_f32(r), vget_low_f32(r));
return vget_lane_f32(vpmax_f32(v, v), 0);
}
#endif
int32x4_t vrnd_towards_zero(float32x4_t r) { return vcvtq_s32_f32(r); }
int32x4_t vrnd_away_zero(float32x4_t r) {
float32x4_t plus = vdupq_n_f32(0.5);
float32x4_t minus = vdupq_n_f32(-0.5);
float32x4_t zero = vdupq_n_f32(0);
uint32x4_t more_than_zero = vcgtq_f32(r, zero);
float32x4_t temp = vbslq_f32(more_than_zero, plus, minus);
temp = vaddq_f32(r, temp);
int32x4_t ret = vcvtq_s32_f32(temp);
return ret;
}
int32x4_t vrnd_to_even(float32x4_t r) {
#if 0
int32x4_t ret;
float value[4];
vst1q_f32(value, r);
for (int i = 0; i < 4; ++i) {
float v = round(value[i]);
int32_t q = (int32_t)v;
if (abs(abs(v - value[i]) - 0.5) > 0) {
ret[i] = q;
} else {
if (abs(q) % 2 == 0) {
ret[i] = q;
} else {
ret[i] = q + ((q > 0) ? -1 : 1);
}
}
}
return ret;
#else
float32x4_t point5 = vdupq_n_f32(0.5);
int32x4_t one = vdupq_n_s32(1);
int32x4_t zero = vdupq_n_s32(0);
int32x4_t rnd = vrnd_away_zero(r);
float32x4_t frnd = vcvtq_f32_s32(rnd);
frnd = vsubq_f32(frnd, r);
frnd = vabsq_f32(frnd);
uint32x4_t equal_point5 = vceqq_f32(frnd, point5);
int32x4_t abs_rnd = vabsq_s32(rnd);
abs_rnd = vandq_s32(abs_rnd, one);
uint32x4_t not_mod2 = vreinterpretq_u32_s32(abs_rnd);
uint32x4_t mask = vandq_u32(equal_point5, not_mod2);
uint32x4_t more_than_zero = vcgtq_s32(rnd, zero);
more_than_zero = vandq_u32(more_than_zero, vreinterpretq_u32_s32(one));
mask = veorq_u32(more_than_zero, mask);
more_than_zero = veorq_u32(more_than_zero, vreinterpretq_u32_s32(one));
mask = vaddq_u32(more_than_zero, mask);
int32x4_t smask = vreinterpretq_s32_u32(mask);
smask = vsubq_s32(smask, one);
rnd = vaddq_s32(rnd, smask);
return rnd;
#endif
}
#endif
namespace paddle_mobile {
namespace operators {
static float find_abs_max(const Tensor *input) {
float max_abs = 0.f;
const float *x = input->data<const float>();
size_t size = input->numel();
#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) {
float32x4_t max;
float32x4_t r0 = vld1q_f32(x);
float32x4_t r1 = vld1q_f32(x + 4);
float32x4_t r2 = vld1q_f32(x + 8);
float32x4_t r3 = vld1q_f32(x + 12);
r0 = vabsq_f32(r0);
r1 = vabsq_f32(r1);
r2 = vabsq_f32(r2);
r3 = vabsq_f32(r3);
max[0] = vmaxvq_f32(r0);
max[1] = vmaxvq_f32(r1);
max[2] = vmaxvq_f32(r2);
max[3] = vmaxvq_f32(r3);
max[0] = vmaxvq_f32(max);
if (max[0] > max_abs) {
max_abs = max[0];
}
x += 16;
}
size = remain;
#endif
for (size_t i = 0; i < size; ++i) {
float value = std::abs(x[i]);
if (value > max_abs) {
max_abs = value;
}
}
return max_abs;
}
static void quantize_round_to_even(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();
#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) {
float32x4_t r0 = vld1q_f32(x);
float32x4_t r1 = vld1q_f32(x + 4);
float32x4_t r2 = vld1q_f32(x + 8);
float32x4_t r3 = vld1q_f32(x + 12);
r0 = vmulq_n_f32(r0, scale);
r1 = vmulq_n_f32(r1, scale);
r2 = vmulq_n_f32(r2, scale);
r3 = vmulq_n_f32(r3, scale);
int32x4_t q0 = vrnd_to_even(r0);
int32x4_t q1 = vrnd_to_even(r1);
int32x4_t q2 = vrnd_to_even(r2);
int32x4_t q3 = vrnd_to_even(r3);
int16x4_t d0 = vmovn_s32(q0);
int16x4_t d1 = vmovn_s32(q1);
int16x4_t d2 = vmovn_s32(q2);
int16x4_t d3 = vmovn_s32(q3);
int16x8_t q5 = vcombine_s16(d0, d1);
int16x8_t q6 = vcombine_s16(d2, d3);
int8x8_t d5 = vmovn_s16(q5);
int8x8_t d6 = vmovn_s16(q6);
vst1_s8(y, d5);
vst1_s8(y + 8, d6);
