diff options
author | Giorgio Arena <giorgio.arena@arm.com> | 2019-10-15 11:09:33 +0100 |
---|---|---|
committer | Giorgio Arena <giorgio.arena@arm.com> | 2019-10-21 10:14:20 +0000 |
commit | d93e263e70e3101422402c95946e520fef34c4c7 (patch) | |
tree | f79d3b325ed6881fb9252cb7ee0b7573739e00be /src/core/NEON/kernels | |
parent | ab5b1a279284bed350d3bb75f3d9d3aec6edca0e (diff) | |
download | ComputeLibrary-d93e263e70e3101422402c95946e520fef34c4c7.tar.gz |
COMPMID-2708 NEDepthwiseConvolution Generic: support for QUANT8_PER_CHANNEL_SYMM
COMPMID-2470 Implement a new and generic depthwise convolution for NEON QASYMM8 NHWC
COMPMID-2477 Enable FP16 data type for the new generic convolution on NEON for NHWC
COMPMID-2625 Remove old implementation files for the generic NEDepthwiseConvolution
Change-Id: I8f6deda4fc69dd7e472fba3228b1ed5dad172f3e
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2094
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels')
5 files changed, 316 insertions, 540 deletions
diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp index c9d4e9be50..a0d45afd2a 100644 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp @@ -24,19 +24,30 @@ #include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h" #include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/CPP/Validate.h" #include "arm_compute/core/NEON/wrapper/traits.h" #include "arm_compute/core/NEON/wrapper/wrapper.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" - -#include "support/ToolchainSupport.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "src/core/NEON/kernels/convolution/depthwise/impl_qa8_qa8.hpp" namespace arm_compute { namespace { +void pad_vectors(std::vector<int> &mult, std::vector<int> &shift, int vec_size) +{ + ARM_COMPUTE_ERROR_ON(mult.size() != shift.size()); + while(mult.size() % vec_size != 0) + { + mult.push_back(0); + shift.push_back(0); + } +} + template <typename T, int S, bool has_biases> -void depthwise_loop_multiplier1(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - const Size2D &dilation, const Window &window) +void depthwise_loop_multiplier1_fp(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + const Size2D &dilation, const Window &window) { using VectorType = typename wrapper::traits::neon_vector<T, S>::type; using TagType = typename wrapper::traits::neon_vector<T, S>::tag_type; @@ -108,8 +119,8 @@ void depthwise_loop_multiplier1(const ITensor *input, const ITensor *weights, co } template <typename T, bool has_biases> -void depthwise_loop_generic(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - const Size2D &dilation, unsigned int depth_multiplier, const Window &window) +void depthwise_loop_generic_fp(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + const Size2D &dilation, unsigned int depth_multiplier, const Window &window) { const size_t input_stride_y = input->info()->strides_in_bytes().y(); const size_t input_stride_z = input->info()->strides_in_bytes().z(); @@ -191,21 +202,243 @@ void depthwise_loop_generic(const ITensor *input, const ITensor *weights, const input_it, weights_it, biases_it, output_it); } +template <typename T, typename TW, int S, bool has_biases, bool is_per_channel> +void depthwise_loop_multiplier1_quantized(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + const Size2D &dilation, std::vector<int> output_multiplier, std::vector<int> output_shift, const Window &window) +{ + using VectorType = typename wrapper::traits::neon_vector<T, S>::type; + using TagType = typename wrapper::traits::neon_vector<T, S>::tag_type; + + const size_t input_stride_y = input->info()->strides_in_bytes().y(); + const size_t input_stride_z = input->info()->strides_in_bytes().z(); + const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * + input->info()->strides_in_bytes().y(); + const size_t weights_width = weights->info()->dimension(1); + const size_t weights_height = weights->info()->dimension(2); + const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); + const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); + const size_t conv_stride_x = conv_info.stride().first; + const size_t conv_stride_y = conv_info.stride().second; + const size_t conv_pad_left = conv_info.pad_left(); + const size_t conv_pad_top = conv_info.pad_top(); + + const int32_t input_qoffset = input->info()->quantization_info().uniform().offset; + const int32_t weights_qoffset = weights->info()->quantization_info().uniform().offset; + const