/* * Copyright (c) 2021-2024 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/cpu/operators/CpuGemmLowpMatrixMultiplyCore.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "arm_compute/runtime/TensorAllocator.h" #include "src/common/utils/Log.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/MemoryHelpers.h" #include "src/cpu/kernels/CpuConvertQuantizedSignednessKernel.h" #include "src/cpu/kernels/CpuGemmInterleave4x4Kernel.h" #include "src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.h" #include "src/cpu/kernels/CpuGemmLowpMatrixReductionKernel.h" #include "src/cpu/kernels/CpuGemmLowpOffsetContributionKernel.h" #include "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.h" #include "src/cpu/kernels/CpuGemmTranspose1xWKernel.h" #include "src/cpu/operators/CpuActivation.h" #include "src/cpu/operators/internal/CpuGemmAssemblyDispatch.h" #include "src/cpu/utils/CpuAuxTensorHandler.h" using namespace arm_compute::misc::shape_calculator; using namespace arm_compute::experimental; namespace arm_compute { namespace cpu { namespace { cpu::AsmGemmInfo init_assembly_metadata(const GEMMInfo &info) { cpu::AsmGemmInfo asm_info; asm_info.method = cpu::AsmConvMethod::Im2Col; asm_info.reinterpret_input_as_3d = info.reinterpret_input_as_3d(); asm_info.depth_output_gemm3d = info.depth_output_gemm3d(); asm_info.activation_info = info.activation_info(); asm_info.output_stage = info.gemmlowp_output_stage(); asm_info.fast_mode = info.fast_math(); asm_info.accumulate = info.accumulate(); return asm_info; } } // namespace CpuGemmLowpMatrixMultiplyCore::CpuGemmLowpMatrixMultiplyCore() : _asm_glue(std::make_unique()), _mm_kernel(), _mtx_a_reshape_kernel(), _mtx_b_reshape_kernel(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _offset_contribution_kernel(), _offset_contribution_output_stage_kernel(), _activation_func(), _convert_to_signed_asymm(), _convert_from_signed_asymm(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _mm_result_s32(), _signed_a(), _signed_output(), _a_offset(0), _b_offset(0), _run_vector_matrix_multiplication(false), _assembly_path(false), _fused_assembly_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false), _fuse_output_stage(false), _run_activation(false), _flip_signedness(false), _gemm_info(), _aux_mem(Count) { } CpuGemmLowpMatrixMultiplyCore::~CpuGemmLowpMatrixMultiplyCore() = default; void CpuGemmLowpMatrixMultiplyCore::configure( const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *dst, const GEMMInfo &gemm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, dst); ARM_COMPUTE_ERROR_THROW_ON(CpuGemmLowpMatrixMultiplyCore::validate(a, b, c, dst, gemm_info)); ARM_COMPUTE_LOG_PARAMS(a, b, c, dst, gemm_info); const ITensorInfo *matrix_a = a; const ITensorInfo *matrix_b = b; GEMMInfo info = gemm_info; // Set internal variables _a_offset = a->quantization_info().uniform().offset; _b_offset = b->quantization_info().uniform().offset; _run_vector_matrix_multiplication = a->dimension(1) < 2; _reshape_b_only_on_first_run = b->are_values_constant(); _is_prepared = false; _fused_assembly_path = false; _flip_signedness = is_data_type_quantized_per_channel(b->data_type()) && (a->data_type() == DataType::QASYMM8) && _reshape_b_only_on_first_run; _gemm_info = gemm_info; // Offset kernel is need if offset is non-zero or it may change (i.e. dynamic). // It is not needed if the datatype is symmetric, because there is no offset bool a_offset_kernel_needed = _a_offset != 0 || a->quantization_info().is_dynamic(); bool b_offset_kernel_needed = _b_offset != 0 || b->quantization_info().is_dynamic(); _asm_glue = std::make_unique(); const ITensorInfo *a_to_use = a; // Convert to QASYMM8 -> QASYMM8_SIGNED and