From bd9097db81f229c2d7bbafc2bcf392b7c1c49b58 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Fri, 26 Jul 2019 15:31:02 +0100 Subject: COMPMID-2336: Rename the new generic depthwise convolution on NEON Change-Id: I45cacf75b08bb9d867343037507e56f200ad6ac0 Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/1637 Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena Comments-Addressed: Arm Jenkins --- arm_compute/core/NEON/NEKernels.h | 2 +- .../kernels/NEDepthwiseConvolutionLayerKernel.h | 109 ------- .../NEDepthwiseConvolutionLayerNativeKernel.h | 109 +++++++ .../NEON/functions/NEDepthwiseConvolutionLayer.h | 6 +- .../kernels/NEDepthwiseConvolutionLayerKernel.cpp | 330 --------------------- .../NEDepthwiseConvolutionLayerNativeKernel.cpp | 330 +++++++++++++++++++++ .../NEON/functions/NEDepthwiseConvolutionLayer.cpp | 2 +- .../NEON/DepthwiseConvolutionLayerKernel.cpp | 180 ----------- .../NEON/DepthwiseConvolutionNativeLayer.cpp | 209 +++++++++++++ .../fixtures/DepthwiseConvolutionLayerFixture.h | 2 +- 10 files changed, 654 insertions(+), 625 deletions(-) delete mode 100644 arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h create mode 100644 arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h delete mode 100644 src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp create mode 100644 src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp delete mode 100644 tests/validation/NEON/DepthwiseConvolutionLayerKernel.cpp create mode 100644 tests/validation/NEON/DepthwiseConvolutionNativeLayer.cpp diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h index 8fddec2c5f..bf77e576ad 100644 --- a/arm_compute/core/NEON/NEKernels.h +++ b/arm_compute/core/NEON/NEKernels.h @@ -53,7 +53,7 @@ #include "arm_compute/core/NEON/kernels/NEDepthConvertLayerKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthToSpaceLayerKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h" -#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h" +#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h" diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h deleted file mode 100644 index 63635b3a6c..0000000000 --- a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h +++ /dev/null @@ -1,109 +0,0 @@ -/* - * Copyright (c) 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. - */ -#ifndef __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONKERNEL_H__ -#define __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONKERNEL_H__ - -#include "arm_compute/core/NEON/INEKernel.h" - -namespace arm_compute -{ -// Forward declarations -class ITensor; - -/** Interface for the kernel to run a depthwise convolution on a tensor. */ -class NEDepthwiseConvolutionLayerKernel : public INEKernel -{ -public: - const char *name() const override - { - return "NEDepthwiseConvolutionLayerKernel"; - } - /** Default constructor */ - NEDepthwiseConvolutionLayerKernel(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEDepthwiseConvolutionLayerKernel(const NEDepthwiseConvolutionLayerKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEDepthwiseConvolutionLayerKernel &operator=(const NEDepthwiseConvolutionLayerKernel &) = delete; - /** Default Move Constructor. */ - NEDepthwiseConvolutionLayerKernel(NEDepthwiseConvolutionLayerKernel &&) = default; - /** Default move assignment operator */ - NEDepthwiseConvolutionLayerKernel &operator=(NEDepthwiseConvolutionLayerKernel &&) = default; - /** Initialize the function's source, destination and parameters. - * - * @note Supported data layouts: NHWC - * - * @param[in] input Source tensor. DataType supported: F32. - * @param[in] weights Weights tensor. This is a 3D tensor with dimensions [IFM, W, H]. Data type supported: Same as @p input. - * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. Data type supported: Same as @p input. - * @param[out] output Destination tensor. Data type supported: Same as @p input. - * @param[in] conv_info Padding and stride information to use for the convolution. - * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. - * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * - */ - void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, - const Size2D &dilation = Size2D(1U, 1U)); - /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerKernel - * - * @note Supported data layouts: NHWC - * - * @param[in] input Source tensor info. DataType supported: F32. - * @param[in] weights Weights tensor info. This is a 3D tensor with dimensions [IFM, W, H]. Data type supported: Same as @p input. - * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. Data type supported: Same as @p input. - * @param[in] output Destination tensor info. Data type supported: Same as @p input. - * @param[in] conv_info Padding and stride information to use for the convolution. - * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. - * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, - const Size2D &dilation = Size2D(1U, 1U)); - - // Inherited methods overridden: - void run(const Window &window, const ThreadInfo &info) override; - BorderSize border_size() const override; - -private: - template - void run_depthwise(const Window &window); - - /** Common signature for all the specialised depthwise convolution functions - * - * @param[in] window Region on which to execute the kernel. - */ - using DepthwiseFunctionPtr = void (NEDepthwiseConvolutionLayerKernel::*)(const Window &window); - - DepthwiseFunctionPtr _func; - BorderSize _border_size; - const ITensor *_input; - const ITensor *_weights; - const ITensor *_biases; - ITensor *_output; - PadStrideInfo _conv_info; - unsigned int _depth_multiplier; - Size2D _dilation; -}; -} // namespace arm_compute -#endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONKERNEL_H__ */ diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h new file mode 100644 index 0000000000..5db79f8bf7 --- /dev/null +++ b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h @@ -0,0 +1,109 @@ +/* + * Copyright (c) 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. + */ +#ifndef __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H__ +#define __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H__ + +#include "arm_compute/core/NEON/INEKernel.h" + +namespace arm_compute +{ +// Forward declarations +class ITensor; + +/** Interface for the kernel to run a depthwise convolution native on a tensor. */ +class NEDepthwiseConvolutionLayerNativeKernel : public INEKernel +{ +public: + const char *name() const override + { + return "NEDepthwiseConvolutionLayerNativeKernel"; + } + /** Default constructor */ + NEDepthwiseConvolutionLayerNativeKernel(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDepthwiseConvolutionLayerNativeKernel(const NEDepthwiseConvolutionLayerNativeKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDepthwiseConvolutionLayerNativeKernel &operator=(const NEDepthwiseConvolutionLayerNativeKernel &) = delete; + /** Default Move Constructor. */ + NEDepthwiseConvolutionLayerNativeKernel(NEDepthwiseConvolutionLayerNativeKernel &&) = default; + /** Default move assignment operator */ + NEDepthwiseConvolutionLayerNativeKernel &operator=(NEDepthwiseConvolutionLayerNativeKernel &&) = default; + /** Initialize the function's source, destination and parameters. + * + * @note Supported data layouts: NHWC + * + * @param[in] input Source tensor. DataType supported: F32. + * @param[in] weights Weights tensor. This is a 3D tensor with dimensions [IFM, W, H]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. Data type supported: Same as @p input. + * @param[out] output Destination tensor. Data type supported: Same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). + * + */ + void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, + const Size2D &dilation = Size2D(1U, 1U)); + /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerNativeKernel + * + * @note Supported data layouts: NHWC + * + * @param[in] input Source tensor info. DataType supported: F32. + * @param[in] weights Weights tensor info. This is a 3D tensor with dimensions [IFM, W, H]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. Data type supported: Same as @p input. + * @param[in] output Destination tensor info. Data type supported: Same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, + const Size2D &dilation = Size2D(1U, 1U)); + + // Inherited methods overridden: + void run(const Window &window, const ThreadInfo &info) override; + BorderSize border_size() const override; + +private: + template + void run_depthwise(const Window &window); + + /** Common signature for all the specialised depthwise convolution native functions + * + * @param[in] window Region on which to execute the kernel. + */ + using DepthwiseFunctionPtr = void (NEDepthwiseConvolutionLayerNativeKernel::*)(const Window &window); + + DepthwiseFunctionPtr _func; + BorderSize _border_size; + const ITensor *_input; + const ITensor *_weights; + const ITensor *_biases; + ITensor *_output; + PadStrideInfo _conv_info; + unsigned int _depth_multiplier; + Size2D _dilation; +}; +} // namespace arm_compute +#endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h index 5b0d1bafcd..87405fdb14 100644 --- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h @@ -25,7 +25,7 @@ #define __ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H__ #include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h" -#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h" +#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h" #include "arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h" @@ -282,7 +282,7 @@ private: /** Basic function to execute a generic depthwise convolution. This function calls the following NEON kernels: * * If data type is F32 and data layout is NHWC: - * -# @ref NEDepthwiseConvolutionLayerKernel + * -# @ref NEDepthwiseConvolutionLayerNativeKernel * * Otherwise: * -# @ref NEDepthwiseIm2ColKernel @@ -344,7 +344,7 @@ private: NEDepthwiseIm2ColKernel _im2col_kernel; NEDepthwiseWeightsReshapeKernel _weights_reshape_kernel; NEGEMMMatrixVectorMultiplyKernel _v2mm_kernel; - NEDepthwiseConvolutionLayerKernel _depthwise_conv_kernel; + NEDepthwiseConvolutionLayerNativeKernel _depthwise_conv_kernel; NEDepthwiseVectorToTensorKernel _vector_to_tensor_kernel; NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel; NEFillBorderKernel _fill_border; diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp deleted file mode 100644 index feb2071d47..0000000000 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp +++ /dev/null @@ -1,330 +0,0 @@ -/* - * Copyright (c) 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/NEDepthwiseConvolutionLayerKernel.h" - -#include "arm_compute/core/AccessWindowStatic.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" - -namespace arm_compute -{ -namespace -{ -template -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) -{ - using VectorType = typename wrapper::traits::neon_vector::type; - using TagType = typename wrapper::traits::neon_vector::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(); - - 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) - { - VectorType acc = wrapper::vdup_n(static_cast(0), TagType{}); - - 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(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); - const auto weights_vals = wrapper::vload(reinterpret_cast(weights_ptr + w * weights_stride_y)); - - acc = wrapper::vmla(acc, weights_vals, input_vals); - offs += dilation.x() * input_stride_y; - } - - weights_ptr += weights_stride_z; - input_offset += dilation.y() * input_stride_z; - } - - if(has_biases) - { - const auto biases_vals = wrapper::vload(reinterpret_cast(biases_it.ptr())); - acc = wrapper::vadd(acc, biases_vals); - } - - wrapper::vstore(reinterpret_cast(output_it.ptr()), acc); - }, - input_it, weights_it, biases_it, output_it); -} - -template -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) -{ - 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(); - - 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 acc(depth_multiplier, static_cast(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(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); - - for(size_t m = 0; m < depth_multiplier; ++m) - { - const auto weights_val = *(reinterpret_cast(weights_ptr + m * sizeof(T) + w * weights_stride_y)); - acc.at(m) = std::fma(weights_val, input_val, acc.at(m)); - } - - offs += dilation.x() * input_stride_y; - } - - weights_ptr += weights_stride_z; - input_offset += dilation.y() * input_stride_z; - } - - if(has_biases) - { - for(size_t m = 0; m < depth_multiplier; ++m) - { - const auto biases_val = *(reinterpret_cast(biases_it.ptr() + m * sizeof(T))); - *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m) + biases_val; - } - } - else - { - for(size_t m = 0; m < depth_multiplier; ++m) - { - *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m); - } - } - }, - 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(depth_multiplier == 0); - 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(biases != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0)); - } - - if(output->total_size() != 0) - { - const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); - } - - return Status{}; -} -} // namespace - -std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *biases, - ITensorInfo *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier, const Size2D &dilation) -{ - // Get convolved dimensions - 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)); - - // 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; - const unsigned int num_elems_written_per_iteration = num_elems_read_per_iteration * depth_multiplier; - - // Configure kernel window - Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration)); - - AccessWindowStatic input_access(input, 0, -conv_info.pad_left(), ceil_to_multiple(num_elems_read_per_iteration, input->dimension(0)), - input->dimension(1) + std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top())); - AccessWindowHorizontal weights_access(weights, 0, num_elems_written_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_written_per_iteration); - - bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access); - - if(biases != nullptr) - { - AccessWindowHorizontal biases_access(biases, 0, num_elems_written_per_iteration); - window_changed |= update_window_and_padding(win, biases_access); - } - - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} - -NEDepthwiseConvolutionLayerKernel::NEDepthwiseConvolutionLayerKernel() - : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation() -{ -} - -BorderSize NEDepthwiseConvolutionLayerKernel::border_size() const -{ - return _border_size; -} - -void NEDepthwiseConvolutionLayerKernel::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, - const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation)); - - _input = input; - _weights = weights; - _biases = biases; - _output = output; - _conv_info = conv_info; - _depth_multiplier = depth_multiplier; - _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()) - { - case DataType::F32: - _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerKernel::run_depthwise : &NEDepthwiseConvolutionLayerKernel::run_depthwise; - break; - default: - ARM_COMPUTE_ERROR("Data type not supported"); - break; - } - - auto win_config = validate_and_configure_window(_input->info(), _weights->info(), (biases != nullptr) ? biases->info() : nullptr, _output->info(), _conv_info, _depth_multiplier, dilation); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - INEKernel::configure(win_config.second); -} - -Status NEDepthwiseConvolutionLayerKernel::validate(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_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, dilation)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), (biases != nullptr) ? biases->clone().get() : nullptr, output->clone().get(), conv_info, - depth_multiplier, dilation) - .first); - return Status{}; -} - -void NEDepthwiseConvolutionLayerKernel::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); -} - -template -void NEDepthwiseConvolutionLayerKernel::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(_input, _weights, _biases, _output, _conv_info, _dilation, window); - } - else - { - depthwise_loop_generic(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); - } -} -} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp new file mode 100644 index 0000000000..aafdb2e8a4 --- /dev/null +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp @@ -0,0 +1,330 @@ +/* + * Copyright (c) 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/NEDepthwiseConvolutionLayerNativeKernel.h" + +#include "arm_compute/core/AccessWindowStatic.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" + +namespace arm_compute +{ +namespace +{ +template +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) +{ + using VectorType = typename wrapper::traits::neon_vector::type; + using TagType = typename wrapper::traits::neon_vector::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(); + + 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) + { + VectorType acc = wrapper::vdup_n(static_cast(0), TagType{}); + + 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(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); + const auto weights_vals = wrapper::vload(reinterpret_cast(weights_ptr + w * weights_stride_y)); + + acc = wrapper::vmla(acc, weights_vals, input_vals); + offs += dilation.x() * input_stride_y; + } + + weights_ptr += weights_stride_z; + input_offset += dilation.y() * input_stride_z; + } + + if(has_biases) + { + const auto biases_vals = wrapper::vload(reinterpret_cast(biases_it.ptr())); + acc = wrapper::vadd(acc, biases_vals); + } + + wrapper::vstore(reinterpret_cast(output_it.ptr()), acc); + }, + input_it, weights_it, biases_it, output_it); +} + +template +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) +{ + 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(); + + 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 