From 900289936c458eff95499e0a0eaba989a27aaa4d Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Wed, 30 Jun 2021 18:29:18 +0100 Subject: Port NEIm2ColKernel Resolves: COMPMID-4510 Change-Id: Ia3e588f599449d975dabad4afafb2974dd44d0ad Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5899 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- src/core/cpu/kernels/CpuIm2ColKernel.cpp | 448 +++++++++++++++++++++++++++++++ src/core/cpu/kernels/CpuIm2ColKernel.h | 123 +++++++++ 2 files changed, 571 insertions(+) create mode 100644 src/core/cpu/kernels/CpuIm2ColKernel.cpp create mode 100644 src/core/cpu/kernels/CpuIm2ColKernel.h (limited to 'src/core/cpu/kernels') diff --git a/src/core/cpu/kernels/CpuIm2ColKernel.cpp b/src/core/cpu/kernels/CpuIm2ColKernel.cpp new file mode 100644 index 0000000000..a5dbcc29c8 --- /dev/null +++ b/src/core/cpu/kernels/CpuIm2ColKernel.cpp @@ -0,0 +1,448 @@ +/* + * Copyright (c) 2017-2021 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/core/cpu/kernels/CpuIm2ColKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Size2D.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.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 "arm_compute/core/utils/misc/ShapeCalculator.h" + +#include +#include +#include +#include +#include + +namespace arm_compute +{ +using namespace misc::shape_calculator; +namespace cpu +{ +namespace kernels +{ +namespace +{ +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, + bool has_bias, const Size2D &dilation, unsigned int num_groups) +{ + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::BFLOAT16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized(input->data_type()) && has_bias); + ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups > 1, "Number of groups greater than one are not supported on Neon"); + + // Since there's no implicit padding added, check the total input spatial dimensions (with conv paddings) are big enough for the kernel dimensions + const unsigned int width_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); + const unsigned int height_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); + const unsigned total_width = input->dimension(width_idx) + conv_info.pad_left() + conv_info.pad_right(); + const unsigned total_height = input->dimension(height_idx) + conv_info.pad_top() + conv_info.pad_bottom(); + ARM_COMPUTE_RETURN_ERROR_ON((total_width < kernel_dims.width) || (total_height < kernel_dims.height)); + + if(output->total_size() > 0) + { + TensorInfo expected_output = output->clone()->set_tensor_shape(compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation, false)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); + } + + return Status{}; +} + +template +inline void linearize_volume_nchw(const uint8_t *const in_ptr, + T *out_ptr, + bool has_bias, + int top_left_x, + int top_left_y, + int kernel_width, + int kernel_height, + int kernel_depth, + int input_w, + int input_h, + int input_stride_x, + int input_stride_y, + int input_stride_z, + int pad_value, + int dilation_x, + int dilation_y) +{ + const int kernel_size2 = kernel_width * kernel_height; + const int x_e = top_left_x + kernel_width * dilation_x; + const int y_e = top_left_y + kernel_height * dilation_y; + + // Linearize volume + int d = 0; + // This for loop linearize a volume with 3 slices. This allows: + // 1) to reduce the iterations of the outer for loop "d" + // 2) to have an optimized im2col for the first convolution layer where usually we have 3 IFMs + for(; d <= (kernel_depth - 3); d += 3) + { + for(int y = top_left_y; y < y_e; y += dilation_y) + { + if((y < 0 || y >= input_h) && has_pads) + { + // All the values will be the offset (will be zeros when not quantized) + for(int x = top_left_x; x < x_e; x += dilation_x, ++out_ptr) + { + *(out_ptr + 0 * kernel_size2) = pad_value; + *(out_ptr + 1 * kernel_size2) = pad_value; + *(out_ptr + 2 * kernel_size2) = pad_value; + } + } + else + { + for(int x = top_left_x; x < x_e; x += dilation_x, ++out_ptr) + { + if((x < 0 || x >= input_w) && has_pads) + { + *(out_ptr + 0 * kernel_size2) = pad_value; + *(out_ptr + 1 * kernel_size2) = pad_value; + *(out_ptr + 2 * kernel_size2) = pad_value; + } + else + { + *(out_ptr + 0 * kernel_size2) = *(reinterpret_cast(in_ptr + ((d + 0) * input_stride_z + y * input_stride_y + x * input_stride_x))); + *(out_ptr + 1 * kernel_size2) = *(reinterpret_cast(in_ptr + ((d + 1) * input_stride_z + y * input_stride_y + x * input_stride_x))); + *(out_ptr + 2 * kernel_size2) = *(reinterpret_cast(in_ptr + ((d + 2) * input_stride_z + y * input_stride_y + x * input_stride_x))); + } + } + } + } + out_ptr += 2 * kernel_size2; + } + + // Left over + for(; d < kernel_depth; d++) + { + for(int y = top_left_y; y < y_e; y += dilation_y) + { + if((y < 0 || y >= input_h) && has_pads) + { + // All the values will be the offset (will be zeros when not quantized) + memset(static_cast(out_ptr), pad_value, kernel_width * sizeof(T)); + out_ptr += kernel_width; + } + else + { + for(int x = top_left_x; x < x_e; x += dilation_x, ++out_ptr) + { + if((x < 0 || x >= input_w) && has_pads) + { + *out_ptr = pad_value; + } + else + { + *out_ptr = *(reinterpret_cast(in_ptr + (d * input_stride_z + y * input_stride_y + x * input_stride_x))); + } + } + } + } + } + + // Append 1 if the convolution layer has biases + if(has_bias) + { + *out_ptr = static_cast(1); + } +} + +template +inline void linearize_volume_nhwc(const uint8_t *const in_ptr, + T *out_ptr, + bool has_bias, + int start_x, + int start_y, + int kernel_width, + int kernel_height, + int input_w, + int input_h, + int input_c, + int input_stride_y, + int input_stride_z, + int pad_value, + int dilation_x, + int dilation_y) +{ + const int end_x = start_x + kernel_width * dilation_x; + const int end_y = start_y + kernel_height * dilation_y; + const int pad_quant = kernel_width * input_c; + const int element_size = static_cast(sizeof(T)); + if((start_y >= 0) && (end_y < input_h) && (start_x >= 0) && (end_x < input_w) && (dilation_x == 1) && (input_stride_y == input_c * element_size)) + { + for(int y = start_y; y < end_y; y += dilation_y) + { + //optimized for no dilation and no boundary pixels + memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + start_x * input_stride_y)), input_c * kernel_width * element_size); + out_ptr += input_c * kernel_width; + } + } + else + { + for(int y = start_y; y < end_y; y += dilation_y) + { + if(y < 0 || y >= input_h) + { + memset(static_cast(out_ptr), pad_value, pad_quant * element_size); + out_ptr += pad_quant; + } + else if(dilation_x > 1 || start_x < 0 || end_x >= input_w || input_stride_y != input_c * element_size) + { + for(int x = start_x; x < end_x; x += dilation_x) + { + if(x < 0 || x >= input_w) + { + memset(static_cast(out_ptr), pad_value, input_c * element_size); + out_ptr += input_c; + } + else + { + memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + x * input_stride_y)), input_c * element_size); + out_ptr += input_c; + } + } + } + else + { + //optimized for no dilation and no boundary pixels + memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + start_x * input_stride_y)), input_c * kernel_width * element_size); + out_ptr += input_c * kernel_width; + } + } + } + // Append 1 if the convolution layer has biases + if(has_bias) + { + *out_ptr = static_cast(1); + } +} +} // namespace + +template +void CpuIm2ColKernel::run_im2col(const ITensor *src, ITensor *dst, const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); + + const unsigned int width_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); + const unsigned int height_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); + const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); + + const int input_w = src->info()->dimension(width_idx); + const int input_h = src->info()->dimension(height_idx); + const int input_c = src->info()->dimension(channel_idx); + const int input_stride_x = src->info()->strides_in_bytes().x(); + const int input_stride_y = src->info()->strides_in_bytes().y(); + const int input_stride_z = src->info()->strides_in_bytes().z(); + const int pad_left = _conv_info.pad_left(); + const int pad_top = _conv_info.pad_top(); + const int stride_x = _conv_info.stride().first; + const int stride_y = _conv_info.stride().second; + const int pad_value = is_data_type_quantized(src->info()->data_type()) ? src->info()->quantization_info().uniform().offset : 0; + + Window window_in_out(window); + // The first three dimensions of the input and output are increased by the inner loops + window_in_out.set(Window::DimX, Window::Dimension(0, 0, 0)); + window_in_out.set(Window::DimY, Window::Dimension(0, 0, 0)); + window_in_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); + + // Create iterators + Iterator in(src, window_in_out); + Iterator out(dst, window_in_out); + + execute_window_loop(window, [&](const Coordinates & id) + { + const int start_w = id[width_idx] * stride_x - pad_left; + const int start_h = id[height_idx] * stride_y - pad_top; + + // Get pointers + const uint8_t *const input_ptr = in.ptr(); + auto output_ptr = reinterpret_cast(out.ptr() + (id[width_idx] + id[height_idx] * _convolved_dims.first) * dst->info()->strides_in_bytes().y()); + + // Linearize volume + if(is_nchw) + { + linearize_volume_nchw(input_ptr, + output_ptr, + _has_bias, + start_w, + start_h, + _kernel_width, + _kernel_height, + input_c, + input_w, + input_h, + input_stride_x, + input_stride_y, + input_stride_z, + pad_value, + _dilation.x(), + _dilation.y()); + } + else + { + linearize_volume_nhwc(input_ptr, + output_ptr, + _has_bias, + start_w, + start_h, + _kernel_width, + _kernel_height, + input_w, + input_h, + input_c, + input_stride_y, + input_stride_z, + pad_value, + _dilation.x(), + _dilation.y()); + } + }, + in, out); +} + +void CpuIm2ColKernel::configure(ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, + bool has_bias, const Size2D &dilation, unsigned int num_groups) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, kernel_dims, conv_info, has_bias, dilation, num_groups)); + ARM_COMPUTE_UNUSED(num_groups); + + _data_layout = src->data_layout(); + const unsigned int width_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); + const unsigned int height_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); + const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); + + _conv_info = conv_info; + _kernel_width = kernel_dims.width; + _kernel_height = kernel_dims.height; + _dilation = dilation; + _convolved_dims = scaled_dimensions(src->dimension(width_idx), dst->dimension(height_idx), + _kernel_width, _kernel_height, + _conv_info, _dilation); + _has_bias = has_bias; + + if(_data_layout == DataLayout::NCHW) + { + switch(src->data_type()) + { + case DataType::F32: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; +#if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) + case DataType::BFLOAT16: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; +#endif /* defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) */ +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + case DataType::QASYMM8_SIGNED: + case DataType::QASYMM8: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; + default: + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + } + else + { + switch(src->data_type()) + { + case DataType::F32: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; +#if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) + case DataType::BFLOAT16: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; +#endif /* defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) */ +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + case DataType::QASYMM8: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; + case DataType::QASYMM8_SIGNED: + _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col : &CpuIm2ColKernel::run_im2col; + break; + default: + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + } + + // Output tensor auto initialization if not yet initialized + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_im2col_conv_shape(src, kernel_dims, conv_info, has_bias, dilation, false))); + + std::pair convolved_dims = scaled_dimensions(src->dimension(width_idx), src->dimension(height_idx), + kernel_dims.width, kernel_dims.height, + conv_info, dilation); + + Window win = calculate_max_window(*src, Steps()); + win.set(width_idx, Window::Dimension(0, convolved_dims.first, 1)); + win.set(height_idx, Window::Dimension(0, convolved_dims.second, 