Skip to content

Commit

Permalink
[OPENCL] Add write tuned param to file & improve lws generator. test=…
Browse files Browse the repository at this point in the history 
…develop (PaddlePaddle#5485)

* [OPENCL] improve local work size generate. test=develop
* [OPENCL][API] add api for saving local work size to device
  • Loading branch information
@ysh329
ysh329 authored Mar 3, 2021
1 parent 08dca04 commit 68e358b
Show file tree
Hide file tree
Showing 20 changed files with 732 additions and 185 deletions.
1 change: 1 addition & 0 deletions CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -193,6 +193,7 @@ else()
COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CMAKE_SOURCE_DIR}/third-party/flatbuffers/pre-build/framework_generated.h" "${CMAKE_SOURCE_DIR}/lite/model_parser/flatbuffers/"
COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CMAKE_SOURCE_DIR}/third-party/flatbuffers/pre-build/param_generated.h" "${CMAKE_SOURCE_DIR}/lite/model_parser/flatbuffers/"
COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CMAKE_SOURCE_DIR}/third-party/flatbuffers/pre-build/cache_generated.h" "${CMAKE_SOURCE_DIR}/lite/backends/opencl/utils/"
COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CMAKE_SOURCE_DIR}/third-party/flatbuffers/pre-build/tune_cache_generated.h" "${CMAKE_SOURCE_DIR}/lite/backends/opencl/utils/"
)
endif()

Expand Down
4 changes: 3 additions & 1 deletion cmake/external/flatbuffers.cmake
View file
Original file line number Diff line number Diff line change
Expand Up @@ -107,12 +107,14 @@ set(FRAMEWORK_FBS_DIR "lite/model_parser/flatbuffers")
set(FRAMEWORK_SCHEMA_PATH "lite/model_parser/flatbuffers/framework.fbs")
set(PARAM_SCHEMA_PATH "lite/model_parser/flatbuffers/param.fbs")
set(CL_CACHE_SCHEMA_PATH "lite/backends/opencl/utils/cache.fbs")
set(CL_TUNE_CACHE_SCHEMA_PATH "lite/backends/opencl/utils/tune_cache.fbs")
compile_flatbuffers_schema_to_cpp_opt(framework_fbs_header ${FRAMEWORK_SCHEMA_PATH} "--no-includes;--gen-compare;--force-empty")
compile_flatbuffers_schema_to_cpp_opt(param_fbs_header ${PARAM_SCHEMA_PATH} "--no-includes;--gen-compare;--force-empty")
compile_flatbuffers_schema_to_cpp_opt(cl_cache_fbs_header ${CL_CACHE_SCHEMA_PATH} "--no-includes;--gen-compare;--force-empty")
compile_flatbuffers_schema_to_cpp_opt(cl_tune_cache_fbs_header ${CL_TUNE_CACHE_SCHEMA_PATH} "--no-includes;--gen-compare;--force-empty")

# All header files generated by flatbuffers must be declared here to avoid compilation failure.
add_custom_target(fbs_headers ALL DEPENDS framework_fbs_header param_fbs_header cl_cache_fbs_header)
add_custom_target(fbs_headers ALL DEPENDS framework_fbs_header param_fbs_header cl_cache_fbs_header cl_tune_cache_fbs_header)

file(GENERATE OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/.fbs_dummy.cc CONTENT "")
add_library(fbs_headers_dummy STATIC ${CMAKE_CURRENT_BINARY_DIR}/.fbs_dummy.cc)
Expand Down
13 changes: 9 additions & 4 deletions lite/api/paddle_api.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -287,15 +287,20 @@ void ConfigBase::set_opencl_binary_path_name(const std::string &path,
#endif
}

