blob: d031076ee72d0b873aab58a6a72bf3b5f1f208d7 [file] [log] [blame]
Pablo Tellof5f34bb2017-08-22 13:34:13 +01001/*
giuros01a69a88b2019-01-31 16:29:19 +00002 * Copyright (c) 2017-2019 ARM Limited.
Pablo Tellof5f34bb2017-08-22 13:34:13 +01003 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
Michalis Spyrouf4643372019-11-29 16:17:13 +000024#ifndef ARM_COMPUTE_NEDECONVOLUTIONLAYER_H
25#define ARM_COMPUTE_NEDECONVOLUTIONLAYER_H
Pablo Tellof5f34bb2017-08-22 13:34:13 +010026
Michalis Spyrou33a69902018-02-23 15:01:52 +000027#include "arm_compute/runtime/CPP/functions/CPPUpsample.h"
Georgios Pinitasced7a8d2018-02-01 16:31:33 +000028#include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
Pablo Tellof5f34bb2017-08-22 13:34:13 +010029#include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h"
30
Michalis Spyrouafbc5ff2018-10-03 14:18:19 +010031#include "arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h"
Pablo Tellof5f34bb2017-08-22 13:34:13 +010032#include "arm_compute/core/Types.h"
33#include "arm_compute/runtime/IFunction.h"
34#include "arm_compute/runtime/IMemoryManager.h"
35#include "arm_compute/runtime/MemoryGroup.h"
36#include "arm_compute/runtime/Tensor.h"
37
38#include <memory>
39
40namespace arm_compute
41{
42/** Function to run the deconvolution layer.
43 *
Michalis Spyrou780db4e2017-11-23 09:49:51 +000044 * 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
45 * 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
46 * specified value where a < stride - 1 that increases the padding top and right of the input image.
Pablo Tellof5f34bb2017-08-22 13:34:13 +010047 *
Michalis Spyrou780db4e2017-11-23 09:49:51 +000048 * The relation between input to output is as follows:
Michele Di Giorgioed5a4922018-09-13 16:22:01 +010049 * \f[
50 * width\_output = (width\_input - 1) \cdot stride\_x - 2 \cdot padding\_x + kernel\_x
51 * \f]
52 * \f[
53 * height\_output = (height\_input - 1) \cdot stride\_y - 2 \cdot padding\_y + kernel\_y
54 * \f]
Pablo Tellof5f34bb2017-08-22 13:34:13 +010055 *
56 * where
57 * width is the size of the first input dimension.
58 * height is the size of the second input dimension.
59 * width_output is the size of the first output dimension.
60 * height_output is the size of the second output dimension.
61 * kernel_x and kernel_y are the convolution sizes in x and y.
Michalis Spyrou780db4e2017-11-23 09:49:51 +000062 * stride_x and stride_y is the input stride of the first and second dimension.
Pablo Tellof5f34bb2017-08-22 13:34:13 +010063 *
Michele Di Giorgioed5a4922018-09-13 16:22:01 +010064 * The weights used by Deconvolution are supposed to be the same as the ones used for Convolution. Therefore, it will be necessary to use the weights in the
65 * reverse order to perform an actual convolution. This is achieved by using the @ref CPPFlipWeightsKernel.
Pablo Tellof5f34bb2017-08-22 13:34:13 +010066 *
Michele Di Giorgioed5a4922018-09-13 16:22:01 +010067 * This function calls the following NEON kernels/functions:
68 *
69 * -# @ref CPPUpsample
70 * -# @ref NEConvolutionLayer
Pablo Tellof5f34bb2017-08-22 13:34:13 +010071 *
72 */
73class NEDeconvolutionLayer : public IFunction
74{
75public:
Michalis Spyrou1a569a32019-09-10 17:20:34 +010076 /** Constructor */
Pablo Tellof5f34bb2017-08-22 13:34:13 +010077 NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
Michalis Spyrou780db4e2017-11-23 09:49:51 +000078
79 /** Prevent instances of this class from being copied (As this class contains pointers) */
80 NEDeconvolutionLayer(const NEDeconvolutionLayer &) = delete;
81 /** Prevent instances of this class from being copied (As this class contains pointers) */
82 NEDeconvolutionLayer &operator=(const NEDeconvolutionLayer &) = delete;
83 /** Allow instances of this class to be moved */
84 NEDeconvolutionLayer(NEDeconvolutionLayer &&) = default;
85 /** Allow instances of this class to be moved */
86 NEDeconvolutionLayer &operator=(NEDeconvolutionLayer &&) = default;
87 /** Default destructor */
88 virtual ~NEDeconvolutionLayer() = default;
Pablo Tellof5f34bb2017-08-22 13:34:13 +010089
giuros01a69a88b2019-01-31 16:29:19 +000090 /** Set the input, weights, biases and output tensors.