back if (_flip_signedness) { const int32_t offset_correction = 128; const DataType dt = DataType::QASYMM8_SIGNED; const UniformQuantizationInfo iqinfo = a_to_use->quantization_info().uniform(); _signed_a = a_to_use->clone()->set_data_type(dt).set_quantization_info( QuantizationInfo(iqinfo.scale, iqinfo.offset + offset_correction)); _convert_to_signed_asymm = std::make_unique(); _convert_to_signed_asymm->configure(a_to_use, &_signed_a); a_to_use = &_signed_a; _a_offset = _signed_a.quantization_info().uniform().offset; const UniformQuantizationInfo oqinfo = dst->quantization_info().uniform(); _signed_output = dst->clone()->set_data_type(dt).set_quantization_info( QuantizationInfo(oqinfo.scale, oqinfo.offset - offset_correction)); // Output stage correction GEMMLowpOutputStageInfo output_stage_corr = info.gemmlowp_output_stage(); output_stage_corr.gemmlowp_offset = _signed_output.quantization_info().uniform().offset; output_stage_corr.gemmlowp_min_bound -= offset_correction; output_stage_corr.gemmlowp_max_bound -= offset_correction; info.set_gemmlowp_output_stage(output_stage_corr); // Update matrix a matrix_a = &_signed_a; } // If GEMMLowpOutputStage != NONE, fuse the offset contribution with the output stage if (info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE) { _fuse_output_stage = true; _mm_result_s32 = TensorInfo(dst->tensor_shape(), 1, DataType::S32); } // Initialize assembly kernel meta-data const cpu::AsmGemmInfo asm_info = init_assembly_metadata(gemm_info); #ifdef __aarch64__ if (!(!b->are_values_constant() && b->tensor_shape().z() > 1)) // Disable batch matmul as optimized GeMM handles batching differently. { switch (a->data_type()) { case DataType::QASYMM8: case DataType::QASYMM8_SIGNED: case DataType::U8: case DataType::S8: { if (is_data_type_quantized_asymmetric(a_to_use->data_type()) && info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { auto c_info_to_use = c == nullptr ? nullptr : c; _asm_glue->configure(a_to_use, b, c_info_to_use, dst, asm_info); _fused_assembly_path = _asm_glue->is_configured(); } else { auto output_to_use = (_fuse_output_stage ? &_mm_result_s32 : dst); _asm_glue->configure(a_to_use, b, nullptr, output_to_use, asm_info); } _assembly_path = _asm_glue->is_configured(); break; } default: { ARM_COMPUTE_ERROR("Datatype not supported"); break; } } } #endif /* __aarch64__ */ if (!(_assembly_path || _run_vector_matrix_multiplication)) { matrix_a = &_tmp_a; matrix_b = &_tmp_b; // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] _tmp_a = TensorInfo(compute_interleaved_shape(*a_to_use), 1, a_to_use->data_type(), a_to_use->quantization_info()); // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] _tmp_b = TensorInfo(compute_transpose1xW_shape(*b), 1, b->data_type(), b->quantization_info()); // Configure interleave kernel _mtx_a_reshape_kernel = std::make_unique(); _mtx_a_reshape_kernel->configure(a_to_use, &_tmp_a); // Configure transpose kernel _mtx_b_reshape_kernel = std::make_unique(); _mtx_b_reshape_kernel->configure(b, &_tmp_b); } if (!_fused_assembly_path) { // Build reduction info const GEMMLowpReductionKernelInfo reduction_info(a_to_use->dimension(0), false, 0, false); if (a_offset_kernel_needed) { _vector_sum_col = TensorInfo(compute_reductionA_shape(*b), 1, DataType::S32); // Configure Matrix B reduction kernel _mtx_b_reduction_kernel = std::make_unique(); _mtx_b_reduction_kernel->configure(b, &_vector_sum_col, reduction_info); } if (b_offset_kernel_needed) { _vector_sum_row = TensorInfo(compute_reductionB_shape(*a_to_use), 1, DataType::S32); // Configure matrix A reduction kernel _mtx_a_reduction_kernel = std::make_unique(); _mtx_a_reduction_kernel->configure(a_to_use, &_vector_sum_row, reduction_info); } if (_fuse_output_stage) { // Configure matrix multiply kernel if (!