Sadik Armagan | 62483be | 2020-10-23 17:14:43 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
Sadik Armagan | 6e36a64 | 2020-11-10 21:18:41 +0000 | [diff] [blame] | 8 | #include "DelegateUtils.hpp" |
| 9 | |
Sadik Armagan | 62483be | 2020-10-23 17:14:43 +0100 | [diff] [blame] | 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 | |
| 15 | namespace armnnDelegate |
| 16 | { |
| 17 | |
| 18 | TfLiteStatus VisitFullyConnectedOperator(DelegateData& delegateData, |
| 19 | TfLiteContext* tfLiteContext, |
| 20 | TfLiteNode* tfLiteNode, |
| 21 | int nodeIndex, |
| 22 | int32_t operatorCode) |
| 23 | { |
Sadik Armagan | 6e36a64 | 2020-11-10 21:18:41 +0000 | [diff] [blame] | 24 | auto numInputs = tfLiteNode->inputs->size; |
| 25 | if (numInputs < 2) |
| 26 | { |
| 27 | TF_LITE_MAYBE_KERNEL_LOG( |
| 28 | tfLiteContext, "TfLiteArmnnDelegate: Minimum number of inputs (%d != %d) in node #%d", |
| 29 | 2, numInputs, nodeIndex); |
| 30 | return kTfLiteError; |
| 31 | } |
| 32 | TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex)); |
| 33 | bool biasEnabled = (numInputs == 3); |
| 34 | |
| 35 | const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors; |
| 36 | const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]]; |
| 37 | if(!IsValid(&tfLiteTensors[tfLiteNode->inputs->data[0]])) |
| 38 | { |
| 39 | TF_LITE_MAYBE_KERNEL_LOG( |
| 40 | tfLiteContext, |
| 41 | "TfLiteArmnnDelegate: Invalid input tensor in operator #%d node #%d: ", |
| 42 | operatorCode, nodeIndex); |
| 43 | return kTfLiteError; |
| 44 | } |
| 45 | if (IsDynamicTensor(tfLiteInputTensor)) |
| 46 | { |
| 47 | TF_LITE_MAYBE_KERNEL_LOG( |
| 48 | tfLiteContext, |
| 49 | "TfLiteArmnnDelegate: Dynamic input tensors are not supported in node #%d: ", |
| 50 | nodeIndex); |
| 51 | return kTfLiteError; |
| 52 | } |
| 53 | const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]]; |
| 54 | if(!IsValid(&tfLiteOutputTensor)) |
| 55 | { |
| 56 | TF_LITE_MAYBE_KERNEL_LOG( |
| 57 | tfLiteContext, |
| 58 | "TfLiteArmnnDelegate: Invalid output tensor in operator #%d node #%d: ", |
| 59 | operatorCode, nodeIndex); |
| 60 | return kTfLiteError; |
| 61 | } |
| 62 | if (IsDynamicTensor(tfLiteOutputTensor)) |
| 63 | { |
| 64 | TF_LITE_MAYBE_KERNEL_LOG( |
| 65 | tfLiteContext, |
| 66 | "TfLiteArmnnDelegate: Dynamic output tensors are not supported in node #%d: ", |
| 67 | nodeIndex); |
| 68 | return kTfLiteError; |
| 69 | } |
| 70 | |
| 71 | const TfLiteTensor& tfLiteWeightsTensor = tfLiteTensors[tfLiteNode->inputs->data[1]]; |
| 72 | if(!IsValid(&tfLiteWeightsTensor)) |
| 73 | { |
| 74 | TF_LITE_MAYBE_KERNEL_LOG( |
| 75 | tfLiteContext, |
| 76 | "TfLiteArmnnDelegate: Invalid weights tensor in operator #%d node #%d: ", |
| 77 | operatorCode, nodeIndex); |
| 78 | return kTfLiteError; |
| 79 | } |
| 80 | if (IsDynamicTensor(tfLiteWeightsTensor)) |
| 81 | { |
| 82 | TF_LITE_MAYBE_KERNEL_LOG( |
| 83 | tfLiteContext, |
| 84 | "TfLiteArmnnDelegate: Dynamic weight tensors are not supported in node #%d: ", |
| 85 | nodeIndex); |
| 86 | return kTfLiteError; |
| 87 | } |
| 88 | |
| 89 | const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor); |
