Port NEIm2ColKernel
Resolves: COMPMID-4510
Change-Id: Ia3e588f599449d975dabad4afafb2974dd44d0ad
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5899
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/NECol2ImKernel.h b/src/core/NEON/kernels/NECol2ImKernel.h
index 397bf5a..1976302 100644
--- a/src/core/NEON/kernels/NECol2ImKernel.h
+++ b/src/core/NEON/kernels/NECol2ImKernel.h
@@ -34,7 +34,7 @@
/** Kernel to perform col2im reshaping.
*
- * Rearranges each matrix column into image blocks. It's the inverse operation of @ref NEIm2ColKernel.
+ * Rearranges each matrix column into image blocks. It's the inverse operation of @ref cpu::kernels::CpuIm2ColKernel.
*
* For example, a vector of 9 elements can be reshaped to a block(image) of 3x3:
*
diff --git a/src/core/NEON/kernels/NEIm2ColKernel.cpp b/src/core/NEON/kernels/NEIm2ColKernel.cpp
deleted file mode 100644
index a28a77a..0000000
--- a/src/core/NEON/kernels/NEIm2ColKernel.cpp
+++ /dev/null
@@ -1,460 +0,0 @@
-/*
- * 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/NEON/kernels/NEIm2ColKernel.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 <arm_neon.h>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <tuple>
-
-using namespace arm_compute;
-using namespace misc::shape_calculator;
-
-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{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
- bool has_bias, const Size2D &dilation)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-
- // Output tensor auto initialization if not yet initialized
- auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation, false)));
-
- const DataLayout data_layout = input->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);
-
- std::pair<unsigned int, unsigned int> convolved_dims = scaled_dimensions(input->dimension(width_idx), input->dimension(height_idx),
- kernel_dims.width, kernel_dims.height,
- conv_info, dilation);
-
- Window win = calculate_max_window(*input, 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));
-
- // The NEIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
-
- return std::make_pair(Status{}, win);
-}
-
-template <typename T, bool has_pads>
-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<const T *>(in_ptr + ((d + 0) * input_stride_z + y * input_stride_y + x * input_stride_x)));
- *(out_ptr + 1 * kernel_size2) = *(reinterpret_cast<const T *>(in_ptr + ((d + 1) * input_stride_z + y * input_stride_y + x * input_stride_x)));
- *(out_ptr + 2 * kernel_size2) = *(reinterpret_cast<const T *>(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<void *>(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<const T *>(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<T>(1);
- }
-}
-
-template <typename T, bool has_pads>
-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<int>(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<const T *>(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<void *>(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<void *>(out_ptr), pad_value, input_c * element_size);
- out_ptr += input_c;
- }
- else
- {
- memcpy(out_ptr, reinterpret_cast<const T *>(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<const T *>(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<T>(1);
- }
-}
-} // namespace
-
-template <typename T, bool has_pads, bool is_nchw>
-void NEIm2ColKernel::run_im2col(const Window &window)
-{
- ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
- ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::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 = _input->info()->dimension(width_idx);
- const int input_h = _input->info()->dimension(height_idx);
- const int input_c = _input->info()->dimension(channel_idx);
- 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 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(_input->info()->data_type()) ? _input->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(_input, window_in_out);
- Iterator out(_output, 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<T *>(out.ptr() + (id[width_idx] + id[height_idx] * _convolved_dims.first) * _output->info()->strides_in_bytes().y());
-
- // Linearize volume
- if(is_nchw)
- {
- linearize_volume_nchw<T, has_pads>(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<T, has_pads>(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);
-}
-
-NEIm2ColKernel::NEIm2ColKernel()
- : _func(), _input(nullptr), _output(nullptr), _convolved_dims(), _conv_info(), _kernel_width(0), _kernel_height(0), _has_bias(false), _dilation(1U, 1U), _data_layout(DataLayout::UNKNOWN)
-{
-}
-
-void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
- bool has_bias, const Size2D &dilation, unsigned int num_groups)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, dilation, num_groups));
- ARM_COMPUTE_UNUSED(num_groups);
-
- _data_layout = input->info()->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);
-
- _input = input;
- _output = output;
- _conv_info = conv_info;
- _kernel_width = kernel_dims.width;
- _kernel_height = kernel_dims.height;
- _dilation = dilation;
- _convolved_dims = scaled_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx),
- _kernel_width, _kernel_height,
- _conv_info, _dilation);
- _has_bias = has_bias;
-
- if(_data_layout == DataLayout::NCHW)
- {
- switch(_input->info()->data_type())
- {
- case DataType::F32:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<float, false, true> : &NEIm2ColKernel::run_im2col<float, true, true>;
- break;
-#if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16)
- case DataType::BFLOAT16:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<bfloat16, false, true> : &NEIm2ColKernel::run_im2col<bfloat16, true, true>;
