1 //
2 // Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ConcatOperator.hpp"
7
ConvertConcatToTosaOperator(const Layer * layer,const std::vector<const TensorInfo * > & inputs,const std::vector<const TensorInfo * > & outputs,const OriginsDescriptor * concatDescriptor)8 TosaSerializationBasicBlock* ConvertConcatToTosaOperator(const Layer* layer,
9 const std::vector<const TensorInfo*>& inputs,
10 const std::vector<const TensorInfo*>& outputs,
11 const OriginsDescriptor* concatDescriptor)
12 {
13 auto numInputs = inputs.size();
14 std::vector<std::string> inputNames;
15 inputNames.reserve(numInputs);
16 std::string outputName = std::string("output0_");
17 std::string blockName = std::string("Op_CONCAT_block_") + GetUniqueTosaMappingID();
18
19 // Set input names for validation purposes only.
20 if (layer == nullptr)
21 {
22 for (uint32_t i = 0; i < numInputs; ++i)
23 {
24 inputNames.push_back("input"+ std::to_string(i) +"_");
25 }
26 }
27 // If a layer is present then the block will be used for execution, so input and output names need to be determined
28 // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter.
29 else
30 {
31 // Get the layers connected to the input slots and determine unique tensor names.
32 for (uint32_t i = 0; i < numInputs; ++i)
33 {
34 Layer& connectedLayer = layer->GetInputSlot(i).GetConnectedOutputSlot()->GetOwningLayer();
35
36 std::string inputName = GenerateUniqueName(connectedLayer, i);
37 inputNames.push_back(inputName);
38 }
39
40 // Determine unique output tensor name.
41 outputName = GenerateUniqueOutputName(*layer, 0);
42 }
43
44 auto axis = static_cast<int32_t>(concatDescriptor->GetConcatAxis());
45 TosaAxisAttribute attribute(axis);
46
47 TosaSerializationOperator* op = new TosaSerializationOperator(Op_CONCAT,
48 Attribute_AxisAttribute,
49 &attribute,
50 inputNames,
51 {outputName});
52
53 std::vector<TosaSerializationTensor*> tensors;
54 tensors.reserve(numInputs);
55
56 for (uint32_t i = 0; i < numInputs; ++i)
57 {
58 // Only add input tensors for validation or when the connected layer is an input layer.
59 // As there can't be duplicate tensors and intermediate or constant tensors are created separately.
60 if(inputNames[i].find("input") != std::string::npos)
61 {
62 std::vector<int32_t> inputShape = GetTosaTensorShape(inputs[i]->GetShape());
63 DType inputDType = ArmNNToDType(inputs[i]->GetDataType());
64 tensors.push_back(new TosaSerializationTensor(inputNames[i], inputShape, inputDType, {}));
65 }
66 }
67
68 std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputs[0]->GetShape());
69 DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType());
70
71 TosaSerializationTensor* outputTensor0 = new TosaSerializationTensor(outputName, outputShape0, outputDType0, {});
72 tensors.push_back(outputTensor0);
73
74 // operatorInputNames/operatorOutputNames ends up being the same as
75 // blockInputNames/blockOutputNames for one-to-one ArmNN to TOSA mappings
76 return new TosaSerializationBasicBlock(blockName, // name
77 {op}, // operators
78 tensors, // tensors
79 inputNames, // inputs
80 {outputName}); // outputs
81 }