/* * Copyright (c) 2021-2022 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/cpu/operators/CpuWinogradConv2d.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/runtime/FunctionDescriptors.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "src/common/utils/Log.h" #include "src/core/CPP/Validate.h" #include "src/core/NEON/kernels/assembly/winograd.hpp" #include "src/core/NEON/kernels/convolution/common/tensor.hpp" #include "src/core/NEON/kernels/convolution/common/utils.hpp" #include "src/core/helpers/MemoryHelpers.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/utils/AssemblyUtils.h" #include "src/cpu/kernels/CpuWinogradConv2dKernel.h" #include "src/cpu/kernels/assembly/arm_gemm.hpp" #include "src/cpu/operators/CpuActivation.h" #include "src/cpu/operators/CpuPermute.h" #include "src/cpu/utils/CpuAuxTensorHandler.h" #include "support/Cast.h" namespace arm_compute { namespace cpu { using namespace arm_compute::experimental; using namespace arm_compute::utils::cast; namespace { inline Tensor4DShape internal_get_shape(const ITensorInfo *in) { const DataLayout data_layout = in->data_layout(); const int in_width = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)); const int in_height = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)); const int in_channels = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)); const int in_batches = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES)); return Tensor4DShape{ in_batches, in_height, in_width, in_channels }; } Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info) { ARM_COMPUTE_UNUSED(dst, weights); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides."); if(biases != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); return Status{}; } bool get_winograd_kernel_implementation(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math, arm_conv::winograd::WinogradImpl *winograd_impl, std::unique_ptr &conv_args) { arm_conv::winograd::WinogradConfig winograd_cfg; arm_gemm::GemmConfig cfg; const DataType data_type = src->data_type(); Tensor4DShape in_shape{ internal_get_shape(src) }; Tensor4DShape out_shape{ internal_get_shape(dst) }; Tensor4DShape kernel_shape{ internal_get_shape(weights) }; uint32_t nthreads = NEScheduler::get().num_threads(); // Get configuration arguments for Winograd winograd_cfg.output_rows = 0; winograd_cfg.output_cols = 0; conv_args = std::make_unique( in_shape.n_batches, arm_conv::Shape2D{ static_cast(in_shape.n_rows), static_cast(in_shape.n_cols) }, in_shape.n_channels, conv_info.pad_top(), conv_info.pad_left(), arm_conv::Shape2D{ static_cast(out_shape.n_rows), static_cast(out_shape.n_cols) }, out_shape.n_channels, arm_conv::Shape2D{ static_cast(kernel_shape.n_rows), static_cast(kernel_shape.n_cols) }, assembly_utils::map_to_arm_gemm_activation(act_info)); bool success = false; if(data_type == DataType::F32) { success = arm_conv::winograd::get_implementation( *winograd_impl, &CPUInfo::get(), *conv_args, nthreads, enable_fast_math, &winograd_cfg, nullptr); } #if defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) else if(data_type == DataType::F16) { success = arm_conv::winograd::get_implementation<__fp16>( *winograd_impl, &CPUInfo::get(), *conv_args, nthreads, enable_fast_math, &winograd_cfg, nullptr); } #endif // defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) else { success = false; } return success; } inline bool fuse_function_supported(const ActivationLayerInfo &act_info) { return act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU || act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU; } } // namespace CpuWinogradConv2d::CpuWinogradConv2d() : _gemm_function(std::make_unique()), _activation_func(std::make_unique()), _transform_input_kernel(nullptr), _transform_output_kernel(nullptr), _permute_input(std::make_unique()), _permute_output(std::make_unique()), _permute_weights(std::make_unique()), _aux_mem(AuxTensorIdx::Count), _conv_args{ nullptr }, _winograd_impl{}, _data_layout(), _winograd_transformed_input{}, _winograd_transformed_output{}, _winograd_transformed_weights{}, _input_workspace(), _output_workspace(), _weights_hwio(), _input_nhwc(), _output_nhwc(), _is_prepared{ false }, _run_activation{ false } { } CpuWinogradConv2d::~CpuWinogradConv2d() = default; void CpuWinogradConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); ARM_COMPUTE_ERROR_THROW_ON(validate(src, weights, biases, dst, conv_info, act_info, enable_fast_math)); ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, act_info, enable_fast_math); ARM_COMPUTE_UNUSED(biases); const DataType data_type = src->data_type(); uint32_t nthreads = NEScheduler::get().num_threads(); _data_layout = src->data_layout(); const Tensor4DShape kernel_shape{ internal_get_shape(weights) }; bool success = get_winograd_kernel_implementation(src, weights, dst, conv_info, act_info, enable_fast_math, &_winograd_impl, _conv_args); ARM_COMPUTE_EXIT_ON_MSG_VAR(!success, "Unsupported kernel size: %d x %d.\n", kernel_shape.n_rows, kernel_shape.n_cols); ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using input transform: %s\n", _winograd_impl.input_transform->get_name().c_str()); ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using weight transform: %s\n", _winograd_impl.input_transform->get_name().c_str()); ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using output transform: %s\n", _winograd_impl.input_transform->get_name().c_str()); const bool has_impl = ((_winograd_impl.input_transform != nullptr) && (_winograd_impl.output_transform != nullptr) && (_winograd_impl.gemm_args != nullptr)); if(has_impl) { // Determine how much working space is required, allocate it. const size_t input_workspace_size = _winograd_impl.input_transform->get_working_space_size(*_conv_args, nthreads); const size_t output_workspace_size = _winograd_impl.output_transform->get_working_space_size(*_conv_args, nthreads); TensorInfo input_workspace_info(TensorShape(input_workspace_size), 1, DataType::U8); TensorInfo output_workspace_info(TensorShape(output_workspace_size), 1, DataType::U8); _input_workspace = input_workspace_info; _output_workspace = output_workspace_info; const auto &wds = _winograd_impl.winograd_spec; // Preparing winograd transformed input tensor const size_t data_type_size = src->element_size(); const uint32_t m = _winograd_impl.gemm_args->_Msize; // Total number of tiles const uint32_t k = _winograd_impl.gemm_args->_Ksize; // Input channels const uint32_t n = _winograd_impl.gemm_args->_Nsize; // Output channels const uint32_t n_gemms = _winograd_impl.gemm_args->_nmulti; const uint32_t n_batches = _winograd_impl.gemm_args->_nbatches; constexpr size_t storage_alignment = 64; const TensorShape a_shape(k, m, n_batches, n_gemms); Strides a_strides(data_type_size); a_strides.set(1, data_type_size * _winograd_impl.winograd_spec.input_ld_row); a_strides.set(2, data_type_size * _winograd_impl.winograd_spec.input_ld_batch); a_strides.set(3, data_type_size * _winograd_impl.winograd_spec.input_ld_matrix); const TensorShape b_shape(n, k, n_gemms); Strides b_strides(data_type_size); b_strides.set(1, data_type_size * _winograd_impl.winograd_spec.weight_ld_row); b_strides.set(2, data_type_size * _winograd_impl.winograd_spec.weight_ld_matrix); const TensorShape d_shape(n, m, n_batches, n_gemms); Strides d_strides(data_type_size); d_strides.set(1, data_type_size * _winograd_impl.winograd_spec.output_ld_row); d_strides.set(2, data_type_size * _winograd_impl.winograd_spec.output_ld_batch); d_strides.set(3, data_type_size * _winograd_impl.winograd_spec.output_ld_matrix); TensorInfo a_info{}; TensorInfo b_info{}; TensorInfo d_info{}; a_info.init(a_shape, 1, data_type, a_strides, 0, wds.input_matrix_size_bytes); b_info.init(b_shape, 1, data_type, b_strides, 0, wds.weight_matrix_size_bytes); d_info.init(d_shape, 1, data_type, d_strides, 0, wds.output_matrix_size_bytes); _winograd_transformed_input = a_info; _winograd_transformed_weights = b_info; _winograd_transformed_output = d_info; PermutationVector weights_permutation_vector(3U, 0U, 1U, 2U); // Configure the kernel to transform the input tensor from NCHW -> NHWC if(_data_layout == DataLayout::NCHW) { _permute_input->configure(src, &_input_nhwc, PermutationVector(2U, 0U, 1U)); weights_permutation_vector = PermutationVector(3U, 2U, 0U, 1U); } // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] _permute_weights->configure(weights, &_weights_hwio, weights_permutation_vector); // Reorder the convoluted output to ACL's ordering NCHW if(_data_layout == DataLayout::NCHW) { // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() TensorInfo