Skip to content

Mace
Mace

项目经理老王 / Mace
与 Fork 源项目一致

Fork自 Xiaomi / Mace

通知 1
Star 0
Fork 0
Mace
Mace

体验新版 GitCode，发现更多精彩内容 >>

提交 fdf938ce 编写于 作者: L Liangliang He
浏览文件
操作

Merge branch 'fc-mali' into 'master' 

FC and CWise ops support opencl 1.1/1.2.

See merge request !343
上级 6ec4a9df 07c10f46
隐藏空白更改
内联 并排
Showing with 211 addition and 178 deletion
mace/benchmark/benchmark_model.cc
浏览文件 @ fdf938ce
...@@ -12,6 +12,7 @@ ...@@ -12,6 +12,7 @@
#include "gflags/gflags.h" #include "gflags/gflags.h"
#include "mace/public/mace.h" #include "mace/public/mace.h"
#include "mace/public/mace_runtime.h"
#include "mace/utils/logging.h" #include "mace/utils/logging.h"
#include "mace/benchmark/stat_summarizer.h" #include "mace/benchmark/stat_summarizer.h"
...@@ -95,9 +96,23 @@ inline int64_t NowMicros() { ...@@ -95,9 +96,23 @@ inline int64_t NowMicros() {
return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec; return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
} }
DeviceType ParseDeviceType(const std::string &device_str) {
if (device_str.compare("CPU") == 0) {
return DeviceType::CPU;
} else if (device_str.compare("NEON") == 0) {
return DeviceType::NEON;
} else if (device_str.compare("OPENCL") == 0) {
return DeviceType::OPENCL;
} else if (device_str.compare("HEXAGON") == 0) {
return DeviceType::HEXAGON;
} else {
return DeviceType::CPU;
}
}
bool RunInference(MaceEngine *engine, bool RunInference(MaceEngine *engine,
const std::vector<mace::MaceInputInfo> &input_infos, const std::map<std::string, mace::MaceTensor> &input_infos,
std::map<std::string, float*> *output_infos, std::map<std::string, mace::MaceTensor> *output_infos,
StatSummarizer *summarizer, StatSummarizer *summarizer,
int64_t *inference_time_us) { int64_t *inference_time_us) {
MACE_CHECK_NOTNULL(output_infos); MACE_CHECK_NOTNULL(output_infos);
...@@ -106,28 +121,16 @@ bool RunInference(MaceEngine *engine, ...@@ -106,28 +121,16 @@ bool RunInference(MaceEngine *engine,
if (summarizer) { if (summarizer) {
run_metadata_ptr = &run_metadata; run_metadata_ptr = &run_metadata;
} }
if (input_infos.size() == 1 && output_infos->size() == 1) {
const int64_t start_time = NowMicros();
bool s = engine->Run(input_infos[0].data, input_infos[0].shape,
output_infos->begin()->second, run_metadata_ptr);
const int64_t end_time = NowMicros();
if (!s) { const int64_t start_time = NowMicros();
LOG(ERROR) << "Error during inference."; mace::MaceStatus s = engine->Run(input_infos, output_infos, run_metadata_ptr);
return s; const int64_t end_time = NowMicros();
}
*inference_time_us = end_time - start_time;
} else {
const int64_t start_time = NowMicros();
bool s = engine->Run(input_infos, *output_infos, run_metadata_ptr);
const int64_t end_time = NowMicros();
if (!s) { if (s != mace::MaceStatus::MACE_SUCCESS) {
LOG(ERROR) << "Error during inference."; LOG(ERROR) << "Error during inference.";
return s; return false;
}
*inference_time_us = end_time - start_time;
} }
*inference_time_us = end_time - start_time;
if (summarizer != nullptr) { if (summarizer != nullptr) {
summarizer->ProcessMetadata(run_metadata); summarizer->ProcessMetadata(run_metadata);
...@@ -137,8 +140,8 @@ bool RunInference(MaceEngine *engine, ...@@ -137,8 +140,8 @@ bool RunInference(MaceEngine *engine,
} }
bool Run(MaceEngine *engine, bool Run(MaceEngine *engine,
const std::vector<mace::MaceInputInfo> &input_infos, const std::map<std::string, mace::MaceTensor> &input_infos,
std::map<std::string, float*> *output_infos, std::map<std::string, mace::MaceTensor> *output_infos,
StatSummarizer *summarizer, StatSummarizer *summarizer,
int num_runs, int num_runs,
double max_time_sec, double max_time_sec,
...@@ -261,12 +264,7 @@ int Main(int argc, char **argv) { ...@@ -261,12 +264,7 @@ int Main(int argc, char **argv) {
stats_options.show_summary = FLAGS_show_summary; stats_options.show_summary = FLAGS_show_summary;
stats.reset(new StatSummarizer(stats_options)); stats.reset(new StatSummarizer(stats_options));
DeviceType device_type = CPU; mace::DeviceType device_type = ParseDeviceType(FLAGS_device);
if (FLAGS_device == "OPENCL") {
device_type = OPENCL;
} else if (FLAGS_device == "NEON") {
device_type = NEON;
}
// config runtime // config runtime
mace::ConfigOmpThreads(FLAGS_omp_num_threads); mace::ConfigOmpThreads(FLAGS_omp_num_threads);
...@@ -302,50 +300,45 @@ int Main(int argc, char **argv) { ...@@ -302,50 +300,45 @@ int Main(int argc, char **argv) {
mace::MACE_MODEL_TAG::LoadModelData(FLAGS_model_data_file.c_str()); mace::MACE_MODEL_TAG::LoadModelData(FLAGS_model_data_file.c_str());
NetDef net_def = mace::MACE_MODEL_TAG::CreateNet(model_data); NetDef net_def = mace::MACE_MODEL_TAG::CreateNet(model_data);
std::vector<mace::MaceInputInfo> input_infos(input_count); std::map<std::string, mace::MaceTensor> inputs;
std::map<std::string, float*> output_infos; std::map<std::string, mace::MaceTensor> outputs;
std::vector<std::unique_ptr<float[]>> input_datas(input_count);
std::vector<std::unique_ptr<float[]>> output_datas(output_count);
for (size_t i = 0; i < input_count; ++i) { for (size_t i = 0; i < input_count; ++i) {
