1 /* Copyright 2019 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_LITE_DELEGATES_GPU_CL_TENSOR_H_ 17 #define TENSORFLOW_LITE_DELEGATES_GPU_CL_TENSOR_H_ 18 19 #include <cstdint> 20 #include <memory> 21 22 #include "tensorflow/lite/delegates/gpu/cl/cl_command_queue.h" 23 #include "tensorflow/lite/delegates/gpu/cl/cl_context.h" 24 #include "tensorflow/lite/delegates/gpu/cl/cl_device.h" 25 #include "tensorflow/lite/delegates/gpu/cl/cl_memory.h" 26 #include "tensorflow/lite/delegates/gpu/cl/gpu_object.h" 27 #include "tensorflow/lite/delegates/gpu/cl/util.h" 28 #include "tensorflow/lite/delegates/gpu/common/data_type.h" 29 #include "tensorflow/lite/delegates/gpu/common/shape.h" 30 #include "tensorflow/lite/delegates/gpu/common/status.h" 31 #include "tensorflow/lite/delegates/gpu/common/task/gpu_tensor.h" 32 #include "tensorflow/lite/delegates/gpu/common/task/tensor_desc.h" 33 #include "tensorflow/lite/delegates/gpu/common/tensor.h" 34 #include "tensorflow/lite/delegates/gpu/common/types.h" 35 36 namespace tflite { 37 namespace gpu { 38 namespace cl { 39 40 class Tensor : public GPUObject, public GpuSpatialTensor { 41 public: Tensor()42 Tensor() 43 : memory_(nullptr), image_buffer_memory_(nullptr), memory_owner_(true) {} 44 Tensor(cl_mem memory, bool memory_owner, const TensorDescriptor& descriptor); 45 Tensor(cl_mem memory, bool memory_owner, cl_mem image_buffer_memory, 46 const TensorDescriptor& descriptor); 47 48 // Move only 49 Tensor(Tensor&& tensor); 50 Tensor& operator=(Tensor&& tensor); 51 Tensor(const Tensor&) = delete; 52 Tensor& operator=(const Tensor&) = delete; 53 ~Tensor()54 ~Tensor() override { Release(); } 55 56 absl::Status GetGPUResources(const GPUObjectDescriptor* obj_ptr, 57 GPUResourcesWithValue* resources) const override; 58 Width()59 int Width() const override { return descriptor_.GetBHWDCShape().w; } Height()60 int Height() const override { return descriptor_.GetBHWDCShape().h; } Depth()61 int Depth() const override { return descriptor_.GetBHWDCShape().d; } Channels()62 int Channels() const override { return descriptor_.GetBHWDCShape().c; } Slices()63 int Slices() const override { 64 return DivideRoundUp(descriptor_.GetBHWDCShape().c, 4); 65 } Batch()66 int Batch() const override { return descriptor_.GetBHWDCShape().b; } 67 GetDescriptor()68 TensorDescriptor GetDescriptor() const override { return descriptor_; } GetDataType()69 DataType GetDataType() const { return descriptor_.GetDataType(); } GetStorageType()70 TensorStorageType GetStorageType() const { 71 return descriptor_.GetStorageType(); 72 } GetMemorySizeInBytes()73 uint64_t GetMemorySizeInBytes() const { 74 return descriptor_.GetMemorySizeInBytes(); 75 } 76 77 cl_mem GetMemoryPtr() const; 78 79 // This function returns buffer memory ptr for IMAGE_BUFFER instead of image 80 // memory ptr. 81 cl_mem GetMemoryPtrForWriting() const; 82 83 absl::Status CreateFromDescriptor(const TensorDescriptor& desc, 84 CLContext* context); 85 absl::Status UploadDescriptorData(const TensorDescriptor& desc, 86 CLCommandQueue* queue); 87 absl::Status ToDescriptor(TensorDescriptor* desc, 88 CLCommandQueue* queue) const; 89 90 private: 91 friend absl::Status CreateTensorSharedImage2DBuffer( 92 const CLContext& context, cl_mem memory, 93 const TensorDescriptor& descriptor, int width_pixel_alignment, 94 Tensor* result); 95 96 absl::Status WriteData(const void* ptr, CLCommandQueue* queue); 97 absl::Status ReadData(void* ptr, CLCommandQueue* queue) const; 98 99 void Release(); 100 101 cl_mem memory_; 102 cl_mem image_buffer_memory_; // for IMAGE_BUFFER/TEXTURE_2D/SINGLE_TEXTURE_2D 103 bool memory_owner_; 104 bool buffer_based_ = false; 105 TensorDescriptor descriptor_; 106 // for use with TEXTURE_2D and when texture created from buffer. 107 int aligned_texture_width_; 108 }; 109 110 using TensorPtr = std::shared_ptr<Tensor>; 111 112 absl::Status AllocateTensorMemory(const CLContext& context, 113 const TensorDescriptor& descriptor, 114 CLMemory* result); 115 116 absl::Status CreateTensor(const CLContext& context, 117 const TensorDescriptor& descriptor, Tensor* result); 118 119 absl::Status CreateTensorShared(const CLContext& context, cl_mem memory, 120 const TensorDescriptor& descriptor, 121 Tensor* result); 122 123 absl::Status CreateTensorSharedImage2DBuffer(const CLContext& context, 124 cl_mem memory, 125 const TensorDescriptor& descriptor, 126 int width_pixel_alignment, 127 Tensor* result); 128 129 130 } // namespace cl 131 } // namespace gpu 132 } // namespace tflite 133 134 #endif // TENSORFLOW_LITE_DELEGATES_GPU_CL_TENSOR_H_ 135