1 /* 2 * Copyright (c) 2019-2021 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_CLFFTCONVOLUTIONLAYER_H 25 #define ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H 26 27 #include "arm_compute/runtime/IFunction.h" 28 29 #include "arm_compute/core/Types.h" 30 #include "arm_compute/runtime/CL/functions/CLActivationLayer.h" 31 #include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h" 32 #include "arm_compute/runtime/CL/functions/CLFFT2D.h" 33 #include "arm_compute/runtime/CL/functions/CLPadLayer.h" 34 #include "arm_compute/runtime/CL/functions/CLPermute.h" 35 #include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h" 36 #include "arm_compute/runtime/CL/functions/CLReductionOperation.h" 37 #include "arm_compute/runtime/CL/functions/CLReshapeLayer.h" 38 #include "arm_compute/runtime/CL/functions/CLReverse.h" 39 #include "arm_compute/runtime/CL/functions/CLSlice.h" 40 41 namespace arm_compute 42 { 43 // Forward declarations 44 class ICLTensor; 45 46 /** Basic function to execute FFT-based convolution on OpenCL. This function calls the following OpenCL functions/kernels: 47 * 48 * -# @ref CLPermute Permute input if NHWC(only NCHW is supported). 49 * -# @ref CLPadLayer Pad input. 50 * -# @ref CLFFT2D Forward transform to the frequency domain. 51 * -# @ref CLComplexPixelWiseMultiplication Complex element-wise product of input and the weights. 52 * -# @ref CLReductionOperation Reduction across channels. 53 * -# @ref CLFFT2D Inverse transform back to the time domain. 54 * -# @ref CLStridedSlice Extract valid output. 55 * -# @ref CLArithmeticAddition Add bias. 56 * -# @ref CLActivationLayer Perform activation. 57 * -# @ref CLPermute Permute output if NHWC(only NCHW is supported). 58 */ 59 class CLFFTConvolutionLayer : public IFunction 60 { 61 public: 62 /** Default constructor */ 63 CLFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); 64 /** Prevent instances of this class from being copied (As this class contains pointers) */ 65 CLFFTConvolutionLayer(const CLFFTConvolutionLayer &) = delete; 66 /** Default move constructor */ 67 CLFFTConvolutionLayer(CLFFTConvolutionLayer &&) = default; 68 /** Prevent instances of this class from being copied (As this class contains pointers) */ 69 CLFFTConvolutionLayer &operator=(const CLFFTConvolutionLayer &) = delete; 70 /** Default move assignment operator */ 71 CLFFTConvolutionLayer &operator=(CLFFTConvolutionLayer &&) = default; 72 /** Set the input and output tensors. 73 * 74 * Valid data layouts: 75 * - All 76 * 77 * Valid data type configurations: 78 * |src |dst | 79 * |:------|:------| 80 * |F32 |F32 | 81 * |F16 |F16 | 82 * 83 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 84 * 85 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 86 * while every optional dimension from 4 and above represent a batch of inputs. 87 * Data types supported: F16/F32. 88 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 89 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 90 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 91 * Data types supported: Same as @p input. 92 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 93 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 94 * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 95 * available which may introduce a drop of accuracy as well. Default is false 96 */ 97 void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 98 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 99 /** Set the input and output tensors. 100 * 101 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 102 * 103 * @param[in] compile_context The compile context to be used. 104 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 105 * while every optional dimension from 4 and above represent a batch of inputs. 106 * Data types supported: F16/F32. 107 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 108 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 109 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 110 * Data types supported: Same as @p input. 111 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 112 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 113 * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 114 * available which may introduce a drop of accuracy as well. Default is false 115 */ 116 void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 117 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 118 /** Static function to check if given info will lead to a valid configuration of @ref CLFFTConvolutionLayer 119 * 120 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 121 * 122 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 123 * while every optional dimension from 4 and above represent a batch of inputs. 124 * Data types supported: F16/F32. 125 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 126 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 127 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 128 * Data types supported: Same as @p input. 129 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 130 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 131 * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 132 * available which may introduce a drop of accuracy as well. Default is false 133 * 134 * @return a status 135 */ 136 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, 137 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 138 139 // Inherited methods overridden: 140 void run() override; 141 void prepare() override; 142 143 private: 144 MemoryGroup _memory_group; 145 CLReverse _flip_weights_func; 146 CLPermute _permute_input_func; 147 CLPermute _permute_output_func; 148 CLPermute _permute_weights_func; 149 CLPermute _permute_bias_func; 150 CLPadLayer _pad_input_func; 151 CLPadLayer _pad_weights_func; 152 CLFFT2D _transform_input_func; 153 std::unique_ptr<CLFFT2D> _transform_weights_func; 154 CLFFT2D _itransform_output_func; 155 CLComplexPixelWiseMultiplication _prod_func; 156 CLReductionOperation _reduce_func; 157 CLSlice _extract_output_func; 158 CLArithmeticAddition _bias_add_func; 159 CLActivationLayer _activation_layer_func; 160 161 CLTensor _permuted_input; 162 CLTensor _permuted_weights; 163 CLTensor _permuted_bias; 164 CLTensor _permuted_output; 165 CLTensor _padded_input; 166 CLTensor _padded_weights; 167 CLTensor _flip_axis; 168 CLTensor _flipped_weights; 169 CLTensor _transformed_input; 170 CLTensor _transformed_weights; 171 CLTensor _input_weights_product; 172 CLTensor _output_product; 173 CLTensor _output_reduced; 174 CLTensor _itransformed_output; 175 CLTensor _reshaped_output; 176 CLTensor _bias_output; 177 178 const ICLTensor *_original_weights; 179 const ICLTensor *_original_bias; 180 bool _is_activationlayer_enabled; 181 bool _needs_permute; 182 bool _has_bias; 183 bool _is_prepared; 184 }; 185 } // namespace arm_compute 186 #endif /* ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H */ 187