acc(depth_multiplier, static_cast(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(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); + + for(size_t m = 0; m < depth_multiplier; ++m) + { + const auto weights_val = *(reinterpret_cast(weights_ptr + m * sizeof(T) + w * weights_stride_y)); + acc.at(m) = std::fma(weights_val, input_val, acc.at(m)); + } + + offs += dilation.x() * input_stride_y; + } + + weights_ptr += weights_stride_z; + input_offset += dilation.y() * input_stride_z; + } + + if(has_biases) + { + for(size_t m = 0; m < depth_multiplier; ++m) + { + const auto biases_val = *(reinterpret_cast(biases_it.ptr() + m * sizeof(T))); + *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m) + biases_val; + } + } + else + { + for(size_t m = 0; m < depth_multiplier; ++m) + { + *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m); + } + } + }, + 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(depth_multiplier == 0); + 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(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0)); + } + + if(output->total_size() != 0) + { + const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); + } + + return Status{}; +} +} // namespace + +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *biases, + ITensorInfo *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, const Size2D &dilation) +{ + // Get convolved dimensions + 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)); + + // 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; + const unsigned int num_elems_written_per_iteration = num_elems_read_per_iteration * depth_multiplier; + + // Configure kernel window + Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration)); + + AccessWindowStatic input_access(input, 0, -conv_info.pad_left(), ceil_to_multiple(num_elems_read_per_iteration, input->dimension(0)), + input->dimension(1) + std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top())); + AccessWindowHorizontal weights_access(weights, 0, num_elems_written_per_iteration); + AccessWindowHorizontal output_access(output, 0, num_elems_written_per_iteration); + + bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access); + + if(biases != nullptr) + { + AccessWindowHorizontal biases_access(biases, 0, num_elems_written_per_iteration); + window_changed |= update_window_and_padding(win, biases_access); + } + + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +NEDepthwiseConvolutionLayerNativeKernel::NEDepthwiseConvolutionLayerNativeKernel() + : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation() +{ +} + +BorderSize NEDepthwiseConvolutionLayerNativeKernel::border_size() const +{ + return _border_size; +} + +void NEDepthwiseConvolutionLayerNativeKernel::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, + const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation)); + + _input = input; + _weights = weights; + _biases = biases; + _output = output; + _conv_info = conv_info; + _depth_multiplier = depth_multiplier; + _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()) + { + case DataType::F32: + _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise : &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise; + break; + default: + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + + auto win_config = validate_and_configure_window(_input->info(), _weights->info(), (biases != nullptr) ? biases->info() : nullptr, _output->info(), _conv_info, _depth_multiplier, dilation); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} + +Status NEDepthwiseConvolutionLayerNativeKernel::validate(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_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, dilation)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), (biases != nullptr) ? biases->clone().get() : nullptr, output->clone().get(), conv_info, + depth_multiplier, dilation) + .first); + return Status{}; +} + +void NEDepthwiseConvolutionLayerNativeKernel::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); +} + +template +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(_input, _weights, _biases, _output, _conv_info, _dilation, window); + } + else + { + depthwise_loop_generic(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); + } +} +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp index c2ed901169..cdd278b2f1 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp @@ -947,7 +947,7 @@ Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe } else { - ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayerKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, dilation)); + ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, dilation)); } // Validate Activation Layer diff --git a/tests/validation/NEON/DepthwiseConvolutionLayerKernel.cpp b/tests/validation/NEON/DepthwiseConvolutionLayerKernel.cpp deleted file mode 100644 index 3af835855b..0000000000 --- a/tests/validation/NEON/DepthwiseConvolutionLayerKernel.cpp +++ /dev/null @@ -1,180 +0,0 @@ -/* - * Copyright (c) 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/NEDepthwiseConvolutionLayerKernel.h" -#include "tests/NEON/Accessor.h" -#include "tests/NEON/Helper.h" -#include "tests/framework/Macros.h" -#include "tests/framework/datasets/Datasets.h" -#include "tests/validation/Validation.h" -#include "tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -using namespace arm_compute::misc::shape_calculator; - -// Create function for NEDepthwiseConvolutionLayerKernel -using NEDepthwiseConvolutionLayer = NESynthetizeFunctionWithZeroConstantKernelBorder; - -// Fixture for NEDepthwiseConvolutionLayerKernel -template -using NEDepthwiseConvolutionLayerKernelFixture = DepthwiseConvolutionLayerKernelValidationFixture; - -namespace -{ -// *INDENT-OFF* -// clang-format off -RelativeTolerance rel_tolerance_f32(0.001f); -constexpr float abs_tolerance_f32(0.0001f); - -/** Width values to test - Precommit */ -const auto width_values = framework::dataset::make("width", { 17U, 47U } ); - -/** Height values to test - Precommit */ -const auto height_values = framework::dataset::make("height", { 19U, 43U } ); - -/** Channel values to test - Precommit */ -const auto channel_values = framework::dataset::make("channels", { 32U, 128U }); - -/** Batch values to test - Precommit */ -const auto batch_values = framework::dataset::make("batch", { 1U, 3U }); - -/** Kernel size values to test - Precommit */ -const auto kernel_sz_values = framework::dataset::make("kernel_size", { Size2D(3U, 5U), Size2D(5U, 3U) }); - -/** Depth multiplier values to test - Precommit */ -const auto depth_multiplier_values = framework::dataset::make("depth_multiplier", { 1U, 3U }); - -/** Dilation values to test - Precommit */ -const auto dilation_values = framework::dataset::make("dilation", { Size2D(1U, 1U), Size2D(3U, 3U) }); - -/** Stride values to test - All */ -const auto stride_values = framework::dataset::make("stride", { Size2D(1U, 1U), Size2D(3U, 2U) }); - -/** Padding values to test - All */ -const auto padding_valid_values = framework::dataset::make("padding_valid", { true, false }); - -/** Data type values to test - All */ -const auto data_type_values = framework::dataset::make("data_type", { DataType::F32 }); - -/** Data layout values to test - All */ -const auto data_layout_values = framework::dataset::make("data_layout", { DataLayout::NHWC }); - -/** Configuration test */ -void validate_configuration(size_t width_value, size_t height_value, size_t channel_value, size_t batch_value, Size2D kernel_sz_value, size_t depth_multiplier_value, Size2D dilation_value, Size2D stride_value, bool padding_valid_value, DataType data_type_value, DataLayout data_layout_value) -{ - TensorShape src_shape(width_value, height_value, channel_value, batch_value); - TensorShape weights_shape(kernel_sz_value.width, kernel_sz_value.height, channel_value * depth_multiplier_value); - TensorShape biases_shape(channel_value * depth_multiplier_value); - - if(data_layout_value == DataLayout::NHWC) - { - permute(src_shape, PermutationVector(2U, 0U, 1U, 3U)); - permute(weights_shape, PermutationVector(2U, 0U, 1U)); - } - - TensorInfo src_info(src_shape, 1, data_type_value); - TensorInfo weights_info(weights_shape, 1, data_type_value); - TensorInfo biases_info(biases_shape, 1, data_type_value); - - src_info.set_data_layout(data_layout_value); - weights_info.set_data_layout(data_layout_value); - biases_info.set_data_layout(data_layout_value); - - PadStrideInfo conv_info; - if(padding_valid_value) - { - conv_info = PadStrideInfo(); - } - else - { - conv_info = calculate_same_pad(src_shape, weights_shape, PadStrideInfo(stride_value.width, stride_value.height), data_layout_value, dilation_value); - } - - const TensorShape dst_shape = compute_depthwise_convolution_shape(src_info, weights_info, conv_info, depth_multiplier_value, dilation_value); - - // Create tensors - Tensor src = create_tensor(src_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); - Tensor weights = create_tensor(weights_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); - Tensor biases = create_tensor(biases_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); - Tensor dst = create_tensor(dst_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); - - ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(biases.