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
ConvertTransposeConv2dToTosaOperator(const Layer * layer,const std::vector<const TensorInfo * > & inputs,const std::vector<const TensorInfo * > & outputs,const TransposeConvolution2dDescriptor * descriptor)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 {
25 // Get the layers connected to the input slots and determine unique tensor names.
26 Layer& connectedInputLayer = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer();
27 input0Name = GenerateUniqueName(connectedInputLayer, 0);
28
29 // Determine unique output tensor name.
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.
99 unsigned int index = (descriptor->m_DataLayout == DataLayout::NHWC) ? 3 : 1;
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
146 TosaTransposeConvAttribute attribute(pad, stride, outputShape, 0, 0, ArmNNToDType(inputs[0]->GetDataType()));
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
158 operators, // operators
159 tensors, // tensors
160 {input0Name, input1Name, input2Name}, // inputs
161 {outputName}); // outputs
162 }