Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[OpenCL]Fix conv PrepareForRun to reduce first frame time #8576

Merged
merged 2 commits into from
Mar 12, 2022
Merged

[OpenCL]Fix conv PrepareForRun to reduce first frame time #8576

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
19e35e0
fix conv PrepareForRun test=develop
sprouteer Mar 4, 2022
6d47c3f
fix conv bias flag bug test=develop
sprouteer Mar 10, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
140 changes: 74 additions & 66 deletions lite/kernels/opencl/conv_image_compute.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@ namespace opencl {
void ConvImageCompute::PrepareForRun() {
ReInitWhenNeeded();

bool bias_buffer_flag = false;
auto& context = ctx_->As<OpenCLContext>();
CHECK(context.cl_context() != nullptr);
is_mali_ = context.cl_context()->IsArmMali();
Expand Down Expand Up @@ -241,6 +242,7 @@ void ConvImageCompute::PrepareForRun() {
if (task_size <= threshold_2) {
CLImageConverterNBlock converter;
kernel_func_names_.push_back("conv2d_1x1_mali_h1w2c1");
bias_buffer_flag = true;
const DDim& filter_image_dims =
converter.InitImageDimInfoWith(filter_dims);
filter_image_h_ = filter_image_dims[1];
Expand All @@ -257,14 +259,10 @@ void ConvImageCompute::PrepareForRun() {
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);

MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
} else if (task_size <= threshold_4) {
CLImageConverterN2Block converter;
kernel_func_names_.push_back("conv2d_1x1_mali_h1w2c2");
bias_buffer_flag = true;
const DDim& filter_image_dims =
converter.InitImageDimInfoWith(filter_dims);
filter_image_h_ = filter_image_dims[1];
Expand All @@ -281,14 +279,10 @@ void ConvImageCompute::PrepareForRun() {
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);

MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
} else {
CLImageConverterN2Block converter;
kernel_func_names_.push_back("conv2d_1x1_mali_h2w2c2");
bias_buffer_flag = true;
const DDim& filter_image_dims =
converter.InitImageDimInfoWith(filter_dims);
filter_image_h_ = filter_image_dims[1];
Expand All @@ -304,11 +298,6 @@ void ConvImageCompute::PrepareForRun() {
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);

MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
}
kernel_func_paths_.push_back("image/conv2d_1x1_default_mali_kernel.cl");
impl_ = &ConvImageCompute::Conv2d1x1opt;
Expand Down Expand Up @@ -382,6 +371,7 @@ void ConvImageCompute::PrepareForRun() {
if (is_mali_) {
kernel_func_names_.push_back("matrix_inner_product_mali");
kernel_func_names_.push_back("transform_to_output_mali");
bias_buffer_flag = true;
} else {
kernel_func_names_.push_back("matrix_inner_product");
kernel_func_names_.push_back("transform_to_output");
Expand All @@ -400,21 +390,24 @@ void ConvImageCompute::PrepareForRun() {
converter.NCHWToImage(filter_cpu, filter_image_data, filter_dims);

// for mali
w_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* w_gpu_data = w_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_filter_image_->memory_size());
TargetWrapperCL::MemcpySync(w_gpu_data,
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);

MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
if (is_mali_) {
w_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* w_gpu_data = w_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_filter_image_->memory_size());
TargetWrapperCL::MemcpySync(w_gpu_data,
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);
} else {
MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
}
} else if (groups_ == 1) {
if (is_mali_ && input_tensor_n_ == 1) {
kernel_func_names_.push_back("conv2d_3x3_opt_mali");
bias_buffer_flag = true;
} else {
kernel_func_names_.push_back(
input_tensor_n_ > 1 ? "conv2d_3x3_multi_batch" : "conv2d_3x3_opt");
Expand All @@ -436,17 +429,20 @@ void ConvImageCompute::PrepareForRun() {
auto* filter_image_data = MUTABLE_DATA_CPU(tensor_hold_filter_image_);
converter.NCHWToImage(filter_cpu, filter_image_data, filter_dims);

