diff options
Diffstat (limited to 'src/runtime/CL/CLTuner.cpp')
-rw-r--r-- | src/runtime/CL/CLTuner.cpp | 154 |
1 files changed, 86 insertions, 68 deletions
diff --git a/src/runtime/CL/CLTuner.cpp b/src/runtime/CL/CLTuner.cpp index e16d6808ed..0d62fe3afe 100644 --- a/src/runtime/CL/CLTuner.cpp +++ b/src/runtime/CL/CLTuner.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,10 +22,12 @@ * SOFTWARE. */ #include "arm_compute/runtime/CL/CLTuner.h" -#include "arm_compute/runtime/CL/tuners/CLTuningParametersList.h" #include "arm_compute/core/Error.h" #include "arm_compute/runtime/CL/CLScheduler.h" +#include "arm_compute/runtime/CL/tuners/CLTuningParametersList.h" + +#include "src/common/utils/Log.h" #include "src/core/CL/ICLKernel.h" #include "support/StringSupport.h" @@ -36,10 +38,36 @@ namespace arm_compute { CLTuner::CLTuner(bool tune_new_kernels, CLTuningInfo tuning_info) - : real_clEnqueueNDRangeKernel(nullptr), _tuning_params_table(), _lws_table(), _kernel_event(), _tune_new_kernels(tune_new_kernels), _tuning_info(tuning_info) + : real_clEnqueueNDRangeKernel(nullptr), + _tuning_params_table(), + _lws_table(), + _kernel_event(), + _tune_new_kernels(tune_new_kernels), + _tuning_info(tuning_info) { } +struct CLTuner::IKernelData +{ + virtual ~IKernelData() = default; + virtual void do_run(ICLKernel &kernel, cl::CommandQueue &queue) = 0; +}; +struct DefaultKernelData : public CLTuner::IKernelData +{ + DefaultKernelData(ITensorPack &tensors) : _tensors{tensors} + { + } + ~DefaultKernelData() override = default; + void do_run(ICLKernel &kernel, cl::CommandQueue &queue) override + { + const bool inject_memory = !_tensors.empty(); + inject_memory ? kernel.run_op(_tensors, kernel.window(), queue) : kernel.run(kernel.window(), queue); + } + +private: + ITensorPack &_tensors; +}; + bool CLTuner::kernel_event_is_set() const { return _kernel_event() != nullptr; @@ -63,11 +91,6 @@ void CLTuner::set_tuner_mode(CLTunerMode mode) _tuning_info.tuner_mode = mode; } -CLTunerMode CLTuner::get_tuner_mode() const -{ - return _tuning_info.tuner_mode; -} - void CLTuner::tune_kernel_static(ICLKernel &kernel) { ARM_COMPUTE_UNUSED(kernel); @@ -79,29 +102,30 @@ void CLTuner::tune_kernel_dynamic(ICLKernel &kernel) tune_kernel_dynamic(kernel, pack); } -void CLTuner::tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) +void CLTuner::do_tune_kernel_dynamic(ICLKernel &kernel, IKernelData *data) { // Get the configuration ID from the kernel and append GPU target name and number of available compute units - const std::string config_id = kernel.config_id() + "_" + string_from_target(kernel.get_target()) + "_MP" + support::cpp11::to_string(CLKernelLibrary::get().get_num_compute_units()); + const std::string config_id = kernel.config_id() + "_" + string_from_target(kernel.get_target()) + "_MP" + + support::cpp11::to_string(CLKernelLibrary::get().get_num_compute_units()); // Check if we need to find the Optimal LWS. If the kernel's config_id is equal to default_config_id, the kernel does not require to be tuned - if(kernel.config_id() != arm_compute::default_config_id) + if (kernel.config_id() != arm_compute::default_config_id) { auto p = _tuning_params_table.find(config_id); - if(p == _tuning_params_table.end()) + if (p == _tuning_params_table.end()) { - if(_tune_new_kernels) + if (_tune_new_kernels) { // Find the optimal LWS for the kernel - CLTuningParams opt_tuning_params = find_optimal_tuning_params(kernel, tensors); + CLTuningParams opt_tuning_params = find_optimal_tuning_params(kernel, data); // Insert the optimal LWS in the table add_tuning_params(config_id, opt_tuning_params); // Set Local-Workgroup-Size kernel.set_lws_hint(opt_tuning_params.get_lws()); - if(_tuning_info.tune_wbsm) + if (_tuning_info.tune_wbsm) { kernel.set_wbsm_hint(opt_tuning_params.get_wbsm()); } @@ -111,17 +135,18 @@ void CLTuner::tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) { // Set Local-Workgroup-Size kernel.set_lws_hint(p->second.get_lws()); - if(_tuning_info.tune_wbsm) + if (_tuning_info.tune_wbsm) { kernel.set_wbsm_hint(p->second.get_wbsm()); } } } } - -void CLTuner::add_lws_to_table(const std::string &kernel_id, cl::NDRange optimal_lws) +void CLTuner::tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) { - add_tuning_params(kernel_id, CLTuningParams(optimal_lws)); + DefaultKernelData