Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 14114fcf 编写于 作者: E eclipsess
浏览文件
操作

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

上级 1f680e4c 663d0fd4
展开全部 显示空白变更内容
内联 并排
Showing with 877 addition and 478 deletion
CMakeLists.txt
浏览文件 @ 14114fcf
cmake_minimum_required(VERSION 3.0)
project(paddle-mobile)
option(DEBUGING "enable debug mode" OFF)
option(DEBUGING "enable debug mode" ON)
option(USE_OPENMP "openmp support" OFF)
option(USE_EXCEPTION "use std exception" OFF)
option(USE_EXCEPTION "use std exception" ON)
option(LOG_PROFILE "log profile" ON)
# select the platform to build
option(CPU "cpu" ON)
option(MALI_GPU "mali gpu" OFF)
option(FPGA "fpga" OFF)
if (CPU)
add_definitions(-DPADDLE_MOBILE_CPU)
elseif (MALI_GPU)
add_definitions(-DPADDLE_MOBILE_MALI_GPU)
elseif(FPGA)
add_definitions(-DPADDLE_MOBILE_FPGA)
endif()
set(CMAKE_CXX_FLAGS "-std=c++14 -O3 -s ${CMAKE_CXX_FLAGS}")
if (DEBUGING)
set(CMAKE_BUILD_TYPE Debug)
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS}")
else()
set(CMAKE_BUILD_TYPE Release)
endif ()
......@@ -24,12 +39,17 @@ else()
endif()
if (USE_EXCEPTION)
message(STATUS "use exception")
add_definitions(-DENABLE_EXCEPTION)
add_definitions(-fexceptions)
else()
add_definitions(-fno-exceptions)
endif ()
if (LOG_PROFILE)
add_definitions(-DPADDLE_MOBILE_PROFILE)
endif()
if(IS_MAC)
add_definitions(-DX86)
elseif(IS_IOS)
......@@ -42,7 +62,6 @@ else ()
add_definitions(-DX86)
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++14")
set(CMAKE_VERBOSE_MAKEFILE ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY build)
......@@ -74,6 +93,7 @@ if (googlenet)
add_definitions(-DFUSION_FC_OP)
add_definitions(-DPOOL_OP)
add_definitions(-DRELU_OP)
add_definitions(-DFUSION_CONVADD_OP)
elseif (mobilenet)
add_definitions(-DCONV_OP)
add_definitions(-DELEMENTWISEADD_OP)
......@@ -112,7 +132,7 @@ else ()
add_definitions(-DCONV_OP)
add_definitions(-DDEPTHWISECONV_OP)
add_definitions(-DELEMENTWISEADD_OP)
add_definitions(-DFUSIONCONVADD_OP)
add_definitions(-DFUSION_CONVADD_OP)
add_definitions(-DCONVADDRELU_OP)
add_definitions(-DFUSION_FC_OP)
add_definitions(-DLRN_OP)
......@@ -127,8 +147,13 @@ else ()
add_definitions(-DTRANSPOSE_OP)
endif()
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
if (IS_IOS)
add_library(paddle-mobile STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
elseif(ANDROID)
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
else()
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
endif ()
if(DEBUGING)
add_subdirectory(test)
......
src/common/types.h
浏览文件 @ 14114fcf
......@@ -99,6 +99,7 @@ static std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
op_input_output_key = {
{G_OP_TYPE_CONV, {{"Input"}, {"Output"}}},
{G_OP_TYPE_CONV_ADD, {{"Input"}, {"Out"}}},
{G_OP_TYPE_RELU, {{"X"}, {"Out"}}},
{G_OP_TYPE_SOFTMAX, {{"X"}, {"Out"}}},
{G_OP_TYPE_MUL, {{"X"}, {"Out"}}},
......
