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
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "WorkloadUtils.hpp"
namespace armnn
{
armnn::ConstTensor PermuteTensor(const ConstCpuTensorHandle* tensor,
const PermutationVector& permutationVector, void* permuteBuffer)
{
BOOST_ASSERT_MSG(tensor, "Invalid input tensor");
BOOST_ASSERT_MSG(permuteBuffer, "Invalid permute buffer");
TensorInfo tensorInfo = tensor->GetTensorInfo();
if (permutationVector.GetSize() > 0)
{
tensorInfo = armnnUtils::Permuted(tensorInfo, permutationVector);
armnnUtils::Permute(tensorInfo.GetShape(), permutationVector,
tensor->GetConstTensor<void>(), permuteBuffer,
GetDataTypeSize(tensorInfo.GetDataType()));
}
else
{
::memcpy(permuteBuffer, tensor->GetConstTensor<void>(), tensorInfo.GetNumBytes());
}
return ConstTensor(tensorInfo, permuteBuffer);
}
void ReshapeWeightsForAcl(TensorInfo& weightInfo, DataLayout dataLayout)
{
// Reshape the weights in-place
const TensorShape& weightShape = weightInfo.GetShape();
switch (dataLayout)
{
case DataLayout::NHWC:
// The data layout is NHWC, reshape from [ H, W, I, M ] to [ 1, H, W, I * M ]
weightInfo.SetShape({ 1,
weightShape[0],
weightShape[1],
weightShape[2] * weightShape[3] });
weightInfo.SetShape({ 1,
weightShape[0] * weightShape[1],
weightShape[2],
weightShape[3] });
break;
case DataLayout::NCHW:
default:
// The data layout is NCHW, reshape from [ M, I, H, W ] to [ 1, I * M, H, W, ]
weightInfo.SetShape({ 1, weightShape[0] * weightShape[1], weightShape[2], weightShape[3] });
break;
}
}
template <typename DataType>
ConstTensor ReorderWeightChannelsForAcl(const ConstTensor& weightHandle, DataLayout dataLayout, void* permuteBuffer)
{
DataType* weight = static_cast<DataType*>(permuteBuffer);
const TensorShape& weightShape = weightHandle.GetShape();
unsigned int multiplier;
unsigned int height;
unsigned int width;
unsigned int inputChannels;
switch (dataLayout)
{
case DataLayout::NHWC: //It actually is [ H, W, I, M ]
height = weightShape[0];
width = weightShape[1];
inputChannels = weightShape[2];
multiplier = weightShape[3];
break;
case DataLayout::NCHW: //It actually is [ M, I, H, W ]
default:
height = weightShape[2];
width = weightShape[3];
inputChannels = weightShape[1];
multiplier = weightShape[0];
break;
}
DataType weightAclOrder[height*width*inputChannels*multiplier];
unsigned int destinationWeightsChannel;
unsigned int totalChannels = inputChannels * multiplier;
unsigned int channelSize = height * width;
for (unsigned int originWeightsChannel = 0; originWeightsChannel < totalChannels; originWeightsChannel++)
{
if (originWeightsChannel % inputChannels == 0)
{
destinationWeightsChannel = originWeightsChannel / inputChannels;
}
else
{
destinationWeightsChannel = (originWeightsChannel - 1) / inputChannels + multiplier;
}
for (unsigned int i = 0; i < channelSize; i++)
{
weightAclOrder[i + destinationWeightsChannel * channelSize] =
weight[i + originWeightsChannel * channelSize];
}
}
::memcpy(permuteBuffer, weightAclOrder, weightHandle.GetInfo().GetNumBytes());
return ConstTensor(weightHandle.GetInfo(), permuteBuffer);
}
TensorInfo ConvertWeightTensorInfoFromArmnnToAcl(const TensorInfo& weightInfo, DataLayout dataLayout)
{
// Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
// [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
// 1. Permute the weights if necessary
// If the data layout is NCHW no permutation is necessary, as a reshape to [ 1, I * M, H, W ] can be better done
// starting from the current shape of [ M, I, H, W ]
TensorInfo weightPermutedInfo(weightInfo);
if (dataLayout == DataLayout::NHWC)
{
// The data layout is NHWC, then permute the weights from [ M, I, H, W ] to [ H, W, I, M ]
PermutationVector permutationVector{ 3, 2, 0, 1 };
weightPermutedInfo = armnnUtils::Permuted(weightInfo, permutationVector);
}
// 2. Reshape the weights
ReshapeWeightsForAcl(weightPermutedInfo, dataLayout);
// 3. Return the permuted weight info
return weightPermutedInfo;
}
armnn::ConstTensor ConvertWeightTensorFromArmnnToAcl(const ConstCpuTensorHandle* weightTensor,
DataLayout dataLayout,
void* permuteBuffer)
{
BOOST_ASSERT_MSG(weightTensor, "Invalid input tensor");
BOOST_ASSERT_MSG(permuteBuffer, "Invalid permute buffer");
auto multiplier = weightTensor->GetTensorInfo().GetShape()[0];
auto inputChannels = weightTensor->GetTensorInfo().GetShape()[1];
// Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
// [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
// 1. Permute the weights if necessary
// If the data layout is NCHW no permutation is necessary, as a reshape to [ 1, I * M, H, W ] can be better done
// starting from the current shape of [ M, I, H, W ]
// If no permutation is necessary, leave the permutation vector empty
PermutationVector permutationVector{};
if (dataLayout == DataLayout::NHWC)
{
// The data layout is NHWC, then permute the weights from [ M, I, H, W ] to [ H, W, I, M ]
permutationVector = { 3, 2, 0, 1 };
}
ConstTensor weightPermuted = PermuteTensor(weightTensor, permutationVector, permuteBuffer);
// Shuffle the weights data to obtain the channel order needed used by Acl
if (multiplier > 1 and inputChannels > 1 and dataLayout == DataLayout::NCHW)
{
switch (weightPermuted.GetDataType())
{
case DataType::Float32:
weightPermuted = ReorderWeightChannelsForAcl<float>(weightPermuted, dataLayout, permuteBuffer);
break;
case DataType::Float16:
weightPermuted =
ReorderWeightChannelsForAcl<half_float::half>(weightPermuted, dataLayout, permuteBuffer);
break;
case DataType::QuantisedAsymm8:
weightPermuted = ReorderWeightChannelsForAcl<uint8_t>(weightPermuted, dataLayout, permuteBuffer);
break;
default:
break;
}
}
// 2. Reshape the weights
ReshapeWeightsForAcl(weightPermuted.GetInfo(), dataLayout);
// 3. Return both the tensor and the allocated storage to ensure that the data stays alive
return weightPermuted;
}
} // namespace armnn
|
