1 /* Copyright 2015 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 #define EIGEN_USE_THREADS 17 18 #include "tensorflow/core/kernels/concat_lib_cpu.h" 19 #include <vector> 20 #include "tensorflow/core/framework/register_types.h" 21 #include "tensorflow/core/kernels/concat_lib.h" 22 23 namespace tensorflow { 24 25 namespace { 26 template <typename T> 27 struct MemCpyCopier { 28 inline void Copy(T* dst, const T* src, int input_index, size_t n) { 29 if (DataTypeCanUseMemcpy(DataTypeToEnum<T>::v())) { 30 memcpy(dst, src, n * sizeof(T)); 31 } else { 32 for (size_t k = 0; k < n; ++k) { 33 *dst++ = *src++; 34 } 35 } 36 } 37 }; 38 template <> 39 struct MemCpyCopier<ResourceHandle> { 40 inline void Copy(ResourceHandle* dst, const ResourceHandle* src, 41 int input_index, size_t n) { 42 for (size_t k = 0; k < n; ++k) { 43 *dst++ = *src++; 44 } 45 } 46 }; 47 48 } // namespace 49 50 template <typename T> 51 void ConcatCPU( 52 DeviceBase* d, 53 const std::vector<std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>& 54 inputs, 55 typename TTypes<T, 2>::Matrix* output) { 56 if (std::is_same<T, string>::value) { 57 // use a large cost here to force strings to be handled by separate threads 58 ConcatCPUImpl<T>(d, inputs, 100000, MemCpyCopier<T>(), output); 59 } else { 60 ConcatCPUImpl<T>(d, inputs, sizeof(T) /* cost_per_unit */, 61 MemCpyCopier<T>(), output); 62 } 63 } 64 65 #define REGISTER(T) \ 66 template void ConcatCPU<T>( \ 67 DeviceBase*, \ 68 const std::vector<std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>&, \ 69 typename TTypes<T, 2>::Matrix* output); 70 TF_CALL_ALL_TYPES(REGISTER) 71 REGISTER(quint8) 72 REGISTER(qint8) 73 REGISTER(quint16) 74 REGISTER(qint16) 75 REGISTER(qint32) 76 77 #if defined(IS_MOBILE_PLATFORM) && !defined(SUPPORT_SELECTIVE_REGISTRATION) && \ 78 !defined(__ANDROID_TYPES_FULL__) 79 // Primarily used for SavedModel support on mobile. Registering it here only 80 // if __ANDROID_TYPES_FULL__ is not defined (which already registers string) 81 // to avoid duplicate registration. 82 REGISTER(string); 83 #endif // defined(IS_MOBILE_PLATFORM) && 84 // !defined(SUPPORT_SELECTIVE_REGISTRATION) && 85 // !defined(__ANDROID_TYPES_FULL__) 86 87 #ifdef TENSORFLOW_USE_SYCL 88 template <typename T> 89 void ConcatSYCL( 90 const Eigen::SyclDevice& d, 91 const std::vector<std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>& 92 inputs, 93 typename TTypes<T, 2>::Matrix* output) { 94 ConcatSYCLImpl<T>(d, inputs, sizeof(T) /* cost_per_unit */, MemCpyCopier<T>(), 95 output); 96 } 97 #define REGISTER_SYCL(T) \ 98 template void ConcatSYCL<T>( \ 99 const Eigen::SyclDevice&, \ 100 const std::vector<std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>&, \ 101 typename TTypes<T, 2>::Matrix* output); 102 103 TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SYCL) 104 105 #undef REGISTER_SYCL 106 #endif // TENSORFLOW_USE_SYCL 107 } // namespace tensorflow 108