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_NEFFTCONVOLUTIONLAYER_H 25 #define ARM_COMPUTE_NEFFTCONVOLUTIONLAYER_H 26 27 #include "arm_compute/runtime/IFunction.h" 28 29 #include "arm_compute/core/Types.h" 30 #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" 31 #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h" 32 #include "arm_compute/runtime/NEON/functions/NEFFT2D.h" 33 #include "arm_compute/runtime/NEON/functions/NEPadLayer.h" 34 #include "arm_compute/runtime/NEON/functions/NEPermute.h" 35 #include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h" 36 #include "arm_compute/runtime/NEON/functions/NEReductionOperation.h" 37 #include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h" 38 #include "arm_compute/runtime/NEON/functions/NEReverse.h" 39 #include "arm_compute/runtime/NEON/functions/NESlice.h" 40 41 namespace arm_compute 42 { 43 // Forward declarations 44 class ITensor; 45 46 /** Basic function to execute FFT-based convolution on CPU. This function calls the following functions/kernels: 47 * 48 * -# @ref NEPermute Permute input if NHWC(only NCHW is supported). 49 * -# @ref NEPadLayer Pad input. 50 * -# @ref NEFFT2D Forward transform to the frequency domain. 51 * -# @ref NEComplexPixelWiseMultiplication Complex element-wise product of input and the weights. 52 * -# @ref NEReductionOperation Reduction across channels. 53 * -# @ref NEFFT2D Inverse transform back to the time domain. 54 * -# @ref NEStridedSlice Extract valid output. 55 * -# @ref NEArithmeticAddition Add bias. 56 * -# @ref NEActivationLayer Perform activation. 57 * -# @ref NEPermute Permute output if NHWC(only NCHW is supported). 58 */ 59 class NEFFTConvolutionLayer : public IFunction 60 { 61 public: 62 /** Default constructor */ 63 NEFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); 64 /** Prevent instances of this class from being copied (As this class contains pointers) */ 65 NEFFTConvolutionLayer(const NEFFTConvolutionLayer &) = delete; 66 /** Prevent instances of this class from being moved (As this class contains non movable objects) */ 67 NEFFTConvolutionLayer(NEFFTConvolutionLayer &&) = delete; 68 /** Prevent instances of this class from being copied (As this class contains pointers) */ 69 NEFFTConvolutionLayer &operator=(const NEFFTConvolutionLayer &) = delete; 70 /** Prevent instances of this class from being moved (As this class contains non movable objects) */ 71 NEFFTConvolutionLayer &operator=(NEFFTConvolutionLayer &&) = delete; 72 /** Default destructor */ 73 ~NEFFTConvolutionLayer(); 74 /** Set the input and output tensors. 75 * 76 * Valid data layouts: 77 * - All 78 * 79 * Valid data type configurations: 80 * |src |dst | 81 * |:------|:------| 82 * |F32 |F32 | 83 * 84 * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout 85 * 86 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 87 * while every optional dimension from 4 and above represent a batch of inputs. 88 * Data types supported: F32. 89 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 90 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 91 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 92 * Data types supported: Same as @p input. 93 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 94 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 95 * @param[in] enable_fast_math (Optional) Enable fast math computation. Unused for CPU backend. 96 */ 97 void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, 98 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 99 /** Static function to check if given info will lead to a valid configuration of @ref NEFFTConvolutionLayer 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] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 104 * while every optional dimension from 4 and above represent a batch of inputs. 105 * Data types supported: F32. 106 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. 107 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input 108 * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 109 * Data types supported: Same as @p input. 110 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 111 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 112 * @param[in] enable_fast_math (Optional) Enable fast math computation. Unused for CPU backend. 113 * 114 * @return a status 115 */ 116 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, 117 const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); 118 119 // Inherited methods overridden: 120 void run() override; 121 void prepare() override; 122 123 private: 124 MemoryGroup _memory_group; 125 NEReverse _flip_weights_func; 126 NEPermute _permute_input_func; 127 NEPermute _permute_output_func; 128 NEPermute _permute_weights_func; 129 NEPermute _permute_bias_func; 130 NEPadLayer _pad_input_func; 131 NEPadLayer _pad_weights_func; 132 NEFFT2D _transform_input_func; 133 std::unique_ptr<NEFFT2D> _transform_weights_func; 134 NEFFT2D _itransform_output_func; 135 NEComplexPixelWiseMultiplication _prod_func; 136 NEReductionOperation _reduce_func; 137 NESlice _extract_output_func; 138 NEArithmeticAddition _bias_add_func; 139 NEActivationLayer _activation_layer_func; 140 141 Tensor _permuted_input; 142 Tensor _permuted_weights; 143 Tensor _permuted_bias; 144 Tensor _permuted_output; 145 Tensor _padded_input; 146 Tensor _padded_weights; 147 Tensor _flip_axis; 148 Tensor _flipped_weights; 149 Tensor _transformed_input; 150 Tensor _transformed_weights; 151 Tensor _input_weights_product; 152 Tensor _output_product; 153 Tensor _output_reduced; 154 Tensor _itransformed_output; 155 Tensor _reshaped_output; 156 Tensor _bias_output; 157 158 const ITensor *_original_weights; 159 const ITensor *_original_bias; 160 bool _is_activationlayer_enabled; 161 bool _needs_permute; 162 bool _has_bias; 163 bool _is_prepared; 164 }; 165 } // namespace arm_compute 166 #endif /* ARM_COMPUTE_NEFFTCONVOLUTIONLAYER_H */ 167