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
|
/*
* Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef __NNFW_CKER_REFERENCE_BATCH_MATMUL_H__
#define __NNFW_CKER_REFERENCE_BATCH_MATMUL_H__
#include "cker/Types.h"
#include "cker/Shape.h"
namespace nnfw
{
namespace cker
{
namespace reference
{
inline void BatchMatMul(const Shape &lhs_shape, const float *lhs_data, const Shape &rhs_shape,
const float *rhs_data, const Shape &, float *output_data)
{
const Shape extended_lhs_shape = Shape::ExtendedShape(5, lhs_shape);
const Shape extended_rhs_shape = Shape::ExtendedShape(5, rhs_shape);
// Determine which dimension is the broadcast dimension.
auto broadcast_dim = [](int lhs_dim, int rhs_dim) {
if (lhs_dim == rhs_dim)
return lhs_dim;
if (lhs_dim == 1)
return rhs_dim;
assert(rhs_dim == 1);
return lhs_dim;
};
// Compute the "extent" for iterating on this dimension.
// If we are broadcasting, then don't advance (i.e return 0).
auto extent = [](const Shape &shape, int x) {
if (shape.Dims(x) == 1)
{
return 0;
}
int prod = 1;
for (int i = x + 1; i < shape.DimensionsCount(); ++i)
{
prod *= shape.Dims(i);
}
return prod;
};
const int batch_dim0 = broadcast_dim(extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
const int batch_dim1 = broadcast_dim(extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
const int batch_dim2 = broadcast_dim(extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));
const int lhs_ext0 = extent(extended_lhs_shape, 0);
const int lhs_ext1 = extent(extended_lhs_shape, 1);
const int lhs_ext2 = extent(extended_lhs_shape, 2);
const int rhs_ext0 = extent(extended_rhs_shape, 0);
const int rhs_ext1 = extent(extended_rhs_shape, 1);
const int rhs_ext2 = extent(extended_rhs_shape, 2);
// Set params for each matrix multiply.
const int lhs_rows = extended_lhs_shape.Dims(3);
const int rhs_cols = extended_rhs_shape.Dims(4);
const int accum_depth = extended_lhs_shape.Dims(4);
for (int b0 = 0; b0 < batch_dim0; ++b0)
{
const float *lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
const float *rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
for (int b1 = 0; b1 < batch_dim1; ++b1)
{
const float *lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
const float *rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
for (int b2 = 0; b2 < batch_dim2; ++b2)
{
const float *lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
const float *rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
float *out_ptr =
output_data +
((b0 * batch_dim1 * batch_dim2) + b1 * batch_dim2 + b2) * lhs_rows * rhs_cols;
for (int j = 0; j < rhs_cols; ++j)
{
for (int i = 0; i < lhs_rows; ++i)
{
float total = 0.f;
for (int k = 0; k < accum_depth; ++k)
{
total += lhs_ptr2[accum_depth * i + k] * rhs_ptr2[j * accum_depth + k];
}
int idx = lhs_rows * j + i;
out_ptr[idx] = total;
}
}
}
}
}
}
} // namespace reference
} // namespace cker
} // namespace nnfw
#endif // __NNFW_CKER_REFERENCE_BATCH_MATMUL_H__
|