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_COMMON_MODEL_BUILDER_H_ 17 #define TENSORFLOW_LITE_DELEGATES_GPU_COMMON_MODEL_BUILDER_H_ 18 19 #include "absl/container/flat_hash_map.h" 20 #include "tensorflow/lite/c/common.h" 21 #include "tensorflow/lite/delegates/gpu/common/model.h" 22 #include "tensorflow/lite/delegates/gpu/common/shape.h" 23 #include "tensorflow/lite/delegates/gpu/common/status.h" 24 #include "tensorflow/lite/delegates/gpu/common/tensor.h" 25 26 namespace tflite { 27 namespace gpu { 28 29 // Validates which operations are supported and returns array of operations to 30 // replace with GPU kernels. The caller must free the pointer on TfLiteIntArray. 31 // 'max_delegated_partitions' limits the maximum number of partitions to 32 // delegate as a graph could possibly have multiple partitions (each partition 33 // consists of a subset of ops) to be replaced. 34 TfLiteIntArray* GetOpsToReplace(TfLiteContext* context, 35 bool allow_quant_ops = false, 36 int max_delegated_partitions = 1); 37 38 // Extracts TFLite delegate execution plan from the input TFLite context and 39 // converts it into generic graph format. 40 // 41 // If model is quantized, quant_conversion_map maps the dequantized tensor 42 // (floating-point) to the original tensor (fixed-point) & vice-versa. 43 // NOTE: Not all of these new tensors will have any data and need memory 44 // allocated for them. We need to do that only for the overall GPU graph inputs 45 // & outputs. This should be done by the delegate, by setting the appropriate 46 // TfLiteNode->temporaries. 47 absl::Status BuildModel( 48 TfLiteContext* context, const TfLiteDelegateParams* delegate_params, 49 GraphFloat32* graph, 50 absl::flat_hash_map<int, int>* quant_conversion_map = nullptr); 51 52 // Same as above but also apply all transformations on the final graph. 53 // Prefer using this method instead of BuildModel. 54 // 55 // If model is quantized, quant_conversion_map maps the dequantized tensor 56 // (floating-point) to the original TFLite tensor (fixed-point) & vice-versa. 57 // NOTE: Not all of these new tensors will have any data and need memory 58 // allocated for them. We need to do that only for the overall GPU graph inputs 59 // & outputs. This should be done by the delegate, by setting the appropriate 60 // TfLiteNode->temporaries. 61 absl::Status BuildFinalModel( 62 TfLiteContext* context, const TfLiteDelegateParams* delegate_params, 63 GraphFloat32* graph, 64 absl::flat_hash_map<int, int>* quant_conversion_map = nullptr); 65 66 // Module-internal converter, exposed for unit testing purpose only. 67 absl::Status ConvertTfLiteTensorToTensorRef(const TfLiteTensor& tflite_tensor, 68 TensorRef<BHWC>* tensor_ref); 69 70 } // namespace gpu 71 } // namespace tflite 72 73 #endif // TENSORFLOW_LITE_DELEGATES_GPU_COMMON_MODEL_BUILDER_H_ 74