/* * Copyright (c) 2017-2022 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 "src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.h" #include "arm_compute/core/Log.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/MemoryHelpers.h" #include "src/gpu/cl/kernels/ClCastKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpOffsetContributionKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpOffsetContributionOutputStageKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpReductionKernel.h" #include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" #include "src/gpu/cl/utils/ClAuxTensorHandler.h" #include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h" namespace arm_compute { namespace opencl { using namespace arm_compute::misc::shape_calculator; using namespace arm_compute::cl_gemm; using namespace arm_compute::opencl::kernels; using namespace arm_compute::experimental; namespace { inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type) { switch(kernel_type) { case CLGEMMKernelType::NATIVE: case CLGEMMKernelType::RESHAPED_ONLY_RHS: case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL: { return true; } default: { return false; } } } //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run) { auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run); if(bool(gemm_kernel)) { if(validate_gemm_kernel(gemm_kernel.gemm_type)) { ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); return gemm_kernel.gemm_type; } } gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run); ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); return gemm_kernel.gemm_type; } // Validate lhs_info and rhs_info for native kernel inline bool validate_lhs_rhs_info_native(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const GEMMReshapeInfo &reshape_info) { // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel TensorInfo mm_result_s32_info{}; // Output tensor auto initialization if not yet initialized auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*a, *b, false, reshape_info)).set_data_type(DataType::S32)); // Validate mm kernel // NOTE: Ignore all other parameters (eg. output stage etc.) and only validate lhs and rhs info // NOTE: This assumes: // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in CLGEMMLowpMatrixMultiplyNativeKernel.cpp validate_arguments). // 2. lhs and rhs info does not cause window and padding issues through side effects (in CLGEMMLowpMatrixMultiplyNativeKernel.cpp validate_and_configure_window). if(!bool(ClGemmLowpMatrixMultiplyNativeKernel::validate(a, b, &mm_result_s32_info, lhs_info, rhs_info, reshape_info))) { return false; } return true; } // Automatically select between mlgo (prioritized) and default heuristics for native kernel configs std::pair auto_select_gemm_config_native(auto_heuristics::CommonQuery query, const ITensorInfo *a, const ITensorInfo *b, const GEMMReshapeInfo &reshape_info) { auto config = auto_heuristics::select_mlgo_gemm_config_native(query); if(config) { if(validate_lhs_rhs_info_native(config.lhs_info, config.rhs_info, a, b, reshape_info)) { ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use native config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); return { config.lhs_info, config.rhs_info }; } } config = auto_heuristics::select_default_gemm_config_native(query); ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use native config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); return { config.lhs_info, config.rhs_info }; } // Validate lhs_info and rhs_info for reshaped only rhs kernel inline bool validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, unsigned int m, unsigned int n, unsigned int k, bool reinterpret_input_as_3d, int depth_output_gemm3d) { // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel TensorInfo tmp_b_info{}; // Validate reshape RHS kernel auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info))) { return false; } // Validate mm kernel // NOTE: Ignore all other parameters (eg. depth_output_gemm3d, output stage etc.) and only validate lhs and rhs info // NOTE: This assumes: // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_arguments). // 2. lhs and rhs info does not cause window and padding issues through side effects (in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_and_configure_window). GEMMKernelInfo