diff options
Diffstat (limited to 'src/cpu/kernels/internal')
4 files changed, 994 insertions, 0 deletions
diff --git a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp new file mode 100644 index 0000000000..296fe88791 --- /dev/null +++ b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp @@ -0,0 +1,397 @@ +/* + * Copyright (c) 2021-2023 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/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.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 "src/core/CPP/Validate.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/kernels/assembly/depthwise.hpp" +#include "src/core/utils/AssemblyUtils.h" + +#include "depthwise_common.hpp" +#include <arm_neon.h> + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +using namespace arm_compute::misc::shape_calculator; + +namespace +{ +constexpr unsigned int idx_width = 1; +constexpr unsigned int idx_height = 2; +constexpr unsigned int idx_channels = 0; +constexpr unsigned int idx_batches = 3; + +template <typename TSrc, typename TWeights, typename TDst> +void create_arm_dwc(const ITensorInfo *src, + const ITensorInfo *weights, + ITensorInfo *dst, + const ConvolutionInfo &info, + const CPUInfo &cpu_info, + std::unique_ptr<arm_conv::depthwise::IDepthwiseCommon> &kernel, + std::string &_name) +{ + unsigned int stride_cols{}; + unsigned int stride_rows{}; + std::tie(stride_cols, stride_rows) = info.pad_stride_info.stride(); + + unsigned int dilation_cols = info.dilation.x(); + unsigned int dilation_rows = info.dilation.y(); + + const arm_conv::PaddingValues padding = assembly_utils::map_to_arm_conv_padding(info.pad_stride_info); + + const unsigned int n_batches = src->dimension(idx_batches); + const unsigned int src_rows = src->dimension(idx_height); + const unsigned int src_cols = src->dimension(idx_width); + const unsigned int n_channels = src->dimension(idx_channels); + const unsigned int dst_rows = dst->dimension(idx_height); + const unsigned int dst_cols = dst->dimension(idx_width); + + const unsigned int kernel_cols = weights->dimension(idx_width); + const unsigned int kernel_rows = weights->dimension(idx_height); + + const arm_gemm::Activation activation = assembly_utils::map_to_arm_gemm_activation(info.act_info); + + arm_conv::depthwise::DepthwiseArgs args(&cpu_info, kernel_rows, kernel_cols, stride_rows, stride_cols, + dilation_rows, dilation_cols, n_batches, src_rows, src_cols, n_channels, + dst_rows, dst_cols, info.depth_multiplier, padding, activation, nullptr); + + // Configure assembly pooling kernel + auto dwc_kernel_asm = arm_conv::depthwise::depthwise<TSrc, TWeights, TDst>(args); + if (dwc_kernel_asm == nullptr) + { + // Configuration not supported: Leave function unconfigured: + return; + } + + _name = dwc_kernel_asm->name(); + kernel = std::move(dwc_kernel_asm); +} + +template <typename TSrc, typename TWeights, typename TDst> +void create_arm_dwc_quant(const ITensorInfo *src, + const ITensorInfo *weights, + ITensorInfo *dst, + const ConvolutionInfo &info, + const CPUInfo &cpu_info, + std::unique_ptr<arm_conv::depthwise::IDepthwiseCommon> &kernel, + std::vector<int32_t> &multipliers, + std::vector<int32_t> &right_shifts, + std::vector<int32_t> &left_shifts, + std::string &_name) +{ + unsigned int stride_cols{}; + unsigned int stride_rows{}; + std::tie(stride_cols, stride_rows) = info.pad_stride_info.stride(); + + unsigned int dilation_cols = info.dilation.x(); + unsigned int dilation_rows = info.dilation.y(); + + const arm_conv::PaddingValues padding = assembly_utils::map_to_arm_conv_padding(info.pad_stride_info); + + const unsigned int n_batches = src->dimension(idx_batches); + const unsigned int src_rows = src->dimension(idx_height); + const unsigned int src_cols = src->dimension(idx_width); + const unsigned int n_channels = src->dimension(idx_channels); + const unsigned int dst_rows = dst->dimension(idx_height); + const unsigned int dst_cols = dst->dimension(idx_width); + + const unsigned