int32_t output_qoffset = output->info()->quantization_info().uniform().offset; + const int32_t k_offset = weights_width * weights_height * input_qoffset * weights_qoffset; + + Window win_input = window; + win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); + + Window win_weights = win_input; + win_weights.set(3, Window::Dimension(0, 0, 0)); + + Iterator input_it(input, win_input); + Iterator weights_it(weights, win_weights); + Iterator output_it(output, window); + Iterator biases_it{}; + + if(has_biases) + { + biases_it = Iterator(biases, win_weights); + } + + execute_window_loop(window, [&](const Coordinates & id) + { + std::vector<int32_t> acc(S, 0); + std::vector<int32_t> in_sum(S, 0); + std::vector<int32_t> we_sum(S, 0); + + const int input_y = id.y() * conv_stride_x - conv_pad_left; + const int input_z = id.z() * conv_stride_y - conv_pad_top; + int input_offset = input_y * input_stride_y + input_z * input_stride_z; + + auto weights_ptr = weights_it.ptr(); + for(size_t h = 0; h < weights_height; ++h) + { + int offs = input_offset; + for(size_t w = 0; w < weights_width; ++w) + { + const auto input_vals = wrapper::vload(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), input_max_offset))); + const auto weights_vals = wrapper::vload(reinterpret_cast<TW *>(weights_ptr + w * weights_stride_y)); + + for(int i = 0; i < S; ++i) + { + acc.at(i) += input_vals[i] * weights_vals[i]; + in_sum.at(i) += input_vals[i]; + we_sum.at(i) += weights_vals[i]; + } + + offs += dilation.x() * input_stride_y; + } + + weights_ptr += weights_stride_z; + input_offset += dilation.y() * input_stride_z; + } + + VectorType out_vals = wrapper::vdup_n(static_cast<T>(0), TagType{}); + for(int i = 0; i < S; ++i) + { + acc.at(i) -= in_sum.at(i) * weights_qoffset; + acc.at(i) -= we_sum.at(i) * input_qoffset; + acc.at(i) += k_offset; + + if(has_biases) + { + acc.at(i) += *reinterpret_cast<int32_t *>(biases_it.ptr() + i * sizeof(int32_t)); + } + + acc.at(i) = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(i), output_multiplier.at(id.x() + i)), output_shift.at(id.x() + i)) + output_qoffset; + out_vals[i] = static_cast<T>(utility::clamp<int32_t, uint8_t>(acc.at(i))); + } + + wrapper::vstore(reinterpret_cast<T *>(output_it.ptr()), out_vals); + }, + input_it, weights_it, biases_it, output_it); +} + +template <typename T, typename TW, bool has_biases, bool is_per_channel> +void depthwise_loop_generic_quantized(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + const Size2D &dilation, unsigned int depth_multiplier, std::vector<int> output_multiplier, std::vector<int> output_shift, const Window &window) +{ + const size_t input_stride_y = input->info()->strides_in_bytes().y(); + const size_t input_stride_z = input->info()->strides_in_bytes().z(); + const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * + input->info()->strides_in_bytes().y(); + const size_t weights_width = weights->info()->dimension(1); + const size_t weights_height = weights->info()->dimension(2); + const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); + const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); + const size_t conv_stride_x = conv_info.stride().first; + const size_t conv_stride_y = conv_info.stride().second; + const size_t conv_pad_left = conv_info.pad_left(); + const size_t conv_pad_top = conv_info.pad_top(); + + const int32_t input_qoffset = input->info()->quantization_info().uniform().offset; + const int32_t weights_qoffset = weights->info()->quantization_info().uniform().offset; + const int32_t output_qoffset = output->info()->quantization_info().uniform().offset; + const int32_t k_offset = weights_width * weights_height * input_qoffset * weights_qoffset; + + Window win_input = window; + win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); + + Window win_weights = win_input; + win_weights.set(3, Window::Dimension(0, 0, 0)); + + win_input.set_dimension_step(Window::DimX, 1); + + Iterator input_it(input, win_input); + Iterator weights_it(weights, win_weights); + Iterator output_it(output, window); + Iterator biases_it{}; + + if(has_biases) + { + biases_it = Iterator(biases, win_weights); + } + + execute_window_loop(window, [&](const Coordinates & id) + { + std::vector<int32_t> acc(depth_multiplier, 0); + std::vector<int32_t> we_sum(depth_multiplier, 0); + int32_t in_sum = 0; + + const int input_y = id.y() * conv_stride_x - conv_pad_left; + const int input_z = id.z() * conv_stride_y - conv_pad_top; + int input_offset = input_y * input_stride_y + input_z * input_stride_z; + + auto weights_ptr = weights_it.ptr(); + for(size_t h = 0; h < weights_height; ++h) + { + int offs = input_offset; + for(size_t w = 0; w < weights_width; ++w) + { + const auto input_val = *(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), input_max_offset))); + + for(size_t m = 0; m < depth_multiplier; ++m) + { + const auto weights_val = *(reinterpret_cast<TW *>(weights_ptr + m * sizeof(T) + w * weights_stride_y)); + acc.at(m) += input_val * weights_val; + + we_sum.at(m) += weights_val; + } + + offs += dilation.x() * input_stride_y; + in_sum += input_val; + } + + weights_ptr += weights_stride_z; + input_offset += dilation.y() * input_stride_z; + } + + for(size_t m = 0; m < depth_multiplier; ++m) + { + acc.at(m) -= in_sum * weights_qoffset; + acc.at(m) -= we_sum.at(m) * input_qoffset; + acc.at(m) += k_offset; + + if(has_biases) + { + const auto biases_val = *(reinterpret_cast<int32_t *>(biases_it.ptr() + m * sizeof(int32_t))); + + int32_t out_val = acc.at(m) + biases_val; + out_val = rounding_divide_by_exp2(saturating_doubling_high_mul(out_val, output_multiplier.at(id.x() + m)), + output_shift.at(id.x() + m)) + + output_qoffset; + *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = static_cast<T>(utility::clamp<int32_t, uint8_t>(out_val)); + } + else + { + int32_t out_val = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(m), output_multiplier.at(id.x() + m)), + output_shift.at(id.x() + m)) + + output_qoffset; + *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = static_cast<T>(utility::clamp<int32_t, uint8_t>(out_val)); + } + } + }, + input_it, weights_it, biases_it, output_it); +} + Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier == 0); + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) + (weights->dimension(1) - 1) * (dilation.x() - 1) > input->dimension(1) + conv_info.pad_left() + conv_info.pad_right()); + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) + (weights->dimension(2) - 1) * (dilation.y() - 1) > input->dimension(2) + conv_info.pad_top() + conv_info.pad_bottom()); ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(0) * depth_multiplier) != weights->dimension(0)); ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); ARM_COMPUTE_RETURN_ERROR_ON((conv_info.stride().first < 1) || (conv_info.stride().second < 1)); + 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_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != weights->quantization_info().scale().size()); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); + } + if(biases != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0)); + + if(is_data_type_quantized_asymmetric(input->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + } } if(output->total_size() != 0) @@ -216,7 +449,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, return Status{}; } -} // namespace std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &conv_info, @@ -226,7 +458,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output, input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape)); + auto_init_if_empty(*output, input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape).set_quantization_info(output->quantization_info())); // Configure kernel window (generic) const unsigned int num_elems_read_per_iteration = (depth_multiplier == 1) ? 8 / element_size_from_data_type(input->data_type()) : 1; @@ -253,9 +485,10 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; return std::make_pair(err, win); } +} // namespace NEDepthwiseConvolutionLayerNativeKernel::NEDepthwiseConvolutionLayerNativeKernel() - : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation() + : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation(), _output_multiplier(), _output_shift() { } @@ -279,10 +512,56 @@ void NEDepthwiseConvolutionLayerNativeKernel::configure(const ITensor *input, co _border_size = BorderSize(_conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0); _dilation = dilation; - switch(_input->info()->data_type()) + if(is_data_type_quantized(_input->info()->data_type())) + { + const auto input_scale = input->info()->quantization_info().uniform().scale; + const auto output_scale = output->info()->quantization_info().uniform().scale; + + auto weights_scale = weights->info()->quantization_info().scale(); + if(!is_data_type_quantized_per_channel(_weights->info()->data_type())) + { + for(size_t