src/framework/op_registry.h
浏览文件 @ 14114fcf
......@@ -96,24 +96,39 @@ class OpRegistry {
}
};
#define REGISTER_OPERATOR(op_type, op_class) \
#define REGISTER_OPERATOR(op_type, op_class, device_name, device_type) \
template <typename Dtype, typename T> \
class _OpClass_##op_type##_ : public op_class<Dtype, T> { \
class _OpClass_##op_type##_##device_name : public op_class<Dtype, T> { \
public: \
DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \
DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_##device_name, op_class); \
}; \
static paddle_mobile::framework::OperatorRegistrar< \
paddle_mobile::CPU, _OpClass_##op_type##_<paddle_mobile::CPU, float>> \
__op_registrar_##op_type##__(#op_type); \
int TouchOpRegistrar_##op_type() { \
__op_registrar_##op_type##__.Touch(); \
device_type, _OpClass_##op_type##_##device_name<device_type, float>> \
__op_registrar_##op_type##_##device_name(#op_type); \
int TouchOpRegistrar_##op_type##_##device_name() { \
__op_registrar_##op_type##_##device_name.Touch(); \
return 0; \
}
#define USE_OP(op_type) \
extern int TouchOpRegistrar_##op_type(); \
static int use_op_itself_##op_type##_ __attribute__((unused)) = \
TouchOpRegistrar_##op_type()
#define REGISTER_OPERATOR_CPU(op_type, op_class) \
REGISTER_OPERATOR(op_type, op_class, cpu, paddle_mobile::CPU);
#define REGISTER_OPERATOR_MALI_GPU(op_type, op_class) \
REGISTER_OPERATOR(op_type, op_class, mali_gpu, paddle_mobile::GPU_MALI);
#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/operator.cpp
浏览文件 @ 14114fcf
......@@ -58,7 +58,12 @@ void OperatorBase<Dtype>::Run() const {
}
template class OperatorBase<CPU>;
template class OperatorBase<FPGA>;
template class OperatorBase<GPU_MALI>;
template class OperatorWithKernel<CPU>;
template class OperatorWithKernel<FPGA>;
template class OperatorWithKernel<GPU_MALI>;
} // namespace framework
} // namespace paddle_mobile
src/framework/operator.h
浏览文件 @ 14114fcf
......@@ -153,6 +153,7 @@ class FusionOpMatcher {
std::string BeginType() { return node_.Type(); }
// virtual bool Fusion();
protected:
Node node_;
std::string type_;
......
src/framework/tensor.h
浏览文件 @ 14114fcf
......@@ -131,7 +131,6 @@ class Tensor {
}
PADDLE_MOBILE_ENFORCE(numel() >= 0, "the Tensor'snumel must >=0.")
int64_t size = numel() * SizeOfType(type);
/* some versions of boost::variant don't have operator!= */
if (holder_ == nullptr || holder_->size() < size + offset_) {
holder_.reset(new PlaceholderImpl(size, type));
offset_ = 0;
......
src/io/io.cpp
浏览文件 @ 14114fcf
......@@ -14,6 +14,10 @@ limitations under the License. */
#include "io.h"
#include <vector>
#ifdef PADDLE_MOBILE_PROFILE
#include <ctime>
#include <map>
#endif
#include "common/enforce.h"
#include "common/log.h"
......@@ -336,10 +340,34 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
feed_tensor->ShareDataWith(t);
std::shared_ptr<framework::BlockDesc> to_predict_block =
to_predict_program_->Block(0);
#ifdef PADDLE_MOBILE_PROFILE
std::map<std::string, clock_t> _profile;
#endif
for (int j = 0; j < ops_of_block_[*to_predict_block.get()].size(); ++j) {
auto op = ops_of_block_[*to_predict_block.get()][j];
#ifdef PADDLE_MOBILE_PROFILE
_profile[op->Type()] = clock();
#endif
op->Run();
#ifdef PADDLE_MOBILE_PROFILE
_profile[op->Type()] = clock() - _profile[op->Type()];
#endif
}
#ifdef PADDLE_MOBILE_PROFILE
{
DLOG << "========================[ profile ]==========================";
clock_t _ptotal = 0;
for (auto const &p : _profile) {
_ptotal += p.second;
}
for (auto const &p : _profile) {
DLOG << p.first << std::string(16 - p.first.size(), ' ') << "\t"
<< (float)p.second << "\t\t"
<< (float)p.second / (float)_ptotal * 100.0;
}
DLOG << "========================[ ]==========================";
}
#endif
auto ops = ops_of_block_[*to_predict_program_->Block(0)];
auto last_op = ops.rbegin();
auto output_map = (*last_op)->Outputs();
......