int64_t input_size = std::accumulate(input_shape_vec[i].begin(), // Allocate input and output
input_shape_vec[i].end(), 1, int64_t input_size =
std::multiplies<int64_t>()); std::accumulate(input_shape_vec[i].begin(), input_shape_vec[i].end(), 1,
input_datas[i].reset(new float[input_size]); std::multiplies<int64_t>());
auto buffer_in = std::shared_ptr<float>(new float[input_size],
std::default_delete<float[]>());
// load input // load input
std::ifstream in_file(FLAGS_input_file + "_" + FormatName(input_names[i]), std::ifstream in_file(FLAGS_input_file + "_" + FormatName(input_names[i]),
std::ios::in | std::ios::binary); std::ios::in | std::ios::binary);
if (in_file.is_open()) { if (in_file.is_open()) {
in_file.read(reinterpret_cast<char *>(input_datas[i].get()), in_file.read(reinterpret_cast<char *>(buffer_in.get()),
input_size * sizeof(float)); input_size * sizeof(float));
in_file.close(); in_file.close();
} else { } else {
LOG(INFO) << "Open input file failed"; LOG(INFO) << "Open input file failed";
return -1; return -1;
} }
inputs[input_names[i]] = mace::MaceTensor(input_shape_vec[i], buffer_in);
input_infos[i].name = input_names[i];
input_infos[i].shape = input_shape_vec[i];
input_infos[i].data = input_datas[i].get();
} }
for (size_t i = 0; i < output_count; ++i) { for (size_t i = 0; i < output_count; ++i) {
int64_t output_size = std::accumulate(output_shape_vec[i].begin(), int64_t output_size =
output_shape_vec[i].end(), 1, std::accumulate(output_shape_vec[i].begin(),
std::multiplies<int64_t>()); output_shape_vec[i].end(), 1,
output_datas[i].reset(new float[output_size]); std::multiplies<int64_t>());
output_infos[output_names[i]] = output_datas[i].get(); auto buffer_out = std::shared_ptr<float>(new float[output_size],
std::default_delete<float[]>());
outputs[output_names[i]] = mace::MaceTensor(output_shape_vec[i],
buffer_out);
} }
// Init model // Init model
LOG(INFO) << "Run init"; LOG(INFO) << "Run init";
std::unique_ptr<mace::MaceEngine> engine_ptr; std::unique_ptr<mace::MaceEngine> engine_ptr(
if (input_count == 1 && output_count == 1) { new mace::MaceEngine(&net_def, device_type, input_names, output_names));
engine_ptr.reset(new mace::MaceEngine(&net_def, device_type)); if (device_type == DeviceType::OPENCL || device_type == DeviceType::HEXAGON) {
} else {
engine_ptr.reset(new mace::MaceEngine(&net_def, device_type,
input_names, output_names));
}
if (device_type == DeviceType::OPENCL) {
mace::MACE_MODEL_TAG::UnloadModelData(model_data); mace::MACE_MODEL_TAG::UnloadModelData(model_data);
} }
...@@ -355,7 +348,7 @@ int Main(int argc, char **argv) { ...@@ -355,7 +348,7 @@ int Main(int argc, char **argv) {
int64_t num_warmup_runs = 0; int64_t num_warmup_runs = 0;
if (FLAGS_warmup_runs > 0) { if (FLAGS_warmup_runs > 0) {
bool status = bool status =
Run(engine_ptr.get(), input_infos, &output_infos, nullptr, Run(engine_ptr.get(), inputs, &outputs, nullptr,
FLAGS_warmup_runs, -1.0, FLAGS_warmup_runs, -1.0,
inter_inference_sleep_seconds, &warmup_time_us, &num_warmup_runs); inter_inference_sleep_seconds, &warmup_time_us, &num_warmup_runs);
if (!status) { if (!status) {
...@@ -370,7 +363,7 @@ int Main(int argc, char **argv) { ...@@ -370,7 +363,7 @@ int Main(int argc, char **argv) {
int64_t no_stat_time_us = 0; int64_t no_stat_time_us = 0;
int64_t no_stat_runs = 0; int64_t no_stat_runs = 0;
bool status = bool status =
Run(engine_ptr.get(), input_infos, &output_infos, Run(engine_ptr.get(), inputs, &outputs,
nullptr, FLAGS_max_num_runs, max_benchmark_time_seconds, nullptr, FLAGS_max_num_runs, max_benchmark_time_seconds,
inter_inference_sleep_seconds, &no_stat_time_us, &no_stat_runs); inter_inference_sleep_seconds, &no_stat_time_us, &no_stat_runs);
if (!status) { if (!status) {
...@@ -379,7 +372,7 @@ int Main(int argc, char **argv) { ...@@ -379,7 +372,7 @@ int Main(int argc, char **argv) {
int64_t stat_time_us = 0; int64_t stat_time_us = 0;
int64_t stat_runs = 0; int64_t stat_runs = 0;
status = Run(engine_ptr.get(), input_infos, &output_infos, status = Run(engine_ptr.get(), inputs, &outputs,
stats.get(), FLAGS_max_num_runs, max_benchmark_time_seconds, stats.get(), FLAGS_max_num_runs, max_benchmark_time_seconds,
inter_inference_sleep_seconds, &stat_time_us, &stat_runs); inter_inference_sleep_seconds, &stat_time_us, &stat_runs);
if (!status) { if (!status) {
......
mace/core/runtime/opencl/opencl_runtime.cc
浏览文件 @ fdf938ce
...@@ -480,12 +480,12 @@ uint64_t OpenCLRuntime::GetKernelWaveSize(const cl::Kernel &kernel) { ...@@ -480,12 +480,12 @@ uint64_t OpenCLRuntime::GetKernelWaveSize(const cl::Kernel &kernel) {
} }
const bool OpenCLRuntime::IsNonUniformWorkgroupsSupported() { const bool OpenCLRuntime::IsNonUniformWorkgroupsSupported() {
if (gpu_type_ == GPUType::QUALCOMM_ADRENO && return (gpu_type_ == GPUType::QUALCOMM_ADRENO &&
opencl_version_ == "2.0") { opencl_version_ == "2.0");
return true; }
} else {
return false; const GPUType OpenCLRuntime::gpu_type() const {
} return gpu_type_;
} }
const GPUType OpenCLRuntime::ParseGPUTypeFromDeviceName( const GPUType OpenCLRuntime::ParseGPUTypeFromDeviceName(
......