i = 1; i < _weights->info()->dimension(0); ++i) + { + weights_scale.push_back(weights_scale.front()); + } + } + + for(size_t i = 0; i < weights_scale.size(); ++i) + { + int out_mult = 0; + int out_shift = 0; + const float multiplier = input_scale * weights_scale.at(i) / output_scale; + ARM_COMPUTE_ERROR_ON(multiplier > 1.f); + arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &out_mult, &out_shift); + + _output_multiplier.push_back(out_mult); + _output_shift.push_back(out_shift); + } + } + + switch(_weights->info()->data_type()) { + case DataType::QASYMM8: + _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<uint8_t, uint8_t, 8, true, false> : + &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<uint8_t, uint8_t, 8, false, false>; + pad_vectors(_output_multiplier, _output_shift, 8); + break; + case DataType::QSYMM8_PER_CHANNEL: + _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<uint8_t, int8_t, 8, true, true> : + &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<uint8_t, int8_t, 8, false, true>; + pad_vectors(_output_multiplier, _output_shift, 8); + break; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float16_t, float16_t, 4, true, false> : + &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float16_t, float16_t, 4, false, false>; + pad_vectors(_output_multiplier, _output_shift, 4); + break; +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F32: - _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float, 2, true> : &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float, 2, false>; + _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float, float, 2, true, false> : + &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float, float, 2, false, false>; + pad_vectors(_output_multiplier, _output_shift, 2); break; default: ARM_COMPUTE_ERROR("Data type not supported"); @@ -314,7 +593,28 @@ void NEDepthwiseConvolutionLayerNativeKernel::run(const Window &window, const Th (this->*_func)(window); } -template <typename T, int S, bool has_biases> +template < typename T, typename TW, int S, bool has_biases, bool is_per_channel, typename std::enable_if < std::is_same<T, float>::value +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + || std::is_same<T, float16_t>::value +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + , + int >::type > +void NEDepthwiseConvolutionLayerNativeKernel::run_depthwise(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + + if(_depth_multiplier == 1) + { + depthwise_loop_multiplier1_fp<T, S, has_biases>(_input, _weights, _biases, _output, _conv_info, _dilation, window); + } + else + { + depthwise_loop_generic_fp<T, has_biases>(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); + } +} + +template <typename T, typename TW, int S, bool has_biases, bool is_per_channel, typename std::enable_if<std::is_same<T, uint8_t>::value, int>::type> void NEDepthwiseConvolutionLayerNativeKernel::run_depthwise(const Window &window) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); @@ -322,11 +622,11 @@ void NEDepthwiseConvolutionLayerNativeKernel::run_depthwise(const Window &window if(_depth_multiplier == 1) { - depthwise_loop_multiplier1<T, S, has_biases>(_input, _weights, _biases, _output, _conv_info, _dilation, window); + depthwise_loop_multiplier1_quantized<T, TW, S, has_biases, is_per_channel>(_input, _weights, _biases, _output, _conv_info, _dilation, _output_multiplier, _output_shift, window); } else { - depthwise_loop_generic<T, has_biases>(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); + depthwise_loop_generic_quantized<T, TW, has_biases, is_per_channel>(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, _output_multiplier, _output_shift, window); } } } // namespace arm_compute diff --git a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp deleted file mode 100644 index 53789e2472..0000000000 --- a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp +++ /dev/null @@ -1,197 +0,0 @@ -/* - * Copyright (c) 2017-2019 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 "arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h" - -#include "arm_compute/core/Coordinates.h" -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/NEON/INEKernel.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/TensorShape.