data{tensors}; + + do_tune_kernel_dynamic(kernel, &data); } void CLTuner::add_tuning_params(const std::string &kernel_id, CLTuningParams optimal_tuning_params) @@ -129,13 +154,13 @@ void CLTuner::add_tuning_params(const std::string &kernel_id, CLTuningParams opt _tuning_params_table.emplace(kernel_id, optimal_tuning_params); } -CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPack &tensors) +CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, IKernelData *data) { // Profiling queue cl::CommandQueue queue_profiler; // Extract real OpenCL function to intercept - if(real_clEnqueueNDRangeKernel == nullptr) + if (real_clEnqueueNDRangeKernel == nullptr) { real_clEnqueueNDRangeKernel = CLSymbols::get().clEnqueueNDRangeKernel_ptr; } @@ -146,7 +171,7 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac // Check if we can use the OpenCL timer with the default queue cl_command_queue_properties props = default_queue.getInfo<CL_QUEUE_PROPERTIES>(); - if((props & CL_QUEUE_PROFILING_ENABLE) == 0) + if ((props & CL_QUEUE_PROFILING_ENABLE) == 0) { // Set the queue for profiling queue_profiler = cl::CommandQueue(CLScheduler::get().context(), props | CL_QUEUE_PROFILING_ENABLE); @@ -157,21 +182,23 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac } // Start intercepting enqueues: - auto interceptor = [this](cl_command_queue command_queue, cl_kernel kernel, cl_uint work_dim, const size_t *gwo, const size_t *gws, const size_t *lws, cl_uint num_events_in_wait_list, - const cl_event * event_wait_list, cl_event * event) + auto interceptor = [this](cl_command_queue command_queue, cl_kernel kernel, cl_uint work_dim, const size_t *gwo, + const size_t *gws, const size_t *lws, cl_uint num_events_in_wait_list, + const cl_event *event_wait_list, cl_event *event) { - if(this->kernel_event_is_set()) + if (this->kernel_event_is_set()) { // If the event is already set it means the kernel enqueue is sliced: given that we only time the first slice we can save time by skipping the other enqueues. return CL_SUCCESS; } cl_event tmp; - cl_int retval = this->real_clEnqueueNDRangeKernel(command_queue, kernel, work_dim, gwo, gws, lws, num_events_in_wait_list, event_wait_list, &tmp); + cl_int retval = this->real_clEnqueueNDRangeKernel(command_queue, kernel, work_dim, gwo, gws, lws, + num_events_in_wait_list, event_wait_list, &tmp); // Set OpenCL event this->set_cl_kernel_event(tmp); - if(event != nullptr) + if (event != nullptr) { //return cl_event from the intercepted call clRetainEvent(tmp); @@ -181,11 +208,19 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac }; CLSymbols::get().clEnqueueNDRangeKernel_ptr = interceptor; - cl::NDRange gws = ICLKernel::gws_from_window(kernel.window()); - // Run the kernel with default lws to be used as baseline - const bool inject_memory = !tensors.empty(); - inject_memory ? kernel.run_op(tensors, kernel.window(), queue_profiler) : kernel.run(kernel.window(), queue_profiler); + data->do_run(kernel, queue_profiler); + + /// Get the cached gws used by the kernel + /// NOTE: The window configured inside configure() is usually changed in run(). Thus we should not calculate gws + /// from this static window. Instead we get the real gws used (and cached) by run() in the previous step. + /// This is only a temporary workaround. An ideal solution involves decoupling the execution window from run() / run_op() + /// Please see COMPMID-5934 + cl::NDRange gws = kernel.get_cached_gws(); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL( + arm_compute::logging::LogLevel::INFO, + "[CLTuner] Kernel with config_id '%s' uses %s as the upper-bound for lws search", kernel.config_id().c_str(), + to_string(gws).c_str()); queue_profiler.finish(); @@ -198,7 +233,7 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac // Construct the list of tuning parameters values to be tested based on the tuner mode. auto tuning_list = cl_tuner::get_tuning_parameters_list(_tuning_info, gws); - for(size_t i = 0; i < tuning_list->size(); ++i) + for (size_t i = 0; i < tuning_list->size(); ++i) { CLTuningParams tuning_test = (*tuning_list)[i]; // Setting the lws @@ -208,20 +243,22 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac auto z = lws_test[2]; const bool invalid_lws = (x * y * z > kernel.get_max_workgroup_size()) || (x == 1 && y == 1 && z == 1); - if(invalid_lws) + if (invalid_lws) { continue; } kernel.set_lws_hint(lws_test); - if(_tuning_info.tune_wbsm && CLKernelLibrary::get().is_wbsm_supported()) + if (_tuning_info.tune_wbsm && CLKernelLibrary::get().is_wbsm_supported()) { cl_int wbsm_test = tuning_test.get_wbsm(); kernel.set_wbsm_hint(wbsm_test); } + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "[CLTuner] Trying LWS: %s, WBSM: %d", + to_string(kernel.lws_hint()).c_str(), kernel.wbsm_hint()); // Run the kernel - inject_memory ? kernel.run_op(tensors, kernel.window(), queue_profiler) : kernel.run(kernel.window(), queue_profiler); + data->do_run(kernel, queue_profiler); queue_profiler.finish(); @@ -231,11 +268,11 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac _kernel_event = nullptr; // Check the execution time - if(diff < min_exec_time) + if (diff < min_exec_time) { min_exec_time = diff; opt_tuning_params.set_lws(tuning_test.get_lws()); - if(_tuning_info.tune_wbsm) + if (_tuning_info.tune_wbsm) { opt_tuning_params.set_wbsm(tuning_test.get_wbsm()); } @@ -247,25 +284,6 @@ CLTuningParams CLTuner::find_optimal_tuning_params(ICLKernel &kernel, ITensorPac return opt_tuning_params; } -void CLTuner::import_lws_table(const std::unordered_map<std::string, cl::NDRange> &lws_table) -{ - _tuning_params_table.clear(); - for(auto && params : lws_table) - { - add_tuning_params(params.first, CLTuningParams(params.second)); - } -} - -const std::unordered_map<std::string, cl::NDRange> &CLTuner::lws_table() -{ - _lws_table.clear(); - for(auto && params : _tuning_params_table) - { - _lws_table.emplace(params.first, params.second.get_lws()); - } - return _lws_table; -} - const std::unordered_map<std::string, CLTuningParams> &CLTuner::tuning_params_table() const { return _tuning_params_table; @@ -282,30 +300,30 @@ void CLTuner::load_from_file(const std::string &filename) std::ifstream fs; fs.exceptions(std::ifstream::badbit); fs.open(filename, std::ios::in); - if(!fs.is_open()) + if (!fs.is_open()) { ARM_COMPUTE_ERROR_VAR("Failed to open '%s' (%s [%d])", filename.c_str(), strerror(errno), errno); } std::string line; bool header_line = true; - while(!std::getline(fs, line).fail()) + while (!std::getline(fs, line).fail()) { - if(header_line) + if (header_line) { header_line = false; size_t pos_lws = line.find("lws"); size_t pos_wbsm = line.find("wbsm"); _tuning_info.tune_wbsm = false; - if(pos_lws != std::string::npos || pos_wbsm != std::string::npos) + if (pos_lws != std::string::npos || pos_wbsm != std::string::npos) { // The file has in the first line the parameters it has been tuned on - if(pos_wbsm != std::string::npos) + if (pos_wbsm != std::string::npos) { _tuning_info.tune_wbsm = true; } // Once the line with the tuning parameter is read we can // read the next one to start collecting the values - if(std::getline(fs, line).fail()) + if (std::getline(fs, line).fail()) { break; } @@ -314,13 +332,13 @@ void CLTuner::load_from_file(const std::string &filename) CLTuningParams tuning_params; size_t pos = line.find(";"); - if(pos == std::string::npos) + if (pos == std::string::npos) { ARM_COMPUTE_ERROR_VAR("Malformed row '%s' in %s", line.c_str(), filename.c_str()); } std::string kernel_id = line.substr(0, pos); line.erase(0, pos + 1); - if(!tuning_params.from_string(_tuning_info, line)) + if (!tuning_params.from_string(_tuning_info, line)) { ARM_COMPUTE_ERROR_VAR("Malformed row '%s' in %s", line.c_str(), filename.c_str()); } @@ -331,7 +349,7 @@ void CLTuner::load_from_file(const std::string &filename) bool CLTuner::save_to_file(const std::string &filename) const { - if(!_tune_new_kernels || _tuning_params_table.empty() || filename.empty()) + if (!_tune_new_kernels || _tuning_params_table.empty() || filename.empty()) { return false; } @@ -340,16 +358,16 @@ bool CLTuner::save_to_file(const std::string &filename) const fs.open(filename, std::ios::out); std::string header_string = ""; header_string += "lws"; - if(_tuning_info.tune_wbsm) + if (_tuning_info.tune_wbsm) { - if(!header_string.empty()) + if (!header_string.empty()) { header_string += " "; } header_string += "wbsm"; } fs << header_string << std::endl; - for(auto const &kernel_data : _tuning_params_table) + for (auto const &kernel_data : _tuning_params_table) { CLTuningParams tun_pams(kernel_data.second); fs << kernel_data.first << tun_pams.to_string(_tuning_info) << std::endl; |