src/operators/batchnorm_op.cpp
浏览文件 @ 14114fcf
......@@ -31,7 +31,13 @@ template class BatchNormOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(batch_norm);
REGISTER_OPERATOR(batch_norm, ops::BatchNormOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(batch_norm);
REGISTER_OPERATOR_CPU(batch_norm, ops::BatchNormOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/box_coder_op.cpp
浏览文件 @ 14114fcf
......@@ -52,7 +52,13 @@ template class BoxCoderOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(box_coder);
REGISTER_OPERATOR(box_coder, ops::BoxCoderOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(box_coder);
REGISTER_OPERATOR_CPU(box_coder, ops::BoxCoderOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/concat_op.cpp
浏览文件 @ 14114fcf
......@@ -62,7 +62,13 @@ template class ConcatOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(concat);
REGISTER_OPERATOR(concat, ops::ConcatOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(concat);
REGISTER_OPERATOR_CPU(concat, ops::ConcatOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/conv_op.cpp
浏览文件 @ 14114fcf
......@@ -53,7 +53,17 @@ template class ConvOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(conv2d);
REGISTER_OPERATOR(conv2d, ops::ConvOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv2d);
REGISTER_OPERATOR_CPU(conv2d, ops::ConvOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv2d);
REGISTER_OPERATOR_MALI_GPU(conv2d, ops::ConvOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(conv2d);
REGISTER_OPERATOR_FPGA(conv2d, ops::ConvOp);
#endif
#endif
src/operators/depthwise_conv_op.cpp
浏览文件 @ 14114fcf
......@@ -54,7 +54,13 @@ template class DepthwiseConvOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(depthwise_conv2d);
REGISTER_OPERATOR(depthwise_conv2d, ops::DepthwiseConvOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(depthwise_conv2d);
REGISTER_OPERATOR_CPU(depthwise_conv2d, ops::DepthwiseConvOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/elementwise_add_op.cpp
浏览文件 @ 14114fcf
......@@ -29,7 +29,13 @@ template class ElementwiseAddOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(elementwise_add);
REGISTER_OPERATOR(elementwise_add, ops::ElementwiseAddOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(elementwise_add);
REGISTER_OPERATOR_CPU(elementwise_add, ops::ElementwiseAddOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/feed_op.h
浏览文件 @ 14114fcf
......@@ -43,8 +43,14 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
};
namespace ops = paddle_mobile::operators;
USE_OP(feed);
REGISTER_OPERATOR(feed, ops::FeedOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(feed);
REGISTER_OPERATOR_CPU(feed, ops::FeedOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/fetch_op.h
浏览文件 @ 14114fcf
......@@ -43,8 +43,14 @@ class FetchOp : public framework::OperatorBase<DeviceType> {
};
namespace ops = paddle_mobile::operators;
USE_OP(fetch);
REGISTER_OPERATOR(fetch, ops::FetchOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fetch);
REGISTER_OPERATOR_CPU(fetch, ops::FetchOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/fusion_conv_add.cpp
浏览文件 @ 14114fcf
......@@ -12,20 +12,49 @@ 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 FUSIONCONVADD_OP