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" - -using namespace arm_compute; - -namespace -{ -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, - const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier, const Size2D &dilation) -{ - ARM_COMPUTE_UNUSED(conv_info); - //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && has_bias); - ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(2) * depth_multiplier) != output->dimension(2)); - ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0))); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); - ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || dilation.y() < 1); - - return Status{}; -} -} // namespace - -template <typename T> -void NEDepthwiseIm2ColKernel::run_generic(const Window &window) -{ - const int input_w = _input->info()->dimension(0); - const int input_h = _input->info()->dimension(1); - const int input_stride_x = _input->info()->strides_in_bytes().x(); - const int input_stride_y = _input->info()->strides_in_bytes().y(); - const int input_stride_z = _input->info()->strides_in_bytes().z(); - const int stride_x = _conv_info.stride().first; - const int stride_y = _conv_info.stride().second; - - const int pad_left = _conv_info.pad_left(); - const int pad_right = _conv_info.pad_right(); - const int pad_top = _conv_info.pad_top(); - - Window window_in(window); - // The first three dimensions of the input are increased by the inner loops - window_in.set(Window::DimX, Window::Dimension(0, 0, 0)); - window_in.set(Window::DimY, Window::Dimension(0, 0, 0)); - window_in.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - // Setup output window - Window window_out(window); - window_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _output->info()->dimension(0))); - window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1)); - window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), 1)); - - Iterator in(_input, window_in); - Iterator out(_output, window_out); - - const int full_length = input_w + pad_left + pad_right; - const int max_initial_x = stride_x * (((full_length - (_kernel_dims.width + (_kernel_dims.width - 1) * (_dilation.x() - 1))) / stride_x) + 1); - - // Define pad value - auto zero = static_cast<T>(0); - if(std::is_same<T, uint8_t>::value) - { - zero = _input->info()->quantization_info().uniform().offset; - } - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int src_pixel_linear = id.y() * stride_x; - - const int src_x = -pad_left + src_pixel_linear % max_initial_x; - const int src_y = -pad_top + src_pixel_linear / max_initial_x * stride_y; - - // Get pointers - const uint8_t *const input_ptr = in.ptr() + id.z() / _depth_multiplier * input_stride_z; - auto output_ptr = reinterpret_cast<T *>(out.ptr()); - const int height = src_y + (_kernel_dims.height + (_kernel_dims.height - 1) * (_dilation.y() - 1)); - const int width = src_x + (_kernel_dims.width + (_kernel_dims.width - 1) * (_dilation.x() - 1)); - - for(int y = src_y; y < height; y += _dilation.y()) - { - for(int x = src_x; x < width; x += _dilation.x(), ++output_ptr) - { - if(x < 0 || x >= input_w || y < 0 || y >= input_h) - { - *output_ptr = zero; - } - else - { - *output_ptr = *(reinterpret_cast<const T *>(input_ptr + x * input_stride_x + y * input_stride_y)); - } - } - } - - if(_has_bias) - { - *output_ptr = static_cast<T>(1); - } - }, - in, out); -} - -NEDepthwiseIm2ColKernel::NEDepthwiseIm2ColKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias(), _depth_multiplier(1), _dilation() -{ -} - -void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier, - const Size2D &dilation) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier, dilation)); - - _input = input; - _output = output; - _kernel_dims = kernel_dims; - _conv_info = conv_info; - _has_bias = has_bias; - _depth_multiplier = depth_multiplier; - _dilation = dilation; - - // Configure kernel window - Window win = calculate_max_window(*input->info(), Steps()); - - // Set appropriate function to run - switch(input->info()->data_type()) - { - case DataType::QASYMM8: - _func = &NEDepthwiseIm2ColKernel::run_generic<uint8_t>; - break; - case DataType::F16: - _func = &NEDepthwiseIm2ColKernel::run_generic<half>; - break; - case DataType::F32: - _func = &NEDepthwiseIm2ColKernel::run_generic<float>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported data type"); - } - - // The NEDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped - output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); - - INEKernel::configure(win); -} - -Status NEDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier, - const