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
|
/*
* Copyright (c) 2017-2020 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.
*/
#pragma once
#include "utils.hpp"
namespace arm_gemm {
/*
* Generic transform.
*
* Assuming the untransposed case, this works by first reading <BlockBy>
* consecutive values from the first input row. This same number of values
* are then read from the next <IntBy-1> rows. Now return to the first
* input row and repeat.
*
* Need to cope with the work requested in either dimension not actually
* being a multiple of the block sizes.
*/
template <unsigned int tIntBy, unsigned int BlockBy, bool Transposed, size_t TOutSize, size_t TInSize, VLType vlt>
struct TransformImpl {
template <typename TOut, typename TIn>
static void Transform(TOut* out, const TIn* const in, const int stride,
const int y0, const int ymax, const int x0, const int xmax) {
// For SVE cases we multiply the interleave factor by the vector length.
const unsigned int IntBy = tIntBy * (vlt == VLType::SVE ? get_vector_length<TOut>() / BlockBy : 1);
const int n_whole_y_blocks = (ymax - y0) / IntBy;
const int y_remainders = (ymax - y0) % IntBy;
const int n_y_blocks = n_whole_y_blocks + (y_remainders ? 1 : 0);
const int n_whole_x_blocks = (xmax - x0) / BlockBy;
const int x_remainders = (xmax - x0) % BlockBy;
const int n_x_blocks = n_whole_x_blocks + (x_remainders ? 1 : 0);
// "Y" loop: advance down the rows of the source IntBy rows at a time.
// Set up fill_rows to show the number rows to copy from, and blank_rows
// for the number of blank rows to add.
for (int y_block=0 ; y_block < n_y_blocks; y_block++) {
int fill_rows = (y_block < n_whole_y_blocks) ? IntBy : y_remainders;
int blank_rows = IntBy - fill_rows;
int y_base = y0 + (y_block * IntBy);
// So now advance along this block of rows, BlockBy columns at a time.
for (int x_block=0 ; x_block < n_x_blocks; x_block++) {
int fill_cols = (x_block < n_whole_x_blocks) ? BlockBy : x_remainders;
int blank_cols = BlockBy - fill_cols;
int x_base = x0 + (x_block * BlockBy);
for (int row = 0; row < fill_rows; row++) {
for (int col = 0; col < fill_cols; col++) {
// In-range copy. If it's transposed, we reverse the sense of rows and columns here.
if (Transposed) {
*out++ = static_cast<TOut>(in[(x_base + col) * stride + y_base + row]);
} else {
*out++ = static_cast<TOut>(in[(y_base + row) * stride + x_base + col]);
}
}
// "col" tail - row is in range but column is out of range.
for (int col=0; col < blank_cols; col++) {
*out++ = static_cast<TOut>(0);
}
}
// "row" tail - row is out of range so fill with zeros always.
TOut zeroval = static_cast<TOut>(0);
int pads = blank_rows * (fill_cols + blank_cols);
for (int i=0; i<pads; i++) {
out[i] = zeroval;
}
out += pads;
}
}
}
template <typename T>
static inline void Transform(T* out, const T* const in, const int stride,
const int k0, const int kmax, const int x0, const int xmax) {
Transform<T, T>(out, in, stride, k0, kmax, x0, xmax);
}
};
/*****************************************************************************/
template <unsigned int IntBy, unsigned int BlockBy, bool Transposed, VLType vlt=VLType::None, typename TOut, typename TIn>
void Transform(
TOut* out, const TIn* const in, const int stride,
const int k0, const int kmax, const int x0, const int xmax
) {
// Redirect to a specialised implementation predicated on argument size.
TransformImpl<IntBy, BlockBy, Transposed, sizeof(TOut), sizeof(TIn), vlt>::Transform(
out, in, stride, k0, kmax, x0, xmax
);
}
/*****************************************************************************/
#include "transforms/list.hpp"
} // namespace arm_gemm
|
