Matthew Sloyan | fc9d5e7 | 2022-12-08 13:38:23 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "TransposeConv2dOperator.hpp" |
| 7 | |
| 8 | #include "layers/TransposeConvolution2dLayer.hpp" |
| 9 | |
| 10 | TosaSerializationBasicBlock* ConvertTransposeConv2dToTosaOperator(const Layer* layer, |
| 11 | const std::vector<const TensorInfo*>& inputs, |
| 12 | const std::vector<const TensorInfo*>& outputs, |
| 13 | const TransposeConvolution2dDescriptor* descriptor) |
| 14 | { |
| 15 | std::string input0Name = std::string("input0_"); |
| 16 | std::string input1Name = std::string("constant_") + GetUniqueTosaMappingID(); |
| 17 | std::string input2Name = std::string("constant_") + GetUniqueTosaMappingID(); |
| 18 | std::string outputName = std::string("output0_"); |
| 19 | std::string blockName = std::string("Op_TRANSPOSE_CONV2D_block_") + GetUniqueTosaMappingID(); |
| 20 | |
| 21 | // If a layer is present then the block will be used for execution, so input and output names need to be determined |
| 22 | // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter. |
| 23 | if(layer != nullptr) |
| 24 | { |
Kevin May | 5b58e31 | 2022-12-15 10:15:21 +0000 | [diff] [blame] | 25 | // Get the layers connected to the input slots and determine unique tensor names. |
Matthew Sloyan | fc9d5e7 | 2022-12-08 13:38:23 +0000 | [diff] [blame] | 26 | Layer& connectedInputLayer = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer(); |
| 27 | input0Name = GenerateUniqueName(connectedInputLayer, 0); |
| 28 | |
Kevin May | 5b58e31 | 2022-12-15 10:15:21 +0000 | [diff] [blame] | 29 | // Determine unique output tensor name. |
Matthew Sloyan | fc9d5e7 | 2022-12-08 13:38:23 +0000 | [diff] [blame] | 30 | outputName = GenerateUniqueOutputName(*layer, 0); |
| 31 | } |
| 32 | |
| 33 | std::vector<TosaSerializationTensor*> tensors; |
| 34 | std::vector<TosaSerializationOperator*> operators; |
| 35 | |
| 36 | // Setup input tensor |
| 37 | // Only add tensor if connected layer is an input layer. |
| 38 | // As intermediate or constant tensors will be created separately. |
| 39 | // There also can't be duplicate tensors. |
| 40 | if(input0Name.find("input0_") != std::string::npos) |
| 41 | { |
| 42 | std::vector<int32_t> inputShape0 = GetTosaTensorShape(inputs[0]->GetShape()); |
| 43 | DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType()); |
| 44 | |
| 45 | tensors.push_back(new TosaSerializationTensor(input0Name, inputShape0, inputDType0, {})); |
| 46 | } |
| 47 | |
| 48 | // Setup weights tensor, constant data will get copied during SetConstantTensorData |
| 49 | operators.push_back(new TosaSerializationOperator(Op_CONST, Attribute_NONE, nullptr, {}, {input1Name})); |
| 50 | |
| 51 | // During validation the TensorInfo can be retrieved from the inputs. |
| 52 | // During execution, it is only available through the layer so use m_Weight. |
| 53 | if(layer == nullptr) |
| 54 | { |
| 55 | std::vector<int32_t> inputShape1 = GetTosaTensorShape(inputs[1]->GetShape()); |
| 56 | DType inputDType1 = ArmNNToDType(inputs[1]->GetDataType()); |
| 57 | |
| 58 | tensors.push_back(new TosaSerializationTensor(input1Name, inputShape1, inputDType1, {})); |
| 59 | } |
| 60 | else |
| 61 | { |
| 62 | auto transposeConv2dLayer = PolymorphicDowncast<const TransposeConvolution2dLayer*>(layer); |
| 63 | |
| 64 | std::vector<int32_t> inputShape1 = GetTosaTensorShape( |
| 65 | transposeConv2dLayer->m_Weight->GetTensorInfo().GetShape()); |
| 66 | DType inputDType1 = ArmNNToDType(transposeConv2dLayer->m_Weight->GetTensorInfo().GetDataType()); |
| 67 | |
| 68 | std::vector<uint8_t> uint8Data = ConvertConstantTensorDataToBuffer(transposeConv2dLayer->m_Weight); |
| 69 | tensors.push_back(new TosaSerializationTensor(input1Name, inputShape1, inputDType1, uint8Data)); |
| 70 | } |
| 71 | |
| 72 | // Setup bias operator and tensor, constant data will get copied during SetConstantTensorData |
| 73 | operators.push_back(new TosaSerializationOperator(Op_CONST, Attribute_NONE, nullptr, {}, {input2Name})); |
| 74 | |
| 75 | // During validation the TensorInfo can be retrieved from the inputs. |
| 76 | // During execution, it is only available through the layer so use m_Bias. |
| 77 | if(layer == nullptr && descriptor->m_BiasEnabled) |
| 78 | { |
| 79 | std::vector<int32_t> inputShape2 = GetTosaTensorShape(inputs[2]->GetShape()); |
