Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 eb2837c6 编写于 作者: D dolphin8 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into develop

上级 1fa1fce4 6db73b7a
显示空白变更内容
内联 并排
Showing with 99 addition and 27 deletion
CMakeLists.txt
浏览文件 @ eb2837c6
......@@ -4,7 +4,8 @@ project(paddle-mobile)
option(DEBUGING "enable debug mode" ON)
option(USE_OPENMP "openmp support" 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)
......@@ -45,6 +46,10 @@ else()
add_definitions(-fno-exceptions)
endif ()
if (LOG_PROFILE)
add_definitions(-DPADDLE_MOBILE_PROFILE)
endif()
if(IS_MAC)
add_definitions(-DX86)
elseif(IS_IOS)
......
src/framework/program/program-optimize/program_optimize.cpp
浏览文件 @ eb2837c6
......@@ -106,11 +106,14 @@ std::shared_ptr<ProgramDesc> ProgramOptimize::FushionOptimize(
}
std::vector<std::shared_ptr<framework::OpDesc>> op_descs;
if (add_split) {
GenerateOps(&op_descs, begin_node.get(), add_split);
} else {
for (int m = 0; m < nodes.size(); ++m) {
auto &node = nodes[m];
op_descs.push_back(node->op_desc_);
}
// GenerateOps(&op_descs, begin_node.get());
}
block->ops_ = op_descs;
}
......@@ -267,12 +270,12 @@ void ProgramOptimize::GenerateOps(
}
void ProgramOptimize::GenerateOps(
std::vector<std::shared_ptr<framework::OpDesc>> *op_descs,
Node *begin_node) {
std::vector<std::shared_ptr<framework::OpDesc>> *op_descs, Node *begin_node,
bool can_add_split) {
// std::vector<std::shared_ptr<framework::OpDesc>> *op_desc,
// Node *input_node, Node *current_node, bool adding_thread, int
// thread_num
if (false) {
if (can_add_split) {
this->GenerateOps(op_descs, begin_node, begin_node, false, -1, nullptr);
} else {
this->GenerateOps(op_descs, begin_node, begin_node);
......
src/framework/program/program-optimize/program_optimize.h
浏览文件 @ eb2837c6
......@@ -34,7 +34,7 @@ class ProgramOptimize {
int current_block_;
std::vector<std::shared_ptr<BlockDesc>> new_blocks_;
void GenerateOps(std::vector<std::shared_ptr<framework::OpDesc>> *op_descs,
Node *begin_node);
Node *begin_node, bool can_add_split);
void GenerateOps(std::vector<std::shared_ptr<framework::OpDesc>> *op_desc,
Node *input_node, Node *current_node);
void GenerateOps(std::vector<std::shared_ptr<framework::OpDesc>> *op_desc,
......
src/io/io.cpp
浏览文件 @ eb2837c6
......@@ -76,8 +76,9 @@ static size_t ReadBuffer(const char *file_name, uint8_t **out) {
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
const std::string &dirname, bool optimize) {
auto program = this->LoadProgram(dirname + "/__model__", optimize);
const std::string &dirname, bool optimize, bool can_add_split) {
auto program =
this->LoadProgram(dirname + "/__model__", optimize, can_add_split);
program.model_path = dirname;
return program;
}
......@@ -94,7 +95,7 @@ const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::LoadProgram(
const std::string &model_path, bool optimize) {
const std::string &model_path, bool optimize, bool can_add_split) {
std::string model_filename = model_path;
PaddleMobile__Framework__Proto__ProgramDesc *c_program;
uint8_t *buf = NULL;
......@@ -146,7 +147,7 @@ const framework::Program<Dtype, P> Loader<Dtype, P>::LoadProgram(
if (optimize) {
framework::ProgramOptimize program_optimize;
program.optimizeProgram =
program_optimize.FushionOptimize(originProgramDesc);
program_optimize.FushionOptimize(originProgramDesc, can_add_split);
}
if (optimize) {
program.optimizeProgram->Description("optimize: ");
......@@ -310,6 +311,7 @@ void Executor<Dtype, P>::InitMemory() {
template <typename Dtype, Precision P>
void Executor<Dtype, P>::InitCombineMemory() {
LOG(kLOG_INFO) << " begin init combine memory";
char *origin_data = Get_binary_data(program_.para_path);
char *data = origin_data;
for (const auto &block : to_predict_program_->Blocks()) {
......@@ -330,6 +332,7 @@ void Executor<Dtype, P>::InitCombineMemory() {
}
}
delete origin_data;
LOG(kLOG_INFO) << " end init combine memory ";
}
template <typename Dtype, Precision P>
......
src/io/io.h
浏览文件 @ eb2837c6
......@@ -35,7 +35,8 @@ class Loader {
* @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 optimize = false,
bool can_add_split = false);
/*
* @b load combine format fluid mode
......@@ -47,7 +48,8 @@ class Loader {
private:
const framework::Program<Dtype, P> LoadProgram(const std::string &model_path,
bool optimize = false);
bool optimize = false,
bool can_add_split = false);
};
template <typename Dtype = CPU, Precision P = Precision::FP32>
......
src/operators/kernel/arm/relu_kernel.cpp
浏览文件 @ eb2837c6
......@@ -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
......
test/framework/test_load.cpp
浏览文件 @ eb2837c6
......@@ -19,9 +19,10 @@ int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
// ../../../test/models/googlenet
// ../../../test/models/mobilenet
auto program = loader.Load(g_resnet, true);
loader.Load(g_googlenet_combine + "/model", g_googlenet_combine + "/params",
true);
auto program = loader.Load(g_googlenet, true, true);
// loader.Load(g_googlenet_combine + "/model", g_googlenet_combine +
// "/params",
// true);
program.originProgram->Description("program desc: ");
return 0;
......
test/net/test_googlenet.cpp
浏览文件 @ eb2837c6
......@@ -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);
......
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