#ifdef FUSION_CONVADD_OP
#include "operators/fusion_conv_add.h"
namespace paddle_mobile {
namespace operators {
template <typename Dtype, typename T>
void FushionConvAddOp<Dtype, T>::InferShape() const {}
void FushionConvAddOp<Dtype, T>::InferShape() const {
auto in_dims = param_.Input()->dims();
auto filter_dims = param_.Filter()->dims();
const std::vector<int> &strides = param_.Strides();
std::vector<int> paddings = param_.Paddings();
int groups = param_.Groups();
std::vector<int> dilations = param_.Dilations();
PADDLE_MOBILE_ENFORCE((in_dims.size() == filter_dims.size() &&
dilations.size() == paddings.size() &&
paddings.size() == strides.size()),
"ConvParam is not suitable");
std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
for (size_t i = 0; i < strides.size(); ++i) {
output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
dilations[i], paddings[i],
strides[i]));
}
framework::DDim ddim = framework::make_ddim(output_shape);
param_.Output()->Resize(ddim);
}
template class FushionConvAddOp<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(conv_add);
REGISTER_OPERATOR(conv_add, ops::FushionConvAddOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv_add);
REGISTER_OPERATOR_CPU(conv_add, ops::FushionConvAddOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add.h
浏览文件 @ 14114fcf
......@@ -11,16 +11,17 @@ 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 FUSIONCONVADD_OP
#define FUSION_CONVADD_OP
#ifdef FUSION_CONVADD_OP
#pragma once
#include <string>
#include <vector>
#include "framework/operator.h"
#include "framework/program/program-optimize/fusion_op_register.h"
#include "op_param.h"
#include "operators/kernel/conv_add_kernel.h"
namespace paddle_mobile {
namespace operators {
......@@ -53,18 +54,37 @@ class FushionConvAddOp : public framework::OperatorWithKernel<DeviceType> {
const framework::AttributeMap &attrs,
std::shared_ptr<framework::Scope> scope)
: framework::OperatorWithKernel<DeviceType>(type, inputs, outputs, attrs,
scope) {}
scope),
param_(inputs, outputs, attrs, *scope) {}
void RunImpl() const {}
void RunImpl() const {
operators::ConvAddKernel<DeviceType, T> kernel;
kernel.Compute(param_);
this->ClearVariables({"Filter", "Input", "Y"});
}
using framework::OperatorWithKernel<DeviceType>::OperatorWithKernel;
void InferShape() const override;
protected:
// FushionFcParam param_;
FushionConvAddParam param_;
};
// static framework::FusionOpRegistrar fc_registrar(new FusionConvAddMatcher());
inline int ConvOutputSize(int input_size, int filter_size, int dilation,
int padding, int stride) {
const int dkernel = dilation * (filter_size - 1) + 1;
int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
return output_size;
}
#ifdef PADDLE_MOBILE_CPU
static framework::FusionOpRegistrar convadd_registrar(
new FusionConvAddMatcher());
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
......
src/operators/fusion_conv_add_relu_op.h
浏览文件 @ 14114fcf
......@@ -46,8 +46,14 @@ class ConvAddReluOp {
private:
};
// static framework::FusionOpRegistrar fc_registrar(
#ifdef PADDLE_MOBILE_CPU
// static framework::FusionOpRegistrar fusion_conv_add_relu_registrar(
// new FushionConvAddReluOpMatcher());
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
......