_assembly_path) { _mm_kernel = std::make_unique(); _mm_kernel->configure(matrix_a, matrix_b, &_mm_result_s32); } _offset_contribution_output_stage_kernel = std::make_unique(); _offset_contribution_output_stage_kernel->configure( &_mm_result_s32, a_offset_kernel_needed ? &_vector_sum_col : nullptr, b_offset_kernel_needed ? &_vector_sum_row : nullptr, c, _flip_signedness ? &_signed_output : dst, a->dimension(0), _a_offset, _b_offset, info.gemmlowp_output_stage()); if (_flip_signedness) { _convert_from_signed_asymm = std::make_unique(); _convert_from_signed_asymm->configure(&_signed_output, dst); } } else { // This scale is needed for the s8_f32 kernel where the multiplication output is dequantized to F32. const float dequantize_scale = (dst->data_type() == DataType::F32) ? a->quantization_info().uniform().scale * b->quantization_info().uniform().scale : 1.0f; // Configure matrix multiply kernel if (!_assembly_path) { _mm_kernel = std::make_unique(); _mm_kernel->configure(matrix_a, matrix_b, dst); } // Configure offset contribution kernel _offset_contribution_kernel = std::make_unique(); _offset_contribution_kernel->configure(dst, a_offset_kernel_needed ? &_vector_sum_col : nullptr, b_offset_kernel_needed ? &_vector_sum_row : nullptr, a_to_use->dimension(0), _a_offset, _b_offset, dequantize_scale); } } // Configure activation const ActivationLayerInfo &activation = gemm_info.activation_info(); _run_activation = activation.enabled() && (!_assembly_path || !cpu::CpuGemmAssemblyDispatch::is_activation_supported(activation)); if (_run_activation) { _activation_func = std::make_unique(); _activation_func->configure(dst, nullptr, activation); } if (_assembly_path) { const auto asm_mem_req = _asm_glue->workspace(); for (unsigned int slot = 0; slot < asm_mem_req.size(); ++slot) { _aux_mem[slot] = asm_mem_req[slot]; } } // Request memory for LHS and RHS reshape matrix _aux_mem[VectorSumCol] = MemoryInfo(offset_int_vec(VectorSumCol), !_fused_assembly_path && a_offset_kernel_needed && _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _vector_sum_col.total_size()); _aux_mem[VectorSumRow] = MemoryInfo(offset_int_vec(VectorSumRow), MemoryLifetime::Temporary, _vector_sum_row.total_size()); _aux_mem[TmpA] = MemoryInfo(offset_int_vec(TmpA), MemoryLifetime::Temporary, _tmp_a.total_size()); _aux_mem[TmpB] = MemoryInfo(offset_int_vec(TmpB), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); _aux_mem[MMResultS32] = MemoryInfo(offset_int_vec(MMResultS32), MemoryLifetime::Temporary, _mm_result_s32.total_size()); _aux_mem[SignedA] = MemoryInfo(offset_int_vec(SignedA), MemoryLifetime::Temporary, _signed_a.total_size()); _aux_mem[SignedOutput] = MemoryInfo(offset_int_vec(SignedOutput), MemoryLifetime::Temporary, _signed_output.total_size()); } Status CpuGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info) { 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_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(c != nullptr && output->data_type() != DataType::F32 && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::NONE, "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore for output S32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG( (a)->dimension(0) != (b)->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); 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"); // When using accumulation(in place summation), for now, the only supported DataType for output is S32. if (gemm_info.accumulate()) { #ifdef __arm__ ARM_COMPUTE_RETURN_ERROR_MSG("Accumulation is not supported for armv7"); #endif /* __arm__ */ ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE, "Accumulation is not supported for output QASYMM8/QASYMM8_SIGNED"); } GEMMInfo info = gemm_info; const ITensorInfo *matrix_a_info = a; const ITensorInfo *matrix_b_info = b; const ITensorInfo *a_to_use = a; TensorInfo tmp_a_info{}; TensorInfo tmp_b_info{}; TensorInfo mm_result_s32_info{}; int32_t a_offset = a->quantization_info().uniform().offset; int32_t b_offset = b->quantization_info().uniform().offset; // Offset kernel is need if offset is non-zero or it may change (i.e. dynamic). bool a_offset_kernel_needed = a_offset != 0 || a->quantization_info().is_dynamic(); bool b_offset_kernel_needed = b_offset != 0 || b->quantization_info().is_dynamic(); bool fuse_output_stage = info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE; if (fuse_output_stage) { auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(output->tensor_shape()).set_data_type(DataType::S32)); } // Convert QASYMM8->QASYMM8_SIGNED TensorInfo signed_a{}; TensorInfo signed_output{}; bool flip_signedness = is_data_type_quantized_per_channel(b->data_type()) && (a->data_type() == DataType::QASYMM8) && info.reshape_b_only_on_first_run(); if (flip_signedness) { const int32_t offset_correction = 128; const DataType dt = DataType::QASYMM8_SIGNED; const UniformQuantizationInfo iqinfo = a_to_use->quantization_info().uniform(); signed_a = a_to_use->clone()->set_data_type(dt).set_quantization_info( QuantizationInfo(iqinfo.scale, iqinfo.offset + offset_correction)); ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuConvertQuantizedSignednessKernel::validate(a_to_use, &signed_a)); a_to_use = &signed_a; a_offset = signed_a.quantization_info().uniform().offset; const UniformQuantizationInfo oqinfo = output->quantization_info().uniform(); signed_output = output->clone()->set_data_type(dt).set_quantization_info( QuantizationInfo(oqinfo.scale, oqinfo.offset - offset_correction)); // Output stage correction GEMMLowpOutputStageInfo output_stage_corr = info.gemmlowp_output_stage(); output_stage_corr.gemmlowp_offset = signed_output.quantization_info().uniform().offset; output_stage_corr.gemmlowp_min_bound -= offset_correction; output_stage_corr.gemmlowp_max_bound -= offset_correction; info.set_gemmlowp_output_stage(output_stage_corr); // Update matrix a matrix_a_info = &signed_a; } // Initialize assembly kernel meta-data const AsmGemmInfo asm_info = init_assembly_metadata(info); // Check if we need to run the optimized assembly kernel bool run_optimised = false; bool run_optimised_requantized = false; if (!(!b->are_values_constant() && b->tensor_shape().z() > 1)) // Disable batch matmul as optimized GeMM handles batching differently. { if (is_data_type_quantized_asymmetric(a_to_use->data_type()) && info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { run_optimised = bool(CpuGemmAssemblyDispatch::validate(a_to_use, b, c, output, asm_info)); run_optimised_requantized = run_optimised; } else { run_optimised = bool(CpuGemmAssemblyDispatch::validate( a_to_use, b, nullptr, fuse_output_stage ? &mm_result_s32_info : output, asm_info)); } } if (run_optimised) { ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0)); if (info.depth_output_gemm3d() != 0) { if (info.reinterpret_input_as_3d()) { ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2)); } else { ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2)); } } else { ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); } } else { ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.reinterpret_input_as_3d(), "NEGEMM cannot reinterpret the input tensor as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.depth_output_gemm3d() != 0, "NEGEMM cannot reinterpret the output tensor