int kernel_cols = weights->dimension(idx_width); + const unsigned int kernel_rows = weights->dimension(idx_height); + + const arm_gemm::Activation activation = assembly_utils::map_to_arm_gemm_activation(info.act_info); + + arm_conv::depthwise::DepthwiseArgs args(&cpu_info, kernel_rows, kernel_cols, stride_rows, stride_cols, + dilation_rows, dilation_cols, n_batches, src_rows, src_cols, n_channels, + dst_rows, dst_cols, info.depth_multiplier, padding, activation, nullptr); + + const auto src_qinfo = src->quantization_info().uniform(); + const auto weights_qinfo = weights->quantization_info(); + const auto dst_qinfo = dst->quantization_info().uniform(); + + const unsigned int num_filters = weights_qinfo.scale().size(); + + multipliers.resize(num_filters); + std::vector<int32_t> dst_shifts(num_filters); + quantization::compute_quantized_multipliers_and_shifts(src, weights, dst, multipliers.data(), dst_shifts.data()); + + // Quantize activation bounds + int32_t min_activation = std::numeric_limits<TSrc>::lowest(); + int32_t max_activation = std::numeric_limits<TSrc>::max(); + if (info.act_info.enabled()) + { + std::tie(min_activation, max_activation) = + get_quantized_activation_min_max(info.act_info, src->data_type(), dst_qinfo); + } + + // Set quantization parameters for assembly kernels + arm_gemm::Requantize32 requant_args{}; + if (is_data_type_quantized_per_channel(weights->data_type())) + { + left_shifts.resize(num_filters); + right_shifts.resize(num_filters); + bool need_left_shift = false; // Select more optimized path if left shift is not needed + for (unsigned int i = 0; i < num_filters; ++i) + { + left_shifts[i] = std::max(-dst_shifts[i], static_cast<int32_t>(0)); + right_shifts[i] = std::min(-dst_shifts[i], static_cast<int32_t>(0)); + if (dst_shifts[i] < 0 && !need_left_shift) + { + need_left_shift = true; + } + } + + requant_args = arm_gemm::Requantize32(nullptr, 0, src_qinfo.offset, weights_qinfo.uniform().offset, + dst_qinfo.offset, (need_left_shift) ? left_shifts.data() : nullptr, + right_shifts.data(), multipliers.data(), + static_cast<TSrc>(min_activation), static_cast<TSrc>(max_activation)); + } + else + { + requant_args = arm_gemm::Requantize32(nullptr, 0, src_qinfo.offset, weights_qinfo.uniform().offset, + dst_qinfo.offset, -dst_shifts[0], multipliers[0], + static_cast<TSrc>(min_activation), static_cast<TSrc>(max_activation)); + } + + // Configure assembly pooling kernel with requantization + auto dwc_kernel_asm = + arm_conv::depthwise::depthwise<TSrc, TWeights, TDst, arm_gemm::Requantize32>(args, requant_args); + if (dwc_kernel_asm == nullptr) + { + // Configuration not supported: Leave function unconfigured: + return; + } + _name = dwc_kernel_asm->name(); + kernel = std::move(dwc_kernel_asm); +} +} // namespace + +CpuDepthwiseConv2dAssemblyWrapperKernel::CpuDepthwiseConv2dAssemblyWrapperKernel() + : _kernel_asm(nullptr), _multipliers(), _left_shifts(), _right_shifts(), _name() +{ +} + +CpuDepthwiseConv2dAssemblyWrapperKernel::~CpuDepthwiseConv2dAssemblyWrapperKernel() = default; + +void CpuDepthwiseConv2dAssemblyWrapperKernel::configure(const ITensorInfo *src, + const ITensorInfo *weights, + const ITensorInfo *, + ITensorInfo *dst, + const ConvolutionInfo &info, + const CPUInfo &cpu_info) +{ + ARM_COMPUTE_UNUSED(cpu_info); + ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); + + // Destination initialization if not yet initialized + const TensorShape dst_shape = compute_depthwise_convolution_shape(*src, *weights, info); + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(dst_shape)); + _name = "CpuDepthwiseConv2dAssemblyWrapperKernel"; + std::string asm_kernel_name(""); +#if defined(__aarch64__) + switch (src->data_type()) + { + case DataType::QASYMM8: + if (is_data_type_quantized_per_channel(weights->data_type())) + { + create_arm_dwc_quant<uint8_t, int8_t, uint8_t>(src, weights, dst, info, cpu_info, _kernel_asm, + _multipliers, _right_shifts, _left_shifts, + asm_kernel_name); + } + else + { + create_arm_dwc_quant<uint8_t, uint8_t, uint8_t>(src, weights, dst, info, cpu_info, _kernel_asm, + _multipliers, _right_shifts, _left_shifts, + asm_kernel_name); + } + break; + case DataType::QASYMM8_SIGNED: + create_arm_dwc_quant<int8_t, int8_t, int8_t>(src, weights, dst, info, cpu_info, _kernel_asm, _multipliers, + _right_shifts, _left_shifts, asm_kernel_name); + break; +#if defined(ENABLE_FP16_KERNELS) + case DataType::F16: + create_arm_dwc<float16_t, float16_t, float16_t>(src, weights, dst, info, cpu_info, _kernel_asm, + asm_kernel_name); + break; +#endif // defined(ENABLE_FP16_KERNELS) + case DataType::F32: + create_arm_dwc<float, float, float>(src, weights, dst, info, cpu_info, _kernel_asm, asm_kernel_name); + break; + default: + break; + } +#endif // defined(__aarch64__) + + Window win = calculate_max_window(*dst, Steps()); + ICpuKernel::configure(win); + if (_kernel_asm != nullptr) + { + _name += "/" + asm_kernel_name; + } +} + +Status CpuDepthwiseConv2dAssemblyWrapperKernel::validate(const ITensorInfo *src, + const ITensorInfo *weights, + const ITensorInfo *bias, + const ITensorInfo *dst, + const ConvolutionInfo &info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + +#if !defined(__aarch64__) + ARM_COMPUTE_RETURN_ERROR_MSG("32-bit is not supported by assembly kernels"); +#endif // !defined(__aarch64__) + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->data_layout() != DataLayout::NHWC, + "Only NHWC is supported by assembly kernels"); + + if (is_data_type_quantized_per_channel(weights->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QSYMM8_PER_CHANNEL); + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != weights->quantization_info().scale().size()); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); + } + + if (bias != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(0) != weights->dimension(0)); + + if (is_data_type_quantized(src->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias); + } + } + + if (dst->total_size() > 0) + { + const TensorShape dst_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*src, *weights, info); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), dst_shape); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + } + + // Assembly kernels cannot work with padding greater than the kernel. + const auto &padding = info.pad_stride_info; + const auto &dilation = info.dilation; + const auto &wei_shape = weights->tensor_shape(); + + const auto dilated_wei_w = wei_shape[1] + (wei_shape[1] - 1) * (dilation.x() - 1); + const auto dilated_wei_h = wei_shape[2] + (wei_shape[2] - 1) * (dilation.y() - 1); + + ARM_COMPUTE_RETURN_ERROR_ON(padding.pad_left() >= dilated_wei_w || padding.pad_right() >= dilated_wei_w || + padding.pad_top() >= dilated_wei_h || padding.pad_bottom() >= dilated_wei_h); + + return Status{}; +} + +void CpuDepthwiseConv2dAssemblyWrapperKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(_kernel_asm.get()); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_UNUSED(info); + + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + ITensor *dst = tensors.get_tensor(TensorType::ACL_DST); + ITensor *workspace = tensors.get_tensor(TensorType::ACL_INT_0); + ITensor *storage = tensors.get_tensor(TensorType::ACL_INT_1); + + const auto src_ptr = src->buffer() + src->info()->offset_first_element_in_bytes(); + auto dst_ptr = dst->buffer() + dst->info()->offset_first_element_in_bytes(); + auto working_space = workspace->buffer() + workspace->info()->offset_first_element_in_bytes(); + auto parameters_ptr = storage->buffer() + storage->info()->offset_first_element_in_bytes(); + + const auto src_shape = src->info()->tensor_shape(); + const auto dst_shape = dst->info()->tensor_shape(); + const auto src_padding = src->info()->padding(); + const auto dst_padding = dst->info()->padding(); + + const size_t ld_src_col = src_shape[0] + src_padding.left + src_padding.right; + const size_t ld_src_row = ld_src_col * (src_shape[1] + src_padding.top + src_padding.bottom); + const size_t ld_src_batch = ld_src_row * src_shape[2]; + const size_t ld_dst_col = dst_shape[0] + dst_padding.left + dst_padding.right; + const size_t ld_dst_row = ld_dst_col * (dst_shape[1] + dst_padding.top + dst_padding.bottom); + const size_t ld_dst_batch = ld_dst_row * dst_shape[2]; + + _kernel_asm->execute(src_ptr, ld_src_col, ld_src_row, ld_src_batch, parameters_ptr, dst_ptr, ld_dst_col, ld_dst_row, + ld_dst_batch, working_space, info.thread_id, info.num_threads); +} + +void CpuDepthwiseConv2dAssemblyWrapperKernel::pack_parameters( + void *parameters_ptr, void *bias_ptr, void *weights_ptr, size_t ld_weights_col, size_t ld_weight_row) +{ + _kernel_asm->pack_parameters(parameters_ptr, bias_ptr, weights_ptr, ld_weights_col, ld_weight_row); +} + +size_t CpuDepthwiseConv2dAssemblyWrapperKernel::get_storage_size() const +{ + return _kernel_asm->get_storage_size(); +} + +size_t CpuDepthwiseConv2dAssemblyWrapperKernel::get_working_size(unsigned int num_threads) const +{ + return _kernel_asm->get_working_size(num_threads); +} + +bool CpuDepthwiseConv2dAssemblyWrapperKernel::is_configured() const +{ + return _kernel_asm != nullptr; +} + +const char *CpuDepthwiseConv2dAssemblyWrapperKernel::name() const +{ + return _name.c_str(); +} + +size_t CpuDepthwiseConv2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const +{ + ARM_COMPUTE_UNUSED(thread_count); + ARM_COMPUTE_UNUSED(platform); + + return ICPPKernel::default_mws; +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h new file mode 100644 index 0000000000..fadaefb999 --- /dev/null +++ b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h @@ -0,0 +1,143 @@ +/* + * Copyright (c) 2019-2023 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. + */ +#ifndef ARM_COMPUTE_CPU_DEPTHWISE_CONV2D_ASSEMBLY_WRAPPER_KERNEL_H +#define ARM_COMPUTE_CPU_DEPTHWISE_CONV2D_ASSEMBLY_WRAPPER_KERNEL_H + +#include "arm_compute/core/Types.h" + +#include "src/core/common/Macros.h" +#include "src/cpu/ICpuKernel.h" +#include "src/cpu/kernels/CpuKernelSelectionTypes.h" + +namespace arm_conv +{ +namespace depthwise +{ +// Forward declarations +class IDepthwiseCommon; +} // namespace depthwise +} // namespace arm_conv + +namespace arm_compute +{ +struct ConvolutionInfo; + +namespace cpu +{ +namespace kernels +{ +/** This class is a wrapper for the depthwise convolution assembly kernels. */ +class CpuDepthwiseConv2dAssemblyWrapperKernel final : public ICpuKernel<CpuDepthwiseConv2dAssemblyWrapperKernel> +{ +public: + /** Default constructor */ + CpuDepthwiseConv2dAssemblyWrapperKernel(); + ~CpuDepthwiseConv2dAssemblyWrapperKernel(); + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuDepthwiseConv2dAssemblyWrapperKernel); + + /** Initialise the kernel's src and dst. + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] weights Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. + * Data type supported: same as @p src or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p src is QASYMM8/QASYMM8_SIGNED. + * @param[in] bias Bias tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: same as @p src, S32 when @p src is QASYMM8/QASYMM8_SIGNED. + * @param[out] dst Destination tensor info. Data type supported: same as @p input. + * @param[in] info Depthwise convolution layer meta-data. + * @param[in] cpu_info CPU information needed to select the most appropriate kernel. + */ + void configure(const ITensorInfo *src, + const ITensorInfo *weights, + const ITensorInfo *bias, + ITensorInfo *dst, + const ConvolutionInfo &info, + const CPUInfo &cpu_info); + + /** Indicates whether or not this function can be used to process the given parameters. + * + * Similar to @ref CpuDepthwiseConv2dAssemblyWrapperKernel::configure() + * + * @return a status. + */ + static Status validate(const ITensorInfo *src, + const ITensorInfo *weights, + const ITensorInfo *bias, + const ITensorInfo *dst, + const ConvolutionInfo &info); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + const char *name() const override; + + /** Pack bias and weights in a storage space for the assembly kernel + * + * @param[in] parameters_ptr Pointer to storage space. + * @param[in] bias_ptr Pointer to bias buffer. + * @param[in] weights_ptr Pointer to weights buffer. + * @param[in] ld_weights_col Columns displacement for the weights tensor. + * @param[in] ld_weights_row Rows displacement for the weights tensor. + */ + void pack_parameters( + void *parameters_ptr, void *bias_ptr, void *weights_ptr, size_t ld_weights_col, size_t ld_weights_row); + + /** Get the amount of storage space required for the rearranged weights and bias. + * + * @return size of workspace + */ + size_t get_storage_size() const; + + /** Get size of the workspace needed by the assembly kernel. + * + * @param[in] num_threads Maximum number of threads that are going to be spawned. + * + * @return size of workspace + */ + size_t get_working_size(unsigned int num_threads) const; + + /** Was the asm kernel successfully configured? + * + * @return True if the asm kernel is configured and ready to run + */ + bool is_configured() const; + + /** Return minimum workload size of the relevant kernel + * + * @param[in] platform The CPU platform used to create the context. + * @param[in] thread_count Number of threads in the execution. + * + * @return[out] small_network_mws Minimum workload size for requsted configuration. + */ + size_t get_mws(const CPUInfo &platform, size_t thread_count) const override; + +private: + std::unique_ptr<arm_conv::depthwise::IDepthwiseCommon> _kernel_asm; + std::vector<int32_t> _multipliers{}; + std::vector<int32_t> _left_shifts{}; + std::vector<int32_t> _right_shifts{}; + std::string _name{}; +}; +} // namespace kernels +} // namespace cpu +} // namespace arm_compute +#endif /* ARM_COMPUTE_CPU_DEPTHWISE_CONV2D_ASSEMBLY_WRAPPER_KERNEL_H */ diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp new file mode 100644 index 0000000000..2c1cb15786 --- /dev/null +++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp @@ -0,0 +1,320 @@ +/* + * 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/kernels/internal/CpuPool2dAssemblyWrapperKernel.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 "src/core/CPP/Validate.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/INEKernel.h" + +#include <arm_neon.h> + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +using namespace arm_compute::misc::shape_calculator; + +void CpuPool2dAssemblyWrapperKernel::configure(const ITensorInfo *src, + ITensorInfo *dst, + const PoolingLayerInfo &info, + const CPUInfo &cpu_info) +{ + ARM_COMPUTE_UNUSED(cpu_info); + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + + // dst initialization if not yet initialized + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_pool_shape(*src, info))); + +#if defined(__aarch64__) + const bool requantize = src->quantization_info() != dst->quantization_info(); + + switch (src->data_type()) + { + case DataType::QASYMM8: + if (requantize) + { + create_arm_pooling_requant<uint8_t, uint8_t>(src, dst, info, cpu_info); + } + else + { + create_arm_pooling<uint8_t, uint8_t>(src, dst, info, cpu_info); + } + break; + case DataType::QASYMM8_SIGNED: + if (requantize) + { + create_arm_pooling_requant<int8_t, int8_t>(src, dst, info, cpu_info); + } + else + { + create_arm_pooling<int8_t, int8_t>(src, dst, info, cpu_info); + } + break; +#if defined(ENABLE_FP16_KERNELS) + case DataType::F16: + create_arm_pooling<float16_t, float16_t>(src, dst, info, cpu_info); + break; +#endif // defined(ENABLE_FP16_KERNELS) + case DataType::F32: + create_arm_pooling<float, float>(src, dst, info, cpu_info); + break; + default: + break; + } +#endif // defined(__aarch64__) + + Window win = calculate_max_window(*dst, Steps()); + INEKernel::configure(win); +} + +Status +CpuPool2dAssemblyWrapperKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); +#ifndef __aarch64__ + ARM_COMPUTE_RETURN_ERROR_MSG("32-bit