blob: 60cbf27bde8bacbd2c7464dd5c864406d54f32b0 (
plain)
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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "Graph.hpp"
namespace armnn
{
inline std::vector<ConvertFp16ToFp32Layer*> InsertConvertFp16ToFp32LayersBefore(Graph& graph, Layer& layer)
{
std::vector<ConvertFp16ToFp32Layer*> convertLayers;
convertLayers.reserve(layer.GetNumInputSlots());
for (auto&& inputSlot = layer.BeginInputSlots(); inputSlot != layer.EndInputSlots(); ++inputSlot)
{
// Insert FP16 to FP32 converter layer before the layer
const std::string name =
std::string("convert_fp16_to_fp32-" + std::to_string(inputSlot->GetSlotIndex()) + "-") + layer.GetName();
ConvertFp16ToFp32Layer* convertLayer =
graph.InsertNewLayer<ConvertFp16ToFp32Layer>(*inputSlot, name.c_str());
// Sets output tensor info for the convert layer
TensorInfo convertInfo = convertLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
convertInfo.SetDataType(DataType::Float32);
convertLayer->GetOutputSlot().SetTensorInfo(convertInfo);
convertLayers.emplace_back(convertLayer);
}
// Sets the output tensor info for the unsupported layer
auto UpdateTensorInfo = [](auto& outputSlot)
{
// Copy original tensor info and change data type to FP32
TensorInfo newTensorInfo = outputSlot.GetTensorInfo();
newTensorInfo.SetDataType(DataType::Float32);
outputSlot.SetTensorInfo(newTensorInfo);
};
std::for_each(layer.BeginOutputSlots(), layer.EndOutputSlots(), UpdateTensorInfo);
return convertLayers;
}
inline std::vector<ConvertFp32ToFp16Layer*> InsertConvertFp32ToFp16LayersAfter(Graph& graph, Layer& layer)
{
std::vector<ConvertFp32ToFp16Layer*> convertLayers;
convertLayers.reserve(layer.GetNumOutputSlots());
int index = 0;
// Change outputs to DataType::Float16
for (auto&& outputSlot = layer.BeginOutputSlots(); outputSlot != layer.EndOutputSlots(); ++outputSlot)
{
BOOST_ASSERT(outputSlot->GetTensorInfo().GetDataType() == DataType::Float32);
// Insert FP32 to FP16 converter layer after the layer
const std::string name =
std::string("convert_fp32_to_fp16-" + std::to_string(index++) + "-") + layer.GetName();
ConvertFp32ToFp16Layer* convertLayer =
graph.InsertNewLayer<ConvertFp32ToFp16Layer>(*outputSlot, name.c_str());
// Sets output tensor info for the convert layer.
TensorInfo convertInfo = convertLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
convertInfo.SetDataType(DataType::Float16);
convertLayer->GetOutputSlot().SetTensorInfo(convertInfo);
convertLayers.emplace_back(convertLayer);
}
return convertLayers;
}
} //namespace armnn
|