| 90 | const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor); |
| 91 | |
| 92 | armnn::TensorInfo weightsTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteWeightsTensor); |
| 93 | // Fully Connected Layer accepts two dimensional weights input |
| 94 | int32_t weightsDimension = static_cast<int32_t>(weightsTensorInfo.GetNumDimensions()); |
| 95 | if (weightsDimension != 2) |
| 96 | { |
| 97 | TF_LITE_MAYBE_KERNEL_LOG( |
| 98 | tfLiteContext, |
| 99 | "TfLiteArmnnDelegate: Dimension #$d for Fully Connected weights is not supported by Armnn" |
| 100 | " in operator #%d node #%d: ", weightsDimension, operatorCode, nodeIndex); |
| 101 | return kTfLiteError; |
| 102 | } |
| 103 | |
| 104 | armnn::TensorInfo biasTensorInfo; |
| 105 | if (biasEnabled) |
| 106 | { |
| 107 | const TfLiteTensor& tfLiteBiasTensor = tfLiteTensors[tfLiteNode->inputs->data[2]]; |
| 108 | if(!IsValid(&tfLiteBiasTensor)) |
| 109 | { |
| 110 | TF_LITE_MAYBE_KERNEL_LOG( |
| 111 | tfLiteContext, |
| 112 | "TfLiteArmnnDelegate: Invalid bias tensor in operator #%d node #%d: ", |
| 113 | operatorCode, nodeIndex); |
| 114 | return kTfLiteError; |
| 115 | } |
| 116 | if (IsDynamicTensor(tfLiteBiasTensor)) |
| 117 | { |
| 118 | TF_LITE_MAYBE_KERNEL_LOG( |
| 119 | tfLiteContext, |
| 120 | "TfLiteArmnnDelegate: Dynamic bias tensors are not supported in node #%d: ", |
| 121 | nodeIndex); |
| 122 | return kTfLiteError; |
| 123 | } |
| 124 | biasTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteBiasTensor); |
| 125 | } |
| 126 | else |
| 127 | { |
| 128 | biasTensorInfo = armnn::TensorInfo(armnn::TensorShape({1}), GetDataType(tfLiteInputTensor)); |
| 129 | } |
| 130 | |
| 131 | armnn::FullyConnectedDescriptor descriptor; |
| 132 | descriptor.m_TransposeWeightMatrix = true; |
| 133 | descriptor.m_BiasEnabled = biasEnabled; |
| 134 | |
| 135 | bool isSupported = false; |
| 136 | auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported) |
| 137 | { |
| 138 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 139 | tfLiteContext, |
| 140 | IsFullyConnectedSupported, |
| 141 | delegateData.m_Backends, |
| 142 | isSupported, |
| 143 | inputTensorInfo, |
| 144 | outputTensorInfo, |
| 145 | weightsTensorInfo, |
| 146 | biasTensorInfo, |
| 147 | descriptor); |
| 148 | }; |
| 149 | |
| 150 | if (!delegateData.m_Network) |
| 151 | { |
| 152 | validateFunc(outputTensorInfo, isSupported); |
| 153 | return isSupported ? kTfLiteOk : kTfLiteError; |
| 154 | } |
| 155 | |
| 156 | auto weightsTensor = CreateConstTensor(&tfLiteWeightsTensor, |
| 157 | weightsTensorInfo, |
| 158 | armnn::Optional<armnn::PermutationVector&>()); |
| 159 | |
| 160 | armnn::IConnectableLayer* layer = nullptr; |
| 161 | if (biasEnabled) |
| 162 | { |
| 163 | const TfLiteTensor& tfLiteBiasTensor = tfLiteTensors[tfLiteNode->inputs->data[2]]; |
| 164 | auto biasTensor = CreateConstTensor(&tfLiteBiasTensor, |
| 165 | biasTensorInfo, |
| 166 | armnn::Optional<armnn::PermutationVector&>()); |
| 167 | layer = delegateData.m_Network->AddFullyConnectedLayer(descriptor, |
Sadik Armagan | 4189cc5 | 2020-11-11 18:01:48 +0000 | [diff] [blame] | 168 | weightsTensor, |