info(TensorShape(dst->dimension(2), dst->dimension(0), dst->dimension(1), dst->dimension(3)), 1, dst->data_type()); _output_nhwc = info; _permute_output->configure(&_output_nhwc, dst, PermutationVector(1U, 2U, 0U)); } // Configure input transform kernel _transform_input_kernel = std::make_unique(_winograd_impl, *_conv_args, nthreads); // Configure GEMM function _gemm_function->configure(&_winograd_transformed_input, &_winograd_transformed_weights, nullptr, &_winograd_transformed_output, 1.0f, 0.f); // Configure output transform kernel _transform_output_kernel = std::make_unique(_winograd_impl, *_conv_args, nthreads); //Configure Activation Layer _run_activation = act_info.enabled() && !fuse_function_supported(act_info); if(_run_activation) { _activation_func->configure(dst, nullptr, act_info); } auto asm_mem_req = _gemm_function->workspace(); _aux_mem[GemmWorkspace] = asm_mem_req[GemmWorkspace]; _aux_mem[Pretranspose] = asm_mem_req[Pretranspose]; _aux_mem[InterleavedLHS] = asm_mem_req[InterleavedLHS]; _aux_mem[TransposedRHS] = asm_mem_req[TransposedRHS]; _aux_mem[TempResult] = asm_mem_req[TempResult]; // Request temporary memory. Overlap memory needed for Input/Output transformations as they run on different non-overlapping time-steps. _aux_mem[TransformedInput] = MemoryInfo(offset_int_vec(TransformedInput), MemoryLifetime::Temporary, wds.input_matrix_size_bytes, storage_alignment); _aux_mem[TransformedOutput] = MemoryInfo(offset_int_vec(TransformedOutput), MemoryLifetime::Temporary, wds.output_matrix_size_bytes, storage_alignment); _aux_mem[WorkspaceIO] = MemoryInfo(offset_int_vec(WorkspaceIO), MemoryLifetime::Temporary, std::max(input_workspace_size, output_workspace_size)); _aux_mem[PermutedWeights] = MemoryInfo(offset_int_vec(PermutedWeights), MemoryLifetime::Prepare, _weights_hwio.total_size()); _aux_mem[TransformedWeights] = MemoryInfo(offset_int_vec(TransformedWeights), MemoryLifetime::Persistent, wds.weight_matrix_size_bytes, storage_alignment); if(_data_layout == DataLayout::NCHW) { _aux_mem[PermutedInput].merge(offset_int_vec(PermutedInput), src->total_size()); _aux_mem[PermutedOutput].merge(offset_int_vec(PermutedOutput), dst->total_size()); } } } Status CpuWinogradConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info)); // Disable winograd for fp16 if fast math is false. if(!enable_fast_math) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); } const Tensor4DShape kernel_shape{ internal_get_shape(weights) }; arm_conv::winograd::WinogradImpl winograd_impl{}; std::unique_ptr conv_args; const bool success = get_winograd_kernel_implementation(src, weights, dst, conv_info, act_info, enable_fast_math, &winograd_impl, conv_args); ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(success == false, "Unsupported kernel size: %d x %d.\n", kernel_shape.n_rows, kernel_shape.n_cols); ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using input transform: %s\n", winograd_impl.input_transform->get_name().c_str()); ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using weight transform: %s\n", winograd_impl.input_transform->get_name().c_str()); ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using output transform: %s\n", winograd_impl.input_transform->get_name().c_str()); return Status{}; } void CpuWinogradConv2d::run(ITensorPack &tensors) { prepare(tensors); auto src = tensors.get_const_tensor(ACL_SRC_0); auto biases = tensors.get_const_tensor(ACL_SRC_2); auto output = tensors.get_tensor(ACL_DST); Window win; const uint32_t nthreads = NEScheduler::get().num_threads(); // The Winograd transform implementation does fine-grain threading inside the transforms. Just pass thread_id and nthreads. win.set(Window::DimX, Window::Dimension(0, nthreads, 1)); // Wrap the winograd-domain tensorInfos created in configuration in tensors and allocate the required memory. CpuAuxTensorHandler input_nhwc(offset_int_vec(PermutedInput), _input_nhwc, tensors, true); CpuAuxTensorHandler winograd_input_transformed(offset_int_vec(TransformedInput), _winograd_transformed_input, tensors, true); CpuAuxTensorHandler input_workspace(offset_int_vec(WorkspaceIO), _input_workspace, tensors, true); const bool is_nchw = _data_layout == DataLayout::NCHW; if(is_nchw) { //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC ITensorPack pack{ { ACL_SRC, src }, { ACL_DST, input_nhwc.get() } }; _permute_input->run(pack); } CpuAuxTensorHandler winograd_output_transformed(offset_int_vec(TransformedOutput), _winograd_transformed_output, tensors, true); CpuAuxTensorHandler output_workspace(offset_int_vec(WorkspaceIO), _output_workspace, tensors, true); CpuAuxTensorHandler output_nhwc(offset_int_vec(PermutedOutput), _output_nhwc, tensors, true); ITensorPack transform_input_pack{ { ACL_SRC, is_nchw ? input_nhwc.get() : src }, { ACL_DST, winograd_input_transformed.get() }, { ACL_INT, input_workspace.get() } }; NEScheduler::get().schedule_op(_transform_input_kernel.get(), Window::DimX, win, transform_input_pack); CpuAuxTensorHandler winograd_weights_transformed(offset_int_vec(TransformedWeights), _winograd_transformed_weights, tensors, true); // Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs ITensorPack gemm_pack = tensors; gemm_pack.add_const_tensor(ACL_SRC, winograd_input_transformed.get()); gemm_pack.add_const_tensor(ACL_SRC_1, winograd_weights_transformed.get()); gemm_pack.add_const_tensor(ACL_BIAS, nullptr); gemm_pack.add_tensor(ACL_DST, winograd_output_transformed.get()); _gemm_function->run(gemm_pack); // Output transform ITensorPack transform_output_pack{ { ACL_SRC_0, winograd_output_transformed.get() }, { ACL_DST, is_nchw ? output_nhwc.get() : output }, { ACL_SRC_1, biases }, { ACL_INT, output_workspace.get() } }; NEScheduler::get().schedule_op(_transform_output_kernel.get(), Window::DimX, win, transform_output_pack); if(is_nchw) { // Reorder the convoluted output to ACL's ordering NCHW ITensorPack pack{ { ACL_SRC, output_nhwc.get() }, { ACL_DST, output } }; _permute_output->run(pack); } if(_run_activation) { ITensorPack pack{ { ACL_SRC, output }, { ACL_DST, output } }; _activation_func->run(pack); } } void CpuWinogradConv2d::prepare(ITensorPack &tensors) { if(!_is_prepared) { const ITensor *weights = tensors.get_const_tensor(ACL_SRC_1); ITensor *weights_aux = utils::cast::polymorphic_cast(tensors.get_tensor(offset_int_vec(PermutedWeights))); CpuAuxTensorHandler permuted_weights(_weights_hwio, *weights_aux); ITensorPack permute_tensors{ { ACL_SRC, weights }, { ACL_DST, permuted_weights.get() } }; _permute_weights->run(permute_tensors); const int element_size_in_bytes = permuted_weights.get()->info()->element_size(); // Weights were in OHWI format, before being permuted "permuted_weights" to be in HWIO format. const unsigned int height_idx = 3; // H in HWIO const unsigned int width_idx = 2; // W in HWIO const unsigned int channel_idx = 1; // I in HWIO const int permuted_weight_row_stride = permuted_weights.get()->info()->strides_in_bytes()[height_idx] / element_size_in_bytes; const int permuted_weight_col_stride = permuted_weights.get()->info()->strides_in_bytes()[width_idx] / element_size_in_bytes; const int permuted_weight_channel_stride = permuted_weights.get()->info()->strides_in_bytes()[channel_idx] / element_size_in_bytes; // Wrap the winograd-domain transformed weight TensorInfo in Auxiliary tensor and allocate the required memory. ITensor *weights_transf = utils::cast::polymorphic_cast(tensors.get_tensor(offset_int_vec(TransformedWeights))); ARM_COMPUTE_ERROR_ON_NULLPTR(weights_transf); CpuAuxTensorHandler winograd_transformed_weights(_winograd_transformed_weights, *weights_transf); const void *permuted_weights_ptr; void *win_wght_transf_ptr; permuted_weights_ptr = reinterpret_cast(permuted_weights.get()->buffer() + permuted_weights.get()->info()->offset_first_element_in_bytes()); win_wght_transf_ptr = reinterpret_cast(winograd_transformed_weights.get()->buffer() + winograd_transformed_weights.get()->info()->offset_first_element_in_bytes()); // Prepare Weights _winograd_impl.weight_transform->execute( *_conv_args, permuted_weights_ptr, permuted_weight_row_stride, permuted_weight_col_stride, permuted_weight_channel_stride, win_wght_transf_ptr, _winograd_impl.winograd_spec, 0, 1 // Thread 1 of 1 ); ITensorPack gemm_pack = tensors; gemm_pack.add_const_tensor(ACL_SRC_1, winograd_transformed_weights.get()); _gemm_function->prepare(gemm_pack); _is_prepared = 1; } } experimental::MemoryRequirements CpuWinogradConv2d::workspace() const { return _aux_mem; } } // namespace cpu } // namespace arm_compute