as 3D"); const bool run_vector_matrix_multiplication = a->dimension(1) < 2; if (!run_vector_matrix_multiplication) { matrix_a_info = &tmp_a_info; matrix_b_info = &tmp_b_info; // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] TensorShape shape_tmp_a = a->tensor_shape(); shape_tmp_a.set(0, a->dimension(0) * 4); shape_tmp_a.set(1, std::ceil(a->dimension(1) / 4.f)); // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] TensorShape shape_tmp_b = b->tensor_shape(); shape_tmp_b.set(0, b->dimension(1) * 16); shape_tmp_b.set(1, std::ceil(b->dimension(0) / 16.f)); // Validate interleave kernel auto_init_if_empty(tmp_a_info, a_to_use->clone()->set_tensor_shape(shape_tmp_a)); auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(shape_tmp_b)); ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuGemmInterleave4x4Kernel::validate(a_to_use, &tmp_a_info)); ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuGemmTranspose1xWKernel::validate(b, &tmp_b_info)); } } if (!run_optimised_requantized) { TensorInfo info_vector_sum_col{}; TensorInfo info_vector_sum_row{}; const GEMMLowpReductionKernelInfo reduction_info(a_to_use->dimension(0), false, 0, false); // Validate matrix B reduction kernel only if _a_offset is not equal to 0 if (a_offset_kernel_needed) { info_vector_sum_col = TensorInfo(compute_reductionA_shape(*b), 1, DataType::S32); // Configure Matrix B reduction kernel ARM_COMPUTE_RETURN_ON_ERROR( kernels::CpuGemmLowpMatrixBReductionKernel::validate(b, &info_vector_sum_col, reduction_info)); } // Validate Matrix A reduction kernel only if _b_offset is not equal to 0 if (b_offset_kernel_needed) { info_vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32); // Configure matrix A reduction kernel ARM_COMPUTE_RETURN_ON_ERROR( kernels::CpuGemmLowpMatrixAReductionKernel::validate(a_to_use, &info_vector_sum_row, reduction_info)); } if (fuse_output_stage) { if (!run_optimised) { ARM_COMPUTE_RETURN_ERROR_ON_MSG( info.reinterpret_input_as_3d(), "CpuGemmLowpMatrixMultiplyKernel cannot reinterpret the input tensor as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG( info.depth_output_gemm3d() != 0, "CpuGemmLowpMatrixMultiplyKernel cannot reinterpret the output tensor as 3D"); ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuGemmLowpMatrixMultiplyKernel::validate( matrix_a_info, matrix_b_info, &mm_result_s32_info)); } // Validate offset contribution kernel ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuGemmLowpOffsetContributionOutputStageKernel::validate( &mm_result_s32_info, a_offset_kernel_needed ? &info_vector_sum_col : nullptr, b_offset_kernel_needed ? &info_vector_sum_row : nullptr, c, flip_signedness ? &signed_output : output, a_offset, b_offset, info.gemmlowp_output_stage())); } else { if (!run_optimised) { ARM_COMPUTE_RETURN_ERROR_ON_MSG( info.reinterpret_input_as_3d(), "CpuGemmLowpMatrixMultiplyKernel cannot reinterpret the input tensor as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG( info.depth_output_gemm3d() != 0, "CpuGemmLowpMatrixMultiplyKernel cannot reinterpret the output tensor as 3D"); ARM_COMPUTE_RETURN_ON_ERROR( kernels::CpuGemmLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output)); } // Validate offset contribution kernel ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuGemmLowpOffsetContributionKernel::validate( output, a_offset_kernel_needed ? &info_vector_sum_col : nullptr, b_offset_kernel_needed ? &info_vector_sum_row : nullptr, a_offset, b_offset)); } } // Validate activation const ActivationLayerInfo &activation = gemm_info.activation_info(); if (activation.enabled()) { ARM_COMPUTE_RETURN_ON_ERROR(CpuActivation::validate(output, nullptr, activation)); } return Status{}; } void CpuGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors) { prepare(tensors); auto a = tensors.get_const_tensor(TensorType::ACL_SRC_0); auto b = tensors.get_const_tensor(TensorType::ACL_SRC_1); auto c = tensors.get_const_tensor(TensorType::ACL_SRC_2); auto dst = tensors.get_tensor(TensorType::ACL_DST); auto a_to_use = a; auto matrix_a = a; auto matrix_b = b; CpuAuxTensorHandler vector_sum_col(offset_int_vec(VectorSumCol), _vector_sum_col, tensors, false); CpuAuxTensorHandler vector_sum_row(offset_int_vec(VectorSumRow), _vector_sum_row, tensors, false); CpuAuxTensorHandler tmp_a(offset_int_vec(TmpA), _tmp_a, tensors, false); CpuAuxTensorHandler tmp_b(offset_int_vec(TmpB), _tmp_b, tensors, true); CpuAuxTensorHandler mm_result_s32(offset_int_vec(MMResultS32), _mm_result_s32, tensors, false); CpuAuxTensorHandler signed_a(offset_int_vec(SignedA), _signed_a, tensors, false); CpuAuxTensorHandler signed_output(offset_int_vec(SignedOutput), _signed_output, tensors, false); const QuantizationInfo a_qinfo = a->info()->quantization_info(); const QuantizationInfo b_qinfo = b->info()->quantization_info(); if (a_qinfo.is_dynamic()) _a_offset = a_qinfo.uniform().offset; if (b_qinfo.is_dynamic()) _b_offset = b_qinfo.uniform().offset; // Convert QASYMM8->QASYMM8_SIGNED if (_flip_signedness) { ITensorPack pack = {{TensorType::ACL_SRC, a}, {TensorType::ACL_DST, signed_a.get()}}; NEScheduler::get().schedule_op(_convert_to_signed_asymm.get(), Window::DimY, _convert_to_signed_asymm->window(), pack); a_to_use = signed_a.get(); matrix_a = signed_a.get(); } // Run GEMM if (_asm_glue->is_configured()) { ITensorPack asm_glue_tensors = tensors; auto output_to_use = (_fuse_output_stage ? mm_result_s32.get() : dst); if (is_data_type_quantized_asymmetric(a_to_use->info()->data_type()) && _gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_0, a_to_use); asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_1, b); asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_2, c); asm_glue_tensors.add_tensor(TensorType::ACL_DST, dst); } else { asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_0, a_to_use); asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_1, b); asm_glue_tensors.add_tensor(TensorType::ACL_DST, output_to_use); } _asm_glue->run(asm_glue_tensors); } else { if (!_run_vector_matrix_multiplication) { matrix_a = tmp_a.get(); matrix_b = tmp_b.get(); // Run interleave kernel ITensorPack pack_a = {{TensorType::ACL_SRC, a_to_use}, {TensorType::ACL_DST, tmp_a.get()}}; NEScheduler::get().schedule_op(_mtx_a_reshape_kernel.get(), Window::DimY, _mtx_a_reshape_kernel->window(), pack_a); if (!_reshape_b_only_on_first_run) { ITensorPack pack_b = {{TensorType::ACL_SRC, b}, {TensorType::ACL_DST, tmp_b.get()}}; // Run transpose kernel NEScheduler::get().schedule_op(_mtx_b_reshape_kernel.get(), Window::DimY, _mtx_b_reshape_kernel->window(), pack_b); } } ITensorPack pack_mm = {{TensorType::ACL_SRC_0, matrix_a}, {TensorType::ACL_SRC_1, matrix_b}}; if (_fuse_output_stage) { pack_mm.add_tensor(TensorType::ACL_DST, mm_result_s32.get()); } else { pack_mm.add_tensor(TensorType::ACL_DST, dst); } NEScheduler::get().schedule_op(_mm_kernel.get(), Window::DimY, _mm_kernel->window(), pack_mm); } if (!