w_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* w_gpu_data = w_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_filter_image_->memory_size());
TargetWrapperCL::MemcpySync(w_gpu_data,
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);
MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
if (is_mali_ && input_tensor_n_ == 1) {
w_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* w_gpu_data = w_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_filter_image_->memory_size());
TargetWrapperCL::MemcpySync(w_gpu_data,
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);
} else {
MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
}
} else { // groups_ > 1
kernel_func_names_.push_back("conv2d_3x3");
kernel_func_paths_.push_back("image/conv2d_3x3_kernel.cl");
Expand Down Expand Up @@ -533,6 +529,7 @@ void ConvImageCompute::PrepareForRun() {
// conv2d_7x7
if (is_mali_ && input_tensor_n_ == 1) {
kernel_func_names_.push_back("conv2d_7x7_opt_mali");
bias_buffer_flag = true;
} else {
kernel_func_names_.push_back(
input_tensor_n_ > 1 ? "conv2d_7x7_multi_batch" : "conv2d_7x7_opt");
Expand All @@ -547,15 +544,20 @@ void ConvImageCompute::PrepareForRun() {

auto* filter_image_data = MUTABLE_DATA_CPU(tensor_hold_filter_image_);
converter.NCHWToImage(filter_cpu, filter_image_data, filter_dims);
w_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* w_gpu_data = w_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_filter_image_->memory_size());
TargetWrapperCL::MemcpySync(w_gpu_data,
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);
MUTABLE_DATA_GPU(
filter_gpu_image_, filter_image_w_, filter_image_h_, filter_image_data);
if (is_mali_ && input_tensor_n_ == 1) {
w_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* w_gpu_data = w_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_filter_image_->memory_size());
TargetWrapperCL::MemcpySync(w_gpu_data,
tensor_hold_filter_image_->raw_data(),
tensor_hold_filter_image_->memory_size(),
IoDirection::HtoD);
} else {
MUTABLE_DATA_GPU(filter_gpu_image_,
filter_image_w_,
filter_image_h_,
filter_image_data);
}

impl_ = &ConvImageCompute::Conv2d7x7opt;
#endif
Expand Down Expand Up @@ -722,31 +724,37 @@ void ConvImageCompute::PrepareForRun() {
bias_converter.NCHWToImage(
bias_cpu_data, bias_image_data, conv_param_->bias->dims());

bias_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* f_gpu_data = bias_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_bias_image_->memory_size());
TargetWrapperCL::MemcpySync(f_gpu_data,
tensor_hold_bias_image_->raw_data(),
tensor_hold_bias_image_->memory_size(),
IoDirection::HtoD);
MUTABLE_DATA_GPU(bias_gpu_image_,
bias_image_dims[0],
bias_image_dims[1],
bias_image_data);
if (bias_buffer_flag) {
bias_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* f_gpu_data = bias_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_bias_image_->memory_size());
TargetWrapperCL::MemcpySync(f_gpu_data,
tensor_hold_bias_image_->raw_data(),
tensor_hold_bias_image_->memory_size(),
IoDirection::HtoD);
} else {
MUTABLE_DATA_GPU(bias_gpu_image_,
bias_image_dims[0],
bias_image_dims[1],
bias_image_data);
}
} else {
bias_gpu_image_ = std::unique_ptr<Tensor>(new Tensor);
CLImageConverterFolder bias_converter;
tensor_hold_bias_image_->Resize({1, 1, 1, 4});
auto* bias_image_data = DATA_GPU(tensor_hold_bias_image_);

bias_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* f_gpu_data = bias_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_bias_image_->memory_size());
TargetWrapperCL::MemcpySync(f_gpu_data,
tensor_hold_bias_image_->raw_data(),
tensor_hold_bias_image_->memory_size(),
IoDirection::HtoD);
MUTABLE_DATA_GPU(bias_gpu_image_, 1, 1, bias_image_data);
if (bias_buffer_flag) {
bias_gpu_t_ = std::unique_ptr<Tensor>(new Tensor);
auto* f_gpu_data = bias_gpu_t_->mutable_data(
TARGET(kOpenCL), tensor_hold_bias_image_->memory_size());
TargetWrapperCL::MemcpySync(f_gpu_data,
tensor_hold_bias_image_->raw_data(),
tensor_hold_bias_image_->memory_size(),
IoDirection::HtoD);
} else {
MUTABLE_DATA_GPU(bias_gpu_image_, 1, 1, bias_image_data);
}
}

// scale options
Expand Down