diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2017-11-23 09:49:51 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:42:33 +0000 |
commit | 780db4eb6a9e3dee565d14f36d772038cd3253da (patch) | |
tree | 53490d6a03bdeb26d77bc8840d1dbf6027e81f5c /arm_compute/runtime/NEON/functions | |
parent | d7ba5397b676c966cb5069c7187a12a0c35305f5 (diff) | |
download | ComputeLibrary-780db4eb6a9e3dee565d14f36d772038cd3253da.tar.gz |
COMPMID-471 Implement Deconvolution on OpenCL
Change-Id: Ie00c6b08a51d30c5ce2637d40ee3d165b8a68686
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110311
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'arm_compute/runtime/NEON/functions')
-rw-r--r-- | arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h | 61 | ||||
-rw-r--r-- | arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h | 72 |
2 files changed, 35 insertions, 98 deletions
diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h index 8757bc63aa..091a928db6 100644 --- a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017, 2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -24,7 +24,6 @@ #ifndef __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__ #define __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__ -#include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h" #include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h" #include "arm_compute/core/Types.h" @@ -39,13 +38,13 @@ namespace arm_compute { /** Function to run the deconvolution layer. * - * The operation is similar to convolution but it's implemented by up-sampling the inputs with zeros insertions between the inputs and convolving - * the kernels on the up-sampled result. + * Deconvolution Layer is the backward pass of Convolution Layer. First we transform the input depending on the stride and pad info and then perfrom a 1x1 + * convolution pass. Input stride defines how many zeroes we should put between each element of the input, pad is the amount of padding and finaly a is a user + * specified value where a < stride - 1 that increases the padding top and right of the input image. * - * Before the Deconvolution is done, up-scaling the first 2D with zeros is performed. The relation between input to - * output is as follows: - * width_output = round((width_input − 1) ∗ upscale_x − 2 ∗ padding_x + kernel_x + a_x ) - * height_output = round((height_input − 1) ∗ upscale_y − 2 ∗ padding_y + kernel_y + a_y ) + * The relation between input to output is as follows: + * width_output = round((width_input − 1) ∗ (stride_x - 1) − 2 ∗ padding_x + kernel_x + inner_border_right ) + * height_output = round((height_input − 1) ∗ (stride_y - 1) − 2 ∗ padding_y + kernel_y + inner_border_top ) * * where * width is the size of the first input dimension. @@ -53,44 +52,54 @@ namespace arm_compute * width_output is the size of the first output dimension. * height_output is the size of the second output dimension. * kernel_x and kernel_y are the convolution sizes in x and y. - * ax and ay the number of zeros added to the top and right edges of the input. - * upscale_x and upscale_y how much to scale the X and Y axis. + * inner_border_right and inner_border_top the number of zeros added to the top and right edges of the input. + * stride_x and stride_y is the input stride of the first and second dimension. * * This function calls the following NEON kernels: * - * -# @ref NEDeconvolutionLayerUpsampleKernel * -# @ref NEDirectConvolutionLayer * */ class NEDeconvolutionLayer : public IFunction { public: - /** Constructor */ + /** Default constructor */ NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); + + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDeconvolutionLayer(const NEDeconvolutionLayer &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDeconvolutionLayer &operator=(const NEDeconvolutionLayer &) = delete; + /** Allow instances of this class to be moved */ + NEDeconvolutionLayer(NEDeconvolutionLayer &&) = default; + /** Allow instances of this class to be moved */ + NEDeconvolutionLayer &operator=(NEDeconvolutionLayer &&) = default; + /** Default destructor */ + virtual ~NEDeconvolutionLayer() = default; /** Set the input, weights, biases and output tensors. * - * @param[in,out] input Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32. - * @param[in] weights The 4d weights with dimensions [width, height, OFM, IFM]. Data type supported: Same as @p input. - * @param[in] bias Optional, ignored if NULL. The biases have one dimension. Data type supported: Same as @p input. - * @param[out] output Output tensor. The output has the same number of dimensions as the @p input. - * @param[in] info Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo. - * @param[in] ax The number of zeros added to right edge of the input. - * @param[in] ay The number of zeros added to top edge of the input. - * @param[in] upscalex How much to scale the X axis. - * @param[in] upscaley How much to scale the Y axis. + * @param[in,out] input Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32. + * @param[in] weights The 4d weights with dimensions [width, height, OFM, IFM]. Data type supported: Same as @p input. + * @param[in] bias Optional, ignored if NULL. The biases have one dimension. Data type supported: Same as @p input. + * @param[out] output Output tensor. The output has the same number of dimensions as the @p input. + * @param[in] info Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo. + * @param[in] inner_border_right The number of zeros added to right edge of the input. + * @param[in] inner_border_top The number of zeros added to top edge of the input. * */ void configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info, - unsigned int ax, unsigned int ay, float upscalex, float upscaley); + unsigned int inner_border_right, unsigned int inner_border_top); // Inherited methods overridden: void run() override; private: - MemoryGroup _memory_group; - NEDeconvolutionLayerUpsample _scale_f; - NEDirectConvolutionLayer _conv_f; - Tensor _scaled_output; + MemoryGroup _memory_group; + NEDirectConvolutionLayer _conv_f; + Tensor _scaled_output; + ITensor *_input; + PadStrideInfo _info; + std::pair<unsigned int, unsigned int> _inner_border; }; } // arm_compute #endif /* __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h deleted file mode 100644 index d2ac12a58a..0000000000 --- a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h +++ /dev/null @@ -1,72 +0,0 @@ -/* - * Copyright (c) 2016, 2017 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_NEDECONVOLUTIONUPSAMPLE_H__ -#define __ARM_COMPUTE_NEDECONVOLUTIONUPSAMPLE_H__ - -#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h" -#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/runtime/IFunction.h" -#include "arm_compute/runtime/IMemoryManager.h" -#include "arm_compute/runtime/MemoryGroup.h" -#include "arm_compute/runtime/Tensor.h" - -#include <cstdint> -#include <memory> - -namespace arm_compute -{ -class ITensor; - -/** Basic function to run @ref NEDeconvolutionLayerUpsampleKernel */ -class NEDeconvolutionLayerUpsample : public IFunction -{ -public: - /** Constructor - * - * Initialize NEDeconvolutionLayerUpsample - */ - NEDeconvolutionLayerUpsample(std::shared_ptr<IMemoryManager> memory_manager = nullptr); - /** Initialize the function's source, destination, interpolation type and border_mode. - * - * @param[in, out] input Source tensor. Data type supported: F32. - * @param[out] output Destination tensor. Data type supported: F32. - * @param[in] a Top and right inner border sizes. These rows and columns will be filled with zero. - * @param[in] iz The number of zeros to be inserted between each input sample - * @param[in] info Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo. - */ - void configure(ITensor *input, ITensor *output, const std::pair<unsigned int, unsigned int> &a, - const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info); - - // Inherited methods overridden: - void run() override; - -private: - MemoryGroup _memory_group; - Tensor _offsets; - NEFillBorderKernel _border_handler; - NEDeconvolutionLayerUpsampleKernel _upsample; -}; -} // arm_compute -#endif /*__ARM_COMPUTE_NEDECONVOLUTIONUPSAMPLE_H__ */ |