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
|
/*
* 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.
*/
#pragma once
namespace winograd {
/* Transform a kernel into the Winograd domain.
*
* NOTE: It is assumed that the kernel is in the form [height x width x
* input_channels x output_channel].
*/
template <typename T>
struct winograd2x2_3x3_gemm_kernel_transform_impl{
static void execute(
const KernelShape &shape,
const T* const kernel,
T* const matrix_base,
const int matrix_stride,
const int matrix_row_stride
);
protected:
template <const int output_channel_tail>
static void transform_kernel(
const T* const kernel,
const int n_input_channels,
const int n_output_channels,
T* const matrix_base,
const int matrix_stride,
const int matrix_row_stride
);
};
}
/*****************************************************************************/
/* Transform a fp32 kernel into the Winograd domain.
*/
#include "kernel_2x2_3x3/a64_float.hpp" // AArch64 specialisations
namespace winograd
{
template <>
inline void winograd2x2_3x3_gemm_kernel_transform_impl<float>::execute(
const KernelShape &shape,
const float* const kernel,
float* const matrix_base,
const int matrix_stride,
const int matrix_row_stride
) {
// Delegate based on tail size
const int n_input_channels = shape.n_input_channels;
const int n_output_channels = shape.n_output_channels;
switch (n_output_channels % 4) {
case 0:
transform_kernel<0>(
kernel, n_input_channels, n_output_channels,
matrix_base, matrix_stride, matrix_row_stride
);
break;
case 1:
transform_kernel<1>(
kernel, n_input_channels, n_output_channels,
matrix_base, matrix_stride, matrix_row_stride
);
break;
case 2:
transform_kernel<2>(
kernel, n_input_channels, n_output_channels,
matrix_base, matrix_stride, matrix_row_stride
);
break;
case 3:
transform_kernel<3>(
kernel, n_input_channels, n_output_channels,
matrix_base, matrix_stride, matrix_row_stride
);
break;
default:
ARM_COMPUTE_ERROR("Cannot happen");
break;
}
}
template <>
template<const int output_channel_tail>
inline void winograd2x2_3x3_gemm_kernel_transform_impl<float>::transform_kernel(
const float* const kernel,
const int n_input_channels,
const int n_output_channels,
float* const matrix_base,
const int mstride,
const int matrix_row_stride
) {
// Use one input pointer for each row of the kernel, use two additional
// offsets to extract columns.
const int kernel_col_stride = n_input_channels * n_output_channels;
const int kernel_row_stride = 3 * kernel_col_stride;
const float *inptr0 = kernel;
const float *inptr1 = kernel + kernel_row_stride;
const float *inptr2 = kernel + kernel_row_stride*2;
// Use four output pointers, for output matrices 0, 4, 8 and 12. Use three
// offsets to extract further matrices.
float *outptr0 = matrix_base;
float *outptr4 = matrix_base + mstride * 4;
float *outptr8 = matrix_base + mstride * 8;
float *outptr12 = matrix_base + mstride * 12;
// For every input channel
for (int in_c = 0; in_c < n_input_channels; in_c++) {
// For every output channel
for (int c = 0; c < n_output_channels; c++) {
// Read in the kernel
float w11 = inptr0[0], w12 = inptr0[kernel_col_stride], w13 = inptr0[kernel_col_stride*2];
float w21 = inptr1[0], w22 = inptr1[kernel_col_stride], w23 = inptr1[kernel_col_stride*2];
float w31 = inptr2[0], w32 = inptr2[kernel_col_stride], w33 = inptr2[kernel_col_stride*2];
// Progress input pointers
inptr0++;
inptr1++;
inptr2++;
// Compute the kernel W w, note we need only compute the middle two rows
// (2 and 3) because the first and last rows are merely copies of values
// from the matrix w.
float Ww11 = w11, Ww12 = w12, Ww13 = w13;
float Ww21 = 0.5*(w11 + w21 + w31), Ww22 = 0.5*(w12 + w22 + w32), Ww23 = 0.5*(w13 + w23 + w33);
float Ww31 = 0.5*(w11 - w21 + w31), Ww32 = 0.5*(w12 - w22 + w32), Ww33 = 0.5*(w13 - w23 + w33);
float Ww41 = w31, Ww42 = w32, Ww43 = w33;
// Hence compute W w W.T; again note we need compute only the middle two
// columns since the first and last columns are copies of the first and
// last columns of the previous matrix.
float WwWT11 = Ww11, WwWT12 = 0.5*(Ww11 + Ww12 + Ww13), WwWT13 = 0.5*(Ww11 - Ww12 + Ww13), WwWT14 = Ww13;
float WwWT21 = Ww21, WwWT22 = 0.5*(Ww21 + Ww22 + Ww23), WwWT23 = 0.5*(Ww21 - Ww22 + Ww23), WwWT24 = Ww23;
float WwWT31 = Ww31, WwWT32 = 0.5*(Ww31 + Ww32 + Ww33), WwWT33 = 0.5*(Ww31 - Ww32 + Ww33), WwWT34 = Ww33;
float WwWT41 = Ww41, WwWT42 = 0.5*(Ww41 + Ww42 + Ww43), WwWT43 = 0.5*(Ww41 - Ww42 + Ww43), WwWT44 = Ww43;
// Store the computed weights
outptr0[0 * mstride] = WwWT11;
outptr0[1 * mstride] = WwWT12;
outptr0[2 * mstride] = WwWT13;
outptr0[3 * mstride] = WwWT14;
outptr4[0 * mstride] = WwWT21;
outptr4[1 * mstride] = WwWT22;
outptr4[2 * mstride] = WwWT23;
outptr4[3 * mstride] = WwWT24;
outptr8[0 * mstride] = WwWT31;
outptr8[1 * mstride] = WwWT32;
outptr8[2 * mstride] = WwWT33;
outptr8[3 * mstride] = WwWT34;
outptr12[0 * mstride] = WwWT41;
outptr12[1 * mstride] = WwWT42;
outptr12[2 * mstride] = WwWT43;
outptr12[3 * mstride] = WwWT44;
// Progress output pointers
outptr0++;
outptr4++;
outptr8++;
outptr12++;
}
// Progression to complete stride
outptr0 += matrix_row_stride - n_output_channels;
outptr4 += matrix_row_stride - n_output_channels;
outptr8 += matrix_row_stride - n_output_channels;
outptr12 += matrix_row_stride - n_output_channels;
}
}
}
|