x += 16;
y += 16;
}
size = remain;
#endif
for (size_t i = 0; i < size; ++i) {
float value = x[i] * scale;
float v = round(value);
int32_t q = (int32_t)v;
if (abs(abs(q - value) - 0.5) > 0) {
y[i] = q;
} else {
if (abs(q) % 2 == 0) {
y[i] = q;
} else {
y[i] = q + ((q > 0) ? -1 : 1);
}
}
}
}
static void quantize_round_to_zero(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();
#ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
size_t loop = size >> 4;
size_t remain = size & 0xF;
for (size_t i = 0; i < loop; ++i) {
float32x4_t r0 = vld1q_f32(x);
float32x4_t r1 = vld1q_f32(x + 4);
float32x4_t r2 = vld1q_f32(x + 8);
float32x4_t r3 = vld1q_f32(x + 12);
r0 = vmulq_n_f32(r0, scale);
r1 = vmulq_n_f32(r1, scale);
r2 = vmulq_n_f32(r2, scale);
r3 = vmulq_n_f32(r3, scale);
int32x4_t q0 = vrnd_towards_zero(r0);
int32x4_t q1 = vrnd_towards_zero(r1);
int32x4_t q2 = vrnd_towards_zero(r2);
int32x4_t q3 = vrnd_towards_zero(r3);
int16x4_t d0 = vmovn_s32(q0);
int16x4_t d1 = vmovn_s32(q1);
int16x4_t d2 = vmovn_s32(q2);
int16x4_t d3 = vmovn_s32(q3);
int16x8_t q5 = vcombine_s16(d0, d1);
int16x8_t q6 = vcombine_s16(d2, d3);
int8x8_t d5 = vmovn_s16(q5);
int8x8_t d6 = vmovn_s16(q6);
vst1_s8(y, d5);
vst1_s8(y + 8, d6);
x += 16;
y += 16;
}
size = remain;
#endif
for (size_t i = 0; i < size; ++i) {
y[i] = trunc(x[i] * 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();
#ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
size_t loop = size >> 4;
size_t remain = size & 0xF;
for (size_t i = 0; i < loop; ++i) {
float32x4_t r0 = vld1q_f32(x);
float32x4_t r1 = vld1q_f32(x + 4);
float32x4_t r2 = vld1q_f32(x + 8);
float32x4_t r3 = vld1q_f32(x + 12);
r0 = vmulq_n_f32(r0, scale);
r1 = vmulq_n_f32(r1, scale);
r2 = vmulq_n_f32(r2, scale);
r3 = vmulq_n_f32(r3, scale);
int32x4_t q0 = vrnd_away_zero(r0);
int32x4_t q1 = vrnd_away_zero(r1);
int32x4_t q2 = vrnd_away_zero(r2);
int32x4_t q3 = vrnd_away_zero(r3);
int16x4_t d0 = vmovn_s32(q0);
int16x4_t d1 = vmovn_s32(q1);
int16x4_t d2 = vmovn_s32(q2);
int16x4_t d3 = vmovn_s32(q3);
int16x8_t q5 = vcombine_s16(d0, d1);
int16x8_t q6 = vcombine_s16(d2, d3);
int8x8_t d5 = vmovn_s16(q5);
int8x8_t d6 = vmovn_s16(q6);
vst1_s8(y, d5);
vst1_s8(y + 8, d6);
x += 16;
y += 16;
}
size = remain;
#endif
for (size_t i = 0; i < size; ++i) {
y[i] = round(x[i] * scale);
}
}
template <>
bool QuantizeKernel<CPU, float>::Init(QuantizeParam<CPU> *param) {
return true;
}
template <>
void QuantizeKernel<CPU, float>::Compute(
const QuantizeParam<CPU> &param) const {
float max_abs = 0.f;
const Tensor *input = param.input_;
Tensor *output = param.out_;
Tensor *output_scale = param.online_scale_;
if (param.is_static_) {
max_abs = param.static_scale_;
} else {
max_abs = find_abs_max(input);
}
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;
switch (param.round_type_) {
case ROUND_NEAREST_TO_EVEN:
quantize_round_to_even(input, online_scale, output);
break;
case ROUND_NEAREST_TOWARDS_ZERO:
quantize_round_to_zero(input, online_scale, output);
break;
case ROUND_NEAREST_AWAY_ZERO:
quantize_round_to_nearest(input, online_scale, output);
default:
LOG(kLOG_ERROR) << "round type is not supported.";
break;
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/arm/shape_kernel.cpp
浏览文件 @ 1d76df93
......@@ -11,7 +11,6 @@ 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
#ifdef SHAPE_OP
......