mace/core/runtime/opencl/opencl_runtime.h
浏览文件 @ fdf938ce
...@@ -66,6 +66,7 @@ class OpenCLRuntime { ...@@ -66,6 +66,7 @@ class OpenCLRuntime {
uint64_t GetKernelWaveSize(const cl::Kernel &kernel); uint64_t GetKernelWaveSize(const cl::Kernel &kernel);
const bool IsNonUniformWorkgroupsSupported(); const bool IsNonUniformWorkgroupsSupported();
const GPUType ParseGPUTypeFromDeviceName(const std::string &device_name); const GPUType ParseGPUTypeFromDeviceName(const std::string &device_name);
const GPUType gpu_type() const;
cl::Kernel BuildKernel(const std::string &program_name, cl::Kernel BuildKernel(const std::string &program_name,
const std::string &kernel_name, const std::string &kernel_name,
const std::set<std::string> &build_options); const std::set<std::string> &build_options);
......
mace/kernels/cwise.h
浏览文件 @ fdf938ce
...@@ -114,6 +114,7 @@ struct CWiseFunctor<DeviceType::OPENCL, T> : CWiseFunctorBase { ...@@ -114,6 +114,7 @@ struct CWiseFunctor<DeviceType::OPENCL, T> : CWiseFunctorBase {
StatsFuture *future); StatsFuture *future);
cl::Kernel kernel_; cl::Kernel kernel_;
uint32_t kwg_size_;
std::vector<index_t> input_shape_; std::vector<index_t> input_shape_;
}; };
......
mace/kernels/opencl/cl/activation.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void activation( __kernel void activation(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
#ifdef USE_PRELU #ifdef USE_PRELU
__read_only image2d_t alpha, __read_only image2d_t alpha,
......
mace/kernels/opencl/cl/addn.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void addn( __kernel void addn(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__read_only image2d_t input0, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input0, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t input1, __read_only image2d_t input1,
#if INPUT_NUM > 2 #if INPUT_NUM > 2
......
mace/kernels/opencl/cl/batch_norm.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
// Supported data types: half/float // Supported data types: half/float
__kernel void batch_norm( __kernel void batch_norm(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__read_only image2d_t scale, __read_only image2d_t scale,
__read_only image2d_t offset, __read_only image2d_t offset,
......
mace/kernels/opencl/cl/bias_add.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
// Supported data types: half/float // Supported data types: half/float
__kernel void bias_add( __kernel void bias_add(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__read_only image2d_t bias, __read_only image2d_t bias,
__write_only image2d_t output) { __write_only image2d_t output) {
......
mace/kernels/opencl/cl/buffer_to_image.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void filter_buffer_to_image( __kernel void filter_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, /* h, w, oc, ic */ __global const DATA_TYPE *input, /* h, w, oc, ic */
__private const int input_offset, __private const int input_offset,
__private const int filter_h, __private const int filter_h,
...@@ -53,8 +52,7 @@ __kernel void filter_buffer_to_image( ...@@ -53,8 +52,7 @@ __kernel void filter_buffer_to_image(
WRITE_IMAGET(output, coord, values); WRITE_IMAGET(output, coord, values);
} }
__kernel void filter_image_to_buffer( __kernel void filter_image_to_buffer(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global DATA_TYPE *output, /* h, w, oc, ic */ __global DATA_TYPE *output, /* h, w, oc, ic */
__private const int filter_h, __private const int filter_h,
__private const int filter_w, __private const int filter_w,
...@@ -102,8 +100,7 @@ __kernel void filter_image_to_buffer( ...@@ -102,8 +100,7 @@ __kernel void filter_image_to_buffer(
} }
} }
__kernel void dw_filter_buffer_to_image( __kernel void dw_filter_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, /* h, w, ic, m */ __global const DATA_TYPE *input, /* h, w, ic, m */
__private const int input_offset, __private const int input_offset,
__private const int filter_w, __private const int filter_w,
...@@ -160,8 +157,7 @@ __kernel void dw_filter_buffer_to_image( ...@@ -160,8 +157,7 @@ __kernel void dw_filter_buffer_to_image(
WRITE_IMAGET(output, coord, values); WRITE_IMAGET(output, coord, values);
} }
__kernel void in_out_buffer_to_image( __kernel void in_out_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, /* nhwc */ __global const DATA_TYPE *input, /* nhwc */
__private const int input_offset, __private const int input_offset,
__private const int height, __private const int height,
...@@ -202,8 +198,7 @@ __kernel void in_out_buffer_to_image( ...@@ -202,8 +198,7 @@ __kernel void in_out_buffer_to_image(
WRITE_IMAGET(output, coord, values); WRITE_IMAGET(output, coord, values);
} }
__kernel void in_out_image_to_buffer( __kernel void in_out_image_to_buffer(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global DATA_TYPE *output, /* nhwc */ __global DATA_TYPE *output, /* nhwc */
__private const int height, __private const int height,
__private const int width, __private const int width,
...@@ -242,8 +237,7 @@ __kernel void in_out_image_to_buffer( ...@@ -242,8 +237,7 @@ __kernel void in_out_image_to_buffer(
} }
} }
__kernel void arg_buffer_to_image( __kernel void arg_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, /* nhwc */ __global const DATA_TYPE *input, /* nhwc */
__private const int input_offset, __private const int input_offset,
__private const int count, __private const int count,
...@@ -278,8 +272,7 @@ __kernel void arg_buffer_to_image( ...@@ -278,8 +272,7 @@ __kernel void arg_buffer_to_image(
WRITE_IMAGET(output, coord, values); WRITE_IMAGET(output, coord, values);
} }
__kernel void arg_image_to_buffer( __kernel void arg_image_to_buffer(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global DATA_TYPE *output, /* nhwc */ __global DATA_TYPE *output, /* nhwc */
__private const int count, __private const int count,
__read_only image2d_t input) { __read_only image2d_t input) {
...@@ -312,8 +305,7 @@ __kernel void arg_image_to_buffer( ...@@ -312,8 +305,7 @@ __kernel void arg_image_to_buffer(
} }
__kernel void in_out_height_buffer_to_image( __kernel void in_out_height_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, //nhwc __global const DATA_TYPE *input, //nhwc
__private const int input_offset, __private const int input_offset,
__private const int height, __private const int height,
...@@ -355,8 +347,7 @@ __kernel void in_out_height_buffer_to_image( ...@@ -355,8 +347,7 @@ __kernel void in_out_height_buffer_to_image(
