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_METAL_COMPUTE_TASK_H_ 17 #define TENSORFLOW_LITE_DELEGATES_GPU_METAL_COMPUTE_TASK_H_ 18 19 #import <Metal/Metal.h> 20 21 #include <map> 22 #include <set> 23 #include <string> 24 #include <vector> 25 26 #include "tensorflow/lite/delegates/gpu/common/precision.h" 27 #include "tensorflow/lite/delegates/gpu/common/shape.h" 28 #include "tensorflow/lite/delegates/gpu/common/status.h" 29 #include "tensorflow/lite/delegates/gpu/common/task/gpu_operation.h" 30 #include "tensorflow/lite/delegates/gpu/common/task/tuning_type.h" 31 #include "tensorflow/lite/delegates/gpu/metal/common.h" 32 #include "tensorflow/lite/delegates/gpu/metal/metal_arguments.h" 33 #include "tensorflow/lite/delegates/gpu/metal/metal_device.h" 34 #include "tensorflow/lite/delegates/gpu/metal/metal_spatial_tensor.h" 35 36 namespace tflite { 37 namespace gpu { 38 namespace metal { 39 40 class ComputeTask { 41 public: 42 ComputeTask() = default; 43 ~ComputeTask(); 44 45 // Move only 46 ComputeTask(ComputeTask&& task); 47 ComputeTask& operator=(ComputeTask&& task); 48 ComputeTask(const ComputeTask&) = delete; 49 ComputeTask& operator=(const ComputeTask&) = delete; 50 51 void Init(std::unique_ptr<GPUOperation>&& operation); 52 53 const OperationDef& GetDefinition() const; GetGpuOperation()54 const GPUOperation& GetGpuOperation() const { return *operation_; } 55 56 absl::Status Compile(MetalDevice* device); 57 58 // should be called after changes of inputs/outputs. 59 absl::Status UpdateParams(); 60 61 void Encode(id<MTLComputeCommandEncoder> encoder); 62 63 API_AVAILABLE(ios(13.0), macos(11.00), tvos(13.0)) 64 void EncodeToICB(id<MTLIndirectComputeCommand> icb_command); 65 API_AVAILABLE(ios(11.0), macos(10.13), tvos(11.0)) 66 void AddResourcesToEncoder(id<MTLComputeCommandEncoder> encoder) const; 67 68 void Update(); 69 70 void SetSrcTensor(MetalSpatialTensor* tensor, int index); 71 72 void SetDstTensor(MetalSpatialTensor* tensor, int index); 73 74 absl::Status Tune(TuningType tuning_type, MetalDevice* device); 75 GetWorkGroupSize()76 int3 GetWorkGroupSize() const { return operation_->work_group_size_; } 77 void SetWorkGroupSize(const int3& work_group_size); 78 GetCode()79 const std::string& GetCode() const { return operation_->code_; } GetDefines()80 const std::map<std::string, std::string>& GetDefines() const { 81 return defines_; 82 } 83 84 absl::Status Init(MetalDevice* device, const std::string& code, 85 const std::map<std::string, std::string>& defines); 86 absl::Status RestoreDeserialized(MetalDevice* device); 87 88 private: 89 absl::Status CompileProgram( 90 MetalDevice* device, const std::string& code, 91 const std::map<std::string, std::string>& defines); 92 void Release(); 93 94 std::unique_ptr<GPUOperation> operation_; 95 id<MTLComputePipelineState> program_ = nullptr; 96 MetalArguments metal_args_; 97 98 bool use_arguments_buffer_ = false; // optional 99 bool need_icb_support_ = false; // optional 100 id<MTLArgumentEncoder> arguments_encoder_ = nullptr; 101 id<MTLBuffer> arg_buffer_ = nullptr; 102 103 // for serialization 104 std::map<std::string, std::string> defines_; 105 }; 106 107 } // namespace metal 108 } // namespace gpu 109 } // namespace tflite 110 111 #endif // TENSORFLOW_LITE_DELEGATES_GPU_METAL_COMPUTE_TASK_H_ 112