1 /*
2 * Copyright (c) 2022 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25 #include <cstddef>
26
27 namespace arm_conv {
28 namespace winograd {
29 namespace weight_transform {
30
cpp_fp32_1x4_1x5(unsigned int n_channels,const float * inptr,size_t,size_t ld_weight_col,float * outptr,size_t matrix_stride)31 void cpp_fp32_1x4_1x5(
32 unsigned int n_channels,
33 const float *inptr,
34 size_t, // ld_weight_row
35 size_t ld_weight_col,
36 float *outptr,
37 size_t matrix_stride
38 )
39 {
40 constexpr auto kernel_cols = 5u, inner_tile_cols = 8u;
41
42 // For each output channel
43 for (; n_channels; n_channels--)
44 {
45 // Matrices used and computed in this kernel
46 float w[kernel_cols], V[inner_tile_cols];
47
48 // Read weights
49 for (auto j = 0u; j < kernel_cols; j++)
50 {
51 w[j] = *(inptr + j * ld_weight_col);
52 }
53
54 // Compute V = w WT
55 V[0] = (w[0]*-1) / 36;
56 V[1] = (w[1]*-1 + w[3]*-1 + w[0]*1 + w[2]*1 + w[4]*1) / 48;
57 V[2] = (w[0]*1 + w[1]*1 + w[2]*1 + w[3]*1 + w[4]*1) / 48;
58 V[3] = (w[0]*-1 + w[4]*-16 + w[2]*-4 + w[1]*2 + w[3]*8) / 120;
59 V[4] = (w[0]*-1 + w[4]*-16 + w[3]*-8 + w[2]*-4 + w[1]*-2) / 120;
60 V[5] = (w[3]*-27 + w[1]*-3 + w[2]*9 + w[4]*81 + w[0]*1) / 720;
61 V[6] = (w[1]*3 + w[2]*9 + w[3]*27 + w[4]*81 + w[0]*1) / 720;
62 V[7] = (w[4]*1) / 1;
63
64 // Store the transformed weights
65 for (auto j = 0u; j < inner_tile_cols; j++)
66 {
67 *(outptr + j*matrix_stride) = V[j];
68 }
69
70 inptr++;
71 outptr++;
72 }
73 }
74
75 } // namespace weight_transform
76 } // namespace winograd
77 } // namespace arm_conv
78