WRITE_IMAGET(output, coord, values); WRITE_IMAGET(output, coord, values);
} }
__kernel void in_out_height_image_to_buffer( __kernel void in_out_height_image_to_buffer(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global DATA_TYPE *output, //nhwc __global DATA_TYPE *output, //nhwc
__private const int height, __private const int height,
__private const int width, __private const int width,
...@@ -394,8 +385,7 @@ __kernel void in_out_height_image_to_buffer( ...@@ -394,8 +385,7 @@ __kernel void in_out_height_image_to_buffer(
} }
__kernel void in_out_width_buffer_to_image( __kernel void in_out_width_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, /* nhwc */ __global const DATA_TYPE *input, /* nhwc */
__private const int input_offset, __private const int input_offset,
__private const int height, __private const int height,
...@@ -437,8 +427,7 @@ __kernel void in_out_width_buffer_to_image( ...@@ -437,8 +427,7 @@ __kernel void in_out_width_buffer_to_image(
} }
// only support 3x3 now // only support 3x3 now
__kernel void winograd_filter_buffer_to_image( __kernel void winograd_filter_buffer_to_image(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global const DATA_TYPE *input, //Oc, Ic, H, W __global const DATA_TYPE *input, //Oc, Ic, H, W
__private const int input_offset, __private const int input_offset,
__private const int in_channels, __private const int in_channels,
...@@ -529,8 +518,7 @@ __kernel void winograd_filter_buffer_to_image( ...@@ -529,8 +518,7 @@ __kernel void winograd_filter_buffer_to_image(
} }
// only support 3x3 now // only support 3x3 now
__kernel void winograd_filter_image_to_buffer( __kernel void winograd_filter_image_to_buffer(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__global DATA_TYPE *output, //Oc, Ic, H, W __global DATA_TYPE *output, //Oc, Ic, H, W
__private const int height, __private const int height,
__private const int width, __private const int width,
......
mace/kernels/opencl/cl/channel_shuffle.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
// assume channes_per_group mod 4 = 0 && groups mod 4 == 0 // assume channes_per_group mod 4 = 0 && groups mod 4 == 0
__kernel void channel_shuffle( __kernel void channel_shuffle(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__private const int groups, __private const int groups,
__private const int channels_per_group, __private const int channels_per_group,
......
mace/kernels/opencl/cl/common.h
浏览文件 @ fdf938ce
...@@ -19,18 +19,18 @@ ...@@ -19,18 +19,18 @@
#ifndef NON_UNIFORM_WORK_GROUP #ifndef NON_UNIFORM_WORK_GROUP
#define UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2 \ #define GLOBAL_WORK_GROUP_SIZE_DIM2 \
__private const int global_size_dim0, \ __private const int global_size_dim0, \
__private const int global_size_dim1, __private const int global_size_dim1,
#define UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3 \ #define GLOBAL_WORK_GROUP_SIZE_DIM3 \
__private const int global_size_dim0, \ __private const int global_size_dim0, \
__private const int global_size_dim1, \ __private const int global_size_dim1, \
__private const int global_size_dim2, __private const int global_size_dim2,
#else #else
#define UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2 #define GLOBAL_WORK_GROUP_SIZE_DIM2
#define UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3 #define GLOBAL_WORK_GROUP_SIZE_DIM3
#endif #endif
......
mace/kernels/opencl/cl/concat.cl
浏览文件 @ fdf938ce
...@@ -22,8 +22,7 @@ DATA_TYPE4 stitch_vector(DATA_TYPE4 left, ...@@ -22,8 +22,7 @@ DATA_TYPE4 stitch_vector(DATA_TYPE4 left,
} }
// Supported data type: half/float // Supported data type: half/float
__kernel void concat_channel( __kernel void concat_channel(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input0, __read_only image2d_t input0,
__read_only image2d_t input1, __read_only image2d_t input1,
__private const int input0_chan, __private const int input0_chan,
...@@ -84,8 +83,7 @@ __kernel void concat_channel( ...@@ -84,8 +83,7 @@ __kernel void concat_channel(
} }
// Required: All input channels are divisible by 4 // Required: All input channels are divisible by 4
__kernel void concat_channel_multi( __kernel void concat_channel_multi(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__private const int chan_blk_offset, __private const int chan_blk_offset,
__write_only image2d_t output) { __write_only image2d_t output) {
......
mace/kernels/opencl/cl/conv_2d.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void conv_2d( __kernel void conv_2d(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t filter, /* cout%4 * cin, kh * kw * cout/4 */ __read_only image2d_t filter, /* cout%4 * cin, kh * kw * cout/4 */
#ifdef BIAS #ifdef BIAS
......
mace/kernels/opencl/cl/conv_2d_1x1.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void conv_2d_1x1( __kernel void conv_2d_1x1(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t filter, /* cout%4 * cin, cout/4 */ __read_only image2d_t filter, /* cout%4 * cin, cout/4 */
#ifdef BIAS #ifdef BIAS
......
mace/kernels/opencl/cl/conv_2d_3x3.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void conv_2d_3x3( __kernel void conv_2d_3x3(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t filter, /* cout%4 * cin , kh * kw * cout/4 */ __read_only image2d_t filter, /* cout%4 * cin , kh * kw * cout/4 */
#ifdef BIAS #ifdef BIAS
......
mace/kernels/opencl/cl/cwise.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void cwise(__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __kernel void cwise(GLOBAL_WORK_GROUP_SIZE_DIM2
__private const float value, __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__write_only image2d_t output) { __private const float value,
__write_only image2d_t output) {
const int w = get_global_id(0); const int w = get_global_id(0);
const int hb = get_global_id(1); const int hb = get_global_id(1);
#ifndef NON_UNIFORM_WORK_GROUP
if (w >= global_size_dim0 || hb >= global_size_dim1) return;
#endif
DATA_TYPE4 in0 = READ_IMAGET(input, SAMPLER, (int2)(w, hb)); DATA_TYPE4 in0 = READ_IMAGET(input, SAMPLER, (int2)(w, hb));
DATA_TYPE4 in1 = (DATA_TYPE4){value, value, value, value}; DATA_TYPE4 in1 = (DATA_TYPE4){value, value, value, value};
DATA_TYPE4 out; DATA_TYPE4 out;
......