gemm_kernel_info; gemm_kernel_info.m = m; gemm_kernel_info.n = n; gemm_kernel_info.k = k; gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; gemm_kernel_info.lhs_info = lhs_info; gemm_kernel_info.rhs_info = rhs_info; // Since we ignore the output stage, output data type has to be S32 to pass the validation TensorInfo output_info_copy(*output); output_info_copy.set_data_type(DataType::S32); if(!bool(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, &output_info_copy, gemm_kernel_info))) { return false; } return true; } // Validate lhs_info and rhs_info for reshaped only rhs kernel inline bool validate_lhs_rhs_info_reshaped_only_rhs_mmul(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, unsigned int m, unsigned int n, unsigned int k, bool reinterpret_input_as_3d, int depth_output_gemm3d) { // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel TensorInfo tmp_b_info{}; // Validate reshape RHS kernel auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info))) { return false; } // Validate mm kernel // NOTE: Ignore all other parameters (eg. depth_output_gemm3d, output stage etc.) and only validate lhs and rhs info // NOTE: This assumes: // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_arguments). // 2. lhs and rhs info does not cause window and padding issues through side effects (in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_and_configure_window). GEMMKernelInfo gemm_kernel_info; gemm_kernel_info.m = m; gemm_kernel_info.n = n; gemm_kernel_info.k = k; gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; gemm_kernel_info.lhs_info = lhs_info; gemm_kernel_info.rhs_info = rhs_info; // Since we ignore the output stage, output data type has to be S32 to pass the validation TensorInfo output_info_copy(*output); output_info_copy.set_data_type(DataType::S32); if(!bool(ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel::validate(a, &tmp_b_info, &output_info_copy, gemm_kernel_info))) { return false; } return true; } // Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs std::pair auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, bool reinterpret_input_as_3d, int depth_output_gemm3d, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output) { auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query); if(config) { if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, output, query.m, query.n, query.k, reinterpret_input_as_3d, depth_output_gemm3d)) { ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); return { config.lhs_info, config.rhs_info }; } } config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query); ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); return { config.lhs_info, config.rhs_info }; } // Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs std::pair auto_select_gemm_config_reshaped_only_rhs_mmul(auto_heuristics::CommonQuery query, bool reinterpret_input_as_3d, int depth_output_gemm3d, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output) { ARM_COMPUTE_UNUSED(a, b, output, reinterpret_input_as_3d, depth_output_gemm3d); auto config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query); validate_lhs_rhs_info_reshaped_only_rhs_mmul(config.lhs_info, config.rhs_info, a, b, output, query.m, query.n, query.k, reinterpret_input_as_3d, depth_output_gemm3d); ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs_mmul config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); return { config.lhs_info, config.rhs_info }; } inline bool is_gemm_reshaped(CLGEMMKernelType kernel_type) { switch(kernel_type) { case CLGEMMKernelType::NATIVE: return false; case CLGEMMKernelType::RESHAPED_ONLY_RHS: case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL: return true; default: ARM_COMPUTE_ERROR("Not supported gemmlowp kernel!"); } } } // namespace ClGemmLowpMatrixMultiplyCore::ClGemmLowpMatrixMultiplyCore() : _weights_to_qasymm8(std::make_unique()), _mm_native_kernel(std::make_unique()), _mm_reshaped_only_rhs_kernel(std::make_unique()), _mm_reshaped_only_rhs_mmul_kernel(std::make_unique()), _mtx_b_reshape_kernel(std::make_unique()), _mtx_a_reduction_kernel(std::make_unique()), _mtx_b_reduction_kernel(std::make_unique()), _offset_contribution_kernel(std::make_unique()), _offset_contribution_output_stage_kernel(std::make_unique()), _aux_mem(AuxTensorIdx::Count) { } ClGemmLowpMatrixMultiplyCore::~ClGemmLowpMatrixMultiplyCore() = default; void