src/operators/kernel/arm/split_kernel.cpp
浏览文件 @ 1d76df93
......@@ -11,7 +11,6 @@ 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
#ifdef SPLIT_OP
......
src/operators/kernel/central-arm-func/conv_transpose_arm_func.h
浏览文件 @ 1d76df93
......@@ -12,18 +12,17 @@ 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_TRANSPOSE
#pragma once
#include <vector>
#ifdef CONV_TRANSPOSE_OP
#include <vector>
#include "framework/ddim.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
#include "operators/op_param.h"
#pragma once
namespace paddle_mobile {
namespace operators {
......
src/operators/kernel/central-arm-func/sigmoid_arm_func.h
浏览文件 @ 1d76df93
......@@ -24,7 +24,9 @@ limitations under the License. */
namespace paddle_mobile {
namespace operators {
using framework::DDim;
void sigmoid(const Tensor *X, Tensor *Y) {
#ifdef __ARM_NEON
const float *input = X->data<float>();
......
src/operators/kernel/conv_transpose_kernel.h
浏览文件 @ 1d76df93
......@@ -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 CONV_TRANSPOSE
#ifdef CONV_TRANSPOSE_OP
#pragma once
......
src/operators/kernel/dequantize_kernel.h 0 → 100644
浏览文件 @ 1d76df93
/* 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/operator.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
template <typename DeviceType, typename T>
class DequantizeKernel
: public framework::OpKernelBase<DeviceType, DequantizeParam<DeviceType>> {
public:
void Compute(const DequantizeParam<DeviceType> &param) const;
bool Init(DequantizeParam<DeviceType> *param);
};
} // namespace operators
} // namespace paddle_mobile
src/operators/kernel/fpga/concat_kernel.cpp
浏览文件 @ 1d76df93
......@@ -24,10 +24,12 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
auto inputs = param->Inputs();
auto out = param->Out();
auto image_num = inputs.size();
auto images_in = (half **)fpga::fpga_malloc(image_num * sizeof(int *));
auto scales_in = (float **)fpga::fpga_malloc(image_num * sizeof(float *));
auto images_in =
(half **)fpga::fpga_malloc(image_num * sizeof(int *)); // NOLINT
auto scales_in =
(float **)fpga::fpga_malloc(image_num * sizeof(float *)); // NOLINT
auto channel_num =
(uint32_t *)fpga::fpga_malloc(image_num * sizeof(uint32_t));
(uint32_t *)fpga::fpga_malloc(image_num * sizeof(uint32_t)); // NOLINT
auto height = inputs[0]->dims()[2];
auto width = inputs[0]->dims()[3];
......@@ -36,22 +38,21 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
PADDLE_MOBILE_ENFORCE(
input->dims()[2] == height && input->dims()[3] == width,
"Image height & width should be unified");
images_in[i] = (half *)input->data<float>();
channel_num[i] = (uint32_t)inputs[i]->dims()[1];
images_in[i] = (half *)input->data<float>(); // NOLINT
channel_num[i] = (uint32_t)inputs[i]->dims()[1]; // NOLINT
scales_in[i] = input->scale;
}
fpga::format_concat_output(out, (int)height, (int)width, (int)image_num,
channel_num);
fpga::format_concat_output(out, height, width, image_num, channel_num);
fpga::ConcatArgs concatArgs = {0};
concatArgs.image_num = (uint32_t)image_num;
concatArgs.image_num = image_num;
concatArgs.images_in = images_in;
concatArgs.scales_in = scales_in;
concatArgs.image_out = (half *)out->data<float>();
concatArgs.image_out = (half *)out->data<float>(); // NOLINT
concatArgs.scale_out = out->scale;
concatArgs.channel_num = channel_num;
concatArgs.height = (uint32_t)height;
concatArgs.width = (uint32_t)width;
concatArgs.height = height;
concatArgs.width = width;
param->SetFpgaArgs(concatArgs);
return true;
}
......
