1 /* 2 * Copyright (c) 2018-2021 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 #ifndef ARM_COMPUTE_CLWINOGRADCONVOLUTIONLAYER_H 25 #define ARM_COMPUTE_CLWINOGRADCONVOLUTIONLAYER_H 26 27 #include "arm_compute/core/Types.h" 28 #include "arm_compute/runtime/IFunction.h" 29 #include "arm_compute/runtime/IMemoryManager.h" 30 31 #include <memory> 32 33 namespace arm_compute 34 { 35 class CLCompileContext; 36 class ICLTensor; 37 class ITensorInfo; 38 39 /** Basic function to execute Winograd-based convolution on OpenCL. This function calls the following OpenCL functions/kernels: 40 * 41 * -# @ref opencl::ClWinogradConv2d 42 * 43 */ 44 class CLWinogradConvolutionLayer : public IFunction 45 { 46 public: 47 /** Default Constructor */ 48 CLWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); 49 /** Default Destructor */ 50 ~CLWinogradConvolutionLayer(); 51 /** Prevent instances of this class from being copied (As this class contains pointers) */ 52 CLWinogradConvolutionLayer(const CLWinogradConvolutionLayer &) = delete; 53 /** Default move constructor */ 54 CLWinogradConvolutionLayer(CLWinogradConvolutionLayer &&) = default; 55 /** Prevent instances of this class from being copied (As this class contains pointers) */ 56 CLWinogradConvolutionLayer &operator=(const CLWinogradConvolutionLayer &) = delete; 57 /** Default move assignment operator */ 58 CLWinogradConvolutionLayer &operator=(CLWinogradConvolutionLayer &&) = default; 59 /** Set the input and output tensors. 60 * 61 * Valid data layouts: 62 * - NHWC 63 * - NCHW 64 * 65 * Valid data type configurations: 66 * |src0 |src1 |src2 |dst | 67 * |:--------------|:--------------|:------|:--------------| 68 * |F16 |F16 |F16 |F16 | 69 * |F32 |F32 |F32 |F32 | 70 * 71 * @note: This function only works with 3x3,3x1,1x3,5x5,5x1,1x5,7x1 and 1x7 kernels along with unit strides for both NCHW and NHWC data layout 72 * @note Some Winograd configurations (i.e. F(4x4, 5x5)) are supported only with enable_fast_math = true 73 * 74 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 75 * while every optional dimension from 4 and above represent a batch of inputs. 76 * Data types supported: F16/F32. 77 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 78 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 79 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 80 * Data types supported: Same as @p input. 81 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 82 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 83 * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 84 * available which may introduce a drop of accuracy as well. Default is false 85 */ 86 void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 87 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 88 /** Set the input and output tensors. 89 * 90 * @note: This function only works with 3x3,3x1,1x3,5x5,5x1,1x5,7x1 and 1x7 kernels along with unit strides for both NCHW and NHWC data layout 91 * @note Some Winograd configurations (i.e. F(4x4, 5x5)) are supported only with enable_fast_math = true 92 * 93 * @param[in] compile_context The compile context to be used. 94 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 95 * while every optional dimension from 4 and above represent a batch of inputs. 96 * Data types supported: F16/F32. 97 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 98 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 99 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 100 * Data types supported: Same as @p input. 101 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 102 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 103 * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 104 * available which may introduce a drop of accuracy as well. Default is false 105 */ 106 void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 107 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 108 /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradConvolutionLayer 109 * 110 * @note: This function only works with 3x3,3x1,1x3,5x5,5x1 and 1x5 kernels along with unit strides for both NCHW and NHWC data layout 111 * @note Some Winograd configurations (i.e. F(4x4, 5x5)) are supported only with enable_fast_math = true 112 * 113 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 114 * while every optional dimension from 4 and above represent a batch of inputs. 115 * Data types supported: F16/F32. 116 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 117 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 118 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 119 * Data types supported: Same as @p input. 120 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 121 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 122 * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 123 * available which may introduce a drop of accuracy as well. Default is false 124 * 125 * @return a status 126 */ 127 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, 128 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 129 130 // Inherited methods overridden: 131 void run() override; 132 void prepare() override; 133 134 private: 135 struct Impl; 136 std::unique_ptr<Impl> _impl; 137 }; 138 } // namespace arm_compute 139 #endif /* ARM_COMPUTE_CLWINOGRADCONVOLUTIONLAYER_H */ 140