1 /* 2 * Copyright (c) 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_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H 26 #define ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H 27 28 #include "arm_compute/core/Types.h" 29 #include "arm_compute/runtime/IFunction.h" 30 31 namespace arm_compute 32 { 33 namespace graph 34 { 35 namespace backends 36 { 37 /** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}DepthwiseConvolutionLayer with the modified weights */ 38 template <typename TargetInfo, typename FusedLayerTypes> 39 class FusedDepthwiseConvolutionBatchNormalizationFunction : public IFunction 40 { 41 public: 42 using TensorType = typename TargetInfo::TensorType; 43 using TensorConcreteType = typename TargetInfo::TensorConcreteType; 44 45 FusedDepthwiseConvolutionBatchNormalizationFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr) _depth_conv_layer(memory_manager)46 : _depth_conv_layer(memory_manager), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) 47 { 48 } 49 50 /** Set the input and output tensors. 51 * 52 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 53 * while every optional dimension from 4 and above represent a batch of inputs. 54 * Data types supported: F16/F32. 55 * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input. 56 * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [IFM]. 57 * Data type supported: Should match @p input data type. 58 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 59 * Data types supported: Same as @p input. 60 * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input 61 * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input 62 * @param[in] beta Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input 63 * @param[in] gamma Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input 64 * @param[in] epsilon Small value to avoid division with zero. Default value is 0.001f. 65 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 66 * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. 67 * @param[in] fused_act Activation layer information in case of a fused activation. 68 * 69 */ configure(TensorType * input,TensorType * weights,TensorType * bias,TensorType * output,const TensorType * mean,const TensorType * var,const TensorType * beta,const TensorType * gamma,float epsilon,const PadStrideInfo & conv_info,unsigned int depth_multiplier,ActivationLayerInfo const & fused_act)70 void configure(TensorType *input, 71 TensorType *weights, 72 TensorType *bias, 73 TensorType *output, 74 const TensorType *mean, 75 const TensorType *var, 76 const TensorType *beta, 77 const TensorType *gamma, 78 float epsilon, const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo const &fused_act) 79 { 80 // We don't run any validate, as we assume that the layers have been already validated 81 const bool has_bias = (bias != nullptr); 82 const TensorType *bias_to_use; 83 84 // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one 85 // as batch normalization might end up with a bias != 0 86 if(has_bias) 87 { 88 _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION); 89 bias_to_use = bias; 90 } 91 else 92 { 93 _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION); 94 bias_to_use = &_fused_bias; 95 } 96 97 _depth_conv_layer.configure(input, weights, bias_to_use, output, conv_info, depth_multiplier, fused_act.enabled() ? fused_act : ActivationLayerInfo()); 98 99 if(!has_bias) 100 { 101 _fused_bias.allocator()->allocate(); 102 } 103 } 104 105 // Inherited methods overridden: run()106 void run() 107 { 108 prepare(); 109 _depth_conv_layer.run(); 110 } 111 prepare()112 void prepare() 113 { 114 if(!_is_prepared) 115 { 116 _fused_batch_norm_layer.run(); 117 _is_prepared = true; 118 } 119 } 120 121 private: 122 typename FusedLayerTypes::DepthwiseConvolutionLayer _depth_conv_layer; 123 typename FusedLayerTypes::FuseBatchNormalization _fused_batch_norm_layer; 124 TensorConcreteType _fused_bias; 125 bool _is_prepared; 126 }; 127 } // namespace backends 128 } // namespace graph 129 } // namespace arm_compute 130 131 #endif /* ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H */ 132