is not supported by assembly kernels"); +#endif /* __aarch64__ */ + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((src->data_layout() != DataLayout::NHWC) || (info.data_layout != DataLayout::NHWC), + "Only NHWC is supported by assembly kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((info.pool_type != PoolingType::AVG) && (info.pool_type != PoolingType::MAX), + "Only AVG and MAX pooling are supported by assembly kernels"); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + is_pool_region_entirely_outside_input(info), + "Pooling region that is entirely outside input tensor is unsupported by assembly kernels"); + + if (dst->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + + const TensorInfo out_info(compute_pool_shape(*src, info), 1, dst->data_type()); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &out_info); + const auto src_qinfo = src->quantization_info().uniform(); + const auto dst_qinfo = dst->quantization_info().uniform(); + + if (src_qinfo != dst_qinfo) + { + const float multiplier = src_qinfo.scale / dst_qinfo.scale; + int32_t dst_multiplier{}; + int32_t dst_shift{}; + ARM_COMPUTE_RETURN_ERROR_ON( + quantization::calculate_quantized_multiplier(multiplier, &dst_multiplier, &dst_shift)); + } + else + { + if (src->data_type() == DataType::QASYMM8) + { + const bool has_padding = info.pad_stride_info.has_padding(); + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + !info.exclude_padding && has_padding, + "Assembly kernels do not support padding for QASYMM8 with same src/dst quantization info"); + } + } + } + else + { + if (src->data_type() == DataType::QASYMM8) + { + // If dst is not configured, the quantization info are the same + const bool has_padding = info.pad_stride_info.has_padding(); + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + !info.exclude_padding && has_padding, + "Assembly kernels do not support padding for QASYMM8 with same src/dst quantization info"); + } + } + return Status{}; +} + +void CpuPool2dAssemblyWrapperKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(_kernel_asm.get()); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_UNUSED(info); + + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC); + ITensor *dst = tensors.get_tensor(TensorType::ACL_DST); + ITensor *workspace = tensors.get_tensor(TensorType::ACL_INT_0); + + const auto in_ptr = src->buffer() + src->info()->offset_first_element_in_bytes(); + auto out_ptr = dst->buffer() + dst->info()->offset_first_element_in_bytes(); + auto working_space = + (workspace == nullptr) ? nullptr : workspace->buffer() + workspace->info()->offset_first_element_in_bytes(); + + const auto src_shape = src->info()->tensor_shape(); + const auto dst_shape = dst->info()->tensor_shape(); + const auto src_padding = src->info()->padding(); + const auto dst_padding = dst->info()->padding(); + + const size_t ld_src_col = src_shape[0] + src_padding.left + src_padding.right; + const size_t ld_src_row = ld_src_col * (src_shape[1] + src_padding.top + src_padding.bottom); + const size_t ld_src_batch = ld_src_row * src_shape[2]; + const size_t ld_dst_col = dst_shape[0] + dst_padding.left + dst_padding.right; + const size_t ld_dst_row = ld_dst_col * (dst_shape[1] + dst_padding.top + dst_padding.bottom); + const size_t ld_dst_batch = ld_dst_row * dst_shape[2]; + + _kernel_asm->execute(in_ptr, ld_src_col, ld_src_row, ld_src_batch, out_ptr, ld_dst_col, ld_dst_row, ld_dst_batch, + working_space, info.thread_id, info.num_threads); +} + +size_t CpuPool2dAssemblyWrapperKernel::get_working_size(unsigned int num_threads) const +{ + return _kernel_asm->get_working_size(num_threads); +} + +bool CpuPool2dAssemblyWrapperKernel::is_configured() const +{ + return _kernel_asm != nullptr; +} + +template <typename Typesrc, typename Typedst> +void CpuPool2dAssemblyWrapperKernel::create_arm_pooling(const ITensorInfo *src, + ITensorInfo *dst, + const PoolingLayerInfo &info, + const CPUInfo &cpu_info) +{ + const