apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-31 20:02:53 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

[benchmark] support lite light api & optimize benchmark flags (#1950)

Browse Source 
* [benchmark] support lite light api & optimize benchmark flags

* [backend] remove un-used option member

* [backend] remove un-used option member
This commit is contained in:
DefTruth
2023-05-17 18:51:32 +08:00
committed by GitHub
parent 09ec386e8d
commit 1f66d5d9d3
6 changed files with 198 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
fastdeploy/runtime/backends/lite/lite_backend.cc
View File

@@ -94,20 +94,39 @@ bool LiteBackend::Init(const RuntimeOption& runtime_option) {
return false;
}
config_.set_model_file(runtime_option.model_file);
config_.set_param_file(runtime_option.params_file);
BuildOption(runtime_option.paddle_lite_option);
predictor_ =
paddle::lite_api::CreatePaddlePredictor<paddle::lite_api::CxxConfig>(
config_);
if (option_.optimized_model_dir != "") {
FDINFO << "Optimzed model dir is not empty, will save optimized model to: "
<< option_.optimized_model_dir << std::endl;
predictor_->SaveOptimizedModel(
option_.optimized_model_dir,
paddle::lite_api::LiteModelType::kNaiveBuffer);
if (runtime_option.params_file == "") {
// Use light api for Arm CPU via MobileConfig.
FDASSERT(runtime_option.device == Device::CPU,
"In FastDeploy, Paddle Lite light API is only support for Arm CPU now!")
mobile_config_.set_model_from_file(runtime_option.model_file);
mobile_config_.set_threads(runtime_option.paddle_lite_option.cpu_threads);
mobile_config_.set_power_mode(static_cast<paddle::lite_api::PowerMode>(
runtime_option.paddle_lite_option.power_mode));
// TODO(qiuyanjun): Add OpenCL support for mobile gpu.
// Paddle-Lite/blob/develop/lite/api/tools/benchmark/benchmark.h#L265
// mobile_config_.set_opencl_tune(
// tune_mode, opencl_cache_dir, opencl_tuned_file);
// mobile_config_.set_opencl_precision(gpu_precision);
predictor_ =
paddle::lite_api::CreatePaddlePredictor<paddle::lite_api::MobileConfig>(
mobile_config_);
} else {
// Use full api for many hardwares via CxxConfig.
config_.set_model_file(runtime_option.model_file);
config_.set_param_file(runtime_option.params_file);
BuildOption(runtime_option.paddle_lite_option);
predictor_ =
paddle::lite_api::CreatePaddlePredictor<paddle::lite_api::CxxConfig>(
config_);
if (option_.optimized_model_dir != "") {
FDINFO << "Optimzed model dir is not empty, will save optimized model to: "
<< option_.optimized_model_dir << std::endl;
predictor_->SaveOptimizedModel(
option_.optimized_model_dir,
paddle::lite_api::LiteModelType::kNaiveBuffer);
}
}
inputs_desc_.clear();
outputs_desc_.clear();
inputs_order_.clear();

4
fastdeploy/runtime/backends/lite/lite_backend.h
View File

@@ -48,8 +48,9 @@ class LiteBackend : public BaseBackend {
std::vector<TensorInfo> GetOutputInfos() override;
private:
// Build CxxConfig from option for Paddle Lite full api.
void BuildOption(const LiteBackendOption& option);
// Configure many hardwares for Paddle Lite full api.
void ConfigureCpu(const LiteBackendOption& option);
void ConfigureGpu(const LiteBackendOption& option);
void ConfigureTimvx(const LiteBackendOption& option);
@@ -59,6 +60,7 @@ class LiteBackend : public BaseBackend {
paddle::lite_api::CxxConfig config_;
std::shared_ptr<paddle::lite_api::PaddlePredictor> predictor_;
paddle::lite_api::MobileConfig mobile_config_;
std::vector<TensorInfo> inputs_desc_;
std::vector<TensorInfo> outputs_desc_;
std::map<std::string, int> inputs_order_;

3
fastdeploy/runtime/backends/lite/option.h
View File

@@ -61,6 +61,9 @@ struct LiteBackendOption {
Device device = Device::CPU;
// Index of inference device
int device_id = 0;
// TODO(qiuyanjun): add opencl binary path and cache settings.
std::string opencl_cache_dir = "/data/local/tmp/";
std::string opencl_tuned_file = "/data/local/tmp/opencl_tuned_kernels.bin";
/// kunlunxin_l3_workspace_size
int kunlunxin_l3_workspace_size = 0xfffc00;