_fused_assembly_path) { // Run matrix A reduction kernel only if _b_offset is not equal to 0 if (_b_offset != 0) { ITensorPack pack = {{TensorType::ACL_SRC, a_to_use}, {TensorType::ACL_DST, vector_sum_row.get()}}; NEScheduler::get().schedule_op(_mtx_a_reduction_kernel.get(), Window::DimX, _mtx_a_reduction_kernel->window(), pack); } // 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 pack = {{TensorType::ACL_SRC, b}, {TensorType::ACL_DST, vector_sum_col.get()}}; NEScheduler::get().schedule_op(_mtx_b_reduction_kernel.get(), Window::DimX, _mtx_b_reduction_kernel->window(), pack); } if (_fuse_output_stage) { if (a_qinfo.is_dynamic()) _offset_contribution_output_stage_kernel->set_a_offset(_a_offset); if (b_qinfo.is_dynamic()) _offset_contribution_output_stage_kernel->set_b_offset(_b_offset); ITensorPack pack; pack.add_tensor(TensorType::ACL_SRC_0, mm_result_s32.get()); pack.add_tensor(TensorType::ACL_SRC_1, _a_offset == 0 ? nullptr : vector_sum_col.get()); pack.add_tensor(TensorType::ACL_SRC_2, _b_offset == 0 ? nullptr : vector_sum_row.get()); pack.add_tensor(TensorType::ACL_SRC_3, c); pack.add_tensor(TensorType::ACL_DST, _flip_signedness ? signed_output.get() : dst); // Run offset contribution kernel NEScheduler::get().schedule_op(_offset_contribution_output_stage_kernel.get(), Window::DimY, _offset_contribution_output_stage_kernel->window(), pack); } else { if (a_qinfo.is_dynamic()) _offset_contribution_kernel->set_a_offset(_a_offset); if (b_qinfo.is_dynamic()) _offset_contribution_kernel->set_b_offset(_b_offset); if (a_qinfo.is_dynamic() || b_qinfo.is_dynamic()) { const float dequantize_scale = a_qinfo.uniform().scale * b_qinfo.uniform().scale; _offset_contribution_kernel->set_scale(dequantize_scale); } ITensorPack pack; pack.add_tensor(TensorType::ACL_SRC_0, _a_offset == 0 ? nullptr : vector_sum_col.get()); pack.add_tensor(TensorType::ACL_SRC_1, _b_offset == 0 ? nullptr : vector_sum_row.get()); pack.add_tensor(TensorType::ACL_DST, dst); // Run offset contribution kernel NEScheduler::get().schedule_op(_offset_contribution_kernel.get(), Window::DimY, _offset_contribution_kernel->window(), pack); } } // Convert QASYMM8_SIGNED->QASYMM8 if (!_fused_assembly_path && _fuse_output_stage && _flip_signedness) { ITensorPack pack = {{TensorType::ACL_SRC, signed_output.get()}, {TensorType::ACL_DST, dst}}; NEScheduler::get().schedule_op(_convert_from_signed_asymm.get(), Window::DimY, _convert_from_signed_asymm->window(), pack); } // Run fused activation unless already run in the fused assembly if (_run_activation) { ITensorPack pack = {{TensorType::ACL_SRC, dst}, {TensorType::ACL_DST, dst}}; _activation_func->run(pack); } } void CpuGemmLowpMatrixMultiplyCore::prepare(ITensorPack &tensors) { if (!_is_prepared) { auto original_b = tensors.get_const_tensor(TensorType::ACL_SRC_1); // Run assembly reshape if (_asm_glue->is_configured()) { _asm_glue->prepare(tensors); } // Run non-assembly reshape else if (_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue->is_configured()) { // Run reshape kernel and mark original weights tensor as unused ITensor *tmp_b_p = utils::cast::polymorphic_downcast(tensors.get_tensor(offset_int_vec(TmpB))); CpuAuxTensorHandler tmp_b(_tmp_b, *tmp_b_p); ITensorPack pack = {{TensorType::ACL_SRC, original_b}, {TensorType::ACL_DST, tmp_b.get()}}; NEScheduler::get().schedule_op(_mtx_b_reshape_kernel.get(), Window::DimY, _mtx_b_reshape_kernel->window(), pack); } // Run matrix B reduction kernel only if _a_offset is not equal to 0 if (!_fused_assembly_path && _a_offset != 0 && _reshape_b_only_on_first_run) { ITensor *vector_sum_col_p = utils::cast::polymorphic_downcast(tensors.get_tensor(offset_int_vec(VectorSumCol))); CpuAuxTensorHandler vector_sum_col(_vector_sum_col, *vector_sum_col_p); ITensorPack pack = {{TensorType::ACL_SRC, original_b}, {TensorType::ACL_DST, vector_sum_col.get()}}; NEScheduler::get().schedule_op(_mtx_b_reduction_kernel.get(), Window::DimX, _mtx_b_reduction_kernel->window(), pack); } _is_prepared = true; } } experimental::MemoryRequirements CpuGemmLowpMatrixMultiplyCore::workspace() const { return _aux_mem; } } // namespace cpu } // namespace arm_compute