1 /* 2 * Copyright (c) 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 25 #ifndef ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H 26 #define ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H 27 28 #include "arm_compute/core/Types.h" 29 #include "arm_compute/core/experimental/IPostOp.h" 30 #include "arm_compute/runtime/IFunction.h" 31 32 namespace arm_compute 33 { 34 namespace graph 35 { 36 namespace backends 37 { 38 /** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}ConvolutionLayer with the modified weights */ 39 template <typename TargetInfo, typename FusedLayerTypes> 40 class FusedConvolutionBatchNormalizationWithPostOpsFunction : public IFunction 41 { 42 public: 43 using TensorType = typename TargetInfo::TensorType; 44 using TensorConcreteType = typename TargetInfo::TensorConcreteType; 45 46 FusedConvolutionBatchNormalizationWithPostOpsFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr) _conv_layer(memory_manager)47 : _conv_layer(memory_manager), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) 48 { 49 } 50 51 /** Set the input and output tensors. 52 * 53 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], 54 * while every optional dimension from 4 and above represent a batch of inputs. 55 * Data types supported: QASYMM8/F16/F32. 56 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. 57 * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. 58 * Data type supported: Should match @p input data type. 59 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. 60 * Data types supported: Same as @p input. 61 * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input 62 * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input 63 * @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 64 * @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 65 * @param[in] epsilon Small value to avoid division with zero. Default value is 0.001f. 66 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. 67 * @param[in] num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout 68 * @param[in] fast_math Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation 69 * available which may introduce a drop of accuracy as well. Default is false 70 * @param[in] post_ops A sequence of post operations that are performed after the main operation. 71 * 72 */ 73 void configure(TensorType *input, 74 TensorType *weights, 75 TensorType *bias, 76 TensorType *output, 77 const TensorType *mean, 78 const TensorType *var, 79 const TensorType *beta, 80 const TensorType *gamma, 81 float epsilon, const PadStrideInfo &conv_info, unsigned int num_groups, bool fast_math, 82 const arm_compute::experimental::PostOpList<TensorType *> &post_ops = experimental::PostOpList<TensorType *> {}) 83 { 84 // We don't run any validate, as we assume that the layers have been already validated 85 const bool has_bias = (bias != nullptr); 86 const TensorType *bias_to_use; 87 88 // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one 89 // as batch normalization might end up with a bias != 0 90 if(has_bias) 91 { 92 _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon); 93 bias_to_use = bias; 94 } 95 else 96 { 97 _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon); 98 bias_to_use = &_fused_bias; 99 } 100 101 ActivationLayerInfo fused_act = ActivationLayerInfo(); // Passing an empty ActivationLayerInfo. 102 _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups, post_ops); 103 104 if(!has_bias) 105 { 106 _fused_bias.allocator()->allocate(); 107 } 108 } 109 110 // Inherited methods overridden: run()111 void run() 112 { 113 prepare(); 114 _conv_layer.run(); 115 } 116 prepare()117 void prepare() 118 { 119 if(!_is_prepared) 120 { 121 _fused_batch_norm_layer.run(); 122 _is_prepared = true; 123 } 124 } 125 126 private: 127 typename FusedLayerTypes::ConvolutionLayer _conv_layer; 128 typename FusedLayerTypes::FuseBatchNormalization _fused_batch_norm_layer; 129 TensorConcreteType _fused_bias; 130 bool _is_prepared; 131 }; 132 } // namespace backends 133 } // namespace graph 134 } // namespace arm_compute 135 136 #endif /* ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H */ 137