- 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()) ? &NEIm2ColKernel::run_im2col<float16_t, false, true> : &NEIm2ColKernel::run_im2col<float16_t, true, true>;
- break;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
- case DataType::QASYMM8_SIGNED:
- case DataType::QASYMM8:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<qasymm8_t, false, true> : &NEIm2ColKernel::run_im2col<qasymm8_t, true, true>;
- break;
- default:
- ARM_COMPUTE_ERROR("Data type not supported");
- break;
- }
- }
- else
- {
- switch(_input->info()->data_type())
- {
- case DataType::F32:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<float, false, false> : &NEIm2ColKernel::run_im2col<float, true, false>;
- break;
-#if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16)
- case DataType::BFLOAT16:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<bfloat16, false, false> : &NEIm2ColKernel::run_im2col<bfloat16, true, false>;
- 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()) ? &NEIm2ColKernel::run_im2col<float16_t, false, false> : &NEIm2ColKernel::run_im2col<float16_t, true, false>;
- break;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
- case DataType::QASYMM8:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<uint8_t, false, false> : &NEIm2ColKernel::run_im2col<qasymm8_t, true, false>;
- break;
- case DataType::QASYMM8_SIGNED:
- _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_im2col<int8_t, false, false> : &NEIm2ColKernel::run_im2col<qasymm8_t, true, false>;
- break;
- default:
- ARM_COMPUTE_ERROR("Data type not supported");
- break;
- }
- }
-
- // Configure kernel window
- auto win_config = validate_and_configure_window(input->info(), output->info(), kernel_dims, conv_info, has_bias, dilation);
- ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
- INEKernel::configure(win_config.second);
-}
-
-Status NEIm2ColKernel::validate(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_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, dilation, num_groups));
- ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), kernel_dims, conv_info, has_bias, dilation).first);
- return Status{};
-}
-
-void NEIm2ColKernel::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);
-
- (this->*_func)(window);
-}
diff --git a/src/core/NEON/kernels/NEIm2ColKernel.h b/src/core/NEON/kernels/NEIm2ColKernel.h
deleted file mode 100644
index 6c1c631..0000000
--- a/src/core/NEON/kernels/NEIm2ColKernel.h
+++ /dev/null
@@ -1,139 +0,0 @@
-/*
- * Copyright (c) 2017-2020 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_NEIM2COLKERNEL_H
-#define ARM_COMPUTE_NEIM2COLKERNEL_H
-
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-class Size2D;
-
-/** 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 NEIm2ColKernel : public INEKernel
-{
-public:
- const char *name() const override
- {
- return "NEIm2ColKernel";
- }
- /** Default constructor */
- NEIm2ColKernel();
- /** Prevent instances of this class from being copied (As this class contains pointers) */
- NEIm2ColKernel(const NEIm2ColKernel &) = delete;
- /** Prevent instances of this class from being copied (As this class contains pointers) */
- NEIm2ColKernel &operator=(const NEIm2ColKernel &) = delete;
- /** Allow instances of this class to be moved */
- NEIm2ColKernel(NEIm2ColKernel &&) = default;
- /** Allow instances of this class to be moved */
- NEIm2ColKernel &operator=(NEIm2ColKernel &&) = default;
- /** Default destructor */
- ~NEIm2ColKernel() = default;
-
- /** Set the input and output of the kernel.
- *
- * @param[in] input The input tensor 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] output The output tensor. 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(const ITensor *input, ITensor *output, 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 of @ref NEIm2ColKernel
- *
- * @param[in] input The input tensor 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[in] output The output tensor. 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
- *
- * @return a status
- */
- static Status validate(const ITensorInfo *input, const ITensorInfo *output, 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(const Window &window, const ThreadInfo &info) override;
-
-private:
- /** Template function to run im2col
- *
- * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
- */
- template <typename T, bool has_pads, bool is_nchw>
- void run_im2col(const Window &window);
-
- /** Common signature for all the specialised im2col functions
- *
- * @param[in] window Region on which to execute the kernel.
- */
- using Im2ColFunctionPtr = void (NEIm2ColKernel::*)(const Window &window);
-
- Im2ColFunctionPtr _func;
- const ITensor *_input;
- ITensor *_output;
- std::pair<unsigned int, unsigned int> _convolved_dims;
- PadStrideInfo _conv_info;
- unsigned int _kernel_width;
- unsigned int _kernel_height;
- bool _has_bias;
- Size2D _dilation;
- DataLayout _data_layout;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_NEIM2COLKERNEL_H */
diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.h b/src/core/NEON/kernels/NEWeightsReshapeKernel.h
index 76eca9f..5701c84 100644
--- a/src/core/NEON/kernels/NEWeightsReshapeKernel.h
+++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.h
@@ -33,7 +33,7 @@
/** Kernel to perform reshaping on the weights used by convolution and locally connected layer
*
* Rearranges each 3-dimensional kernel to a single row leading to a matrix with linearized kernels.
- * In combination with the @ref NEIm2ColKernel can transform a convolution to a matrix multiplication.
+ * In combination with the @ref cpu::kernels::CpuIm2ColKernel can transform a convolution to a matrix multiplication.
*
* For example assuming a 3D weight kernel of 3x3 dimensions and depth of 2 we have:
* @f[