src/operators/fusion_fc_op.cpp
浏览文件 @ 14114fcf
......@@ -54,7 +54,13 @@ template class FushionFcOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(fc);
REGISTER_OPERATOR(fc, ops::FushionFcOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fc);
REGISTER_OPERATOR_CPU(fc, ops::FushionFcOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_fc_op.h
浏览文件 @ 14114fcf
......@@ -37,8 +37,6 @@ class FusionFcMatcher : public framework::FusionOpMatcher {
void FolderNodes(
framework::Node *node,
std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
vector<std::shared_ptr<framework::OpDesc>> origin_descs =
node->OpDescs(node_.Depth());
node->Folder(node_.Depth(), Type(),
{{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Z"}}}, removed_nodes);
}
......@@ -69,7 +67,14 @@ class FushionFcOp : public framework::OperatorWithKernel<DeviceType> {
FushionFcParam param_;
};
// static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#ifdef PADDLE_MOBILE_CPU
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/conv_add_kernel.cpp 0 → 100644
浏览文件 @ 14114fcf
/* 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 FUSION_CONVADD_OP
#include "operators/kernel/conv_add_kernel.h"
namespace paddle_mobile {
namespace operators {
void expand_bias(Tensor &bias, int axis, const DDim &dDim) {
auto bias_ptr = bias.data<float>();
const DDim bias_ddim = bias.dims();
PADDLE_MOBILE_ENFORCE(bias.dims().size() == 1,
"the bias tensor's dims size != 1")
DDim outer_ddim = paddle_mobile::framework::slice_ddim(dDim, 0, axis + 1);
DDim inner_ddim =
paddle_mobile::framework::slice_ddim(dDim, axis + 1, dDim.size());
int outer_size = paddle_mobile::framework::product(outer_ddim);
int inner_size = paddle_mobile::framework::product(inner_ddim);
bias.Resize(dDim);
auto new_ptr = bias.mutable_data<float>();
int axis_size = dDim[axis];
for (int i = 0; i < outer_size; ++i) {
float v_bias = bias_ptr[i * axis_size / outer_size];
for (int j = 0; j < inner_size; ++j) {
new_ptr[i * inner_size + j] = v_bias;
}
}
}
template <>
void ConvAddKernel<CPU, float>::Compute(
const FushionConvAddParam &param) const {
DLOG << param;
const Tensor *input = param.Input();
Tensor filter = *param.Filter();
Tensor bias = *param.Bias();
int axis = param.Axis();
Tensor *output = param.Output();
expand_bias(bias, axis, output->dims());
output->ShareDataWith(bias);
int groups = param.Groups();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
std::vector<int> dilations = param.Dilations();
const int batch_size = static_cast<int>(input->dims()[0]);
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
Tensor col_matrix;
if (is_expand) {
col.mutable_data<float>(col_shape);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output->dims()[1],
output->numel() / (output->dims()[0] * output->dims()[1])};
// convolution operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<CPU, float> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
for (int g = 0; g < groups; g++) {
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
// im2col
im2col(in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
// vol2col
vol2col(in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<float>(filter_slice, false, col_matrix, false,
static_cast<float>(1), &out_slice,
static_cast<float>(1));
}
}
}
template class ConvAddKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/arm/relu_kernel.cpp
浏览文件 @ 14114fcf
......@@ -37,13 +37,71 @@ void ReluKernel<CPU, float>::Compute(const ReluParam &param) const {
auto *out = param.Out();
auto *out_ptr = out->mutable_data<float>();
int numel = input_x->numel();
if (numel > 32) {
asm volatile(
"pld [%[input_x_ptr], #0] \n\t"
"vmov.f32 q8, #0.0 \n\t"
"subs %[num], %[num], #32 \n\t"
"blt end_num_%= \n\t"
"loop_num_%=: \n\t"
"pld [%[input_x_ptr], #1024] \n\t"
"vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
"vmax.f32 q0, q0, q8 \n\t"
"vmax.f32 q1, q1, q8 \n\t"
