/* * 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/CpuGemmConv2d.h" #include "arm_compute/core/Size2D.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "src/common/utils/Log.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/MemoryHelpers.h" #include "src/core/helpers/Utils.h" #include "src/cpu/kernels/CpuCol2ImKernel.h" #include "src/cpu/kernels/CpuIm2ColKernel.h" #include "src/cpu/kernels/CpuWeightsReshapeKernel.h" #include "src/cpu/operators/CpuGemm.h" #include "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.h" #include "src/cpu/operators/CpuGemmLowpOutputStage.h" #include "src/cpu/operators/CpuReshape.h" #include "src/cpu/utils/CpuAuxTensorHandler.h" #include #include using namespace arm_compute::misc::shape_calculator; using namespace arm_compute::experimental; namespace arm_compute { namespace cpu { /** @section note_CpuGemmConv2d_weight_transformation Weight Transformations in CpuGemmConv2d * * A. Terminology * Throughout CpuGemmConv2d, we use the following terms in ways that may differ from other operators / kernels: * - "Transform" or "Reshape" of the weights: they both mean all the operations that we perform on the weight * tensor up until they are consumed by gemm (CpuGemm or CpuGemmLowpMatrixMultiplyCore) * Note that the specific gemm operator may perform further transformations on the weights, but the * transformations here only mean those performed in CpuGemmConv2d * - "Transpose" of weights: The @ref CpuTranspose operation. I.e. transpose of the weights' lowest two * dimensions * * B. Gemm-based conv2d * We want to convert the 2d convolution op (ignoring bias): * dst = conv2d(src, weight) * into a matrix multiplication op: * gemm_dst = gemm(lhs, rhs) * * E.g.: For data layout NHWC * 3 (hi) <----------> (lo) 0 * src.shape = [batch, in_h , in_w, in_c] * weight.shape = [out_c, k_h , k_w, in_c] * dst.shape = [batch, out_h, out_w, out_c] * * This requires three transformations: * * src -> lhs, transform conv input to gemm lhs; gemm_lhs is a 2d matrix where each row (or column, * depending on the convention) is a linearized "patch" of the conv_input that corresponds to * the receptive field of the corresponding output element. * The convention is to use "column", but to disambiguate from the column vector of a matrix, * in this documentation we shall use "patch". * This transform is called im2col (for details see @ref CpuIm2ColKernel) * * weight -> rhs, transform conv weight to gemm rhs, known as weight transform/reshape (wt) * * gemm_dst -> dst, transform gemm output back to conv output, known as col2im (for details see * @ref CpuCol2ImKernel) * * This section focuses on the weight transformation and assumes the im2col is already performed * * C. Weight Transformation * After im2col, assume: lhs.shape = [num_patch, patch_size], * where patch_size is the number of elements in a "patch": patch_size = k_h * k_w * in_c * num_patch is the number of patches; we can ignore it here (for details see @ref CpuIm2ColKernel) * * After wt, rhs should have the shape: rhs = [patch_size, out_c] * * Therefore, the weight transformation consists of two steps: * 1. Collapsing all 3 spatial dimensions: [out_c, k_h, k_w, in_c] -> [out_c, patch_size] * 2. Transpose the collapsed shape: [out_c, patch_size] -> [patch_size, out_c] * * D. Implementation * There are 4 paths for weight transformation * * 1. Path 1: Fixed weight format - no transformation * The underlying gemm kernel may adopt fixed weight format (isVarWeightsKernel() == true), which requires * that no weight transformation shall be performed * Note that this no-transform requirement applies both to this op (CpuGemmConv2d) and the constituent ops, up * until the fixed format kernels themselves * * 2. Path 2: Reinterpret then transpose later * If the weight tensor has no "holes" (see @ref has_holes), there are two optimizations we can apply: * - We can ignore the first step (collapsing of spatial dimensions) by simply re-interpreting the shape * in TensorInfo * - Instead of performing transpose here, we can pass the transpose flag to the underlying