Tianle Cheng | 92ce35c | 2023-07-25 16:41:00 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2023 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <armnn/utility/IgnoreUnused.hpp> |
| 9 | |
| 10 | #include <tensorflow/lite/builtin_ops.h> |
| 11 | #include <tensorflow/lite/c/builtin_op_data.h> |
| 12 | #include <tensorflow/lite/c/common.h> |
| 13 | #include <tensorflow/lite/minimal_logging.h> |
| 14 | #include <tensorflow/lite/kernels/internal/tensor_ctypes.h> |
| 15 | #include <tensorflow/lite/schema/schema_generated.h> |
| 16 | |
| 17 | namespace armnnDelegate |
| 18 | { |
| 19 | TfLiteStatus ValidateTileOperator(DelegateData& delegateData, |
| 20 | TfLiteContext* tfLiteContext, |
| 21 | const armnn::TensorInfo& inputInfo, |
| 22 | const armnn::TensorInfo& outputInfo, |
| 23 | const armnn::TileDescriptor& descriptor) |
| 24 | { |
| 25 | bool isSupported = false; |
| 26 | FORWARD_LAYER_SUPPORT_FUNC("TILE", |
| 27 | tfLiteContext, |
| 28 | IsTileSupported, |
| 29 | delegateData.m_Backends, |
| 30 | isSupported, |
| 31 | armnn::BackendId(), |
| 32 | inputInfo, |
| 33 | outputInfo, |
| 34 | descriptor); |
| 35 | return isSupported ? kTfLiteOk : kTfLiteError; |
| 36 | } |
| 37 | |
| 38 | TfLiteStatus VisitTileOperator(DelegateData& delegateData, |
| 39 | TfLiteContext* tfLiteContext, |
| 40 | TfLiteNode* tfLiteNode, |
| 41 | int nodeIndex, |
| 42 | int32_t tileOperatorCode) |
| 43 | { |
| 44 | TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex)); |
| 45 | TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex)); |
| 46 | |
| 47 | const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors; |
| 48 | |
| 49 | // The input contains the data that should be tiled |
| 50 | const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]]; |
| 51 | if (IsDynamicTensor(tfLiteInputTensor)) |
| 52 | { |
| 53 | TF_LITE_MAYBE_KERNEL_LOG( |
| 54 | tfLiteContext, |
| 55 | "TfLiteArmnnDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ", |
| 56 | tileOperatorCode, nodeIndex); |
| 57 | return kTfLiteError; |
| 58 | } |
| 59 | |
| 60 | // The multiples tensor contains the number of copies for each axis |
| 61 | const TfLiteTensor& tfLiteMultiplesTensor = tfLiteTensors[tfLiteNode->inputs->data[1]]; |
| 62 | if (IsDynamicTensor(tfLiteMultiplesTensor)) |
| 63 | { |
| 64 | TF_LITE_MAYBE_KERNEL_LOG( |
| 65 | tfLiteContext, |
| 66 | "TfLiteArmnnDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ", |
| 67 | tileOperatorCode, nodeIndex); |
| 68 | return kTfLiteError; |
| 69 | } |
| 70 | |
| 71 | // The output tensor |
| 72 | const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]]; |
| 73 | if (IsDynamicTensor(tfLiteOutputTensor)) |
| 74 | { |
| 75 | TF_LITE_MAYBE_KERNEL_LOG( |
| 76 | tfLiteContext, |
| 77 | "TfLiteArmnnDelegate: Dynamic output tensors are not supported in operator #%d node #%d: ", |
| 78 | tileOperatorCode, nodeIndex); |
| 79 | return kTfLiteError; |
| 80 | } |
| 81 | |
| 82 | const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor); |
| 83 | const armnn::TensorInfo& multiplesTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteMultiplesTensor); |
| 84 | const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor); |
| 85 | |
| 86 | // Multiples length must be the same as the number of dimension in input tensor |
| 87 | if (multiplesTensorInfo.GetNumElements() != inputTensorInfo.GetNumDimensions()) |
| 88 | { |