91 *
Manuel Bottinif391fff2019-05-15 13:01:26 +010092 * @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/F16/QASYMM8.
giuros01a69a88b2019-01-31 16:29:19 +000093 * @param[in] weights The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input.
Manuel Bottinif391fff2019-05-15 13:01:26 +010094 * @param[in] bias Optional, ignored if NULL. The biases have one dimension. Data type supported: Data types supported: S32 for QASYMM8 input, F32 for F32 input, F16 for F16 input.
giuros01a69a88b2019-01-31 16:29:19 +000095 * @param[out] output Output tensor. The output has the same number of dimensions as the @p input.
96 * @param[in] info Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
97 *
98 */
99 void configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info);
100 /** Static function to check if given info will lead to a valid configuration of @ref NEDeconvolutionLayer
101 *
Manuel Bottinif391fff2019-05-15 13:01:26 +0100102 * @param[in] input Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32/F16/QASYMM8.
giuros01a69a88b2019-01-31 16:29:19 +0000103 * @param[in] weights The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input.
Manuel Bottinif391fff2019-05-15 13:01:26 +0100104 * @param[in] bias (Optional) The biases have one dimension. Data type supported: Data types supported: S32 for QASYMM8 input, F32 for F32 input, F16 for F16 input.
giuros01a69a88b2019-01-31 16:29:19 +0000105 * @param[in] output Output tensor info. The output has the same number of dimensions as the @p input.
106 * @param[in] info Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
107 *
108 * @return a status
109 */
110 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info);
111
Pablo Tellof5f34bb2017-08-22 13:34:13 +0100112 // Inherited methods overridden:
113 void run() override;
Georgios Pinitas72219332018-06-05 14:56:06 +0100114 void prepare() override;
Pablo Tellof5f34bb2017-08-22 13:34:13 +0100115
116private:
Michalis Spyrouafbc5ff2018-10-03 14:18:19 +0100117 MemoryGroup _memory_group;
118 NEConvolutionLayer _conv_f;
119 CPPUpsample _upsample_f;
120 CPPFlipWeightsKernel _flip_weights;
Manuel Bottinid25af672019-07-10 17:06:12 +0100121 NEPermute _permute_input;
122 NEPermute _permute_weights;
123 NEPermute _permute_output;
Michalis Spyrouafbc5ff2018-10-03 14:18:19 +0100124 Tensor _scaled_output;
125 Tensor _weights_flipped;
Manuel Bottinid25af672019-07-10 17:06:12 +0100126 Tensor _permuted_input;
127 Tensor _permuted_weights;
128 Tensor _permuted_output;
129 bool _is_nchw;
Michele Di Giorgio061dd362018-10-17 17:10:27 +0100130 const ITensor *_original_weights;
Michalis Spyrouafbc5ff2018-10-03 14:18:19 +0100131 ITensor *_input;
132 PadStrideInfo _info;
Manuel Bottinic1b76fa2019-06-17 12:04:40 +0100133 bool _is_prepared;
Pablo Tellof5f34bb2017-08-22 13:34:13 +0100134};
135} // arm_compute
Michalis Spyrouf4643372019-11-29 16:17:13 +0000136#endif /* ARM_COMPUTE_NEDECONVOLUTIONLAYER_H */