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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
|
//
// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "BatchMatMulImpl.hpp"
#include <armnn/backends/WorkloadData.hpp>
#include <armnn/Logging.hpp>
namespace armnn
{
void BatchMatMul::BatchMatMulImpl()
{
inputXData = inputXDecoder.DecodeTensor(inputXInfo.GetShape());
inputYData = inputYDecoder.DecodeTensor(inputYInfo.GetShape());
// At this point, we don't touch the input decoders - just the resultant vectors
// Pre-transpose and pre-adjoint if their vectors aren't empty
// and also DataLayouts which may change with permutations/adjoints
// Todo: Have you updated input validation and inferred output shapes to accommodate for these pre-permutes?
auto idx = std::vector<unsigned int>(outputInfo.GetNumDimensions(), 0);
RecurseBMM(idx, 0);
}
void BatchMatMul::RecurseBMM(std::vector<unsigned int>& curIdx, unsigned int curDim)
{
// We're working off of the indexes of the output tensor (the max possible shape)
if(!(curDim < outputInfo.GetNumDimensions()))
{
// We're at the leaf level of this call tree, so we operate here (each leaf is a data point)
auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(params,
inputXInfo.GetShape(),
inputYInfo.GetShape());
AdjustAxesToMulForUnequalRanks(axesToMul);
unsigned int inputXColDim = axesToMul.first.second;
unsigned int inputYRowDim = axesToMul.second.first;
unsigned int inputYRowSize = inputYInfo.GetShape()[inputYRowDim];
float sum = 0.0f;
// You could also use inputXColSize
for (unsigned int inputYRowIdx = 0; inputYRowIdx < inputYRowSize; inputYRowIdx++) {
auto xIdx = curIdx;
xIdx[inputXColDim] = inputYRowIdx;
auto yIdx = curIdx;
yIdx[inputYRowDim] = inputYRowIdx;
sum += (GetValueAt(DataSlot::InputX, xIdx)
* GetValueAt(DataSlot::InputY, yIdx));
}
SetValueAt(sum, DataSlot::Output, curIdx);
return;
}
for (unsigned int i = 0; i < outputInfo.GetShape()[curDim]; i++)
{
curIdx[curDim] = i;
RecurseBMM(curIdx, curDim+1);
}
}
void BatchMatMul::AdjustAxesToMulForUnequalRanks(
std::pair<std::pair<unsigned int, unsigned int>, std::pair<unsigned int, unsigned int>>& axesToMul)
{
int rankDiff = static_cast<int>(inputXInfo.GetNumDimensions()) -
static_cast<int>(inputYInfo.GetNumDimensions());
if(rankDiff == 0)
{
return;
}
else if(rankDiff < 0)
{
// Y is the larger one
axesToMul.first.first += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff));
axesToMul.first.second += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff));
}
else if(rankDiff > 0)
{
// X is the larger one
axesToMul.second.first += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff));
axesToMul.second.second += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff));
}
}
float BatchMatMul::GetValueAt(DataSlot type, std::vector<unsigned int> idx)
{
// This gets the data from the input vector that we have, Not the decoder
// But for the output, it is operating on the encoder itself
AdjustToSafeIdx(type, idx);
unsigned int flatIdx = CalcFlatIdx(type, idx);
float value = 0.0f;
switch(type)
{
case DataSlot::InputX:
value = inputXData[flatIdx];
break;
case DataSlot::InputY:
value = inputYData[flatIdx];
break;
case DataSlot::Output:
outputEncoder[flatIdx];
value = outputEncoder.Get();
break;
default:
break;
}
return value;
}
void BatchMatMul::SetValueAt(float value, DataSlot type, std::vector<unsigned int> idx)
{
AdjustToSafeIdx(type, idx);
unsigned int flatIdx = CalcFlatIdx(type, idx);
switch(type)
{
case DataSlot::InputX:
inputXData[flatIdx] = value;
break;
case DataSlot::InputY:
inputYData[flatIdx] = value;
break;
case DataSlot::Output:
outputEncoder[flatIdx];
outputEncoder.Set(value);
break;
default:
break;
}
}
void BatchMatMul::AdjustToSafeIdx(DataSlot type, std::vector<unsigned int>& idx)
{
for(unsigned int dim = 0; dim < idx.size(); dim++)
{
switch(type)
{
case DataSlot::InputX:
{
auto xRank = inputXInfo.GetNumDimensions();
auto xDiff = outputInfo.GetNumDimensions() - xRank;
if (dim < xDiff ||
idx[dim] > inputXInfo.GetShape()[dim-xDiff]-1)
{
idx[dim] = 0; // Broadcasting
}
break;
}
case DataSlot::InputY:
{
auto yRank = inputYInfo.GetNumDimensions();
auto yDiff = outputInfo.GetNumDimensions() - yRank;
if (dim < yDiff ||
idx[dim] > inputYInfo.GetShape()[dim-yDiff]-1)
{
idx[dim] = 0;
}
break;
}
case DataSlot::Output:
{
// Our indices are based off the output
break;
}
default:
break;
}
}
}
unsigned int BatchMatMul::CalcFlatIdx(DataSlot type, const std::vector<unsigned int>& idx)
{
unsigned int result = idx[idx.size()-1];
unsigned int dimMultiplier = 1;
unsigned int offset;
// -2 because final dim is already accounted for in the multiplier (last dim is just a multiplier of 1x)
for(unsigned int i = static_cast<unsigned int>(idx.size()-2); static_cast<int>(i) >= 0; i--)
{
switch(type)
{
case DataSlot::InputX:
offset = outputInfo.GetNumDimensions() - inputXInfo.GetNumDimensions();
dimMultiplier *= inputXInfo.GetShape()[i + 1 - offset];
break;
case DataSlot::InputY:
offset = outputInfo.GetNumDimensions() - inputYInfo.GetNumDimensions();
dimMultiplier *= inputYInfo.GetShape()[i + 1 - offset];
break;
case DataSlot::Output:
dimMultiplier *= outputInfo.GetShape()[i+1];
break;
default:
break;
}
result += (idx[i] * dimMultiplier);
}
return result;
}
template <typename T>
std::string BatchMatMul::StringifyVec(const std::vector<T>& vec)
{
std::string res = "{ ";
for(auto x : vec)
{
res += std::to_string(x);
res += " ";
}
res += "}";
return res;
}
} // namespace armnn
|