ClGemmLowpMatrixMultiplyCore::configure(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, const GEMMInfo &gemm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); ARM_COMPUTE_ERROR_THROW_ON(ClGemmLowpMatrixMultiplyCore::validate(a, b, c, output, gemm_info)); ARM_COMPUTE_LOG_PARAMS(a, b, c, output, gemm_info); _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); _a_offset = a->quantization_info().uniform().offset; _convert_to_qasymm8 = is_data_type_quantized_per_channel(b->data_type()) && is_data_type_quantized_symmetric(b->data_type()) && a->data_type() == DataType::QASYMM8; _b_offset = _convert_to_qasymm8 ? -128 : b->quantization_info().uniform().offset; _gemm_info = gemm_info; // Get the GPU target const GPUTarget gpu_target = CLScheduler::get().target(); // Set the target for the kernels _mm_native_kernel->set_target(gpu_target); _mm_reshaped_only_rhs_kernel->set_target(gpu_target); _mm_reshaped_only_rhs_mmul_kernel->set_target(gpu_target); GEMMRHSMatrixInfo rhs_info; GEMMLHSMatrixInfo lhs_info; // Arguments used by GEMMReshapeInfo // in order to know how the matrices have been reshaped bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); const unsigned int n = b->dimension(0); const unsigned int k = a->dimension(0); const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); const auto reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d); _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ gpu_target, a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run); if(_convert_to_qasymm8) { // Set data type for converted weights _qasymm8_weights = *b; _qasymm8_weights.set_data_type(DataType::QASYMM8); _weights_to_qasymm8->configure(compile_context, b, &_qasymm8_weights, ConvertPolicy::WRAP); } ITensorInfo *matrix_b = _convert_to_qasymm8 ? &_qasymm8_weights : b; if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS) { matrix_b = &_tmp_b; // Pick up the GEMM configuration // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, reinterpret_input_as_3d, depth_output_gemm3d, a, _convert_to_qasymm8 ? &_qasymm8_weights : b, output); // Configure reshape RHS kernel _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_tmp_b, rhs_info); } if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL) { matrix_b = &_tmp_b; // Pick up the GEMM configuration // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs_mmul(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, reinterpret_input_as_3d, depth_output_gemm3d, a, _convert_to_qasymm8 ? &_qasymm8_weights : b, output); // Configure reshape RHS kernel _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_tmp_b, rhs_info); } // Using default reduction info const GEMMLowpReductionKernelInfo reduction_info {}; // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 if(_a_offset != 0) { _vector_sum_col = TensorInfo(compute_reductionA_shape(*b), 1, DataType::S32); // Configure Matrix B reduction kernel _mtx_b_reduction_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_vector_sum_col, reduction_info); } // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 if(_b_offset != 0) { _vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32); // Configure matrix A reduction kernel _mtx_a_reduction_kernel->configure(compile_context, a, &_vector_sum_row, reduction_info); } GEMMKernelInfo gemm_kernel_info; gemm_kernel_info.m = m; gemm_kernel_info.n = n; gemm_kernel_info.k = k; gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; gemm_kernel_info.lhs_info = lhs_info; gemm_kernel_info.rhs_info = rhs_info; gemm_kernel_info.a_offset = _a_offset; gemm_kernel_info.b_offset = _b_offset; // If GEMMLowpOutputStage != NONE, fuse the offset contribution with the output stage if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE) { // Configure offset contribution kernel const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1; _gemm_output_stage_multipliers = TensorInfo(TensorShape(num_filters), 1, DataType::S32); _gemm_output_stage_shifts = TensorInfo(TensorShape(num_filters), 1, DataType::S32); GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage(); gemmlowp_output_stage.output_data_type = a->data_type(); if(num_filters == 1) { // Per-channel quantization with OFM == 1 is equivalent to uniform