| 169 | armnn::Optional<armnn::ConstTensor>(biasTensor)); |
Sadik Armagan | 6e36a64 | 2020-11-10 21:18:41 +0000 | [diff] [blame] | 170 | } |
| 171 | else |
| 172 | { |
| 173 | layer = delegateData.m_Network->AddFullyConnectedLayer(descriptor, |
Sadik Armagan | 4189cc5 | 2020-11-11 18:01:48 +0000 | [diff] [blame] | 174 | weightsTensor, |
Sadik Armagan | 6e36a64 | 2020-11-10 21:18:41 +0000 | [diff] [blame] | 175 | armnn::EmptyOptional()); |
| 176 | } |
| 177 | ARMNN_ASSERT(layer != nullptr); |
| 178 | |
| 179 | armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0); |
| 180 | outputSlot.SetTensorInfo(outputTensorInfo); |
| 181 | |
| 182 | armnn::IConnectableLayer* reshapeLayer = nullptr; |
| 183 | if (inputTensorInfo.GetNumDimensions() > 2) |
| 184 | { |
| 185 | // Add reshape to flatten to 2D [batch_size, input_size] |
| 186 | std::vector<unsigned int> reshapedDimensions(2); |
| 187 | reshapedDimensions[1] = weightsTensorInfo.GetShape()[1]; |
| 188 | reshapedDimensions[0] = inputTensorInfo.GetNumElements() / reshapedDimensions[1]; |
| 189 | |
| 190 | if (inputTensorInfo.GetNumElements() % reshapedDimensions[1] != 0) |
| 191 | { |
| 192 | TF_LITE_MAYBE_KERNEL_LOG( |
| 193 | tfLiteContext, |
| 194 | "TfLiteArmnnDelegate: Failed to deduce input tensor shape from filter size #%d #%d node #%d: ", |
| 195 | reshapedDimensions[1], operatorCode, nodeIndex); |
| 196 | return kTfLiteError; |
| 197 | } |
| 198 | |
| 199 | armnn::TensorInfo reshapedTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor); |
| 200 | reshapedTensorInfo.SetShape(armnn::TensorShape{ 2, reshapedDimensions.data() }); |
| 201 | |
| 202 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 203 | reshapeDescriptor.m_TargetShape = reshapedTensorInfo.GetShape(); |
| 204 | reshapeLayer = delegateData.m_Network->AddReshapeLayer(reshapeDescriptor); |
| 205 | ARMNN_ASSERT(reshapeLayer != nullptr); |
| 206 | |
| 207 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapedTensorInfo); |
| 208 | |
| 209 | // Connect |
| 210 | delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[0]]->Connect(reshapeLayer->GetInputSlot(0)); |
| 211 | reshapeLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 212 | armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0); |
| 213 | delegateData.m_OutputSlotForNode[tfLiteNode->outputs->data[0]] = &outputSlot; |
| 214 | } |
| 215 | |
| 216 | if (reshapeLayer == nullptr) |
| 217 | { |
| 218 | Connect(layer, tfLiteNode, delegateData); |
| 219 | } |
| 220 | |
| 221 | auto* tfLiteNodeParameters = reinterpret_cast<TfLiteAddParams*>(tfLiteNode->builtin_data); |
| 222 | if (!tfLiteNodeParameters) |
| 223 | { |
| 224 | // No Activation |
| 225 | return kTfLiteOk; |
| 226 | } |
| 227 | |
| 228 | // Check Activation |
| 229 | TfLiteFusedActivation activationType = tfLiteNodeParameters->activation; |
| 230 | return FusedActivation(tfLiteContext, tfLiteNode, activationType, layer, 0, delegateData); |
Sadik Armagan | 62483be | 2020-10-23 17:14:43 +0100 | [diff] [blame] | 231 | } |
| 232 | |
Sadik Armagan | 6e36a64 | 2020-11-10 21:18:41 +0000 | [diff] [blame] | 233 | } // namespace armnnDelegate |