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 #ifndef TENSORFLOW_CORE_KERNELS_CUDA_DEVICE_ARRAY_H_ 16 #define TENSORFLOW_CORE_KERNELS_CUDA_DEVICE_ARRAY_H_ 17 18 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM 19 20 #include "tensorflow/core/common_runtime/gpu/gpu_event_mgr.h" 21 #include "tensorflow/core/framework/op_kernel.h" 22 #include "tensorflow/core/kernels/gpu_device_array_gpu.h" 23 24 namespace tensorflow { 25 26 // Create an array of value on the host, to be sent to kernel using 27 // GpuDeviceArrayStruct. 28 // 29 // Usage: 30 // int size = ...; 31 // GpuDeviceArrayOnHost ptrs(context, size); 32 // OP_REQUIRES_OK(ptrs.Init()); 33 // for (int i = 0; i < size; ++i) { 34 // ptrs.Set(i, ...); 35 // } 36 // OP_REQUIRES_OK(ptrs.Finalize()); 37 // launchKernel(..., ptrs.data, ...); 38 // 39 // ValueType must be memcopyable. 40 template <typename ValueType, int MaxInlineValues = 8> 41 class GpuDeviceArrayOnHost { 42 public: GpuDeviceArrayOnHost(OpKernelContext * context,int32 size)43 GpuDeviceArrayOnHost(OpKernelContext* context, int32 size) 44 : context_(context), 45 total_bytes_(static_cast<int64>(size) * sizeof(ValueType)) { 46 data_.size = size; 47 } 48 Init()49 Status Init() { 50 if (inlined()) { 51 values_ = data_.inline_values; 52 return Status::OK(); 53 } 54 55 // Out-of-line: allocate data that will be memcopied. 56 AllocatorAttributes attr; 57 attr.set_on_host(true); 58 attr.set_gpu_compatible(true); 59 TF_RETURN_IF_ERROR( 60 context_->allocate_temp(DT_INT8, TensorShape{total_bytes_}, 61 &out_of_line_values_on_host_, attr)); 62 values_ = reinterpret_cast<ValueType*>( 63 out_of_line_values_on_host_.flat<int8>().data()); 64 return Status::OK(); 65 } 66 Set(int index,ValueType val)67 void Set(int index, ValueType val) { 68 DCHECK(values_); // ensure Init was called. 69 DCHECK_LT(index, data_.size); 70 *(values_ + index) = val; 71 } 72 Finalize()73 Status Finalize() { 74 if (inlined()) { 75 return Status::OK(); 76 } 77 78 // Out-of-line - copy pointers to device. 79 auto stream = context_->op_device_context()->stream(); 80 TensorReference tensor_ref(out_of_line_values_on_host_); 81 TF_RETURN_IF_ERROR(context_->allocate_temp( 82 DT_INT8, TensorShape{total_bytes_}, &out_of_line_values_on_gpu_)); 83 se::DeviceMemoryBase output_values_base{ 84 out_of_line_values_on_gpu_.flat<int8>().data(), 85 static_cast<uint64>(total_bytes_)}; 86 stream->ThenMemcpy(&output_values_base, 87 out_of_line_values_on_host_.flat<int8>().data(), 88 total_bytes_); 89 context_->device()->tensorflow_gpu_device_info()->event_mgr->ThenExecute( 90 stream, [tensor_ref]() { tensor_ref.Unref(); }); 91 data_.out_of_line_values = reinterpret_cast<ValueType*>( 92 out_of_line_values_on_gpu_.flat<int8>().data()); 93 return Status::OK(); 94 } 95 data()96 const GpuDeviceArrayStruct<ValueType, MaxInlineValues>& data() const { 97 // Ensure Finalize is called. 98 DCHECK(inlined() || out_of_line_values_on_gpu_.IsInitialized()); 99 return data_; 100 } 101 102 private: inlined()103 bool inlined() const { return data_.size <= MaxInlineValues; } 104 105 OpKernelContext* const context_; 106 const int64 total_bytes_; // total size of all pointers. 107 ValueType* values_ = nullptr; 108 GpuDeviceArrayStruct<ValueType, MaxInlineValues> data_; 109 110 Tensor out_of_line_values_on_host_; 111 Tensor out_of_line_values_on_gpu_; 112 113 TF_DISALLOW_COPY_AND_ASSIGN(GpuDeviceArrayOnHost); 114 }; 115 116 } // namespace tensorflow 117 118 #endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM 119 120 #endif // TENSORFLOW_CORE_KERNELS_CUDA_DEVICE_ARRAY_H_ 121