diff options
Diffstat (limited to 'src/runtime/CL/tuners')
-rw-r--r-- | src/runtime/CL/tuners/BifrostTuner.cpp | 304 | ||||
-rw-r--r-- | src/runtime/CL/tuners/CLTuningParametersList.cpp | 77 | ||||
-rw-r--r-- | src/runtime/CL/tuners/MidgardTuner.cpp | 82 |
3 files changed, 43 insertions, 420 deletions
diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp deleted file mode 100644 index 7a06de6d1c..0000000000 --- a/src/runtime/CL/tuners/BifrostTuner.cpp +++ /dev/null @@ -1,304 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/runtime/CL/tuners/BifrostTuner.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "src/core/CL/CLKernels.h" -#include "support/Cast.h" - -#include "src/core/gpu/cl/kernels/ClPoolingKernel.h" - -namespace arm_compute -{ -namespace tuners -{ -namespace -{ -/** Tunes a @ref CLDirectConvolutionLayerKernel for a bifrost target - * - * @param[in] k Kernels to tune - */ -void tune_direct_convolution_kernel(CLDirectConvolutionLayerKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - - const GPUTarget gpu_target = k.get_target(); - const DataType dt = k._input->info()->data_type(); - const TensorShape weights_shape = k._weights->info()->tensor_shape(); - const TensorShape inputs_shape = k._input->info()->tensor_shape(); - const size_t kernel_size = weights_shape.x(); - const unsigned int stride_x = k._conv_stride_x; - const unsigned int stride_y = k._conv_stride_y; - - if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72) && (kernel_size <= 5) && (stride_x == 1) && (stride_y == 1) && (dt == DataType::F32)) - { - // Through extensive experimentation with over 30 representative tensor - // shapes, we found a small number of local work size configurations - // that result in nearly optimal execution times. Selecting the right - // lws for a given shape, however, required a complex decision tree, - // until we constructed a simple feature as described below. - // - // We started from the number of multiply-accumulate operations for a - // convolution layer, which is equal to the product of the input - // dimensions 0..2 and the weights dimensions 0..2. Unfortunately, - // this resulted in ties between distinct shapes that required distinct - // lws configurations. Replacing the width of the input with the kernel - // size, however, resulted in nearly optimal predictions. We use underscores - // in variable names to indicate when they are intentionally misleading. - const size_t product_of_weights_dimensions = weights_shape[0] * weights_shape[1] * weights_shape[2]; - const size_t product_of_input_dimensions_ = inputs_shape[0] * inputs_shape[1] * inputs_shape[2]; - const float mega_ops_ = 1e-6 * product_of_weights_dimensions * product_of_input_dimensions_; - - switch(kernel_size) - { - case 1: - { - if(mega_ops_ < 1.f) - { - lws_hint = cl::NDRange(1, 1, 8); - } - else if(mega_ops_ < 7.f) - { - lws_hint = cl::NDRange(1, 1, 4); - } - else - { - lws_hint = cl::NDRange(1, 1, 2); - } - break; - } - case 3: - { - if(mega_ops_ < 1.f) - { - lws_hint = cl::NDRange(1, 1, 8); - } - else if(mega_ops_ < 13.f) - { - lws_hint = cl::NDRange(2, 1, 4); - } - else if(mega_ops_ < 50.f) - { - lws_hint = cl::NDRange(3, 1, 4); - } - else - { - lws_hint = cl::NDRange(2, 1, 6); - } - break; - } - case 5: - { - if(mega_ops_ < 2.f || mega_ops_ > 80.f) - { - lws_hint = cl::NDRange(2, 1, 4); - } - else - { - lws_hint = cl::NDRange(2, 1, 8); - } - break; - } - default: - break; - } - k.set_lws_hint(lws_hint); - } -} - -void tune_col2im_kernel(CLCol2ImKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - const GPUTarget gpu_target = k.get_target(); - - // Configure the local work size for Bifrost with a value obtained - // via exhaustive autotuning over 30 representative tensor shapes. - if(gpu_target_is_in(gpu_target, - GPUTarget::G71, GPUTarget::G72, GPUTarget::G76, - GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, - GPUTarget::G52, GPUTarget::G52LIT)) - { - if((k._convolved_dims.width == 7) || (k._convolved_dims.width == 14)) - { - lws_hint = cl::NDRange(1, 7, 1); - } - else - { - lws_hint = cl::NDRange(1, 8, 1); - } - } - - k.set_lws_hint(lws_hint); -} - -void tune_im2col_kernel(CLIm2ColKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - const GPUTarget gpu_target = k.get_target(); - - // Local work size optimized for the 11x11 AlexNet convolution on Bifrost. - if(gpu_target_is_in(gpu_target, - GPUTarget::G71, GPUTarget::G72, GPUTarget::G76, - GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, - GPUTarget::G52, GPUTarget::G52LIT) - && k._kernel_dims.width == 11) - { - const bool is_square_kernel = (k._kernel_dims.width == k._kernel_dims.height); - if(!is_square_kernel && k._kernel_dims.width > 1 && !k._conv_info.has_padding()) - { - lws_hint = cl::NDRange(1, 1, 1); - } - } - k.set_lws_hint(lws_hint); -} - -void tune_gemm_kernel(CLGEMMMatrixMultiplyKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - const GPUTarget gpu_target = k.get_target(); - - // Configure LWS hint - switch(gpu_target) - { - case GPUTarget::G71: - case GPUTarget::G72: - case GPUTarget::G51: - case GPUTarget::G51BIG: - case GPUTarget::G51LIT: - case GPUTarget::G52: - case GPUTarget::G52LIT: - case GPUTarget::G76: - if(k._input1->info()->dimension(1) == 24) - { - // LWS optimized for the 11x11 AlexNet convolution on Bifrost. - lws_hint = cl::NDRange(2, 2); - } - else if(k._output->info()->dimension(1) == 196) - { - lws_hint = cl::NDRange(1, 7); - } - else - { - lws_hint = cl::NDRange(8, 8); - } - break; - default: - lws_hint = cl::NullRange; - } - - k.set_lws_hint(lws_hint); -} - -void tune_pooling_kernel(opencl::kernels::ClPoolingKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - const GPUTarget gpu_target = k.get_target(); - - // Configure the local work size (hint) from the first two dimensions of the global work size. - // On Bifrost, this works for up to 35x35xC filters, for which the pooling_layer_3_optimized - // kernel is launched with gws=(9, 33, C). In any case, the hint will be ignored if it is - // invalid (e.g. exceeds the maximum workgroup size that the kernel can be launched with). - if(k._pool_info.data_layout == DataLayout::NCHW) - { - if(gpu_target_is_in(gpu_target, - GPUTarget::G71, GPUTarget::G72, GPUTarget::G76, - GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, - GPUTarget::G52, GPUTarget::G52LIT)) - { - cl::NDRange gws = ICLKernel::gws_from_window(k.window()); - lws_hint = cl::NDRange(gws[0], gws[1], 1); - } - } - - k.set_lws_hint(lws_hint); -} - -void tune_scale_kernel(CLScaleKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - const GPUTarget gpu_target = k.get_target(); - const DataType dt = k.input()->info()->data_type(); - const InterpolationPolicy interpolation = k.get_interpolation_policy(); - - // Configure the local work size for Bifrost, interpolation (bilinear) and datatype F32. - // The value are obtained via exhaustive autotuning. - if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72) && (dt == DataType::F32) && (interpolation == InterpolationPolicy::BILINEAR)) - { - auto dim_0 = k.output()->info()->dimension(0); - if(dim_0 == 480) - { - lws_hint = cl::NDRange(2, 1); - } - else if(dim_0 == 3120) - { - lws_hint = cl::NDRange(2, 8); - } - else if(dim_0 == 4160) - { - lws_hint = cl::NDRange(4, 8); - } - k.set_lws_hint(lws_hint); - } -} -} // namespace - -void BifrostTuner::tune_kernel_static(ICLKernel &kernel) -{ - if(dynamic_cast<CLDirectConvolutionLayerKernel *>(&kernel) != nullptr) - { - tune_direct_convolution_kernel(*utils::cast::polymorphic_downcast<CLDirectConvolutionLayerKernel *>(&kernel)); - } - else if(dynamic_cast<CLCol2ImKernel *>(&kernel) != nullptr) - { - tune_col2im_kernel(*utils::cast::polymorphic_downcast<CLCol2ImKernel *>(&kernel)); - } - else if(dynamic_cast<CLIm2ColKernel *>(&kernel) != nullptr) - { - tune_im2col_kernel(*utils::cast::polymorphic_downcast<CLIm2ColKernel *>(&kernel)); - } - else if(dynamic_cast<CLGEMMMatrixMultiplyKernel *>(&kernel) != nullptr) - { - tune_gemm_kernel(*utils::cast::polymorphic_downcast<CLGEMMMatrixMultiplyKernel *>(&kernel)); - } - else if(dynamic_cast<opencl::kernels::ClPoolingKernel *>(&kernel) != nullptr) - { - tune_pooling_kernel(*utils::cast::polymorphic_downcast<opencl::kernels::ClPoolingKernel *>(&kernel)); - } - else if(dynamic_cast<CLScaleKernel *>(&kernel) != nullptr) - { - tune_scale_kernel(*utils::cast::polymorphic_downcast<CLScaleKernel *>(&kernel)); - } -} - -void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel) -{ - ARM_COMPUTE_UNUSED(kernel); -} - -void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) -{ - ARM_COMPUTE_UNUSED(kernel, tensors); -} -} // namespace tuners -} // namespace arm_compute
\ No newline at end of file diff --git a/src/runtime/CL/tuners/CLTuningParametersList.cpp b/src/runtime/CL/tuners/CLTuningParametersList.cpp index 6cb2212794..5e3907f1ea 100644 --- a/src/runtime/CL/tuners/CLTuningParametersList.cpp +++ b/src/runtime/CL/tuners/CLTuningParametersList.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -27,20 +27,20 @@ namespace arm_compute { namespace cl_tuner { -constexpr unsigned int max_lws_supported_x{ 64u }; -constexpr unsigned int max_lws_supported_y{ 32u }; -constexpr unsigned int max_lws_supported_z{ 32u }; +constexpr unsigned int max_lws_supported_x{64u}; +constexpr unsigned int max_lws_supported_y{32u}; +constexpr unsigned int max_lws_supported_z{32u}; -/** Non instantiable base class for Tuning parameters combinations that use Index2Cooard mapping */ +/** Non instantiable base class for Tuning parameters combinations that use Index2Coord mapping */ class CLTuningParametersList : public ICLTuningParametersList { protected: /* Shape of 4-D search space */ - TensorShape search_space_shape{ 0, 0, 0, 0 }; - std::vector<unsigned int> _lws_x{ 0 }; - std::vector<unsigned int> _lws_y{ 0 }; - std::vector<unsigned int> _lws_z{ 0 }; - std::vector<int> _wbsm{ 0 }; /* Modify the batches size of workgroups distributed to compute units. + TensorShape search_space_shape{0, 0, 0, 0}; + std::vector<unsigned int> _lws_x{0}; + std::vector<unsigned int> _lws_y{0}; + std::vector<unsigned int> _lws_z{0}; + std::vector<int> _wbsm{0}; /* Modify the batches size of workgroups distributed to compute units. The value is in the range [-31,+31]. When 0, the runtime-selected wbs used is unmodified. */ @@ -116,7 +116,8 @@ private: * @param[in] lws_max Max LWS value allowed to be tested * @param[in] mod_let_one True if the results of the modulo operation between gws and the lws can be less than one. */ - void initialize_lws_values(std::vector<unsigned int> &lws, unsigned int gws, unsigned int lws_max, bool