1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
|
//
// Copyright © 2022-2024 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Conv2dOperator.hpp"
TosaSerializationBasicBlock* ConvertConv2dToTosaOperator(const Layer* layer,
const std::vector<const TensorInfo*>& inputs,
const std::vector<const TensorInfo*>& outputs,
const Convolution2dDescriptor* conv2dDescriptor)
{
std::vector<std::string> inputNames;
std::string outputName = std::string("output0_");
std::string blockName = std::string("Op_CONV2D_block_") + GetUniqueTosaMappingID();
// Set input names for validation purposes only.
if(layer == nullptr)
{
inputNames.emplace_back("input0_");
inputNames.emplace_back("input1_");
if(conv2dDescriptor->m_BiasEnabled)
{
inputNames.emplace_back("input2_");
}
}
// If a layer is present then the block will be used for execution, so input and output names need to be
// determined using the previous and following layers so the graph is connected correctly.
// For validation this doesn't matter.
else
{
// Get the layer connected to the input slot and determine unique tensor names.
for (uint32_t i = 0; i < inputs.size(); ++i)
{
Layer& connectedLayer = layer->GetInputSlot(i).GetConnectedOutputSlot()->GetOwningLayer();
std::string inputName = GenerateUniqueName(connectedLayer, i);
inputNames.push_back(inputName);
}
// Determine unique output tensor name.
outputName = GenerateUniqueOutputName(*layer, 0);
}
std::vector<TosaSerializationTensor*> tensors;
std::vector<TosaSerializationOperator*> operators;
// Setup input Tensor
// Only add tensor if connected layer is an input layer.
// As intermediate or constant tensors will be created separately.
// There also can't be duplicate tensors.
if(inputNames[0].find("input0_") != std::string::npos)
{
std::vector<int32_t> inputShape0 = GetTosaTensorShape(inputs[0]->GetShape());
DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType());
tensors.push_back(new TosaSerializationTensor(inputNames[0], inputShape0, inputDType0, {}));
}
// Only add input tensors if weights and bias are not constant or if running validation.
// Constant tensors will be created in the ConvertConstantToTosaOperator function.
if(!inputs[1]->IsConstant() || layer == nullptr)
{
std::vector<int32_t> inputShape1 = GetTosaTensorShape(inputs[1]->GetShape());
DType inputDType1 = ArmNNToDType(inputs[1]->GetDataType());
tensors.push_back(new TosaSerializationTensor(inputNames[1], inputShape1, inputDType1, {}));
}
if(conv2dDescriptor->m_BiasEnabled)
{
if(!inputs[2]->IsConstant() || layer == nullptr)
{
std::vector<int32_t> inputShape2 = GetTosaTensorShape(inputs[2]->GetShape());
DType inputDType2 = ArmNNToDType(inputs[2]->GetDataType());
tensors.push_back(new TosaSerializationTensor(inputNames[2], inputShape2, inputDType2, {}));
}
}
else
{
// If bias is disabled, create a constant bias of 0 as three inputs are required.
std::string constantName = std::string("constant_") + GetUniqueTosaMappingID();
operators.push_back(new TosaSerializationOperator(Op_CONST, Attribute_NONE, nullptr, {}, {constantName}));
// The size of the bias must match the channels dimension, so get the correct index.
unsigned int index = (conv2dDescriptor->m_DataLayout == DataLayout::NHWC) ? 3 : 1;
std::vector<uint8_t> uint8Data;
std::vector<float> data(outputs[0]->GetShape()[index], 0.0f);
TosaSerializationHandler::ConvertF32toU8(data, uint8Data);
tensors.push_back(new TosaSerializationTensor(constantName,
{static_cast<int32_t>(outputs[0]->GetShape()[index])},
DType_FP32,
uint8Data));
inputNames.emplace_back(constantName);
}
// Setup Output Tensor
std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputs[0]->GetShape());
DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType());
tensors.push_back(new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}));
// Set up CONV2D operator
std::vector<int> pad = {static_cast<int>(conv2dDescriptor->m_PadTop),
static_cast<int>(conv2dDescriptor->m_PadBottom),
static_cast<int>(conv2dDescriptor->m_PadLeft),
static_cast<int>(conv2dDescriptor->m_PadRight)};
std::vector<int> stride = {static_cast<int>(conv2dDescriptor->m_StrideY),
static_cast<int>(conv2dDescriptor->m_StrideX)};
std::vector<int> dilation = {static_cast<int>(conv2dDescriptor->m_DilationY),
static_cast<int>(conv2dDescriptor->m_DilationX)};
TosaConvAttribute attribute(pad, stride, dilation, 0, 0, false); // input_zp, weight_zp, local_bound
auto* op = new TosaSerializationOperator(Op_CONV2D,
Attribute_ConvAttribute,
&attribute,
inputNames,
{outputName});
operators.push_back(op);
// operatorInputNames/operatorOutputNames ends up being the same as
// blockInputNames/blockOutputNames for one-to-one ArmNN to TOSA mappings
return new TosaSerializationBasicBlock(blockName, // name
mainName, // region name
operators, // operators
tensors, // tensors
inputNames, // inputs
{outputName}); // outputs
}
|