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); - - // Create and configure function - NEDepthwiseConvolutionLayer dwc; - dwc.configure(&src, &weights, &biases, &dst, conv_info, depth_multiplier_value, dilation_value); -} -} // namespace - -TEST_SUITE(NEON) -TEST_SUITE(DepthwiseConvolutionLayer) -TEST_SUITE(Float) -TEST_SUITE(FP32) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(width_values, - height_values), - channel_values), - batch_values), - kernel_sz_values), - depth_multiplier_values), - dilation_values), - stride_values), - padding_valid_values), - data_type_values), - data_layout_values), -width_value, height_value, channel_value, batch_value, kernel_sz_value, depth_multiplier_value, dilation_value, stride_value, padding_valid_value, data_type_value, data_layout_value) -{ - validate_configuration(width_value, height_value, channel_value, batch_value, kernel_sz_value, depth_multiplier_value, dilation_value, stride_value, padding_valid_value, data_type_value, data_layout_value); -} - -FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerKernelFixture, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(width_values, - height_values), - channel_values), - batch_values), - kernel_sz_values), - depth_multiplier_values), - dilation_values), - stride_values), - padding_valid_values), - data_type_values), - data_layout_values)) -{ - // Validate output - validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); -} - -TEST_SUITE_END() // FP32 -TEST_SUITE_END() // Float -TEST_SUITE_END() // DepthwiseConvolutionLayer -TEST_SUITE_END() // NEON -} // namespace validation -} // namespace test -} // namespace arm_compute \ No newline at end of file diff --git a/tests/validation/NEON/DepthwiseConvolutionNativeLayer.cpp b/tests/validation/NEON/DepthwiseConvolutionNativeLayer.cpp new file mode 100644 index 0000000000..a44c2dc3c9 --- /dev/null +++ b/tests/validation/NEON/DepthwiseConvolutionNativeLayer.cpp @@ -0,0 +1,209 @@ +/* + * Copyright (c) 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/NEDepthwiseConvolutionLayerNativeKernel.h" +#include "tests/NEON/Accessor.h" +#include "tests/NEON/Helper.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::misc::shape_calculator; + +// Create function for NEDepthwiseConvolutionLayerKernel +using NEDepthwiseConvolutionLayerNative = NESynthetizeFunctionWithZeroConstantKernelBorder; + +// Fixture for NEDepthwiseConvolutionLayerKernel +template +using NEDepthwiseConvolutionLayerNativeFixture = DepthwiseConvolutionLayerNativeValidationFixture; + +namespace +{ +// *INDENT-OFF* +// clang-format off +RelativeTolerance rel_tolerance_f32(0.001f); +constexpr float abs_tolerance_f32(0.0001f); + +/** Width values to test - Precommit */ +const auto width_values_precommit = framework::dataset::make("width", { 17U } ); + +/** Width values to test - Nightly */ +const auto width_values_nightly = framework::dataset::make("width", { 53U, 47U } ); + +/** Height values to test - Precommit */ +const auto height_values_precommit = framework::dataset::make("height", { 19U } ); + +/** Height values to test - Nightly */ +const auto height_values_nightly = framework::dataset::make("height", { 39U, 43U } ); + +/** Channel values to test - Precommit */ +const auto channel_values_precommit = framework::dataset::make("channels", { 15U }); + +/** Channel values to test - Nightly */ +const auto channel_values_nightly = framework::dataset::make("channels", { 33U, 19U }); + +/** Batch values to test - Precommit */ +const auto batch_values_precommit = framework::dataset::make("batch", { 1U, 2U }); + +/** Batch values to test - Nightly */ +const auto batch_values_nightly = framework::dataset::make("batch", { 1U, 3U }); + +/** Kernel size values to test - All */ +const auto kernel_sz_values = framework::dataset::make("kernel_size", { Size2D(3U, 5U), Size2D(5U, 3U) }); + +/** Depth multiplier values to test - All */ +const auto depth_multiplier_values = framework::dataset::make("depth_multiplier", { 1U, 3U }); + +/** Dilation values to test - All */ +const auto dilation_values = framework::dataset::make("dilation", { Size2D(1U, 1U), Size2D(3U, 3U) }); + +/** Stride values to test - All */ +const auto stride_values = framework::dataset::make("stride", { Size2D(1U, 