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_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H 26 #define ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_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}ConvolutionLayer with the modified weights */ 38 template <typename TargetInfo, typename FusedLayerTypes> 39 class FusedConvolutionBatchNormalizationFunction : public IFunction 40 { 41 public: 42 using TensorType = typename TargetInfo::TensorType; 43 using TensorConcreteType = typename TargetInfo::TensorConcreteType; 44 45 FusedConvolutionBatchNormalizationFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr) _conv_layer(memory_manager)46 : _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: QASYMM8/F16/F32. 55 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. 56 * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. 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] num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout 67 * @param[in] fast_math Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 68 * available which may introduce a drop of accuracy as well. Default is false 69 * @param[in] fused_act Activation layer information in case of a fused activation. 70 * 71 */ 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 num_groups,bool fast_math,ActivationLayerInfo const & fused_act)72 void configure(TensorType *input, 73 TensorType *weights, 74 TensorType *bias, 75 TensorType *output, 76 const TensorType *mean, 77 const TensorType *var, 78 const TensorType *beta, 79 const TensorType *gamma, 80 float epsilon, const PadStrideInfo &conv_info, unsigned int num_groups, bool fast_math, ActivationLayerInfo const &fused_act) 81 { 82 // We don't run any validate, as we assume that the layers have been already validated 83 const bool has_bias = (bias != nullptr); 84 const TensorType *bias_to_use; 85 86 // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one 87 // as batch normalization might end up with a bias != 0 88 if(has_bias) 89 { 90 _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon); 91 bias_to_use = bias; 92 } 93 else 94 { 95 _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon); 96 bias_to_use = &_fused_bias; 97 } 98 99 _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups); 100 101 if(!has_bias) 102 { 103 _fused_bias.allocator()->allocate(); 104 } 105 } 106 107 // Inherited methods overridden: run()108 void run() 109 { 110 prepare(); 111 _conv_layer.run(); 112 } 113 prepare()114 void prepare() 115 { 116 if(!_is_prepared) 117 { 118 _fused_batch_norm_layer.run(); 119 _is_prepared = true; 120 } 121 } 122 123 private: 124 typename FusedLayerTypes::ConvolutionLayer _conv_layer; 125 typename FusedLayerTypes::FuseBatchNormalization _fused_batch_norm_layer; 126 TensorConcreteType _fused_bias; 127 bool _is_prepared; 128 }; 129 } // namespace backends 130 } // namespace graph 131 } // namespace arm_compute 132 133 #endif /* ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H */ 134