void ConfigBase::set_opencl_tune(CLTuneMode tune_mode, size_t lws_repeats) {
void ConfigBase::set_opencl_tune(CLTuneMode tune_mode,
const std::string &path,
const std::string &name,
size_t lws_repeats) {
#ifdef LITE_WITH_OPENCL
if (paddle::lite_api::IsOpenCLBackendValid()) {
opencl_tune_mode_ = tune_mode;
paddle::lite::CLRuntime::Global()->set_auto_tune(opencl_tune_mode_,
lws_repeats);
paddle::lite::CLRuntime::Global()->set_auto_tune(
opencl_tune_mode_, path, name, lws_repeats);
#ifdef LITE_WITH_LOG
LOG(INFO) << "set opencl_tune_mode: "
<< CLTuneModeToStr(lite::CLRuntime::Global()->auto_tune());
<< CLTuneModeToStr(lite::CLRuntime::Global()->auto_tune())
<< ", lws_repeats:" << lws_repeats;
LOG(INFO) << "tuned file path & name:" << path << "/" << name;
#endif
}
#endif
Expand Down
3 changes: 3 additions & 0 deletions lite/api/paddle_api.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -177,7 +177,10 @@ class LITE_API ConfigBase {
const std::string& name);
// set GPU opencl tune
void set_opencl_tune(CLTuneMode tune_mode = CL_TUNE_NONE,
const std::string& path = "",
const std::string& name = "",
size_t lws_repeats = 4);

// set GPU opencl precision
void set_opencl_precision(CLPrecisionType p = CL_PRECISION_AUTO);
// set subgraph_model_dir
Expand Down
183 changes: 52 additions & 131 deletions lite/backends/opencl/cl_context.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -106,155 +106,76 @@ cl::NDRange CLContext::DefaultGlobalWorkSize(const CLImage &image) {
}
}

std::vector<cl::NDRange> CLContext::GenerateLocalWorkSizes(
cl::NDRange global_work_size, size_t max_work_size) {
size_t generate_lws_type = CLRuntime::Global()->auto_tune();
std::set<cl::NDRange> CLContext::GenerateLocalWorkSizes(cl::NDRange gws,
size_t max_ws) {
size_t tune_type = CLRuntime::Global()->auto_tune();

cl::NDRange tmp_lws = DefaultLocalWorkSize(
global_work_size, max_work_size, /*divisor=*/2, /*tune_reverse=*/false);
cl::NDRange tmp_lws =
DefaultLocalWorkSize(gws, max_ws, /*divisor=*/2, /*tune_reverse=*/false);
cl::NDRange last_lws = cl::NDRange{
static_cast<size_t>(0), static_cast<size_t>(0), static_cast<size_t>(0)};

std::vector<cl::NDRange> lwss{tmp_lws};
VLOG(1) << "generate_lws_type:" << generate_lws_type;
// 0 - None, 1 - Rapid, 2 - Normal, 3 - Exhaustive
if (generate_lws_type == lite_api::CL_TUNE_NONE) {
// 0 - None: nothing to do
} else if (generate_lws_type == lite_api::CL_TUNE_RAPID) {
// 1 - Rapid
for (auto tune_reverse : {true, false}) {
for (size_t divisor = 1; divisor < /*max_divisor=*/17; divisor *= 2) {
tmp_lws = DefaultLocalWorkSize(
global_work_size, max_work_size, divisor, tune_reverse);
if (last_lws[0] == tmp_lws[0] && last_lws[1] == tmp_lws[1] &&
last_lws[2] == tmp_lws[2]) {
// skip tuned lws
continue;
}
lwss.emplace_back(tmp_lws);
}
}
} else if (generate_lws_type == lite_api::CL_TUNE_NORMAL) {
// 2 - Normal
for (auto tune_reverse : {true, false}) {
for (size_t divisor = 1; divisor < /*max_divisor=*/15; divisor += 2) {
tmp_lws = DefaultLocalWorkSize(
global_work_size, max_work_size, divisor, tune_reverse);
if (last_lws[0] == tmp_lws[0] && last_lws[1] == tmp_lws[1] &&
last_lws[2] == tmp_lws[2]) {
// skip tuned lws
continue;
}
lwss.emplace_back(tmp_lws);
}
}
} else if (generate_lws_type == lite_api::CL_TUNE_EXHAUSTIVE) {
// 3 - Exhaustive
for (auto tune_reverse : {true, false}) {
for (size_t divisor = 1; divisor < /*max_divisor=*/15; divisor++) {
tmp_lws = DefaultLocalWorkSize(
global_work_size, max_work_size, divisor, tune_reverse);
if (last_lws[0] == tmp_lws[0] && last_lws[1] == tmp_lws[1] &&
last_lws[2] == tmp_lws[2]) {
// skip tuned lws
continue;
}
lwss.emplace_back(tmp_lws);
std::set<cl::NDRange> lwss{tmp_lws};