1)); + win.set(channel_idx, Window::Dimension(0, 1, 1)); + // Configure kernel window + ICpuKernel::configure(win); +} + +Status CpuIm2ColKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, + bool has_bias, const Size2D &dilation, unsigned int num_groups) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, kernel_dims, conv_info, has_bias, dilation, num_groups)); + return Status{}; +} + +void CpuIm2ColKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); + + auto src = tensors.get_const_tensor(TensorType::ACL_SRC); + auto dst = tensors.get_tensor(TensorType::ACL_DST); + (this->*_func)(src, dst, window); +} +const char *CpuIm2ColKernel::name() const +{ + return "CpuIm2ColKernel"; +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute \ No newline at end of file diff --git a/src/core/cpu/kernels/CpuIm2ColKernel.h b/src/core/cpu/kernels/CpuIm2ColKernel.h new file mode 100644 index 0000000000..4301a237fe --- /dev/null +++ b/src/core/cpu/kernels/CpuIm2ColKernel.h @@ -0,0 +1,123 @@ +/* + * Copyright (c) 2017-2021 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_IM2COL_KERNEL_H +#define ARM_COMPUTE_CPU_IM2COL_KERNEL_H + +#include "arm_compute/core/Size2D.h" +#include "src/core/common/Macros.h" +#include "src/core/cpu/ICpuKernel.h" + +namespace arm_compute +{ +class ITensor; +namespace cpu +{ +namespace kernels +{ +/** Interface for the im2col reshape kernel. + * + * Rearranges image blocks into columns. It is used to strip out each convolution block to a single column. + * It is used to transform a convolution to a plain matrix multiplication. + * + * For example taking into account the image below and assuming 3x3 image blocks with stride of 1 we have: + * + * @f[ + * \left( \begin{array}{cccc} + * a00 & a01 & a02 & a03 \\ + * a10 & a11 & a12 & a13 \\ + * a20 & a21 & a22 & a23 \\ + * a30 & a31 & a32 & a33 \\ + * \end{array} \right) + * \rightarrow + * \left( \begin{array}{ccccccccc} + * a00 & a01 & a02 & a10 & a11 & a12 & a20 & a21 & a22 \\ + * a01 & a02 & a03 & a11 & a12 & a13 & a21 & a22 & a23 \\ + * a10 & a11 & a12 & a20 & a21 & a22 & a30 & a31 & a32 \\ + * a11 & a12 & a13 & a21 & a22 & a23 & a31 & a32 & a33 \\ + * \end{array} \right) + * @f] + */ +class CpuIm2ColKernel : public ICpuKernel +{ +public: + /** Default constructor */ + CpuIm2ColKernel() = default; + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuIm2ColKernel); + /** Set the input and output of the kernel. + * + * @param[in] src The input tensor info to convert. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32 + * Note: QASYMM8/QASYMM8_SIGNED works only for has_bias = false + * @param[out] dst The output tensor info. Data types supported: Same as @p input + * @param[in] kernel_dims The kernel dimensions (width and height). + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] has_bias In case biases are provided expands the matrix with 1. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported + */ + void configure(ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, + bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to CpuIm2ColKernel::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, + bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + const char *name() const override; + +private: + /** Template function to run im2col + * + * @param[in] src The input tensor info + * @param[out] dst The output tensor info + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + template + void run_im2col(const ITensor *src, ITensor *dst, const Window &window); + + /** Common signature for all the specialised im2col functions + * + * @param[in] window Region on which to execute the kernel. + */ + using Im2ColFunctionPtr = void (CpuIm2ColKernel::*)(const ITensor *src, ITensor *dst, const Window &window); + + Im2ColFunctionPtr _func{ nullptr }; + std::pair _convolved_dims{}; + PadStrideInfo _conv_info{}; + unsigned int _kernel_width{ 0 }; + unsigned int _kernel_height{ 0 }; + bool _has_bias{ false }; + Size2D _dilation{ 1U, 1U }; + DataLayout _data_layout{ DataLayout::UNKNOWN }; +}; +} // namespace kernels +} // namespace cpu +} // namespace arm_compute +#endif /*ARM_COMPUTE_CPU_IM2COL_KERNEL_H */ -- cgit v1.2.1