Size2D &dilation) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier, dilation)); - return Status{}; -} - -void NEDepthwiseIm2ColKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - - if(_func != nullptr) - { - (this->*_func)(window); - } -} diff --git a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp deleted file mode 100644 index 37269cafaf..0000000000 --- a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp +++ /dev/null @@ -1,156 +0,0 @@ -/* - * Copyright (c) 2017-2019 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 "arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h" - -#include "arm_compute/core/CPP/Validate.h" -#include "arm_compute/core/Coordinates.h" -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/NEON/INEKernel.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/TensorShape.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" - -using namespace arm_compute; -using namespace arm_compute::misc::shape_calculator; - -namespace -{ -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h) -{ - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32); - - if(output->total_size() != 0) - { - TensorShape output_shape = compute_vector_to_tensor_output_shape(input->tensor_shape(), conv_w, conv_h, output->data_layout()); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); - } - - return Status{}; -} -} // namespace - -template <typename T> -void NEDepthwiseVectorToTensorKernel::vector_to_tensor(const Window &window) -{ - // const int input_w = _input->info()->dimension(0); - const int output_stride_x = _output->info()->strides_in_bytes().x(); - const int output_stride_y = _output->info()->strides_in_bytes().y(); - const int output_stride_z = _output->info()->strides_in_bytes().z(); - - // Setup output window - Window window_out(window); - window_out.set(Window::DimX, Window::Dimension(0, 0, 0)); - window_out.set(Window::DimY, Window::Dimension(0, 0, 0)); - window_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - Iterator in(_input, window); - Iterator out(_output, window_out); - - const int patch_size = _conv_dims.first * _conv_dims.second; - - execute_window_loop(window, [&](const Coordinates & id) - { - const int z = id.x() / patch_size; - const int index2D = id.x() - z * patch_size; - - auto input_ptr = reinterpret_cast<T *>(in.ptr()); - auto output_ptr = reinterpret_cast<T *>(out.ptr() + index2D % _conv_dims.first * output_stride_x + index2D / _conv_dims.first * output_stride_y + z * output_stride_z); - - *output_ptr = *input_ptr; - }, - in, out); -} - -NEDepthwiseVectorToTensorKernel::NEDepthwiseVectorToTensorKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _conv_dims() -{ -} - -void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *output, size_t conv_w, size_t conv_h) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - // Output auto inizialitation if not yet initialized - TensorShape output_shape = compute_vector_to_tensor_output_shape(input->info()->tensor_shape(), conv_w, conv_h, output->info()->data_layout()); - auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), conv_w, conv_h)); - - _input = input; - _output = output; - _conv_dims = std::pair<size_t, size_t>(conv_w, conv_h); - - // Set appropriate function to run - switch(input->info()->data_type()) - { - case DataType::QASYMM8: - _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<uint8_t>; - break; - case DataType::S32: - _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<int32_t>; - break; - case DataType::F16: - _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<half>; - break; - case DataType::F32: - _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<float>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported data type"); - } - - // Configure kernel window - Window win = calculate_max_window(*input->info(), Steps()); - // The NEDepthwisevectorToTensorKernel doesn't need padding so update_window_and_padding() can be skipped - output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); - - INEKernel::configure(win); -} - -Status NEDepthwiseVectorToTensorKernel::validate(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, conv_w, conv_h)); - return Status{}; -} - -void NEDepthwiseVectorToTensorKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - - if(_func != nullptr) - { - (this->*_func)(window); - } -} diff --git a/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp deleted file mode 100644 index b0e1fcb3f8..0000000000 --- a/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp +++ /dev/null @@ -1,165 +0,0 @@ -/* - * Copyright (c) 2017-2019 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 "arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h" - -#include "arm_compute/core/Coordinates.h" -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/NEON/INEKernel.