auto gen_lws = [&](const std::set<bool> &tune_reverses,
const std::set<size_t> &divisors) {
for (bool tune_reverse : tune_reverses) {
for (size_t divisor : divisors) {
tmp_lws = DefaultLocalWorkSize(gws, max_ws, divisor, tune_reverse);
lwss.emplace(tmp_lws);
}
}
};

std::set<bool> tune_reverses{true, false};
std::set<size_t> divisors;
if (tune_type == lite_api::CL_TUNE_NONE) {
// do nothing
} else if (tune_type == lite_api::CL_TUNE_RAPID) {
divisors = {1, 2, 4, 8};
} else if (tune_type == lite_api::CL_TUNE_NORMAL) {
divisors = {1, 3, 5, 7, 9, 11, 13};
} else if (tune_type == lite_api::CL_TUNE_EXHAUSTIVE) {
divisors = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14};
} else {
LOG(FATAL) << "Unsupported opencl tune type:" << generate_lws_type;
LOG(FATAL) << "Unsupported opencl tune type:" << tune_type;
}

gen_lws(tune_reverses, divisors);
return lwss;
}

cl::NDRange CLContext::DefaultLocalWorkSize(
cl::NDRange global_work_size,
size_t max_work_size,
int divisor /*=2*/,
bool tune_reverse /*=false*/,
size_t user_defined_max_work_size /*=0*/) {
int gws0 = global_work_size[0];
int gws1 = global_work_size[1];
int gws2 = global_work_size[2];

if (tune_reverse) {
gws2 = global_work_size[0];
gws1 = global_work_size[1];
gws0 = global_work_size[2];
}

if (divisor > 1) {
max_work_size /= divisor;
}
if (user_defined_max_work_size > 0 &&
user_defined_max_work_size <= max_work_size) {
max_work_size = user_defined_max_work_size;
}

while (gws1 > max_work_size && max_work_size > 0) {
gws1 = gws1 % 2 == 0 ? gws1 / 2 : 1;
}
while (gws2 * gws1 > max_work_size && max_work_size > 0) {
gws2 = gws2 % 2 == 0 ? gws2 / 2 : 1;
}
while (gws0 * gws1 * gws2 > max_work_size && max_work_size > 0) {
gws0 = gws0 % 2 == 0 ? gws0 / 2 : 1;
const cl::NDRange &gws,
register size_t max_ws,
const int &divisor /*=2*/,
const bool &reverse /*=false*/,
const size_t &user_def_max_ws /*=0*/) {
register size_t lx = reverse ? gws[2] : gws[0];
register size_t ly = gws[1];
register size_t lz = reverse ? gws[0] : gws[2];

max_ws = (user_def_max_ws > 0 && user_def_max_ws <= max_ws) ? user_def_max_ws
: max_ws;
max_ws = divisor > 1 ? max_ws / divisor : max_ws;