gemm. The gemm * may then decide to fuse the transpose with any further transformations * * 3. Path 3: Reshape then transpose later * If the weight tensor has holes, then we use a dedicated @ref CpuReshape, followed by transpose later * * 4. Path 4: Fused reshape and transpose * This is only for quantized types for now (TODO: Remove (COMPMID-6596)). We fall back to a legacy * non-optimized kernel @ref CpuWeightsReshapeKernel to perform a fused reshape + transpose * * Path 1 is the long term solution that we shall migrate to once (if) we adopt fixed weight format for all gemm * kernels. * In the short term, Path 2 is the favored, more performant path. */ namespace { /** Initialize reshaped / transformed weight info * * @param[in] weights Input weights * @param[out] reshaped_weights Transformed weights */ void initialize_reshaped_weight_info(const ITensorInfo &weights, ITensorInfo &reshaped_weights) { auto_init_if_empty(reshaped_weights, weights); if (is_data_type_quantized(weights.data_type())) { // WT method: FusedReshapeAndTranspose reshaped_weights.set_tensor_shape(compute_weights_reshaped_shape(weights, /* has_bias */ false)); } else { TensorShape collapsed_weights = weights.tensor_shape(); collapsed_weights.collapse(3); reshaped_weights.set_tensor_shape(collapsed_weights); } } } // namespace CpuGemmConv2d::WeightTransformMethod CpuGemmConv2d::get_wt_method(const ITensorInfo &weights) { // TODO: Extend ReinterpretThenTranspose support for quantized data types COMPMID-6596 if (is_data_type_quantized(weights.data_type())) { return WeightTransformMethod::FusedReshapeAndTranspose; } return has_holes(weights) ? WeightTransformMethod::ReshapeThenTranspose : WeightTransformMethod::ReinterpretThenTranspose; } CpuGemmConv2d::SkipInfo CpuGemmConv2d::skip_im_col_info(const ITensorInfo *src, const ITensorInfo *weights, const PadStrideInfo &conv_info, const Size2D &dilation, const ActivationLayerInfo &act_info) { const DataLayout data_layout = src->data_layout(); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); const unsigned int kernel_width = weights->dimension(idx_width); const unsigned int kernel_height = weights->dimension(idx_height); unsigned int conv_w = 0; unsigned int conv_h = 0; std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), src->dimension(idx_height), kernel_width, kernel_height, conv_info, dilation); const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); if (skip_im2col) { const bool skip_col2im = (data_layout == DataLayout::NHWC && (bool(CpuGemmConv2d::validate_gemm3d(src, weights, act_info, conv_h, /*skip_im2col*/ true)))); if (skip_col2im) { return {true, true}; } } else { const bool skip_col2im = (data_layout == DataLayout::NHWC && (bool(CpuGemmConv2d::validate_gemm3d(src, weights, act_info, conv_h, /*skip_im2col*/ false)))); if (skip_col2im) { return {false, true}; } } // Default case when we cannot reinterpret the input and output as 3D. return {false, false}; } CpuGemmConv2d::CpuGemmConv2d() : _weights_reshape(nullptr), _weights_reshape_and_transpose_kernel(nullptr), _im2col_kernel(), _mm_gemm(), _mm_gemmlowp(), _col2im_kernel(), _reshape(), _im2col_output(), _weights_reshaped(), _gemm_output(), _gemm_output_3d(), _data_layout(DataLayout::NCHW), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _is_prepared(false), _wt_method(WeightTransformMethod::ReshapeThenTranspose), _run_wt(true), _aux_mem(AuxTensorIdx::Count) { } CpuGemmConv2d::~CpuGemmConv2d() = default; void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act_info, bool enable_fast_math, int gemm_3d_depth, bool fixed_format, arm_compute::WeightFormat weight_format) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights); ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, act_info, enable_fast_math, gemm_3d_depth, _skip_im2col, fixed_format, weight_format)); // Supported activations in GEMM const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU}; if (_is_quantized) { TensorInfo tmp_src{*src}; TensorInfo tmp_weights{*weights}; // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() // Extract and negate input