mod_let_one); + void + initialize_lws_values(std::vector<unsigned int> &lws, unsigned int gws, unsigned int lws_max, bool mod_let_one); }; /** A minimal subset of LWS values that only have 1,2 and 4/8 */ @@ -162,14 +163,17 @@ CLTuningParams CLTuningParametersListExhaustive::operator[](size_t index) CLTuningParametersListExhaustive::CLTuningParametersListExhaustive(const cl::NDRange &gws, CLTuningInfo tuning_info) { - ARM_COMPUTE_UNUSED(gws); - search_space_shape[0] = max_lws_supported_x; - search_space_shape[1] = max_lws_supported_y; - search_space_shape[2] = max_lws_supported_z; + const auto lws_x_max = std::min(static_cast<unsigned int>(gws[0]), max_lws_supported_x); + const auto lws_y_max = std::min(static_cast<unsigned int>(gws[1]), max_lws_supported_y); + const auto lws_z_max = std::min(static_cast<unsigned int>(gws[2]), max_lws_supported_z); + + search_space_shape[0] = lws_x_max; + search_space_shape[1] = lws_y_max; + search_space_shape[2] = lws_z_max; search_space_shape[3] = 1; - if(tuning_info.tune_wbsm) + if (tuning_info.tune_wbsm) { - _wbsm = { -3, -2, -1, 0, 1, 2, 3 }; + _wbsm = {-3, -2, -1, 0, 1, 2, 3}; search_space_shape[3] = _wbsm.size(); } } @@ -183,34 +187,39 @@ CLTuningParams CLTuningParametersListNormal::operator[](size_t index) CLTuningParametersListNormal::CLTuningParametersListNormal(const cl::NDRange &gws, CLTuningInfo tuning_info) { - auto lws_x_max = std::min(static_cast<unsigned int>(gws[0]), max_lws_supported_x); - auto lws_y_max = std::min(static_cast<unsigned int>(gws[1]), max_lws_supported_y); - auto lws_z_max = std::min(static_cast<unsigned int>(gws[2]), max_lws_supported_z); + const auto lws_x_max = std::min(static_cast<unsigned int>(gws[0]), max_lws_supported_x); + const auto lws_y_max = std::min(static_cast<unsigned int>(gws[1]), max_lws_supported_y); + const auto lws_z_max = std::min(static_cast<unsigned int>(gws[2]), max_lws_supported_z); // Initialize the tuning parameters values to test _lws_x = {}; _lws_y = {}; _lws_z = {}; - initialize_lws_values(_lws_x, gws[0], lws_x_max, gws[2] > 16); // Explore lws that are not factors of gws only when gws[2] > 16 - initialize_lws_values(_lws_y, gws[1], lws_y_max, gws[2] > 16); // Explore lws that are not factors of gws only when gws[2] > 16 + initialize_lws_values(_lws_x, gws[0], lws_x_max, + gws[2] > 16); // Explore lws that are not factors of gws only when gws[2] > 16 + initialize_lws_values(_lws_y, gws[1], lws_y_max, + gws[2] > 16); // Explore lws that are not factors of gws only when gws[2] > 16 initialize_lws_values(_lws_z, gws[2], lws_z_max, false); search_space_shape[0] = _lws_x.size(); search_space_shape[1] = _lws_y.size(); search_space_shape[2] = _lws_z.size(); search_space_shape[3] = 1; - if(tuning_info.tune_wbsm) + if (tuning_info.tune_wbsm) { - _wbsm = { -2, -1, 0, 1, 2 }; + _wbsm = {-2, -1, 0, 1, 2}; search_space_shape[3] = _wbsm.size(); } } -void CLTuningParametersListNormal::initialize_lws_values(std::vector<unsigned int> &lws, unsigned int gws, unsigned int lws_max, bool mod_let_one) +void CLTuningParametersListNormal::initialize_lws_values(std::vector<unsigned int> &lws, + unsigned int gws, + unsigned int lws_max, + bool mod_let_one) { lws.push_back(1); - for(unsigned int i = 2; i <= lws_max; ++i) + for (unsigned int i = 2; i <= lws_max; ++i) { // Power of two condition const bool is_power_of_two = (i & (i - 1)) == 0; @@ -218,7 +227,7 @@ void CLTuningParametersListNormal::initialize_lws_values(std::vector<unsigned in // Condition for the module accordingly with the mod_let_one