"vmax.f32 q2, q2, q8 \n\t"
"vmax.f32 q3, q3, q8 \n\t"
"vmax.f32 q4, q4, q8 \n\t"
"vmax.f32 q5, q5, q8 \n\t"
"vmax.f32 q6, q6, q8 \n\t"
"vmax.f32 q7, q7, q8 \n\t"
"vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
"vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
"vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
"vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
"subs %[num], %[num], #32 \n\t"
"bge loop_num_%= \n\t"
"end_num_%=: \n\t"
"cmp %[num], #0 \n\t"
"bge end_%= \n\t"
"mov r6, #4 \n\t"
"mul r5, %[num], r6 \n\t"
"add %[input_x_ptr], %[input_x_ptr], r5 \n\t"
"vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
"vmax.f32 q0, q0, q8 \n\t"
"vmax.f32 q1, q1, q8 \n\t"
"vmax.f32 q2, q2, q8 \n\t"
"vmax.f32 q3, q3, q8 \n\t"
"vmax.f32 q4, q4, q8 \n\t"
"vmax.f32 q5, q5, q8 \n\t"
"vmax.f32 q6, q6, q8 \n\t"
"vmax.f32 q7, q7, q8 \n\t"
"add %[out_ptr], %[out_ptr], r5 \n\t"
"vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
"vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
"vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
"vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
"end_%=: \n\t"
:
:
[out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num] "r"(numel)
: "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "r5",
"r6");
} else {
ReluFunctor<float> func_;
math::Transform trans;
trans(input_x_ptr, input_x_ptr + input_x->numel(), out_ptr, func_);
// for (int i = 0; i < input_x->numel(); i++) {
// out_ptr[i] = input_x_ptr[i] > 0 ? input_x_ptr[i] : 0;
// }
trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
}
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/conv_add_kernel.h 0 → 100644
浏览文件 @ 14114fcf
/* 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 FUSION_CONVADD_OP
#pragma once
#include <vector>
#include "framework/ddim.h"
#include "framework/operator.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
using framework::DDim;
using framework::OpKernelBase;
template <typename DeviceType, typename T>
class ConvAddKernel : public OpKernelBase<DeviceType, FushionConvAddParam> {
public:
void Compute(const FushionConvAddParam &param) const;
};
inline bool IsExpand(const std::vector<int64_t> &filter_dim,
const std::vector<int> &strides,
const std::vector<int> &paddings,
const std::vector<int> &dilations) {
bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true;
for (size_t j = 0; j < strides.size(); ++j) {
filter_1 = filter_1 && (static_cast<int>(filter_dim[j + 2]) == 1);
strides_1 = strides_1 && (strides[j] == 1);
padding_0 = padding_0 && (paddings[j] == 0);
dilation_1 = dilation_1 && (dilations[j] == 1);
}
return !(filter_1 && strides_1 && padding_0 && dilation_1);
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/fpga/conv_kernel.cpp
浏览文件 @ 14114fcf
......@@ -14,15 +14,16 @@ limitations under the License. */
#ifdef CONV_OP
#include "operators/kernel/conv_kernel.h"
namespace paddle_mobile {
namespace operators {
// template<>
// void ConvKernel<FPGA, float>::Compute(const ConvParam &param) const
// {}
//
// template class ConvKernel<FPGA, float>;
}
template <>
void ConvKernel<FPGA, float>::Compute(const ConvParam &param) const {}
template class ConvKernel<FPGA, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/lrn_op.cpp
浏览文件 @ 14114fcf
......@@ -29,7 +29,13 @@ template class LrnOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(lrn);
REGISTER_OPERATOR(lrn, ops::LrnOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(lrn);
REGISTER_OPERATOR_CPU(lrn, ops::LrnOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/mul_op.cpp
浏览文件 @ 14114fcf
......@@ -55,7 +55,13 @@ template class MulOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(mul);
REGISTER_OPERATOR(mul, ops::MulOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(mul);
REGISTER_OPERATOR_CPU(mul, ops::MulOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/multiclass_nms_op.cpp
浏览文件 @ 14114fcf