mace/kernels/opencl/cl/depth_to_space.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void depth_to_space( __kernel void depth_to_space(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__private const int block_size, __private const int block_size,
__private const int input_height, __private const int input_height,
...@@ -36,7 +35,7 @@ __kernel void depth_to_space( ...@@ -36,7 +35,7 @@ __kernel void depth_to_space(
} }
__kernel void space_to_depth( __kernel void space_to_depth(
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3 GLOBAL_WORK_GROUP_SIZE_DIM3
__read_only image2d_t input, __read_only image2d_t input,
__private const int block_size, __private const int block_size,
__private const int input_height, __private const int input_height,
......
mace/kernels/opencl/cl/depthwise_conv2d.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
// Only multiplier = 1 is supported // Only multiplier = 1 is supported
__kernel void depthwise_conv2d( __kernel void depthwise_conv2d(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t filter, /* cout%4 * kh * kw * m, cin/4 */ __read_only image2d_t filter, /* cout%4 * kh * kw * m, cin/4 */
#ifdef BIAS #ifdef BIAS
...@@ -138,8 +137,7 @@ __kernel void depthwise_conv2d( ...@@ -138,8 +137,7 @@ __kernel void depthwise_conv2d(
WRITE_IMAGET(output, (int2)(out_x_base + w, out_hb), out3); WRITE_IMAGET(output, (int2)(out_x_base + w, out_hb), out3);
} }
__kernel void depthwise_conv2d_s1( __kernel void depthwise_conv2d_s1(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t filter, /* cout%4 * kh * kw * m, cin/4 */ __read_only image2d_t filter, /* cout%4 * kh * kw * m, cin/4 */
#ifdef BIAS #ifdef BIAS
......
mace/kernels/opencl/cl/eltwise.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void eltwise( __kernel void eltwise(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__read_only image2d_t input0, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input0, /* [c%4 * w * c/4, h * b] */
__read_only image2d_t input1, __read_only image2d_t input1,
#ifdef COEFF_SUM #ifdef COEFF_SUM
......
mace/kernels/opencl/cl/fully_connected.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
// output = weight * input + bias // output = weight * input + bias
__kernel void fully_connected(__read_only image2d_t input, __kernel void fully_connected(GLOBAL_WORK_GROUP_SIZE_DIM2
__read_only image2d_t input,
__read_only image2d_t weight, __read_only image2d_t weight,
#ifdef BIAS #ifdef BIAS
__read_only image2d_t bias, __read_only image2d_t bias,
...@@ -15,6 +16,10 @@ __kernel void fully_connected(__read_only image2d_t input, ...@@ -15,6 +16,10 @@ __kernel void fully_connected(__read_only image2d_t input,
const int out_blk_idx = get_global_id(1); const int out_blk_idx = get_global_id(1);
const int input_chan_blk = (input_channel + 3) >> 2; const int input_chan_blk = (input_channel + 3) >> 2;
#ifndef NON_UNIFORM_WORK_GROUP
if (batch_idx >= global_size_dim0 || out_blk_idx >= global_size_dim1) return;
#endif
float4 input_value; float4 input_value;
float4 w0, w1, w2, w3; float4 w0, w1, w2, w3;
...@@ -57,7 +62,8 @@ __kernel void fully_connected(__read_only image2d_t input, ...@@ -57,7 +62,8 @@ __kernel void fully_connected(__read_only image2d_t input,
} }
// output = weight * input + bias // output = weight * input + bias
__kernel void fully_connected_width(__read_only image2d_t input, __kernel void fully_connected_width(GLOBAL_WORK_GROUP_SIZE_DIM3
__read_only image2d_t input,
__read_only image2d_t weight, __read_only image2d_t weight,
#ifdef BIAS #ifdef BIAS
__read_only image2d_t bias, __read_only image2d_t bias,
...@@ -73,6 +79,7 @@ __kernel void fully_connected_width(__read_only image2d_t input, ...@@ -73,6 +79,7 @@ __kernel void fully_connected_width(__read_only image2d_t input,
const int width_blk_idx = get_global_id(1); const int width_blk_idx = get_global_id(1);
const int width_blk_count = get_global_size(1); const int width_blk_count = get_global_size(1);
const int batch_out_blk_idx = get_global_id(2); const int batch_out_blk_idx = get_global_id(2);
const int batch_idx = batch_out_blk_idx / out_blks; const int batch_idx = batch_out_blk_idx / out_blks;
const int out_blk_idx = batch_out_blk_idx % out_blks; const int out_blk_idx = batch_out_blk_idx % out_blks;
...@@ -115,6 +122,16 @@ __kernel void fully_connected_width(__read_only image2d_t input, ...@@ -115,6 +122,16 @@ __kernel void fully_connected_width(__read_only image2d_t input,
short inter_idx = mad24((short)get_local_id(2), local_size, inter_out_offset); short inter_idx = mad24((short)get_local_id(2), local_size, inter_out_offset);
intermediate_output[inter_idx] = sum; intermediate_output[inter_idx] = sum;
#ifdef NON_QUALCOMM_ADRENO
barrier(CLK_LOCAL_MEM_FENCE);
#endif
#ifndef NON_UNIFORM_WORK_GROUP
if (batch_out_blk_idx >= global_size_dim2) {
return;
}
#endif
if (inter_out_offset == 0) { if (inter_out_offset == 0) {
#ifdef BIAS #ifdef BIAS
DATA_TYPE4 result = READ_IMAGET(bias, SAMPLER, (int2)(out_blk_idx, 0)); DATA_TYPE4 result = READ_IMAGET(bias, SAMPLER, (int2)(out_blk_idx, 0));
...@@ -122,7 +139,7 @@ __kernel void fully_connected_width(__read_only image2d_t input, ...@@ -122,7 +139,7 @@ __kernel void fully_connected_width(__read_only image2d_t input,
DATA_TYPE4 result = (DATA_TYPE4)(0, 0, 0, 0); DATA_TYPE4 result = (DATA_TYPE4)(0, 0, 0, 0);
#endif #endif
for(short i = 0; i < local_width_blk_size; ++i) { for (short i = 0; i < local_width_blk_size; ++i) {
result += vload4(0, intermediate_output+inter_idx); result += vload4(0, intermediate_output+inter_idx);
inter_idx += 4; inter_idx += 4;
} }
......