flag const bool mod_cond = mod_let_one ? (gws % i) <= 1 : (gws % i) == 0; - if(mod_cond || is_power_of_two) + if (mod_cond || is_power_of_two) { lws.push_back(i); } @@ -227,9 +236,9 @@ void CLTuningParametersListNormal::initialize_lws_values(std::vector<unsigned in CLTuningParametersListRapid::CLTuningParametersListRapid(const cl::NDRange &gws, CLTuningInfo tuning_info) { - auto lws_x_max = std::min(static_cast<unsigned int>(gws[0]), 8u); // Limit exploration to 1 - 8 - auto lws_y_max = std::min(static_cast<unsigned int>(gws[1]), 4u); // Limit exploration to 1 - 4 - auto lws_z_max = std::min(static_cast<unsigned int>(gws[2]), 4u); // Limit exploration to 1 - 4 + const auto lws_x_max = std::min(static_cast<unsigned int>(gws[0]), 8u); // Limit exploration to 1 - 8 + const auto lws_y_max = std::min(static_cast<unsigned int>(gws[1]), 4u); // Limit exploration to 1 - 4 + const auto lws_z_max = std::min(static_cast<unsigned int>(gws[2]), 4u); // Limit exploration to 1 - 4 // Initialize the LWS values to test _lws_x = {}; @@ -243,9 +252,9 @@ CLTuningParametersListRapid::CLTuningParametersListRapid(const cl::NDRange &gws, search_space_shape[1] = _lws_y.size(); search_space_shape[2] = _lws_z.size(); search_space_shape[3] = 1; - if(tuning_info.tune_wbsm) + if (tuning_info.tune_wbsm) { - _wbsm = { -1, 0, 1 }; + _wbsm = {-1, 0, 1}; search_space_shape[3] = _wbsm.size(); } } @@ -254,7 +263,7 @@ void CLTuningParametersListRapid::initialize_lws_values(std::vector<unsigned int { lws.push_back(1); - for(unsigned int i = 2; i <= lws_max; i *= 4) + for (unsigned int i = 2; i <= lws_max; i *= 4) { lws.push_back(i); } @@ -262,7 +271,7 @@ void CLTuningParametersListRapid::initialize_lws_values(std::vector<unsigned int std::unique_ptr<ICLTuningParametersList> get_tuning_parameters_list(CLTuningInfo tuning_info, const cl::NDRange &gws) { - switch(tuning_info.tuner_mode) + switch (tuning_info.tuner_mode) { case CLTunerMode::EXHAUSTIVE: return std::make_unique<CLTuningParametersListExhaustive>(gws, tuning_info); diff --git a/src/runtime/CL/tuners/MidgardTuner.cpp b/src/runtime/CL/tuners/MidgardTuner.cpp deleted file mode 100644 index 72734f2207..0000000000 --- a/src/runtime/CL/tuners/MidgardTuner.cpp +++ /dev/null @@ -1,82 +0,0 @@ -/* - * Copyright (c) 2018-2020 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/runtime/CL/tuners/MidgardTuner.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "src/core/CL/CLKernels.h" -#include "support/Cast.h" - -namespace arm_compute -{ -namespace tuners -{ -namespace -{ -void tune_gemm_kernel(CLGEMMMatrixMultiplyKernel &k) -{ - cl::NDRange lws_hint = k.lws_hint(); - const GPUTarget gpu_target = k.get_target(); - - switch(gpu_target) - { - case GPUTarget::MIDGARD: - case GPUTarget::T600: - case GPUTarget::T700: - case GPUTarget::T800: - if(k._output->info()->dimension(1) == 196) - { - lws_hint = cl::NDRange(1, 7); - } - else - { - lws_hint = cl::NDRange(8, 8); - } - break; - default: - lws_hint = cl::NullRange; - } - - k.set_lws_hint(lws_hint); -} -} // namespace - -void MidgardTuner::tune_kernel_static(ICLKernel &kernel) -{ - if(dynamic_cast<CLGEMMMatrixMultiplyKernel *>(&kernel) != nullptr) - { - tune_gemm_kernel(*utils::cast::polymorphic_downcast<CLGEMMMatrixMultiplyKernel *>(&kernel)); - } -} - -void MidgardTuner::tune_kernel_dynamic(ICLKernel &kernel) -{ - ARM_COMPUTE_UNUSED(kernel); -} - -void MidgardTuner::tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) -{ - ARM_COMPUTE_UNUSED(kernel, tensors); -} -} // namespace tuners -} // namespace arm_compute |