Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 1 | #!/bin/sh |
| 2 | # Copyright 2020 Google LLC |
| 3 | # |
| 4 | # This source code is licensed under the BSD-style license found in the |
| 5 | # LICENSE file in the root directory of this source tree. |
| 6 | |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame^] | 7 | ################################## ARM NEON ################################### |
| 8 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=1 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c |
| 9 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=2 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c |
| 10 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=3 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c |
| 11 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=4 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c |
| 12 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=5 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c |
| 13 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=6 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c |
| 14 | |
| 15 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=1 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c |
| 16 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=1 -D ACCUMULATORS=3 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c |
| 17 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=1 -D ACCUMULATORS=4 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c |
| 18 | tools/xngen src/f32-dwconv2d-chw/3x3p1-neonfma.c.in -D ROW_TILE=2 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c |
| 19 | |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 20 | ################################### x86 SSE ################################### |
| 21 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c |
| 22 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=2 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c |
| 23 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=3 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c |
| 24 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=4 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c |
| 25 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=5 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c |
| 26 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=6 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c |
| 27 | |
| 28 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c |
| 29 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=3 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c |
| 30 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=4 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c |
| 31 | tools/xngen src/f32-dwconv2d-chw/3x3p1-sse.c.in -D ROW_TILE=2 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c |
| 32 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 33 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=1 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c |
| 34 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=2 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c |
| 35 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=3 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c |
| 36 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=4 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c |
| 37 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=5 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c |
| 38 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=6 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c |
| 39 | |
| 40 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=1 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c |
| 41 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=1 -D ACCUMULATORS=3 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c |
| 42 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=1 -D ACCUMULATORS=4 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c |
| 43 | tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=2 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c |
| 44 | |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 45 | ################################## Unit tests ################################# |
| 46 | tools/generate-dwconv2d-chw-test.py --spec test/f32-dwconv2d-chw.yaml --output test/f32-dwconv2d-chw.cc |