mace/kernels/opencl/cl/matmul.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
// C = A * B // C = A * B
__kernel void matmul( __kernel void matmul(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__read_only image2d_t A, __read_only image2d_t A,
__read_only image2d_t B, __read_only image2d_t B,
__write_only image2d_t C, __write_only image2d_t C,
......
mace/kernels/opencl/cl/pooling.cl
浏览文件 @ fdf938ce
...@@ -19,8 +19,7 @@ inline int calculate_avg_block_size(const int pool_size, ...@@ -19,8 +19,7 @@ inline int calculate_avg_block_size(const int pool_size,
} }
// Supported data type: half/float // Supported data type: half/float
__kernel void pooling( __kernel void pooling(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__private const int in_height, __private const int in_height,
__private const int in_width, __private const int in_width,
......
mace/kernels/opencl/cl/resize_bilinear.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void resize_bilinear_nocache( __kernel void resize_bilinear_nocache(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, /* [c%4 * w * c/4, h * b] */ __read_only image2d_t input, /* [c%4 * w * c/4, h * b] */
__write_only image2d_t output, __write_only image2d_t output,
__private const float height_scale, __private const float height_scale,
......
mace/kernels/opencl/cl/slice.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void slice( __kernel void slice(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__private const int chan_blk_offset, __private const int chan_blk_offset,
__write_only image2d_t output) { __write_only image2d_t output) {
......
mace/kernels/opencl/cl/softmax.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void softmax( __kernel void softmax(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t input, __read_only image2d_t input,
__private const int channels, __private const int channels,
__private const int remain_channels, __private const int remain_channels,
......
mace/kernels/opencl/cl/space_to_batch.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void space_to_batch( __kernel void space_to_batch(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t space_data, __read_only image2d_t space_data,
__write_only image2d_t batch_data, __write_only image2d_t batch_data,
__private const int block_height, __private const int block_height,
...@@ -48,8 +47,7 @@ __kernel void space_to_batch( ...@@ -48,8 +47,7 @@ __kernel void space_to_batch(
WRITE_IMAGET(batch_data, batch_coord, value); WRITE_IMAGET(batch_data, batch_coord, value);
} }
__kernel void batch_to_space( __kernel void batch_to_space(GLOBAL_WORK_GROUP_SIZE_DIM3
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_3
__read_only image2d_t batch_data, __read_only image2d_t batch_data,
__write_only image2d_t space_data, __write_only image2d_t space_data,
__private const int block_height, __private const int block_height,
......
mace/kernels/opencl/cl/winograd_transform.cl
浏览文件 @ fdf938ce
#include <common.h> #include <common.h>
__kernel void winograd_transform_2x2( __kernel void winograd_transform_2x2(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__read_only image2d_t input, __read_only image2d_t input,
__write_only image2d_t output, __write_only image2d_t output,
__private const int in_height, __private const int in_height,
...@@ -116,8 +115,7 @@ __kernel void winograd_transform_2x2( ...@@ -116,8 +115,7 @@ __kernel void winograd_transform_2x2(
} }
} }
__kernel void winograd_inverse_transform_2x2( __kernel void winograd_inverse_transform_2x2(GLOBAL_WORK_GROUP_SIZE_DIM2
UNIFORM_WORK_GROUP_SIZE_PARAMS_IN_DIM_2
__read_only image2d_t input, __read_only image2d_t input,
#ifdef BIAS #ifdef BIAS
__read_only image2d_t bias, /* cout%4 * cout/4 */ __read_only image2d_t bias, /* cout%4 * cout/4 */
......
mace/kernels/opencl/cwise_opencl.cc
浏览文件 @ fdf938ce
...@@ -23,8 +23,10 @@ void CWiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input, ...@@ -23,8 +23,10 @@ void CWiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input,
const index_t width_pixels = channel_blocks * width; const index_t width_pixels = channel_blocks * width;
const index_t batch_height_pixels = batch * height; const index_t batch_height_pixels = batch * height;
auto runtime = OpenCLRuntime::Global();
const uint32_t gws[2] = {static_cast<uint32_t>(width_pixels),
static_cast<uint32_t>(batch_height_pixels)};
if (kernel_.get() == nullptr) { if (kernel_.get() == nullptr) {
auto runtime = OpenCLRuntime::Global();
std::set<std::string> built_options; std::set<std::string> built_options;
auto dt = DataTypeToEnum<T>::value; auto dt = DataTypeToEnum<T>::value;
std::string kernel_name = MACE_OBFUSCATE_SYMBOL("cwise"); std::string kernel_name = MACE_OBFUSCATE_SYMBOL("cwise");
...@@ -32,19 +34,27 @@ void CWiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input, ...@@ -32,19 +34,27 @@ void CWiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input,
built_options.emplace("-DDATA_TYPE=" + DtToUpstreamCLDt(dt)); built_options.emplace("-DDATA_TYPE=" + DtToUpstreamCLDt(dt));
built_options.emplace("-DCMD_DATA_TYPE=" + DtToUpstreamCLCMDDt(dt)); built_options.emplace("-DCMD_DATA_TYPE=" + DtToUpstreamCLCMDDt(dt));
built_options.emplace(MakeString("-DCWISE_TYPE=", type_)); built_options.emplace(MakeString("-DCWISE_TYPE=", type_));
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
kernel_ = runtime->BuildKernel("cwise", kernel_name, built_options); kernel_ = runtime->BuildKernel("cwise", kernel_name, built_options);
kwg_size_ =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(kernel_));
} }
if (!IsVecEqual(input_shape_, input->shape())) { if (!IsVecEqual(input_shape_, input->shape())) {
uint32_t idx = 0; uint32_t idx = 0;
if (!runtime->IsNonUniformWorkgroupsSupported()) {
kernel_.setArg(idx++, gws[0]);
kernel_.setArg(idx++, gws[1]);
}
kernel_.setArg(idx++, *(input->opencl_image())); kernel_.setArg(idx++, *(input->opencl_image()));
kernel_.setArg(idx++, static_cast<float>(coeff_)); kernel_.setArg(idx++, static_cast<float>(coeff_));
kernel_.setArg(idx++, *(output->opencl_image())); kernel_.setArg(idx++, *(output->opencl_image()));
input_shape_ = input->shape(); input_shape_ = input->shape();
} }
const uint32_t gws[2] = {static_cast<uint32_t>(width_pixels), const std::vector<uint32_t> lws = {kwg_size_ / 16, 16, 1};
static_cast<uint32_t>(batch_height_pixels)};
const std::vector<uint32_t> lws = {64, 16, 1};
std::stringstream ss; std::stringstream ss;
ss << "cwise_opencl_kernel_" << output->dim(0) << "_" << output->dim(1) ss << "cwise_opencl_kernel_" << output->dim(0) << "_" << output->dim(1)
<< "_" << output->dim(2) << "_" << output->dim(3); << "_" << output->dim(2) << "_" << output->dim(3);
......