src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp
浏览文件 @ 1d76df93
......@@ -38,7 +38,8 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
"Output channel should be equal to bias number");
const int channel = out->dims()[1];
auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
auto bs_ptr =
(float *)fpga::fpga_malloc(2 * channel * sizeof(float)); // NOLINT
auto new_scale = new Tensor();
auto new_bias = new Tensor();
auto new_scale_ptr = new_scale->mutable_data<float>({channel});
......
src/operators/kernel/fpga/conv_add_relu_kernel.cpp
浏览文件 @ 1d76df93
......@@ -31,7 +31,8 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
PADDLE_MOBILE_ENFORCE(out->dims()[1] == bias->dims()[0],
"Output channel should be equal to bias number");
int channel = out->dims()[1];
auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
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] = bias_ptr[i];
......
src/operators/kernel/fpga/conv_bn_kernel.cpp
浏览文件 @ 1d76df93
......@@ -33,7 +33,8 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
"Output channel should be equal to bias number");
const int channel = out->dims()[1];
auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
auto bs_ptr =
(float *)fpga::fpga_malloc(2 * channel * sizeof(float)); // // NOLINT
auto new_scale = new Tensor();
auto new_bias = new Tensor();
auto new_scale_ptr = new_scale->mutable_data<float>({channel});
......
src/operators/kernel/fpga/conv_bn_relu_kernel.cpp
浏览文件 @ 1d76df93
......@@ -33,7 +33,8 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
"Output channel should be equal to bias number");
const int channel = out->dims()[1];
auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
auto bs_ptr =
(float *)fpga::fpga_malloc(2 * channel * sizeof(float)); // NOLINT
auto new_scale = new Tensor();
auto new_bias = new Tensor();
auto new_scale_ptr = new_scale->mutable_data<float>({channel});
......
src/operators/kernel/fpga/fc_relu_kernel.cpp
浏览文件 @ 1d76df93
......@@ -28,7 +28,8 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
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));
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] = input_z_ptr[i];
......@@ -61,7 +62,7 @@ template <>
void FusionFcReluKernel<FPGA, float>::Compute(
const FusionFcReluParam<FPGA> &param) const {
fpga::ComputeFpgaConv(param.FpgaArgs());
};
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/fpga/fusion_fc_kernel.cpp
浏览文件 @ 1d76df93
......@@ -30,7 +30,8 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
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));
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] = input_z_ptr[i];
......
src/operators/kernel/fpga/softmax_kernel.cpp
浏览文件 @ 1d76df93
......@@ -54,8 +54,8 @@ void SoftmaxKernel<FPGA, float>::Compute(
fpga::PerformBypass(param.FpgaArgs());
fpga::fpga_invalidate(
(void *)in_x->data<float>(),
(size_t)fpga::get_align_image_cw((int)in_x->dims()[1]) * sizeof(float));
(void *)in_x->data<float>(), // NOLINT
fpga::get_align_image_cw(in_x->dims()[1]) * sizeof(float));
math::SoftmaxFuntor<CPU, float>()(in_x, out);
fpga::fpga_flush(out->data<float>(), out->memory_size());
......
src/operators/kernel/quantize_kernel.h 0 → 100644
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/lookup_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/lrn_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/lrn_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/math/pool_3x3.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/mul_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/mul_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/multiclass_nms_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/multiclass_nms_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/op_param.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/pool_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/prelu_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/prelu_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/prior_box_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/prior_box_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/quantize_op.cpp 0 → 100644
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/quantize_op.h 0 → 100644
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/relu_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/relu_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/reshape_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/reshape_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/resize_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/scale_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/shape_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/shape_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/sigmoid_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/sigmoid_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/slice_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/softmax_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/split_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/split_op.h
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/transpose_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
src/operators/transpose_op.h
浏览文件 @ 1d76df93
此差异已折叠。
test/CMakeLists.txt
浏览文件 @ 1d76df93
此差异已折叠。
test/operators/test_dequantize_op.cpp 0 → 100644
浏览文件 @ 1d76df93
此差异已折叠。
test/operators/test_quantize_op.cpp 0 → 100644
浏览文件 @ 1d76df93
此差异已折叠。
test/operators/test_sigmoid_op.cpp
浏览文件 @ 1d76df93
此差异已折叠。
tools/op.cmake
浏览文件 @ 1d76df93
此差异已折叠。
tools/pre-commit.hooks/cpplint.hook
浏览文件 @ 1d76df93
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或