......@@ -39,7 +39,13 @@ template class MultiClassNMSOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(multiclass_nms);
REGISTER_OPERATOR(multiclass_nms, ops::MultiClassNMSOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(multiclass_nms);
REGISTER_OPERATOR_CPU(multiclass_nms, ops::MultiClassNMSOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/op_param.cpp
浏览文件 @ 14114fcf
......@@ -39,5 +39,31 @@ Print &operator<<(Print &printer, const ConvParam &conv_param) {
}
#endif
#ifdef FUSION_CONVADD_OP
Print &operator<<(Print &printer, const FushionConvAddParam &conv_param) {
printer << "parameter of conv_add: "
<< "\n";
printer << " stride: "
<< " (" << conv_param.Strides()[0] << conv_param.Strides()[1] << ") "
<< "\n";
printer << " paddings: "
<< " (" << conv_param.Paddings()[0] << conv_param.Paddings()[1]
<< ") "
<< "\n";
printer << " dilations: "
<< " (" << conv_param.Dilations()[0] << conv_param.Dilations()[1]
<< ") "
<< "\n";
printer << " groups: " << conv_param.Groups() << "\n";
printer << " input dims: " << conv_param.Input()->dims() << "\n";
printer << " filter dims: " << conv_param.Filter()->dims() << "\n";
printer << " bias dims: " << conv_param.Bias()->dims() << "\n";
printer << " output dims: " << conv_param.Output()->dims();
return printer;
}
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/op_param.h
浏览文件 @ 14114fcf
......@@ -165,6 +165,8 @@ class OpParam {
template <typename T>
static T *GetVarValue(const string &key, const VariableNameMap &var_map,
const Scope &scope) {
PADDLE_MOBILE_ENFORCE(var_map.count(key) > 0,
"%s is not contained in var_map", key.c_str())
auto var_vec = var_map.at(key);
if (!var_vec.empty()) {
auto var = scope.FindVar(var_vec[0]);
......@@ -787,5 +789,54 @@ class FushionFcParam : public OpParam {
};
#endif
#ifdef FUSION_CONVADD_OP
class FushionConvAddParam : public OpParam {
public:
FushionConvAddParam(const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs,
const Scope &scope) {
bias_ = InputYFrom<LoDTensor>(inputs, scope);
axis_ = GetAttr<int>("axis", attrs);
filter_ = FilterFrom<LoDTensor>(inputs, scope);
input_ = InputFrom<LoDTensor>(inputs, scope);
output_ = OutFrom<LoDTensor>(outputs, scope);
strides_ = GetAttr<vector<int>>("strides", attrs);
paddings_ = GetAttr<vector<int>>("paddings", attrs);
dilations_ = GetAttr<vector<int>>("dilations", attrs);
groups = GetAttr<int>("groups", attrs);
}
Tensor *Bias() const { return bias_; }
const int &Axis() const { return axis_; }
const Tensor *Input() const { return input_; }
const Tensor *Filter() const { return filter_; }
Tensor *Output() const { return output_; }
const vector<int> &Strides() const { return strides_; }
const vector<int> &Paddings() const { return paddings_; }
const vector<int> &Dilations() const { return dilations_; }
const int &Groups() const { return groups; }
private:
Tensor *bias_;
int axis_;
Tensor *input_;
Tensor *output_;
Tensor *filter_;
vector<int> strides_;
vector<int> paddings_;
vector<int> dilations_;
int groups;
};
Print &operator<<(Print &printer, const FushionConvAddParam &conv_param);
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/pool_op.cpp
浏览文件 @ 14114fcf
......@@ -59,7 +59,13 @@ template class PoolOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(pool2d);
REGISTER_OPERATOR(pool2d, ops::PoolOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(pool2d);
REGISTER_OPERATOR_CPU(pool2d, ops::PoolOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/prior_box_op.cpp
浏览文件 @ 14114fcf
......@@ -49,7 +49,13 @@ template class PriorBoxOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(prior_box);
REGISTER_OPERATOR(prior_box, ops::PriorBoxOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(prior_box);
REGISTER_OPERATOR_CPU(prior_box, ops::PriorBoxOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/relu_op.cpp
浏览文件 @ 14114fcf
......@@ -33,7 +33,13 @@ template class ReluOp<CPU, float>;