and weights offset const QuantizationInfo iqinfo = src->quantization_info(); const QuantizationInfo wqinfo = weights->quantization_info(); const QuantizationInfo oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info(); const UniformQuantizationInfo uiqinfo = iqinfo.uniform(); const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); const DataType data_type = src->data_type(); tmp_src.set_quantization_info(QuantizationInfo(uiqinfo.scale, -uiqinfo.offset)); if (!is_data_type_quantized_per_channel(tmp_weights.data_type())) { const UniformQuantizationInfo uwqinfo = wqinfo.uniform(); tmp_weights.set_quantization_info(QuantizationInfo(uwqinfo.scale, -uwqinfo.offset)); } // Merge activation with output stage PixelValue type_min{}; PixelValue type_max{}; std::tie(type_min, type_max) = get_min_max(data_type); int32_t min_activation = type_min.get(); int32_t max_activation = type_max.get(); if (supported_acts.count(act_info.activation()) != 0) { std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); } GEMMLowpOutputStageInfo output_info; output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; output_info.gemmlowp_offset = uoqinfo.offset; output_info.gemmlowp_min_bound = min_activation; output_info.gemmlowp_max_bound = max_activation; output_info.is_quantized_per_channel = (tmp_weights.data_type() == DataType::QSYMM8_PER_CHANNEL); quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info); _mm_gemmlowp = std::make_unique(); _mm_gemmlowp->configure(&tmp_src, &tmp_weights, biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info, false, enable_fast_math, false, act_info, fixed_format, weight_format, false /* pretranspose_B. TODO: COMPMID-6596 */)); auto mm_mem_req = _mm_gemmlowp->workspace(); for (unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) { _aux_mem[cont] = mm_mem_req[cont]; } } else { // Create GEMMInfo structure const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info, fixed_format, weight_format, true /*pretranspose_B. For fp gemm (wt path 1 - 3), We always pretranspose B (for wt path 1 this flag is ignored)*/); // Configure matrix multiply function _mm_gemm = std::make_unique(); _mm_gemm->configure(src, weights, biases, dst, 1.0f, 1.0f, gemm_info); auto mm_mem_req = _mm_gemm->workspace(); for (unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) { _aux_mem[cont] = mm_mem_req[cont]; } } } Status CpuGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act_info, bool enable_fast_math, int gemm_3d_depth, bool skip_im2col, bool fixed_format, arm_compute::WeightFormat weight_format) { const DataType data_type = src->data_type(); const bool is_quantized = is_data_type_quantized_asymmetric(data_type); const bool is_activation_enabled = act_info.enabled(); if (is_quantized) { // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() // Extract and negate input and weights offset const QuantizationInfo &iqinfo = src->quantization_info(); const QuantizationInfo &wqinfo = weights->quantization_info(); const QuantizationInfo &oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info(); const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); // Merge activation with output stage PixelValue type_min{}; PixelValue type_max{}; std::tie(type_min, type_max) = get_min_max(data_type); int32_t min_activation = type_min.get(); int32_t max_activation = type_max.get(); const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU}; if (is_activation_enabled && supported_acts.count(act_info.activation()) != 0) { std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); } GEMMLowpOutputStageInfo output_info; output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; output_info.gemmlowp_offset = uoqinfo.offset; output_info.gemmlowp_min_bound = min_activation; output_info.gemmlowp_max_bound = max_activation; output_info.is_quantized_per_channel = (weights->data_type() == DataType::QSYMM8_PER_CHANNEL); ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info)); // Perform validation step on GEMMLowp std::unique_ptr input_qa = src->clone(); std::unique_ptr weights_qa = weights->clone(); input_qa->set_quantization_info(QuantizationInfo(iqinfo.uniform().scale, -iqinfo.uniform().offset)); weights_qa->set_quantization_info(QuantizationInfo(wqinfo.uniform().scale, -wqinfo.uniform().offset)); return CpuGemmLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info, false, enable_fast_math, false, act_info, false /* pretranspose_B. TODO: COMPMID-6596 */)); } else { // Create GEMMInfo structure const GEMMInfo gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info, fixed_format, weight_format, true /*pretranspose_B. For fp gemm (wt path 1 - 3), We always pretranspose B (for wt path 1 this flag is ignored)*/); // Perform validation step on Matrix multiply function return CpuGemm::validate(src, weights, biases, dst, 1.0f, 1.0f, gemm_info); } } Status CpuGemmConv2d::validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col) { const DataType data_type = input_info->data_type(); const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth; const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U; // Set dummy tensor shapes for the validation const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type, input_info->quantization_info()); const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type, weights_info->quantization_info()); const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, data_type, input_info->quantization_info()); return validate_mm(&dummy_input_info, &dummy_weights_info, nullptr, &dummy_output_info, act_info, false, gemm_3d_depth, skip_im2col); } void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); ARM_COMPUTE_UNUSED(num_groups, weights_info); ARM_COMPUTE_ERROR_THROW_ON(CpuGemmConv2d::validate(src, weights, biases, dst, conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups)); ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups); const DataType data_type = src->data_type(); const DataLayout data_layout = src->data_layout(); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); const unsigned int kernel_width = weights->dimension(idx_width); const unsigned int kernel_height = weights->dimension(idx_height); _is_prepared = weights_info.retain_internal_weights(); _is_quantized = is_data_type_quantized_asymmetric(src->data_type()); _data_layout = data_layout; _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); const ITensorInfo *gemm_input_to_use = src; ITensorInfo *gemm_output_to_use = dst; // Get convolved dimensions unsigned int conv_w = 0; unsigned int conv_h = 0; std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), src->dimension(idx_height), kernel_width, kernel_height, conv_info, dilation); ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h), "Output shape does not match the expected one"); // Check if GEMM3D is supported const CpuGemmConv2d::SkipInfo skip_info = CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); _skip_im2col = skip_info.skip_im2col; _skip_col2im = skip_info.skip_col2im; // Get parameters from conv_info unsigned int stride_x = 0; unsigned int stride_y = 0; std::tie(stride_x, stride_y) = conv_info.stride(); // Initialize reshaped weights initialize_reshaped_weight_info(*weights, _weights_reshaped); // Create tensor to store im2col reshaped inputs if (!_skip_im2col) { const int block_by = arm_compute::block_by(weights_info.weight_format()); unsigned int input_pad_right = 0; if (block_by > 1) { input_pad_right = (src->dimension(idx_channel) % block_by) == 0 ? 0 : block_by - (src->dimension(idx_channel) % block_by); } // Configure _im2col_kernel = std::make_unique(); _im2col_kernel->configure(src, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, false, dilation, num_groups, input_pad_right); // Update GEMM input gemm_input_to_use = &_im2col_output; } const unsigned int mat_weights_cols = weights->dimension(idx_kernels); // Create temporary GEMM output tensor in case we cannot skip col2im const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type; if (!_skip_col2im) { TensorShape shape_gemm; // Calculate GEMM output shape shape_gemm = _im2col_output.tensor_shape(); shape_gemm.set(0, mat_weights_cols); shape_gemm.set(1, conv_w * conv_h); _gemm_output = TensorInfo(shape_gemm, 1, output_data_type); _gemm_output.