quantization. // Setting this flag to false prevents the kernel from adding useless padding to the output multipliers and shifts gemmlowp_output_stage.is_quantized_per_channel = false; } gemm_kernel_info.output_stage = gemmlowp_output_stage; if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS && gemmlowp_output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { // Configure and tune matrix multiply kernel with fused output stage _mm_reshaped_only_rhs_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c != nullptr ? c : nullptr, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts); } else if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL && gemmlowp_output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { // Configure and tune matrix multiply kernel with fused output stage _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c != nullptr ? c : nullptr, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts); } else { _run_output_stage = true; if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS) { _mm_reshaped_only_rhs_kernel->configure(compile_context, a, matrix_b, &_mm_result_s32, gemm_kernel_info); } if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL) { _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, matrix_b, &_mm_result_s32, gemm_kernel_info); } else { // Pick up the GEMM configuration // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration std::tie(lhs_info, rhs_info) = auto_select_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, a, _convert_to_qasymm8 ? &_qasymm8_weights : matrix_b, reshape_info); // Configure matrix multiply kernel _mm_native_kernel->configure(compile_context, a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, reshape_info); _offset_contribution_output_stage_kernel->configure(compile_context, &_mm_result_s32, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c != nullptr ? c : nullptr, output, a->dimension(0), _a_offset, _b_offset, gemmlowp_output_stage, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts); } } } else { _run_offset_contribution = true; if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS) { // Configure and tune matrix multiply kernel _mm_reshaped_only_rhs_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info); } else if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL) { // Configure and tune matrix multiply kernel _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info); } else { // Pick up the GEMM configuration // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration std::tie(lhs_info, rhs_info) = auto_select_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, a, _convert_to_qasymm8 ? &_qasymm8_weights : b, reshape_info); // Configure matrix multiply kernel _mm_native_kernel->configure(compile_context, a, matrix_b, output, lhs_info, rhs_info, reshape_info); } // Configure offset contribution kernel _offset_contribution_kernel->configure(compile_context, output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c != nullptr ? c : nullptr, a->dimension(0), _a_offset, _b_offset); } // Request memory _aux_mem[RhsQAsymm8] = MemoryInfo(offset_int_vec(RhsQAsymm8), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _qasymm8_weights.total_size()); if(is_gemm_reshaped(_gemm_kernel_type)) { // Overwrite Rhs as prepare if gemm is reshaped as there will be a two-step transformation _aux_mem[RhsQAsymm8] = MemoryInfo(offset_int_vec(RhsQAsymm8), _reshape_b_only_on_first_run ? MemoryLifetime::Prepare : MemoryLifetime::Temporary, _qasymm8_weights.total_size()); _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); } if(_a_offset != 0) { _aux_mem[VecSumCol] = MemoryInfo(offset_int_vec(VecSumCol), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _vector_sum_col.total_size()); } if(_b_offset != 0) { _aux_mem[VecSumRow] = MemoryInfo(offset_int_vec(VecSumRow), MemoryLifetime::Temporary, _vector_sum_row.total_size()); } _aux_mem[ResultS32] = MemoryInfo(offset_int_vec(ResultS32), MemoryLifetime::Temporary, _mm_result_s32.total_size()); _aux_mem[Multipliers] = MemoryInfo(offset_int_vec(Multipliers), MemoryLifetime::Persistent, _gemm_output_stage_multipliers.total_size()); _aux_mem[Shifts] = MemoryInfo(offset_int_vec(Shifts), MemoryLifetime::Persistent, _gemm_output_stage_shifts.total_size()); } Status ClGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8, DataType::QSYMM8_PER_CHANNEL); ARM_COMPUTE_RETURN_ERROR_ON(a->data_type() == DataType::QASYMM8 && b->data_type() == DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON(a->data_type() == DataType::QASYMM8_SIGNED && b->data_type() == DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); int32_t a_offset = a->quantization_info().uniform().offset; int32_t b_offset = b->quantization_info().uniform().offset; const ITensorInfo *matrix_a_info = a; TensorInfo tmp_b_info{}; GEMMRHSMatrixInfo rhs_info; GEMMLHSMatrixInfo lhs_info; // Get the GPU target const GPUTarget gpu_target = CLScheduler::get().target(); bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); const unsigned int n = b->dimension(0); const unsigned int k = a->dimension(0); const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); bool reshape_matrix_b = is_gemm_reshaped(auto_select_gemm_kernel(auto_heuristics::CommonQuery{ gpu_target, a->data_type(), m, n, k, batch_size }, gemm_info.reshape_b_only_on_first_run())); const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d); bool convert_to_qasymm8 = is_data_type_quantized_per_channel(b->data_type()) && is_data_type_quantized_symmetric(b->data_type()) && is_data_type_quantized_asymmetric(a->data_type()); TensorInfo weights_info(*b); if(convert_to_qasymm8) { b_offset = -128; weights_info.set_data_type(DataType::QASYMM8); ARM_COMPUTE_RETURN_ON_ERROR(ClCastKernel::validate(b, &weights_info, ConvertPolicy::WRAP)); } const ITensorInfo *matrix_b_info = &weights_info; if(reshape_matrix_b) { matrix_b_info = &tmp_b_info; // Pick up the GEMM configuration // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration const auto res = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }); lhs_info = res.lhs_info; rhs_info = res.rhs_info; // Validate reshape RHS kernel auto_init_if_empty(tmp_b_info, weights_info.clone()->set_tensor_shape(compute_rhs_reshaped_shape(weights_info, rhs_info))); ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(&weights_info, &tmp_b_info, rhs_info)); } TensorInfo info_vector_sum_col{}; TensorInfo info_vector_sum_row{}; const GEMMLowpReductionKernelInfo reduction_info; // Validate matrix B reduction kernel only if _a_offset is not equal to 0 if(a_offset != 0) { info_vector_sum_col = TensorInfo(compute_reductionA_shape(weights_info), 1, DataType::S32); // Configure Matrix B reduction kernel ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixBReductionKernel::validate(&weights_info, &info_vector_sum_col, reduction_info)); } // Validate Matrix A reduction kernel only if _b_offset is not equal to 0 if(b_offset != 0) { info_vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32); // Configure matrix A reduction kernel ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row, reduction_info)); } GEMMKernelInfo gemm_kernel_info; gemm_kernel_info.m = m; gemm_kernel_info.n = n; gemm_kernel_info.k = k; gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; gemm_kernel_info.lhs_info = lhs_info; gemm_kernel_info.rhs_info = rhs_info; gemm_kernel_info.a_offset = a_offset; gemm_kernel_info.b_offset = b_offset; if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE) { const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1; const TensorInfo gemm_output_stage_multipliers_shifts_info(TensorInfo(TensorShape(num_filters), 1, DataType::S32)); GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage(); gemmlowp_output_stage.output_data_type = a->data_type(); gemm_kernel_info.output_stage = gemmlowp_output_stage; if(reshape_matrix_b && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info, a_offset == 0 ? nullptr : &info_vector_sum_col, b_offset == 0 ? nullptr : &info_vector_sum_row, c, &gemm_output_stage_multipliers_shifts_info, &gemm_output_stage_multipliers_shifts_info)); } else { TensorInfo mm_result_s32_info{}; if(reshape_matrix_b) { // Output tensor auto inizialitation if not yet initialized auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_info)).set_data_type(DataType::S32)); // Validate matrix multiply ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, gemm_kernel_info)); } else { // Output tensor auto inizialitation if not yet initialized auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, false, reshape_info)).set_data_type(DataType::S32)); // Pick up the GEMM