arm_conv::pooling::PoolingType pool_type = (info.pool_type == PoolingType::AVG) + ? arm_conv::pooling::PoolingType::AVERAGE + : arm_conv::pooling::PoolingType::MAX; + + arm_conv::pooling::PoolingWindow window{}; + window.cols = static_cast<unsigned int>(info.pool_size.x()); + window.rows = static_cast<unsigned int>(info.pool_size.y()); + + arm_conv::pooling::PoolingStride stride{}; + std::tie(stride.cols, stride.rows) = info.pad_stride_info.stride(); + + const arm_conv::pooling::PaddingValues padding{info.pad_stride_info.pad_left(), info.pad_stride_info.pad_top(), + info.pad_stride_info.pad_right(), info.pad_stride_info.pad_bottom()}; + + constexpr unsigned int idx_width = 1; + constexpr unsigned int idx_height = 2; + constexpr unsigned int idx_channels = 0; + constexpr unsigned int idx_batches = 3; + + const unsigned int n_batches = src->dimension(idx_batches); + const unsigned int src_rows = src->dimension(idx_height); + const unsigned int src_cols = src->dimension(idx_width); + const unsigned int n_channels = src->dimension(idx_channels); + const unsigned int dst_rows = dst->dimension(idx_height); + const unsigned int dst_cols = dst->dimension(idx_width); + + arm_conv::pooling::PoolingArgs args(&cpu_info, pool_type, window, stride, info.exclude_padding, n_batches, src_rows, + src_cols, n_channels, dst_rows, dst_cols, padding, nullptr); + + // Configure assembly pooling kernel + auto pooling_kernel_asm = arm_conv::pooling::pooling<Typesrc, Typedst>(args); + if (pooling_kernel_asm == nullptr) + { + // Configuration not supported: Leave function unconfigured: + return; + } + + _kernel_asm = std::move(pooling_kernel_asm); +} + +template <typename Typesrc, typename Typedst> +void CpuPool2dAssemblyWrapperKernel::create_arm_pooling_requant(const ITensorInfo *src, + ITensorInfo *dst, + const PoolingLayerInfo &info, + const CPUInfo &cpu_info) +{ + const arm_conv::pooling::PoolingType pool_type = (info.pool_type == PoolingType::AVG) + ? arm_conv::pooling::PoolingType::AVERAGE + : arm_conv::pooling::PoolingType::MAX; + + arm_conv::pooling::PoolingWindow window{}; + window.cols = static_cast<unsigned int>(info.pool_size.x()); + window.rows = static_cast<unsigned int>(info.pool_size.y()); + + arm_conv::pooling::PoolingStride stride{}; + std::tie(stride.cols, stride.rows) = info.pad_stride_info.stride(); + + const arm_conv::pooling::PaddingValues padding{info.pad_stride_info.pad_left(), info.pad_stride_info.pad_top(), + info.pad_stride_info.pad_right(), info.pad_stride_info.pad_bottom()}; + + constexpr unsigned int idx_width = 1; + constexpr unsigned int idx_height = 2; + constexpr unsigned int idx_channels = 0; + constexpr unsigned int idx_batches = 3; + + const unsigned int n_batches = src->dimension(idx_batches); + const unsigned int src_rows = src->dimension(idx_height); + const unsigned int src_cols = src->dimension(idx_width); + const unsigned int n_channels = src->dimension(idx_channels); + const unsigned int dst_rows = dst->dimension(idx_height); + const unsigned int dst_cols = dst->dimension(idx_width); + + arm_conv::pooling::PoolingArgs args(&cpu_info, pool_type, window, stride, info.exclude_padding, n_batches, src_rows, + src_cols, n_channels, dst_rows, dst_cols, padding, nullptr); + + const auto src_qinfo = src->quantization_info().uniform(); + const auto dst_qinfo = dst->quantization_info().uniform(); + + const float multiplier = src_qinfo.scale / dst_qinfo.scale; + int32_t dst_multiplier{}; + int32_t dst_shift{}; + quantization::calculate_quantized_multiplier(multiplier, &dst_multiplier, &dst_shift); + + const arm_conv::pooling::Requantize32 requant_args(src_qinfo.offset, dst_qinfo.offset, + dst_shift, // left shift + 0, // right shift + dst_multiplier); + + // Configure assembly pooling kernel with requantization + auto pooling_kernel_asm = + arm_conv::pooling::pooling<Typesrc, Typedst, arm_conv::pooling::Requantize32>(args, requant_args); + if (pooling_kernel_asm == nullptr) + { + // Configuration not