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); _gemm_output_3d = TensorInfo(_gemm_output); // Update GEMM output gemm_output_to_use = &_gemm_output; } else { _gemm_output_3d = TensorInfo(*dst); _gemm_output_3d.set_data_type(output_data_type).set_data_layout(src->data_layout()).set_is_resizable(true); _gemm_output = TensorInfo(_gemm_output_3d); // Update GEMM output gemm_output_to_use = &_gemm_output_3d; } // Configure GEMM // In case we need to skip col2im, GEMM3D (gemm_3d_depth != 0) must be called in order to avoid reshaping the output matrix const unsigned int gemm_3d_depth = _skip_col2im ? conv_h : 0; const bool fixed_format = weights_info.weight_format() != arm_compute::WeightFormat::UNSPECIFIED; /** @section note_CpuGemmConv2d_weight_use_in_configure Which weights tensor should we use to configure gemm * * A. The problem: * In principle, we should use the weights tensor corresponding to the weights transformation path. I.e.: * - If no weight transformation (_run_wt == false): Use original weights * - else: Use transformed weights * However in practice we have a dilemma: * - We need to know _run_wt before we can configure gemm with the corresponding weights, but * - _run_wt depends on isVarWeightsKernel(), which is only known after gemm is configured * * B. The decision: * To simplify the matter, we decide to always use the transformed weights, regardless of _run_wt * * This decision requires the following conditions: * 1. The underlying gemm where isVarWeightsKernel() == true, must guarantee that: * A. Ignore the flag to transpose weights (GEMMInfo::pretranspose_B) * B. Use weights/B tensor passed to it at prepare() or run() instead of that passed at configure() * 2. CpuGemmConv2d where isVarWeightsKernel() == true, must guarantee that: * A. Pass original weights instead of reshaped or reinterpreted weights * * C. Future actions: * Condition 2 is a given, based on our implementation. * If condition 1 cannot hold, we must make changes to the underlying gemm to: * 1. Either expose isVarWeightsKernel() before gemm is configured somehow, or * 2. Take in an additional "original_weights" tensor info at configure */ configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, enable_fast_math, gemm_3d_depth, fixed_format, weights_info.weight_format()); // Can only decide isVarWeightsKernel after gemm is configured _run_wt = !isVarWeightsKernel(); if (!_skip_col2im && _data_layout == DataLayout::NCHW) { // Configure col2im _col2im_kernel = std::make_unique(); _col2im_kernel->configure(gemm_output_to_use, dst, Size2D(conv_w, conv_h)); } else { // Configure reshape layer _reshape = std::make_unique(); _reshape->configure(gemm_output_to_use, dst); } // Check lifetime _aux_mem[Im2ColOutput] = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size()); // Add WeightsReshaped memory requirement to workspace // Note that in case of WeightTransformMethod::ReinterpretThenTranspose, we do not need to allocate this memory // However since we cannot determine weight transformation method until prepare (see prepare()), we will have to // settle with allocating more if (_run_wt) { // Check if GEMM transforms weights // If weight is further transformed by underlying gemm after ReshapeThenTranspose then we can free // WeightsReshaped in prepare // Otherwise WeightsReshaped is the final transformation of weights and needs to persist bool gemm_trans_wei = _aux_mem[GemmAsmPretransposedRHS].size > 0; gemm_trans_wei = _mm_gemm != nullptr ? _aux_mem[GemmTransposed1xWRHS].size > 0 : gemm_trans_wei; gemm_trans_wei = _mm_gemmlowp != nullptr ? _aux_mem[GemmLowpTransposed1xWRHS].size > 0 : gemm_trans_wei; _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), gemm_trans_wei ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _weights_reshaped.total_size()); } _aux_mem[GemmOutput] = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size()); } Status CpuGemmConv2d::has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, const bool enable_fast_math) { const