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
|
/*
* Copyright (c) 2017 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "GEMMLowp.h"
#include "arm_compute/core/Types.h"
namespace arm_compute
{
namespace test
{
namespace validation
{
namespace reference
{
template <typename T>
SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset)
{
TensorShape shape(b.shape()[0], a.shape()[1]);
SimpleTensor<int32_t> c(shape, DataType::S32);
const int K = a.shape().x();
const int b_width = b.shape().x();
const int rows = c.shape().y(); //M
const int cols = c.shape().x(); //N
std::vector<int32_t> acc;
acc.resize(cols);
for(int i = 0; i < rows; ++i)
{
for(int j = 0; j < cols; ++j)
{
acc[j] = 0;
}
for(int k = 0; k < K; ++k)
{
const int32_t tmp_a = a_offset + static_cast<int32_t>(a[k + i * K]);
for(int j = 0; j < b_width; ++j)
{
const int32_t tmp_b = b_offset + static_cast<int32_t>(b[j + k * b_width]);
const int32_t mult_as_int = tmp_a * tmp_b;
acc[j] += mult_as_int;
}
}
for(int j = 0; j < cols; ++j)
{
c[j + i * cols] = acc[j];
}
}
return c;
}
// used to validate assembly kernels which don't know anything about offsets
SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b)
{
return gemmlowp_matrix_multiply_core(a, b, 0, 0);
}
template <typename T>
SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
{
SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
for(int i = 0; i < in.num_elements(); ++i)
{
const int32_t result = ((in[i] + result_offset) * result_mult_int) >> result_shift;
dst[i] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
}
return dst;
}
template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, int32_t a_offset, int32_t b_offset);
template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, int32_t result_mult_int, int32_t result_shift);
} // namespace reference
} // namespace validation
} // namespace test
} // namespace arm_compute
|