| 80 | DType inputDType2 = ArmNNToDType(inputs[2]->GetDataType()); |
| 81 | |
| 82 | tensors.push_back(new TosaSerializationTensor(input2Name, inputShape2, inputDType2, {})); |
| 83 | } |
| 84 | else if(descriptor->m_BiasEnabled) |
| 85 | { |
| 86 | auto transposeConv2dLayer = PolymorphicDowncast<const TransposeConvolution2dLayer*>(layer); |
| 87 | |
| 88 | std::vector<int32_t> inputShape2 = GetTosaTensorShape( |
| 89 | transposeConv2dLayer->m_Bias->GetTensorInfo().GetShape()); |
| 90 | DType inputDType2 = ArmNNToDType(transposeConv2dLayer->m_Bias->GetTensorInfo().GetDataType()); |
| 91 | |
| 92 | std::vector<uint8_t> uint8Data = ConvertConstantTensorDataToBuffer(transposeConv2dLayer->m_Bias); |
| 93 | tensors.push_back(new TosaSerializationTensor(input2Name, inputShape2, inputDType2, uint8Data)); |
| 94 | } |
| 95 | else |
| 96 | { |
| 97 | // If bias is disabled, create a constant bias tensor of 0's as three inputs are required. |
| 98 | // The size of the bias must match the channels dimension, so get the correct index. |
Matthew Sloyan | da6bf9e | 2022-12-14 10:16:27 +0000 | [diff] [blame] | 99 | unsigned int index = (descriptor->m_DataLayout == DataLayout::NHWC) ? 3 : 1; |
Matthew Sloyan | fc9d5e7 | 2022-12-08 13:38:23 +0000 | [diff] [blame] | 100 | |
| 101 | std::vector<uint8_t> uint8Data; |
| 102 | std::vector<float> data(outputs[0]->GetShape()[index], 0.0f); |
| 103 | |
| 104 | TosaSerializationHandler::ConvertF32toU8(data, uint8Data); |
| 105 | |
| 106 | tensors.push_back(new TosaSerializationTensor(input2Name, |
| 107 | {static_cast<int32_t>(outputs[0]->GetShape()[index])}, |
| 108 | DType_FP32, |
| 109 | uint8Data)); |
| 110 | } |
| 111 | |
| 112 | // Setup Output Tensor |
| 113 | std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputs[0]->GetShape()); |
| 114 | DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType()); |
| 115 | |
| 116 | tensors.push_back(new TosaSerializationTensor(outputName, outputShape0, outputDType0, {})); |
| 117 | |
| 118 | // Set up TRANSPOSE_CONV2D operator |
| 119 | // The TOSA Reference Model pads the output shape, so it is added to output shape. |
| 120 | // In Arm NN we pad the input shape, so it is taken away. |
| 121 | // To offset this the negative padding value can be used. |
| 122 | std::vector<int> pad = {-static_cast<int>(descriptor->m_PadTop), |
| 123 | -static_cast<int>(descriptor->m_PadBottom), |
| 124 | -static_cast<int>(descriptor->m_PadLeft), |
| 125 | -static_cast<int>(descriptor->m_PadRight)}; |
| 126 | std::vector<int> stride = {static_cast<int>(descriptor->m_StrideY), |
| 127 | static_cast<int>(descriptor->m_StrideX)}; |
| 128 | |
| 129 | std::vector<int> outputShape; |
| 130 | // If available use shape in descriptor otherwise use output shape. |
| 131 | if (descriptor->m_OutputShape.size() == 4) |
| 132 | { |
| 133 | for (uint32_t i = 0; i < descriptor->m_OutputShape.size(); ++i) |
| 134 | { |
| 135 | outputShape.push_back(static_cast<int>(descriptor->m_OutputShape[i])); |
| 136 | } |
| 137 | } |
| 138 | else |
| 139 | { |
| 140 | for (uint32_t i = 0; i < outputs[0]->GetNumDimensions(); ++i) |
| 141 | { |
| 142 | outputShape.push_back(static_cast<int>(outputs[0]->GetShape()[i])); |
| 143 | } |
| 144 | } |
| 145 | |
Narumol Prangnawarat | ad323af | 2023-09-29 17:00:38 +0100 | [diff] [blame] | 146 | TosaTransposeConvAttribute attribute(pad, stride, outputShape, 0, 0); |
Matthew Sloyan | fc9d5e7 | 2022-12-08 13:38:23 +0000 | [diff] [blame] | 147 | |
| 148 | auto* op = new TosaSerializationOperator(Op_TRANSPOSE_CONV2D, |
| 149 | Attribute_TransposeConvAttribute, |
| 150 | &attribute, |
| 151 | {input0Name, input1Name, input2Name}, |
| 152 | {outputName}); |
| 153 | operators.push_back(op); |
| 154 | |
| 155 | // operatorInputNames/operatorOutputNames ends up being the same as |
| 156 | // blockInputNames/blockOutputNames for one-to-one ArmNN to TOSA mappings |
| 157 | return new TosaSerializationBasicBlock(blockName, // name |
Narumol Prangnawarat | ad323af | 2023-09-29 17:00:38 +0100 | [diff] [blame] | 158 | mainName, // region name |
Matthew Sloyan | fc9d5e7 | 2022-12-08 13:38:23 +0000 | [diff] [blame] | 159 | operators, // operators |
| 160 | tensors, // tensors |
| 161 | {input0Name, input1Name, input2Name}, // inputs |
| 162 | {outputName}); // outputs |
| 163 | } |