DataLayout data_layout = src->data_layout(); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); const unsigned int kernel_width = weights->dimension(idx_width); const unsigned int kernel_height = weights->dimension(idx_height); unsigned int conv_w = 0; unsigned int conv_h = 0; std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), src->dimension(idx_height), kernel_width, kernel_height, conv_info, dilation); const CpuGemmConv2d::SkipInfo skip_info = CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); const bool skip_im2col = skip_info.skip_im2col; const bool skip_col2im = skip_info.skip_col2im; const unsigned int gemm_3d_depth = skip_col2im ? conv_h : 0; const bool fixed_format = weights_info.weight_format() != arm_compute::WeightFormat::UNSPECIFIED; /** @section note_CpuGemmConv2d_weight_use_in_has_opt_impl Which weights tensor should we use for has_opt_impl * * For the pretranspose_B flag, this shares a similar problem and thus the same decision as that of * @ref note_CpuGemmConv2d_weight_use_in_configure * * But for the weights, we shall always use the original instead of reshaped weights here */ const GEMMInfo gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info, fixed_format, weights_info.weight_format(), true /* pretranspose_B */); return CpuGemm::has_opt_impl(expected_weight_format, src, weights, biases, dst, gemm_info); } Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!"); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::BFLOAT16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::BFLOAT16, DataType::F16, DataType::F32); if (!is_fixed_format(weights_info.weight_format())) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights); } ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups > 1, "Grouping (num_groups != 1) is not supported"); const DataLayout data_layout = src->data_layout(); const DataType data_type = src->data_type(); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); const unsigned int kernel_width = weights->dimension(idx_width); const unsigned int kernel_height = weights->dimension(idx_height); TensorInfo im2col_reshaped_info{}; TensorInfo info_gemm{}; TensorInfo tmp_info{}; TensorInfo weights_reshaped_info{}; const ITensorInfo *gemm_input_to_use = src; const ITensorInfo *gemm_output_to_use = dst; const ITensorInfo *weights_to_use = weights; const bool append_bias = false; const bool is_quantized = is_data_type_quantized_asymmetric(data_type); const bool is_bf16 = data_type == DataType::BFLOAT16; // Get convolved dimensions unsigned int conv_w = 0; unsigned int conv_h = 0; std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), src->dimension(idx_height), kernel_width, kernel_height, conv_info, dilation); // Check if GEMM3D is supported const CpuGemmConv2d::SkipInfo skip_info = CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); const bool skip_im2col = skip_info.skip_im2col, skip_col2im = skip_info.skip_col2im; ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_channel) != src->dimension(idx_channel)); ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); // Validate biases if (biases != nullptr) { if (is_quantized) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); } else if (is_bf16) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); } else { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); } ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != dst->dimension(idx_channel)); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } unsigned int mat_weights_cols = weights->dimension(idx_kernels); unsigned int mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel); // Initialize reshaped weights initialize_reshaped_weight_info(*weights, weights_reshaped_info); // No need to call CpuReshape::validate() or CpuTranspose::validate() as the dst info is auto-configured from the // src weights_to_use = &weights_reshaped_info; if (!skip_im2col) { const int block_by = arm_compute::block_by(weights_info.weight_format()); int input_pad_right = 0; if (block_by > 1) { input_pad_right = (src->dimension(idx_channel) % block_by) == 0 ? 