* 都是需要和model中类型对应起来的
* */
namespace ops = paddle_mobile::operators;
USE_OP(relu);
REGISTER_OPERATOR(relu, ops::ReluOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(relu);
REGISTER_OPERATOR_CPU(relu, ops::ReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/reshape_op.cpp
浏览文件 @ 14114fcf
......@@ -32,7 +32,13 @@ template class ReshapeOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(reshape);
REGISTER_OPERATOR(reshape, ops::ReshapeOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(reshape);
REGISTER_OPERATOR_CPU(reshape, ops::ReshapeOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/sigmoid_op.cpp
浏览文件 @ 14114fcf
......@@ -27,7 +27,13 @@ template class SigmoidOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(sigmoid);
REGISTER_OPERATOR(sigmoid, ops::SigmoidOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(sigmoid);
REGISTER_OPERATOR_CPU(sigmoid, ops::SigmoidOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/softmax_op.cpp
浏览文件 @ 14114fcf
......@@ -27,7 +27,13 @@ template class SoftmaxOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(softmax);
REGISTER_OPERATOR(softmax, ops::SoftmaxOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(softmax);
REGISTER_OPERATOR_CPU(softmax, ops::SoftmaxOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/transpose_op.cpp
浏览文件 @ 14114fcf
......@@ -52,7 +52,13 @@ template class TransposeOp<CPU, float>;
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
USE_OP(transpose);
REGISTER_OPERATOR(transpose, ops::TransposeOp);
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(transpose);
REGISTER_OPERATOR_CPU(transpose, ops::TransposeOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
test/net/test_googlenet.cpp
浏览文件 @ 14114fcf
......@@ -20,9 +20,9 @@ int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
bool optimize = false;
auto time1 = time();
// auto program = loader.Load(g_googlenet, optimize);
auto program = loader.Load(g_googlenet_combine + "/model",
g_googlenet_combine + "/params", optimize);
auto program = loader.Load(g_googlenet, optimize);
// auto program = loader.Load(g_googlenet_combine + "/model",
// g_googlenet_combine + "/params", optimize);
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time2) << "ms\n";
paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, optimize);
......
tools/build.sh
浏览文件 @ 14114fcf
......@@ -15,7 +15,6 @@ build_for_mac() {
fi
PLATFORM="x86"
MODE="Release"
CXX_FLAGS="-std=c++11 -O3 -s"
BUILD_DIR=../build/release/"${PLATFORM}"
mkdir -p ${BUILD_DIR}/build
......@@ -25,7 +24,6 @@ build_for_mac() {
cmake .. \
-B"${BUILD_DIR}" \
-DCMAKE_BUILD_TYPE="${MODE}" \
-DCMAKE_CXX_FLAGS="${CXX_FLAGS}" \
-DIS_MAC=true
cd ${BUILD_DIR}
......@@ -46,11 +44,11 @@ build_for_android() {
if [ "${PLATFORM}" = "arm-v7a" ]; then
ABI="armeabi-v7a with NEON"
ARM_PLATFORM="V7"
CXX_FLAGS="-O3 -std=c++11 -s -march=armv7-a -mfpu=neon -mfloat-abi=softfp -pie -fPIE -w -Wno-error=format-security"
CXX_FLAGS="-march=armv7-a -mfpu=neon -mfloat-abi=softfp -pie -fPIE -w -Wno-error=format-security"
elif [ "${PLATFORM}" = "arm-v8a" ]; then
ABI="arm64-v8a"
ARM_PLATFORM="V8"
CXX_FLAGS="-O3 -std=c++11 -s -march=armv8-a -pie -fPIE -w -Wno-error=format-security -llog"
CXX_FLAGS="-march=armv8-a -pie -fPIE -w -Wno-error=format-security -llog"
else
echo "unknown platform!"
exit -1
......@@ -98,7 +96,7 @@ build_for_ios() {
BUILD_DIR=../build/release/"${PLATFORM}"
TOOLCHAIN_FILE="./tools/ios-cmake/ios.toolchain.cmake"
C_FLAGS="-fobjc-abi-version=2 -fobjc-arc -isysroot ${CMAKE_OSX_SYSROOT}"
CXX_FLAGS="-fobjc-abi-version=2 -fobjc-arc -std=gnu++11 -stdlib=libc++ -isysroot ${CMAKE_OSX_SYSROOT}"
CXX_FLAGS="-fobjc-abi-version=2 -fobjc-arc -std=gnu++14 -stdlib=libc++ -isysroot ${CMAKE_OSX_SYSROOT}"
mkdir -p "${BUILD_DIR}"
if [ $# -eq 1 ]; then
NET=$1
......
tools/ios-cmake/ios.toolchain.cmake
浏览文件 @ 14114fcf
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