mace/kernels/opencl/fully_connected_opencl.cc
浏览文件 @ fdf938ce
...@@ -27,6 +27,10 @@ void FCWXKernel(cl::Kernel *kernel, ...@@ -27,6 +27,10 @@ void FCWXKernel(cl::Kernel *kernel,
auto runtime = OpenCLRuntime::Global(); auto runtime = OpenCLRuntime::Global();
if (kernel->get() == nullptr) { if (kernel->get() == nullptr) {
const index_t batch = output->dim(0);
const index_t output_size = output->dim(3);
const index_t output_blocks = RoundUpDiv4(output_size);
std::set<std::string> built_options; std::set<std::string> built_options;
auto dt = DataTypeToEnum<T>::value; auto dt = DataTypeToEnum<T>::value;
std::string kernel_name = MACE_OBFUSCATE_SYMBOL("fully_connected"); std::string kernel_name = MACE_OBFUSCATE_SYMBOL("fully_connected");
...@@ -55,28 +59,47 @@ void FCWXKernel(cl::Kernel *kernel, ...@@ -55,28 +59,47 @@ void FCWXKernel(cl::Kernel *kernel,
default: default:
LOG(FATAL) << "Unknown activation type: " << activation; LOG(FATAL) << "Unknown activation type: " << activation;
} }
if (runtime->gpu_type() != GPUType::QUALCOMM_ADRENO) {
built_options.emplace("-DNON_QUALCOMM_ADRENO");
}
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
*kernel = *kernel =
runtime->BuildKernel("fully_connected", kernel_name, built_options); runtime->BuildKernel("fully_connected", kernel_name, built_options);
const index_t batch = output->dim(0); if (runtime->gpu_type() == GPUType::QUALCOMM_ADRENO) {
const index_t output_size = output->dim(3); built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
const index_t output_blocks = RoundUpDiv4(output_size); const uint32_t wave_size =
const uint32_t wave_size = static_cast<uint32_t>(runtime->GetKernelWaveSize(*kernel));
static_cast<uint32_t>(runtime->GetKernelWaveSize(*kernel));
*gws = {4, (wave_size / 4), static_cast<uint32_t>(batch * output_blocks)}; *gws = {4, (wave_size / 4), static_cast<uint32_t>(batch * output_blocks)};
const uint32_t kwg_size = const uint32_t kwg_size =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(*kernel)); static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(*kernel));
const uint32_t inter_local_blks = kwg_size / ((*gws)[0] * (*gws)[1]); const uint32_t inter_local_blks = kwg_size / ((*gws)[0] * (*gws)[1]);
*lws = {(*gws)[0], (*gws)[1], inter_local_blks}; *lws = {(*gws)[0], (*gws)[1], inter_local_blks};
} else {
*gws = {4, 8, static_cast<uint32_t>(batch * output_blocks)};
const uint32_t kwg_size =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(*kernel));
const uint32_t inter_local_blks = kwg_size / ((*gws)[0] * (*gws)[1]);
*lws = {(*gws)[0], (*gws)[1], inter_local_blks};
}
} }
if (!IsVecEqual(*prev_input_shape, input->shape())) { if (!IsVecEqual(*prev_input_shape, input->shape())) {
const index_t batch = output->dim(0); const index_t batch = output->dim(0);
const index_t output_blocks = RoundUpDiv4(output->dim(3)); const index_t output_blocks = RoundUpDiv4(output->dim(3));
(*gws)[2] = static_cast<uint32_t>(batch * output_blocks);
uint32_t idx = 0; uint32_t idx = 0;
if (!runtime->IsNonUniformWorkgroupsSupported()) {
kernel->setArg(idx++, (*gws)[0]);
kernel->setArg(idx++, (*gws)[1]);
kernel->setArg(idx++, (*gws)[2]);
}
kernel->setArg(idx++, *(input->opencl_image())); kernel->setArg(idx++, *(input->opencl_image()));
kernel->setArg(idx++, *(weight->opencl_image())); kernel->setArg(idx++, *(weight->opencl_image()));
if (bias != nullptr) { if (bias != nullptr) {
...@@ -91,15 +114,25 @@ void FCWXKernel(cl::Kernel *kernel, ...@@ -91,15 +114,25 @@ void FCWXKernel(cl::Kernel *kernel,
kernel->setArg(idx++, static_cast<int>(output_blocks)); kernel->setArg(idx++, static_cast<int>(output_blocks));
kernel->setArg(idx++, relux_max_limit); kernel->setArg(idx++, relux_max_limit);
(*gws)[2] = static_cast<uint32_t>(batch * output_blocks);
*prev_input_shape = input->shape(); *prev_input_shape = input->shape();
} }
cl::Event event; cl::Event event;
cl_int error = runtime->command_queue().enqueueNDRangeKernel( cl_int error;
*kernel, cl::NullRange, cl::NDRange((*gws)[0], (*gws)[1], (*gws)[2]), if (runtime->IsNonUniformWorkgroupsSupported()) {
cl::NDRange((*lws)[0], (*lws)[1], (*lws)[2]), nullptr, &event); error = runtime->command_queue().enqueueNDRangeKernel(
MACE_CHECK_CL_SUCCESS(error); *kernel, cl::NullRange, cl::NDRange((*gws)[0], (*gws)[1], (*gws)[2]),
cl::NDRange((*lws)[0], (*lws)[1], (*lws)[2]), nullptr, &event);
} else {
std::vector<uint32_t> roundup_gws(lws->size());
for (size_t i = 0; i < lws->size(); ++i) {
roundup_gws[i] = RoundUp((*gws)[i], (*lws)[i]);
}
error = runtime->command_queue().enqueueNDRangeKernel(
*kernel, cl::NullRange,
cl::NDRange(roundup_gws[0], roundup_gws[1], roundup_gws[2]),
cl::NDRange((*lws)[0], (*lws)[1], (*lws)[2]), nullptr, &event);
}
MACE_CHECK(error == CL_SUCCESS) << "Error code: " << error;
if (future != nullptr) { if (future != nullptr) {
