1 /* 2 * Copyright (c) 2019-2020 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 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 75 * 76 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 77 * while every optional dimension from 4 and above represent a batch of inputs. 78 * Data types supported: F32. 79 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 80 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 81 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 82 * Data types supported: Same as @p input. 83 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 84 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 85 */ 86 void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 87 const ActivationLayerInfo &act_info = ActivationLayerInfo()); 88 /** Set the input and output tensors. 89 * 90 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 91 * 92 * @param[in] compile_context The compile context to be used. 93 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 94 * while every optional dimension from 4 and above represent a batch of inputs. 95 * Data types supported: F32. 96 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 97 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 98 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 99 * Data types supported: Same as @p input. 100 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 101 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 102 */ 103 void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 104 const ActivationLayerInfo &act_info = ActivationLayerInfo()); 105 /** Static function to check if given info will lead to a valid configuration of @ref CLFFTConvolutionLayer 106 * 107 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 108 * 109 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 110 * while every optional dimension from 4 and above represent a batch of inputs. 111 * Data types supported: F32. 112 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 113 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 114 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 115 * Data types supported: Same as @p input. 116 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 117 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 118 * 119 * @return a status 120 */ 121 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, 122 const ActivationLayerInfo &act_info = ActivationLayerInfo()); 123 124 // Inherited methods overridden: 125 void run() override; 126 void prepare() override; 127 128 private: 129 MemoryGroup _memory_group; 130 CLReverse _flip_weights_func; 131 CLPermute _permute_input_func; 132 CLPermute _permute_output_func; 133 CLPermute _permute_weights_func; 134 CLPermute _permute_bias_func; 135 CLPadLayer _pad_input_func; 136 CLPadLayer _pad_weights_func; 137 CLFFT2D _transform_input_func; 138 std::unique_ptr<CLFFT2D> _transform_weights_func; 139 CLFFT2D _itransform_output_func; 140 CLComplexPixelWiseMultiplication _prod_func; 141 CLReductionOperation _reduce_func; 142 CLSlice _extract_output_func; 143 CLArithmeticAddition _bias_add_func; 144 CLActivationLayer _activation_layer_func; 145 146 CLTensor _permuted_input; 147 CLTensor _permuted_weights; 148 CLTensor _permuted_bias; 149 CLTensor _permuted_output; 150 CLTensor _padded_input; 151 CLTensor _padded_weights; 152 CLTensor _flip_axis; 153 CLTensor _flipped_weights; 154 CLTensor _transformed_input; 155 CLTensor _transformed_weights; 156 CLTensor _input_weights_product; 157 CLTensor _output_product; 158 CLTensor _output_reduced; 159 CLTensor _itransformed_output; 160 CLTensor _reshaped_output; 161 CLTensor _bias_output; 162 163 const ICLTensor *_original_weights; 164 const ICLTensor *_original_bias; 165 bool _is_activationlayer_enabled; 166 bool _needs_permute; 167 bool _has_bias; 168 bool _is_prepared; 169 }; 170 } // namespace arm_compute 171 #endif /* ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H */ 172