configuration // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration const auto res = select_default_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }); lhs_info = res.lhs_info; rhs_info = res.rhs_info; // Validate matrix multiply ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info)); } // Validate offset contribution kernel ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpOffsetContributionOutputStageKernel::validate(&mm_result_s32_info, a_offset == 0 ? nullptr : &info_vector_sum_col, b_offset == 0 ? nullptr : &info_vector_sum_row, c, output, a_offset, b_offset, gemmlowp_output_stage, &gemm_output_stage_multipliers_shifts_info, &gemm_output_stage_multipliers_shifts_info)); } } else { if(reshape_matrix_b) { // Validate matrix multiply ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info)); } else { // Pick up the GEMM configuration // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration const auto res = select_default_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }); lhs_info = res.lhs_info; rhs_info = res.rhs_info; // Validate matrix multiply ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info)); } if(output->total_size() != 0) { // Validate offset contribution kernel ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpOffsetContributionKernel::validate(output, a_offset == 0 ? nullptr : &info_vector_sum_col, b_offset == 0 ? nullptr : &info_vector_sum_row, c, a_offset, b_offset)); } } return Status{}; } void ClGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors) { const ITensor *a = tensors.get_const_tensor(ACL_SRC_0); const ITensor *b = tensors.get_const_tensor(ACL_SRC_1); const ITensor *c = tensors.get_const_tensor(ACL_SRC_2); ITensor *dst = tensors.get_tensor(ACL_DST); ARM_COMPUTE_ERROR_ON_NULLPTR(a, dst); CLAuxTensorHandler vec_sum_col(offset_int_vec(VecSumCol), _vector_sum_col, tensors, true); CLAuxTensorHandler vec_sum_row(offset_int_vec(VecSumRow), _vector_sum_row, tensors, true); CLAuxTensorHandler rhs_qasymm8(offset_int_vec(RhsQAsymm8), _qasymm8_weights, tensors, true); CLAuxTensorHandler tmp_b(offset_int_vec(RhsReshape), _tmp_b, tensors, true); CLAuxTensorHandler res32(offset_int_vec(ResultS32), _mm_result_s32, tensors, true); CLAuxTensorHandler shifts(offset_int_vec(Shifts), _gemm_output_stage_shifts, tensors, true); CLAuxTensorHandler multipliers(offset_int_vec(Multipliers), _gemm_output_stage_multipliers, tensors, true); // Prepare the consts if needed prepare(tensors); const ITensor *matrix_a = a; const ITensor *matrix_b = _convert_to_qasymm8 ? rhs_qasymm8.get() : b; if(is_gemm_reshaped(_gemm_kernel_type)) { matrix_b = tmp_b.get(); if(!_reshape_b_only_on_first_run) { // Run reshape matrix B ITensorPack mtx_b_reshape_pack = { { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b }, { TensorType::ACL_DST, tmp_b.get() } }; CLScheduler::get().enqueue_op(*_mtx_b_reshape_kernel, mtx_b_reshape_pack, false); } } // Run matrix B reduction kernel only if _a_offset is not equal to 0 if(_a_offset != 0 && !_reshape_b_only_on_first_run) { ITensorPack mtx_b_red_pack = { { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b }, { TensorType::ACL_DST, vec_sum_col.get() } }; CLScheduler::get().enqueue_op(*_mtx_b_reduction_kernel, mtx_b_red_pack, false); } // Run matrix A reduction kernel only if _b_offset is not equal to 0 if(_b_offset != 0) { ITensorPack mtx_a_red_pack = { { TensorType::ACL_SRC, matrix_a }, { TensorType::ACL_DST, vec_sum_row.get() } }; CLScheduler::get().enqueue_op(*_mtx_a_reduction_kernel, mtx_a_red_pack, false); } // Run matrix multiply if(is_gemm_reshaped(_gemm_kernel_type)) { ITensorPack gemm_reshaped_pack; if(_run_offset_contribution) { gemm_reshaped_pack = ITensorPack({ { TensorType::ACL_SRC_0, matrix_a }, { TensorType::ACL_SRC_1, matrix_b }, { TensorType::ACL_DST, _run_output_stage ? res32.get() : dst } }); } else { gemm_reshaped_pack = ITensorPack( { { TensorType::ACL_SRC, matrix_a }, { TensorType::ACL_SRC_1, matrix_b }, { TensorType::ACL_BIAS, c }, { TensorType::ACL_VEC_ROW_SUM, _b_offset == 0 ? nullptr : vec_sum_row.get() }, { TensorType::ACL_VEC_COL_SUM, _a_offset == 0 ? nullptr : vec_sum_col.get() }, { TensorType::ACL_SHIFTS, shifts.get() }, { TensorType::ACL_MULTIPLIERS, multipliers.get() }, { TensorType::ACL_DST, dst }, }); } if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS) { CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_kernel, gemm_reshaped_pack, false); } else if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL) { CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_mmul_kernel, gemm_reshaped_pack, false); } else { ARM_COMPUTE_ERROR("Invalid reshaped kernel"); } } else { ITensorPack gemm_native_pack = { { TensorType::ACL_SRC_0, matrix_a }, { TensorType::ACL_SRC_1, matrix_b }, { TensorType::ACL_DST, _run_offset_contribution ? dst : res32.get() } }; CLScheduler::get().enqueue_op(*_mm_native_kernel, gemm_native_pack, false); } if(_run_output_stage) { // Run offset contribution/output stage kernel ITensorPack output_stage_pack = { { TensorType::ACL_SRC, res32.get() }, { TensorType::ACL_BIAS, c }, { TensorType::ACL_VEC_ROW_SUM, _b_offset == 0 ? nullptr : vec_sum_row.get() }, { TensorType::ACL_VEC_COL_SUM, _a_offset == 0 ? nullptr : vec_sum_col.get() }, { TensorType::ACL_SHIFTS, shifts.get() }, { TensorType::ACL_MULTIPLIERS, multipliers.get() }, { TensorType::ACL_DST, dst }, }; CLScheduler::get().enqueue_op(*_offset_contribution_output_stage_kernel, output_stage_pack, true); } if(_run_offset_contribution) { // Run offset contribution kernel ITensorPack offset_contrib_pack = { { TensorType::ACL_SRC_DST, dst }, { TensorType::ACL_BIAS, c }, { TensorType::ACL_VEC_ROW_SUM, _b_offset == 0 ? nullptr : vec_sum_row.get() }, { TensorType::ACL_VEC_COL_SUM, _a_offset == 0 ? nullptr : vec_sum_col.get() } }; CLScheduler::get().enqueue_op(*_offset_contribution_kernel, offset_contrib_pack, true); } } void ClGemmLowpMatrixMultiplyCore::prepare(ITensorPack &tensors) { if(!_is_prepared) { auto b = tensors.get_const_tensor(TensorType::ACL_SRC_1); CLAuxTensorHandler tmp_b(offset_int_vec(RhsReshape), _tmp_b, tensors, true); CLAuxTensorHandler vec_sum_col(offset_int_vec(VecSumCol), _vector_sum_col, tensors, true); CLAuxTensorHandler rhs_qasymm8(offset_int_vec(RhsQAsymm8), _qasymm8_weights, tensors, false); ARM_COMPUTE_ERROR_ON_NULLPTR(b); if(_convert_to_qasymm8) { ITensorPack convert_to_qs8_pack = { { ACL_SRC, b }, { ACL_DST, rhs_qasymm8.get() } }; CLScheduler::get().enqueue_op(*_weights_to_qasymm8, convert_to_qs8_pack, false); b->mark_as_unused(); } if(is_gemm_reshaped(_gemm_kernel_type) && _reshape_b_only_on_first_run) { // Run reshape kernel and mark original weights tensor as unused ITensorPack mtx_b_pack = { { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b }, { TensorType::ACL_DST, tmp_b.get() } }; CLScheduler::get().enqueue_op(*_mtx_b_reshape_kernel, mtx_b_pack, false); b->mark_as_unused(); } // Run matrix B reduction kernel only if _a_offset is not equal to 0 if(_a_offset != 0 && _reshape_b_only_on_first_run) { ITensorPack mtx_b_red_pack = { { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b }, { TensorType::ACL_DST, vec_sum_col.get() } }; CLScheduler::get().enqueue_op(*_mtx_b_reduction_kernel, mtx_b_red_pack, false); } // Compute GEMM output multipliers and shifts for output stage { const size_t num_filters = (_gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? _gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1; CLAuxTensorHandler multipliers(offset_int_vec(Multipliers), _gemm_output_stage_multipliers, tensors, false); CLAuxTensorHandler shifts(offset_int_vec(Shifts), _gemm_output_stage_shifts, tensors, false); ICLTensor *multiplier_tensor = multipliers.get(); if(multiplier_tensor != nullptr && multiplier_tensor->info()->total_size() > 0) { multiplier_tensor->map(CLScheduler::get().queue(), true); std::memcpy(multiplier_tensor->ptr_to_element(Coordinates(0)), _gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.data(), num_filters * sizeof(int32_t)); multiplier_tensor->unmap(CLScheduler::get().queue()); } ICLTensor *shifts_tensor = shifts.get(); if(shifts.get() != nullptr && shifts_tensor->info()->total_size() > 0) { shifts_tensor->map(CLScheduler::get().queue(), true); std::memcpy(shifts_tensor->ptr_to_element(Coordinates(0)), _gemm_info.gemmlowp_output_stage().gemmlowp_shifts.data(), num_filters * sizeof(int32_t)); shifts_tensor->unmap(CLScheduler::get().queue()); } } CLScheduler::get().queue().finish(); _is_prepared = true; } } experimental::MemoryRequirements ClGemmLowpMatrixMultiplyCore::workspace() const { return _aux_mem; } } // namespace opencl } // namespace arm_compute