1 /* 2 * Copyright (c) 2017-2019 Arm Limited. 3 * 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 */ 24 25 #ifndef ARM_COMPUTE_GCIM2COLKERNEL_H 26 #define ARM_COMPUTE_GCIM2COLKERNEL_H 27 28 #include "arm_compute/core/GLES_COMPUTE/IGCKernel.h" 29 30 namespace arm_compute 31 { 32 class IGCTensor; 33 class Size2D; 34 35 /** Interface for the im2col reshape kernel. 36 * 37 * Rearranges image blocks into columns. It is used to strip out each convolution block to a single column. 38 * It is used to transform a convolution to a plain matrix multiplication. 39 * 40 * For example taking into account the image below and assuming 3x3 image blocks with stride of 1 we have: 41 * @f[ 42 * \left( \begin{array}{cccc} 43 * a00 & a01 & a02 & a03 \\ 44 * a10 & a11 & a12 & a13 \\ 45 * a20 & a21 & a22 & a23 \\ 46 * a30 & a31 & a32 & a33 \\ 47 * \end{array} \right) 48 * = 49 * \left( \begin{array}{ccccccccc} 50 * a00 & a01 & a02 & a10 & a11 & a12 & a20 & a21 & a22 \\ 51 * a01 & a02 & a03 & a11 & a12 & a13 & a21 & a22 & a23 \\ 52 * a10 & a11 & a12 & a20 & a21 & a22 & a30 & a31 & a32 \\ 53 * a11 & a12 & a13 & a21 & a22 & a23 & a31 & a32 & a33 \\ 54 * \end{array} \right) 55 * @f] 56 */ 57 class GCIm2ColKernel : public IGCKernel 58 { 59 public: 60 /** Default constructor */ 61 GCIm2ColKernel(); 62 /** Prevent instances of this class from being copied (As this class contains pointers) */ 63 GCIm2ColKernel(const GCIm2ColKernel &) = delete; 64 /** Prevent instances of this class from being copied (As this class contains pointers) */ 65 GCIm2ColKernel &operator=(const GCIm2ColKernel &) = delete; 66 /** Allow instances of this class to be moved */ 67 GCIm2ColKernel(GCIm2ColKernel &&) = default; 68 /** Allow instances of this class to be moved */ 69 GCIm2ColKernel &operator=(GCIm2ColKernel &&) = default; 70 /** Set the input and output of the kernel. 71 * 72 * @param[in] input The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM], 73 * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: F16/F32 74 * @param[out] output The output tensor. First 2 lower dimensions represent a transform of each 3D input, 75 * while every dimension above represents a batch. Data types supported: Same as @p input 76 * @param[in] kernel_dims The kernel dimensions (width and height). 77 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 78 * @param[in] has_bias In case biases are provided expands the matrix with 1. 79 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). 80 */ 81 void configure(const IGCTensor *input, IGCTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U)); 82 83 // Inherited methods overridden: 84 void run(const Window &window) override; 85 86 /** Static function to check if given info will lead to a valid configuration of @ref CLIm2ColKernel 87 * 88 * @param[in] input The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM], 89 * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: F16/F32 90 * @param[in] output The output tensor. First 2 lower dimensions represent a transform of each 3D input, 91 * while every dimension above represents a batch. Data types supported: Same as @p input 92 * @param[in] kernel_dims The kernel dimensions (width and height). 93 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 94 * @param[in] has_bias In case biases are provided expands the matrix with 1. 95 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). 96 * 97 * @return a status 98 */ 99 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)); 100 101 private: 102 /** Run the reshape kernel optimised for the special case (stride is 1, padding is 0 and kernel's low 3 dimensions are same as input) 103 * 104 * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). 105 * @param[in,out] queue Command queue on which to enqueue the kernel. 106 */ 107 void run_reduced(const Window &window); 108 /** run the generic convolution layer input reshape kernel 109 * 110 * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). 111 * @param[in,out] queue Command queue on which to enqueue the kernel. 112 */ 113 void run_generic(const Window &window); 114 115 /** Common signature for the kernel to run */ 116 using Im2ColFunction = void (GCIm2ColKernel::*)(const Window &); 117 118 private: 119 const IGCTensor *_input; 120 IGCTensor *_output; 121 std::pair<unsigned int, unsigned int> _convolved_dims; 122 std::pair<unsigned int, unsigned int> _kernel_dims; 123 unsigned int _num_elems_processed_per_iteration; 124 Im2ColFunction _run_func; 125 }; 126 } 127 128 #endif /*ARM_COMPUTE_GCIM2COLKERNEL_H */ 129