1U), Size2D(3U, 2U) }); + +/** Padding values to test - All */ +const auto padding_valid_values = framework::dataset::make("padding_valid", { true, false }); + +/** Data type values to test - All */ +const auto data_type_values = framework::dataset::make("data_type", { DataType::F32 }); + +/** Data layout values to test - All */ +const auto data_layout_values = framework::dataset::make("data_layout", { DataLayout::NHWC }); + +/** Configuration test */ +void validate_configuration(size_t width_value, size_t height_value, size_t channel_value, size_t batch_value, Size2D kernel_sz_value, size_t depth_multiplier_value, Size2D dilation_value, Size2D stride_value, bool padding_valid_value, DataType data_type_value, DataLayout data_layout_value) +{ + TensorShape src_shape(width_value, height_value, channel_value, batch_value); + TensorShape weights_shape(kernel_sz_value.width, kernel_sz_value.height, channel_value * depth_multiplier_value); + TensorShape biases_shape(channel_value * depth_multiplier_value); + + if(data_layout_value == DataLayout::NHWC) + { + permute(src_shape, PermutationVector(2U, 0U, 1U, 3U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + } + + TensorInfo src_info(src_shape, 1, data_type_value); + TensorInfo weights_info(weights_shape, 1, data_type_value); + TensorInfo biases_info(biases_shape, 1, data_type_value); + + src_info.set_data_layout(data_layout_value); + weights_info.set_data_layout(data_layout_value); + biases_info.set_data_layout(data_layout_value); + + PadStrideInfo conv_info; + if(padding_valid_value) + { + conv_info = PadStrideInfo(); + } + else + { + conv_info = calculate_same_pad(src_shape, weights_shape, PadStrideInfo(stride_value.width, stride_value.height), data_layout_value, dilation_value); + } + + const TensorShape dst_shape = compute_depthwise_convolution_shape(src_info, weights_info, conv_info, depth_multiplier_value, dilation_value); + + // Create tensors + Tensor src = create_tensor(src_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); + Tensor weights = create_tensor(weights_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); + Tensor biases = create_tensor(biases_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); + Tensor dst = create_tensor(dst_shape, data_type_value, 1, QuantizationInfo(), data_layout_value); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(biases.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Create and configure function + NEDepthwiseConvolutionLayerNative dwc; + dwc.configure(&src, &weights, &biases, &dst, conv_info, depth_multiplier_value, dilation_value); +} +} // namespace + +TEST_SUITE(NEON) +TEST_SUITE(DepthwiseConvolutionLayerNative) +TEST_SUITE(Float) +TEST_SUITE(FP32) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(width_values_precommit, + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + data_type_values), + data_layout_values), +width_value, height_value, channel_value, batch_value, kernel_sz_value, depth_multiplier_value, dilation_value, stride_value, padding_valid_value, data_type_value, data_layout_value) +{ + validate_configuration(width_value, height_value, channel_value, batch_value, kernel_sz_value, depth_multiplier_value, dilation_value, stride_value, padding_valid_value, data_type_value, data_layout_value); +} + +FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(width_values_precommit, + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + data_type_values), + data_layout_values)) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(width_values_nightly, + height_values_nightly), + channel_values_nightly), + batch_values_nightly), + kernel_sz_values), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + data_type_values), + data_layout_values)) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float +TEST_SUITE_END() // DepthwiseConvolutionLayerNative +TEST_SUITE_END() // NEON +} // namespace validation +} // namespace test +} // namespace arm_compute \ No newline at end of file diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h index 30b8df9da5..a3ac49eef1 100644 --- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h +++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h @@ -193,7 +193,7 @@ public: }; template -class DepthwiseConvolutionLayerKernelValidationFixture : public DepthwiseConvolutionLayerValidationGenericFixture +class DepthwiseConvolutionLayerNativeValidationFixture : public DepthwiseConvolutionLayerValidationGenericFixture { public: template -- cgit v1.2.1