| 89 | TF_LITE_MAYBE_KERNEL_LOG( |
| 90 | tfLiteContext, |
| 91 | "TfLiteArmnnDelegate: The Multiples length must be the same as the number of dimension in input tensor", |
| 92 | "Operator: #%d node #%d: ", |
| 93 | tileOperatorCode, nodeIndex); |
| 94 | return kTfLiteError; |
| 95 | } |
| 96 | |
| 97 | // Get the Multiples data: In armnn, the values of the multiples input tensor is saved in the operator descriptor |
| 98 | // We have to read it from the input tensor and write it the descriptor |
| 99 | auto* multiplesTensorDataPtr = tflite::GetTensorData<int32_t>(&tfLiteMultiplesTensor); |
| 100 | auto multiplesTensorNum = tfLiteMultiplesTensor.dims->data[0]; |
| 101 | std::vector<int32_t> multiplesIntData(multiplesTensorDataPtr, multiplesTensorDataPtr + multiplesTensorNum); |
| 102 | |
| 103 | // The multiples must be positive |
| 104 | for (auto multiple : multiplesIntData) |
| 105 | { |
| 106 | if (multiple < 0) |
| 107 | { |
| 108 | TF_LITE_MAYBE_KERNEL_LOG( |
| 109 | tfLiteContext, |
| 110 | "TfLiteArmnnDelegate: The Multiples must be positive values", |
| 111 | "Operator: #%d node #%d: ", |
| 112 | tileOperatorCode, nodeIndex); |
| 113 | return kTfLiteError; |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | // The original input from TFLite is int32, and we have to make it as uint32 for our descriptor |
| 118 | std::vector<uint32_t> multiplesUintData; |
| 119 | std::transform(multiplesIntData.begin(), |
| 120 | multiplesIntData.end(), |
| 121 | std::back_inserter(multiplesUintData), |
| 122 | [] (const int value) |
| 123 | { |
| 124 | return static_cast<uint32_t>(value); |
| 125 | }); |
| 126 | |
| 127 | armnn::TileDescriptor tileDescriptor; |
| 128 | tileDescriptor.m_Multiples = multiplesUintData; |
| 129 | |
| 130 | // Check output dimensions |
| 131 | if (inputTensorInfo.GetNumDimensions() != outputTensorInfo.GetNumDimensions()) |
| 132 | { |
| 133 | TF_LITE_MAYBE_KERNEL_LOG( |
| 134 | tfLiteContext, |
| 135 | "TfLiteArmnnDelegate: Input tensor dimension and output tensor dimension differ", |
| 136 | "Operator: #%d node #%d: ", |
| 137 | tileOperatorCode, nodeIndex); |
| 138 | return kTfLiteError; |
| 139 | } |
| 140 | |
| 141 | // No network pointer indicates that only support for this operator should be checked |
| 142 | if (!delegateData.m_Network) |
| 143 | { |
| 144 | return ValidateTileOperator(delegateData, |
| 145 | tfLiteContext, |
| 146 | inputTensorInfo, |
| 147 | outputTensorInfo, |
| 148 | tileDescriptor); |
| 149 | } |
| 150 | |
Mike Kelly | 07169c8 | 2023-08-02 13:23:09 +0100 | [diff] [blame] | 151 | auto layerName = GetLayerName(armnn::LayerType::Tile, nodeIndex); |
Tianle Cheng | 92ce35c | 2023-07-25 16:41:00 +0100 | [diff] [blame] | 152 | armnn::IConnectableLayer* layer = delegateData.m_Network->AddTileLayer(tileDescriptor, layerName.c_str()); |
| 153 | |
| 154 | if (layer == nullptr) |
| 155 | { |
| 156 | return kTfLiteError; |
| 157 | } |
| 158 | |
| 159 | layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 160 | |
Mike Kelly | 07169c8 | 2023-08-02 13:23:09 +0100 | [diff] [blame] | 161 | if (ProcessInputs(layer, delegateData, tfLiteContext, tfLiteNode, nodeIndex) != kTfLiteOk) |
Tianle Cheng | 92ce35c | 2023-07-25 16:41:00 +0100 | [diff] [blame] | 162 | { |
| 163 | return kTfLiteError; |
| 164 | } |
| 165 | |
| 166 | return Connect(layer, tfLiteNode, delegateData); |
| 167 | } |
| 168 | |
| 169 | } // namespace armnnDelegate |