0 : block_by - (src->dimension(idx_channel) % block_by); mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * (weights->dimension(idx_channel) + input_pad_right); } // Create tensor info for im2col reshaped inputs // For CPU, the batch size is on the fourth dimension TensorShape shape_im2col = src->tensor_shape(); shape_im2col.set(0, mat_weights_rows); shape_im2col.set(1, conv_w * conv_h); shape_im2col.set(2, 1); im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type); im2col_reshaped_info.set_quantization_info(src->quantization_info()); ARM_COMPUTE_RETURN_ON_ERROR( kernels::CpuIm2ColKernel::validate(src, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation, num_groups, input_pad_right)); gemm_input_to_use = &im2col_reshaped_info; } // Create temporary GEMM output tensor in case we cannot skip col2im const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type; if (!skip_col2im) { TensorShape shape_gemm = gemm_input_to_use->tensor_shape(); shape_gemm.set(0, mat_weights_cols); shape_gemm.set(1, conv_w * conv_h); info_gemm = TensorInfo(shape_gemm, 1, output_data_type); } else { info_gemm = TensorInfo(dst->tensor_shape(), 1, output_data_type); } info_gemm.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); gemm_output_to_use = &info_gemm; const bool fixed_format = weights_info.weight_format() != arm_compute::WeightFormat::UNSPECIFIED; // See note_CpuGemmConv2d_weight_use_in_configure regarding the choice of the weights ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, act_info, enable_fast_math, skip_col2im ? conv_h : 0, skip_im2col, fixed_format, weights_info.weight_format())); // Validate Col2Im/ReshapeLayer if (!skip_col2im && (data_layout == DataLayout::NCHW)) { ARM_COMPUTE_RETURN_ON_ERROR( kernels::CpuCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h))); } return Status{}; } void CpuGemmConv2d::run(ITensorPack &tensors) { prepare(tensors); auto src = tensors.get_const_tensor(ACL_SRC_0); auto dst = tensors.get_tensor(ACL_DST); auto gemm_input_to_use = src; CpuAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false); CpuAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false); bool out_has_padding = _skip_col2im && (dst->info()->padding().bottom != 0 || dst->info()->padding().top != 0); if (!_skip_im2col) { // Run input reshaping unsigned int y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); ITensorPack pack = {{TensorType::ACL_SRC, src}, {TensorType::ACL_DST, im2col_output.get()}}; NEScheduler::get().schedule_op(_im2col_kernel.get(), y_dim, _im2col_kernel->window(), pack); gemm_input_to_use = im2col_output.get(); } // Handle the case where output has top/bottom padding const ITensor *out_to_use = out_has_padding ? gemm_output.get() : dst; Tensor gemm3d; _gemm_output_3d.extend_padding(out_to_use->info()->padding()); gemm3d.allocator()->soft_init(_gemm_output_3d); gemm3d.allocator()->import_memory(out_to_use->buffer()); auto gemm_output_to_use = gemm_output.get(); if (_skip_im2col) { gemm_output_to_use = &gemm3d; } if (_skip_col2im && !out_has_padding) { gemm_output_to_use = dst; } ITensorPack gemm_pack = tensors; gemm_pack.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use); gemm_pack.add_tensor(TensorType::ACL_DST, gemm_output_to_use); // Allocate reshaped weights if required auto weights = gemm_pack.get_const_tensor(TensorType::ACL_SRC_1); ARM_COMPUTE_ERROR_ON_NULLPTR(weights); // Re-interpreted weights. Only tensor shape is changed. Only memory import, no allocation const bool use_reinterpreted_wei = (_run_wt && _wt_method == WeightTransformMethod::ReinterpretThenTranspose); CpuAuxTensorHandler reinterpreted_wei( _weights_reshaped, *weights, /* import only if we chose the ReinterpretThenTranspose path, because otherwise the weight may have been freed */ !use_reinterpreted_wei); const bool use_reshaped_wei = (_run_wt && (_wt_method == WeightTransformMethod::ReshapeThenTranspose || _wt_method == WeightTransformMethod::FusedReshapeAndTranspose)); CpuAuxTensorHandler reshaped_wei(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false /* pack_inject */, !use_reshaped_wei /* bypass_alloc */, !use_reshaped_wei /* bypass_import */ ); // Update the weights to use if it has been reshaped if (use_reinterpreted_wei) { gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, reinterpreted_wei.get()); } else if (use_reshaped_wei) { gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, reshaped_wei.get()); } // Runs CpuGemm or CpuGemmLowpMatrixMultiplyCore functions _is_quantized ? _mm_gemmlowp->run(gemm_pack) : _mm_gemm->run(gemm_pack); // Reshape output matrix if (!