future->wait_fn = [runtime, event](CallStats *stats) { future->wait_fn = [runtime, event](CallStats *stats) {
...@@ -125,8 +158,8 @@ void FCWTXKernel(cl::Kernel *kernel, ...@@ -125,8 +158,8 @@ void FCWTXKernel(cl::Kernel *kernel,
StatsFuture *future) { StatsFuture *future) {
MACE_CHECK_NOTNULL(gws); MACE_CHECK_NOTNULL(gws);
MACE_CHECK_NOTNULL(lws); MACE_CHECK_NOTNULL(lws);
auto runtime = OpenCLRuntime::Global();
if (kernel->get() == nullptr) { if (kernel->get() == nullptr) {
auto runtime = OpenCLRuntime::Global();
std::set<std::string> built_options; std::set<std::string> built_options;
auto dt = DataTypeToEnum<T>::value; auto dt = DataTypeToEnum<T>::value;
std::string kernel_name = MACE_OBFUSCATE_SYMBOL("fully_connected"); std::string kernel_name = MACE_OBFUSCATE_SYMBOL("fully_connected");
...@@ -136,6 +169,9 @@ void FCWTXKernel(cl::Kernel *kernel, ...@@ -136,6 +169,9 @@ void FCWTXKernel(cl::Kernel *kernel,
if (bias != nullptr) { if (bias != nullptr) {
built_options.emplace("-DBIAS"); built_options.emplace("-DBIAS");
} }
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
switch (activation) { switch (activation) {
case NOOP: case NOOP:
break; break;
...@@ -157,10 +193,23 @@ void FCWTXKernel(cl::Kernel *kernel, ...@@ -157,10 +193,23 @@ void FCWTXKernel(cl::Kernel *kernel,
*kernel = *kernel =
runtime->BuildKernel("fully_connected", kernel_name, built_options); runtime->BuildKernel("fully_connected", kernel_name, built_options);
*lws = {16, 64, 1}; uint32_t kwg_size =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(*kernel));
*lws = {16, kwg_size/16, 1};
} }
if (!IsVecEqual(*prev_input_shape, input->shape())) { if (!IsVecEqual(*prev_input_shape, input->shape())) {
const index_t batch = output->dim(0);
const index_t output_blocks = RoundUpDiv4(output->dim(3));
*gws = {
static_cast<uint32_t>(batch), static_cast<uint32_t>(output_blocks),
};
uint32_t idx = 0; uint32_t idx = 0;
if (!runtime->IsNonUniformWorkgroupsSupported()) {
kernel->setArg(idx++, (*gws)[0]);
kernel->setArg(idx++, (*gws)[1]);
}
kernel->setArg(idx++, *(input->opencl_image())); kernel->setArg(idx++, *(input->opencl_image()));
kernel->setArg(idx++, *(weight->opencl_image())); kernel->setArg(idx++, *(weight->opencl_image()));
if (bias != nullptr) { if (bias != nullptr) {
...@@ -173,12 +222,6 @@ void FCWTXKernel(cl::Kernel *kernel, ...@@ -173,12 +222,6 @@ void FCWTXKernel(cl::Kernel *kernel,
// FIXME handle flexable data type: half not supported // FIXME handle flexable data type: half not supported
kernel->setArg(idx++, relux_max_limit); kernel->setArg(idx++, relux_max_limit);
const index_t batch = output->dim(0);
const index_t output_blocks = RoundUpDiv4(output->dim(3));
*gws = {
static_cast<uint32_t>(batch), static_cast<uint32_t>(output_blocks),
};
*prev_input_shape = input->shape(); *prev_input_shape = input->shape();
} }
......
mace/ops/depthwise_conv2d_test.cc
浏览文件 @ fdf938ce
...@@ -57,7 +57,6 @@ void SimpleValidTest() { ...@@ -57,7 +57,6 @@ void SimpleValidTest() {
.AddIntsArg("strides", {1, 1}) .AddIntsArg("strides", {1, 1})
.AddIntArg("padding", Padding::VALID) .AddIntArg("padding", Padding::VALID)
.AddIntsArg("dilations", {1, 1}) .AddIntsArg("dilations", {1, 1})
.AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
.Finalize(net.NewOperatorDef()); .Finalize(net.NewOperatorDef());
// Run // Run
net.RunOp(D); net.RunOp(D);
......
mace/ops/fully_connected_test.cc
浏览文件 @ fdf938ce
...@@ -225,7 +225,7 @@ void TestWXFormat(const index_t batch, ...@@ -225,7 +225,7 @@ void TestWXFormat(const index_t batch,
kernels::BufferType::IN_OUT_CHANNEL); kernels::BufferType::IN_OUT_CHANNEL);
BufferToImage<DeviceType::OPENCL, T>(&net, "Weight", "WeightImage", BufferToImage<DeviceType::OPENCL, T>(&net, "Weight", "WeightImage",
kernels::BufferType::WEIGHT_WIDTH); kernels::BufferType::WEIGHT_WIDTH);
BufferToImage<DeviceType::OPENCL, float>(&net, "Bias", "BiasImage", BufferToImage<DeviceType::OPENCL, T>(&net, "Bias", "BiasImage",
kernels::BufferType::ARGUMENT); kernels::BufferType::ARGUMENT);
OpDefBuilder("FC", "FullyConnectedTest") OpDefBuilder("FC", "FullyConnectedTest")
...@@ -236,7 +236,7 @@ void TestWXFormat(const index_t batch, ...@@ -236,7 +236,7 @@ void TestWXFormat(const index_t batch,
.AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value)) .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
.Finalize(net.NewOperatorDef()); .Finalize(net.NewOperatorDef());
// Run on opencl // Run
net.RunOp(DeviceType::OPENCL); net.RunOp(DeviceType::OPENCL);
ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput", ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
......
tools/benchmark.sh
浏览文件 @ fdf938ce
#!/bin/bash #!/bin/bash
set -x
Usage() { Usage() {
echo "Usage: bash tools/benchmark.sh target_soc model_output_dir option_args" echo "Usage: bash tools/benchmark.sh target_soc model_output_dir option_args"
} }
......
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