1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 16 #ifndef TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_ 17 #define TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_ 18 19 // This file requires the following include because it uses CudaAtomicMax: 20 // #include "tensorflow/core/util/cuda_kernel_helper.h" 21 22 // Unfortunately we can't add the #include, since it breaks compilation for 23 // non-GPU targets. This only breaks in clang, because it's more strict for 24 // template code and CudaAtomicMax is used in template context. 25 26 // This file requires the following include because it uses CudaAtomicMax: 27 // #include "tensorflow/core/util/cuda_kernel_helper.h" 28 29 // Unfortunately we can't add the #include, since it breaks compilation for 30 // non-GPU targets. This only breaks in clang, because it's more strict for 31 // template code and CudaAtomicMax is used in template context. 32 33 // This file requires the following include because it uses CudaAtomicMax: 34 // #include "tensorflow/core/util/cuda_kernel_helper.h" 35 36 // Unfortunately we can't add the #include, since it breaks compilation for 37 // non-GPU targets. This only breaks in clang, because it's more strict for 38 // template code and CudaAtomicMax is used in template context. 39 40 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" 41 #include "tensorflow/core/framework/tensor.h" 42 #include "tensorflow/core/framework/tensor_shape.h" 43 #include "tensorflow/core/framework/tensor_types.h" 44 45 namespace tensorflow { 46 47 class OpKernelContext; 48 49 namespace functor { 50 51 #ifdef GOOGLE_CUDA 52 typedef Eigen::GpuDevice GPUDevice; 53 // Functor for SegmentSumGPUOp. 54 // output_rows: the number of output segments (unique segment ids in 55 // 'segment_ids'). 56 // segment_ids_shape: shape of 'segment_ids' tensor. 57 // segment_ids: unsorted map from input to output segment ids at which to 58 // perform segment sum operation. 59 // data_size: size of input data tensor. 60 // data: input data tensor. 61 // output: output reshaped to {output_rows, output.size/output_rows} 62 template <typename T, typename Index> 63 struct SegmentSumFunctor { 64 void operator()(OpKernelContext* ctx, const GPUDevice& d, 65 const Index output_rows, const TensorShape& segment_ids_shape, 66 typename TTypes<Index>::ConstFlat segment_ids, 67 const Index data_size, const T* data, 68 typename TTypes<T, 2>::Tensor output); 69 }; 70 71 #endif 72 73 template <typename Device, typename T, typename Index, typename InitialValueF, 74 typename ReductionF> 75 struct UnsortedSegmentFunctor { 76 void operator()(OpKernelContext* ctx, const Index num_segments, 77 const TensorShape& segment_ids_shape, 78 typename TTypes<Index>::ConstFlat segment_ids, 79 const Index data_size, const T* data, 80 typename TTypes<T, 2>::Tensor output); 81 }; 82 83 #ifdef GOOGLE_CUDA 84 // reduction functors for the gpu 85 template <typename T> 86 struct SumOpGpu { operatorSumOpGpu87 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(T* dest, 88 const T& value) { 89 CudaAtomicAdd(dest, value); 90 } 91 }; 92 93 template <typename T> 94 struct ProdOpGpu { operatorProdOpGpu95 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(T* dest, 96 const T& value) { 97 CudaAtomicMul(dest, value); 98 } 99 }; 100 101 template <typename T> 102 struct MaxOpGpu { operatorMaxOpGpu103 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(T* dest, 104 const T& value) { 105 CudaAtomicMax(dest, value); 106 } 107 }; 108 109 template <typename T> 110 struct MinOpGpu { operatorMinOpGpu111 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(T* dest, 112 const T& value) { 113 CudaAtomicMin(dest, value); 114 } 115 }; 116 117 #endif // GOOGLE_CUDA 118 119 // initial value functors 120 template <typename T> 121 struct Zero { operatorZero122 EIGEN_STRONG_INLINE T operator()() const { return T(0); } 123 }; 124 125 template <typename T> 126 struct One { operatorOne127 EIGEN_STRONG_INLINE T operator()() const { return T(1); } 128 }; 129 130 template <typename T> 131 struct Lowest { operatorLowest132 EIGEN_STRONG_INLINE T operator()() const { 133 return Eigen::NumTraits<T>::lowest(); 134 } 135 }; 136 137 template <typename T> 138 struct Highest { operatorHighest139 EIGEN_STRONG_INLINE T operator()() const { 140 return Eigen::NumTraits<T>::highest(); 141 } 142 }; 143 144 } // namespace functor 145 } // namespace tensorflow 146 147 #endif // TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_ 148