if (max_ws > 0) {
while (ly > max_ws) {
// replace mod with bit operate
ly = (ly & 0x01) ? 1 : ly >> 1;
}
while (ly * lz > max_ws) {
lz = (lz & 0x01) ? 1 : lz >> 1;
}
while (ly * lz * lx > max_ws) {
lx = (lx & 0x01) ? 1 : lx >> 1;
}
}

if (tune_reverse) {
return cl::NDRange{static_cast<size_t>(gws2),
static_cast<size_t>(gws1),
static_cast<size_t>(gws0)};
} else {
return cl::NDRange{static_cast<size_t>(gws0),
static_cast<size_t>(gws1),
static_cast<size_t>(gws2)};
}
return reverse ? cl::NDRange{lz, ly, lx} : cl::NDRange{lx, ly, lz};
}

bool CLContext::IsArmMali() {
return CLRuntime::Global()->GetGpuType() == GpuType::ARM_MALI;
}

bool CLContext::HasTunedLocalWorkSizeMap(const std::string &key,
cl::NDRange *lws) {
bool has = false;
auto it = tuned_lwss_map_.find(key);
if (it != tuned_lwss_map_.end()) {
*lws = it->second;
has = true;
}
return has;
}

void CLContext::SetTunedLocalWorkSizeMap(const std::string &key,
const cl::NDRange lws) {
auto it = tuned_lwss_map_.find(key);
if (it != tuned_lwss_map_.end()) {
auto lws_old = it->second;
LOG(FATAL) << "===> found lws_old with same key, please add more detailed "
"info to key <==="
<< "\n lws_old:" << lws_old[0] << "," << lws_old[1] << ","
<< lws_old[2] << "\n lws_new:" << lws[0] << "," << lws[1] << ","
<< lws[2];
}
tuned_lwss_map_.insert(std::pair<std::string, cl::NDRange>(key, lws));
}

std::map<std::string, cl::NDRange> CLContext::GetTunedLocalWorkSizeMap() {
return tuned_lwss_map_;
}

cl::NDRange CLContext::GetTunedLocalWorkSizeFromMap(const std::string &key) {
cl::NDRange lws = cl::NullRange;
auto it = tuned_lwss_map_.find(key);
if (it != tuned_lwss_map_.end()) {
lws = it->second;
}
return lws;
}

} // namespace lite
} // namespace paddle
24 changes: 9 additions & 15 deletions lite/backends/opencl/cl_context.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@ limitations under the License. */

#include <map>
#include <memory>
#include <set>
#include <string>
#include <vector>
#include "lite/backends/opencl/cl_image.h"
Expand Down Expand Up @@ -57,24 +58,17 @@ class CLContext {

cl::NDRange DefaultGlobalWorkSize(const CLImage &image);

cl::NDRange DefaultLocalWorkSize(cl::NDRange global_work_size,
size_t max_work_size,
int divitor = 2,
bool tune_reverse = false,
size_t user_defined_max_work_size = 0);
cl::NDRange DefaultLocalWorkSize(
const cl::NDRange &global_work_size,
register size_t max_work_size,
const int &divitor = 2,
const bool &tune_reverse = false,
const size_t &user_defined_max_work_size = 0);

std::vector<cl::NDRange> GenerateLocalWorkSizes(cl::NDRange global_work_size,
size_t max_work_size);
std::set<cl::NDRange> GenerateLocalWorkSizes(cl::NDRange global_work_size,
size_t max_work_size);
bool IsArmMali();

bool HasTunedLocalWorkSizeMap(const std::string &key, cl::NDRange *lws);

void SetTunedLocalWorkSizeMap(const std::string &key, const cl::NDRange lws);

std::map<std::string, cl::NDRange> GetTunedLocalWorkSizeMap();

cl::NDRange GetTunedLocalWorkSizeFromMap(const std::string &key);

private:
std::vector<std::shared_ptr<cl::Kernel>> kernels_;
std::map<std::string, int> kernel_offset_;
Expand Down
Loading

0 comments on commit 68e358b

Please sign in to comment.