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/TensorShape.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" - -using namespace arm_compute; - -namespace -{ -template <typename T> -void weights_reshape(const ITensor *input, const ITensor *bias, ITensor *output, const Window &window) -{ - const int input_w = input->info()->dimension(0); - const int output_stride_x = output->info()->strides_in_bytes().x(); - const int output_stride_y = output->info()->strides_in_bytes().y(); - - Window window_in(window); - // The first three dimensions of the input are increased by the inner loops - window_in.set(Window::DimX, Window::Dimension(0, input->info()->dimension(0), input->info()->dimension(0))); - window_in.set(Window::DimY, Window::Dimension(0, input->info()->dimension(1), 1)); - window_in.set(Window::DimZ, Window::Dimension(0, input->info()->dimension(2), 1)); - - // Setup output window - Window window_out; - window_out.set(Window::DimX, Window::Dimension(0, 0, 0)); - window_out.set(Window::DimY, Window::Dimension(0, 0, 0)); - - Iterator in(input, window_in); - Iterator out(output, window_out); - - execute_window_loop(window_in, [&](const Coordinates & id) - { - auto input_ptr = reinterpret_cast<T *>(in.ptr()); - auto output_ptr = reinterpret_cast<T *>(out.ptr() + id.y() * input_w * output_stride_x + id.z() * output_stride_y); - - for(int i = 0; i < input_w; ++i, ++input_ptr) - { - *(output_ptr + i) = *input_ptr; - } - - if(bias != nullptr) - { - *(output_ptr + input_w) = *(reinterpret_cast<T *>(bias->ptr_to_element(Coordinates(id.z())))); - } - }, - in, out); -} - -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases) -{ - //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (biases != nullptr)); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != output->dimension(1)); - ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(0) * input->dimension(1) + ((biases != nullptr) ? 1 : 0))); - - if(biases != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(2)); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); - } - - return Status{}; -} -} // namespace - -NEDepthwiseWeightsReshapeKernel::NEDepthwiseWeightsReshapeKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _biases(nullptr) -{ -} - -void NEDepthwiseWeightsReshapeKernel::configure(const ITensor *input, ITensor *output, const ITensor *biases) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), (biases != nullptr) ? biases->info() : nullptr)); - - _input = input; - _output = output; - _biases = biases; - - switch(_input->info()->element_size()) - { - case 4: - { - _func = &weights_reshape<uint32_t>; - break; - } - case 2: - { - _func = &weights_reshape<uint16_t>; - break; - } - case 1: - { - _func = &weights_reshape<uint8_t>; - break; - } - default: - { - ARM_COMPUTE_ERROR_ON("Element size not supported"); - break; - } - } - - // Configure kernel window - Window win = calculate_max_window(*input->info(), Steps()); - // The NEDepthwiseWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped - output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); - - INEKernel::configure(win); -} - -Status NEDepthwiseWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, biases)); - return Status{}; -} - -void NEDepthwiseWeightsReshapeKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - - if(_func != nullptr) - { - (*_func)(_input, _biases, _output, window); - } -} diff --git a/src/core/NEON/kernels/NEPermuteKernel.cpp b/src/core/NEON/kernels/NEPermuteKernel.cpp index 1df94aef06..897b764b45 100644 --- a/src/core/NEON/kernels/NEPermuteKernel.cpp +++ b/src/core/NEON/kernels/NEPermuteKernel.cpp @@ -91,12 +91,6 @@ inline bool is_permutation_supported(const PermutationVector &v) Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm) { - //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, - DataType::U16, DataType::S16, - DataType::U32, DataType::S32, - DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_permutation_supported(perm), "PermutationVector not supported."); const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input, perm); |