_skip_col2im) { if (_data_layout == DataLayout::NCHW) { ITensorPack pack = {{TensorType::ACL_SRC, gemm_output.get()}, {TensorType::ACL_DST, dst}}; NEScheduler::get().schedule_op(_col2im_kernel.get(), Window::DimY, _col2im_kernel->window(), pack); } else { ITensorPack pack = {{TensorType::ACL_SRC, gemm_output_to_use}, {TensorType::ACL_DST, dst}}; _reshape->run(pack); } } else if (out_has_padding) { ITensorPack pack = {{TensorType::ACL_SRC, gemm_output_to_use}, {TensorType::ACL_DST, dst}}; _reshape->run(pack); } } void CpuGemmConv2d::prepare(ITensorPack &tensors) { if (!_is_prepared) { auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); // Determine which weights reshape path to take // Note that this decision can only occur at prepare instead of configure because it relies on the presence of // any holes in the weight tensor, which may change after configure (e.g. from extending padding) if (_run_wt) { _wt_method = get_wt_method(*(weights->info())); switch (_wt_method) { case (WeightTransformMethod::FusedReshapeAndTranspose): { ARM_COMPUTE_LOG_INFO_WITH_FUNCNAME_ACL("Perform weight transformation: FusedReshapeAndTranspose"); _weights_reshape_and_transpose_kernel = std::make_unique(); _weights_reshape_and_transpose_kernel->configure(weights->info(), nullptr, &_weights_reshaped); break; } case (WeightTransformMethod::ReshapeThenTranspose): { ARM_COMPUTE_LOG_INFO_WITH_FUNCNAME_ACL("Perform weight transformation: ReshapeThenTranspose"); _weights_reshape = std::make_unique(); _weights_reshape->configure(weights->info(), &_weights_reshaped); break; } case (WeightTransformMethod::ReinterpretThenTranspose): { ARM_COMPUTE_LOG_INFO_WITH_FUNCNAME_ACL("Perform weight transformation: ReinterpretThenTranspose"); // Nothing to configure break; } default: { ARM_COMPUTE_ERROR("Unsupported weight transform method"); } } } else { ARM_COMPUTE_LOG_INFO_WITH_FUNCNAME_ACL("No weight transformation is performed"); } ITensorPack gemm_pack = tensors; // Allocate reshaped weights if required CpuAuxTensorHandler reinterpreted_wei( _weights_reshaped, *weights); // Re-interpreted weights. Only tensor shape is changed. No allocation CpuAuxTensorHandler reshaped_wei(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors); // Run weights reshape if required if (_run_wt) { switch (_wt_method) { case (WeightTransformMethod::FusedReshapeAndTranspose): { ITensorPack pack = {{TensorType::ACL_SRC, weights}, {TensorType::ACL_DST, reshaped_wei.get()}}; NEScheduler::get().schedule_op(_weights_reshape_and_transpose_kernel.get(), Window::DimW, _weights_reshape_and_transpose_kernel->window(), pack); weights->mark_as_unused(); gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, reshaped_wei.get()); break; } case (WeightTransformMethod::ReshapeThenTranspose): { ITensorPack pack = {{TensorType::ACL_SRC, weights}, {TensorType::ACL_DST, reshaped_wei.get()}}; _weights_reshape->run(pack); weights->mark_as_unused(); gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, reshaped_wei.get()); break; } case (WeightTransformMethod::ReinterpretThenTranspose): { gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, reinterpreted_wei.get()); // Nothing to run break; } default: { ARM_COMPUTE_ERROR("Unsupported weight transform method"); } } } _is_quantized ? _mm_gemmlowp->prepare(gemm_pack) : _mm_gemm->prepare(gemm_pack); _is_prepared = true; } } experimental::MemoryRequirements CpuGemmConv2d::workspace() const { return _aux_mem; } bool CpuGemmConv2d::isVarWeightsKernel() const { return _mm_gemm && _mm_gemm->isVarWeightsKernel(); } } // namespace cpu } // namespace arm_compute