supported: Leave function unconfigured: + return; + } + + _kernel_asm = std::move(pooling_kernel_asm); +} + +size_t CpuPool2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const +{ + ARM_COMPUTE_UNUSED(thread_count); + ARM_COMPUTE_UNUSED(platform); + + return ICPPKernel::default_mws; +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h new file mode 100644 index 0000000000..b4ff1e6f2d --- /dev/null +++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h @@ -0,0 +1,134 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_CPU_POOL2D_ASSEMBLY_WRAPPER_KERNEL_H +#define ARM_COMPUTE_CPU_POOL2D_ASSEMBLY_WRAPPER_KERNEL_H + +#include "arm_compute/core/Types.h" + +#include "src/core/common/Macros.h" +#include "src/core/NEON/kernels/assembly/pooling.hpp" +#include "src/cpu/ICpuKernel.h" +#include "src/cpu/kernels/CpuKernelSelectionTypes.h" + +#include "pool_common.hpp" + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +/** This class is a wrapper for the assembly kernels. + * + * Some kernels were written in assembly and highly optimised for specific + * CPUs like A53 or A55. The arm compute library creates an instance of + * CpuPool2dAssemblyWrapperKernel and other auxiliary data structures to + * execute a single assembly kernel in the context of an NEFunction. + * + */ +class CpuPool2dAssemblyWrapperKernel final : public ICpuKernel<CpuPool2dAssemblyWrapperKernel> +{ +public: + /** Constructor + */ + CpuPool2dAssemblyWrapperKernel() = default; + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuPool2dAssemblyWrapperKernel); + + const char *name() const override + { + return "CpuPool2dAssemblyWrapperKernel"; + } + + /** Initialise the kernel's src and dst. + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[out] dst Destination tensor info to store the result of pooling. Data types supported: same as @p src. + * @param[in] info Pooling meta-data. + * @param[in] cpu_info CPU information needed to select the most appropriate kernel. + */ + void configure(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info); + + /** Static function to check if given info will lead to a valid configuration + * + * Similar to CpuPool2dAssemblyWrapperKernel::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + + /** Get size of the workspace needed by the assembly kernel. + * + * @param[in] num_threads Maximum number of threads that are going to be spawned. + * + * @return size of workspace + */ + size_t get_working_size(unsigned int num_threads) const; + + /** Was the asm kernel successfully configured? + * + * @return True if the asm kernel is configured and ready to run + */ + bool is_configured() const; + +private: + /** Helper function to create the assembly kernel. + * + * @param[in] src Source tensor info. + * @param[in] dst Destination tensor info. + * @param[in] info Pooling layer meta-data. + */ + template <typename Typesrc, typename Typedst> + void + create_arm_pooling(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info); + + /** Helper function to create the assembly kernel with requantization support + * + * @param[in] src Source tensor info. + * @param[in] dst Destination tensor info. + * @param[in] info Pooling layer meta-data. + */ + template <typename Typesrc, typename Typedst> + void create_arm_pooling_requant(const ITensorInfo *src, + ITensorInfo *dst, + const PoolingLayerInfo &info, + const CPUInfo &cpu_info); + + std::unique_ptr<arm_conv::pooling::IPoolingCommon> _kernel_asm{nullptr}; + + /** Return minimum workload size of the relevant kernel + * + * @param[in] platform The CPU platform used to create the context. + * @param[in] thread_count Number of threads in the execution. + * + * @return[out] small_network_mws Minimum workload size for requsted configuration. + */ + size_t get_mws(const CPUInfo &platform, size_t thread_count) const override; +}; +} // namespace kernels +} // namespace cpu +} // namespace arm_compute +#endif /* ARM_COMPUTE_CPU_POOL2D_ASSEMBLY_WRAPPER_KERNEL_H */ |