Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1 | #!/bin/sh |
| 2 | # Copyright 2019 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 | |
| 7 | ################################### ARM NEON ################################## |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 8 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c |
| 9 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c |
| 10 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c |
| 11 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c |
| 12 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c |
| 13 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 14 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 15 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c |
| 16 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c |
| 17 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c |
| 18 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c |
| 19 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c |
| 20 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 21 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 22 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c |
| 23 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c |
| 24 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c |
| 25 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c |
| 26 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c |
| 27 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 28 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 29 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c |
| 30 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c |
| 31 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c |
| 32 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c |
| 33 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c |
| 34 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 35 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 36 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=4 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c |
| 37 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c |
| 38 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=12 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c |
| 39 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c |
| 40 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=20 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c |
| 41 | tools/xngen src/f32-sigmoid/neon-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 42 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 43 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c |
| 44 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c |
| 45 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c |
| 46 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c |
| 47 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c |
| 48 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 49 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 50 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c |
| 51 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c |
| 52 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c |
| 53 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c |
| 54 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c |
| 55 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 56 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 57 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c |
| 58 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c |
| 59 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c |
| 60 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c |
| 61 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c |
| 62 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 63 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 64 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c |
| 65 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c |
| 66 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c |
| 67 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c |
| 68 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c |
| 69 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 70 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 71 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=4 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c |
| 72 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=8 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c |
| 73 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=12 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c |
| 74 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=16 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c |
| 75 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=20 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c |
| 76 | tools/xngen src/f32-sigmoid/neon-lut64-p2.c.in -D BATCH_TILE=24 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 77 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 78 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c |
| 79 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c |
| 80 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c |
| 81 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c |
| 82 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c |
| 83 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 84 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 85 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c |
| 86 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c |
| 87 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c |
| 88 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c |
| 89 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c |
| 90 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 91 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 92 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c |
| 93 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c |
| 94 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c |
| 95 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c |
| 96 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c |
| 97 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr1recps1fma -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 98 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 99 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=4 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c |
| 100 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c |
| 101 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=12 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c |
| 102 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c |
| 103 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=20 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c |
| 104 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D FMA=1 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 105 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 106 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=4 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c |
| 107 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=8 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c |
| 108 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=12 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c |
| 109 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=16 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c |
| 110 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=20 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c |
| 111 | tools/xngen src/f32-sigmoid/neon-lut2048-p1.c.in -D BATCH_TILE=24 -D RR_STEPS=2 -D FMA=0 -D DIV_ALGO=nr2recps -o src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 112 | |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 113 | tools/xngen src/f32-sigmoid/neon-frac-p9-p10-nr1recps.c.in -D BATCH_TILE=16 -o src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c |
| 114 | |
| 115 | ################################### x86 SSE ################################### |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 116 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=4 -D BLEND=0 -o src/f32-sigmoid/gen/sse2-p5-div-x4.c |
| 117 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=8 -D BLEND=0 -o src/f32-sigmoid/gen/sse2-p5-div-x8.c |
| 118 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=12 -D BLEND=0 -o src/f32-sigmoid/gen/sse2-p5-div-x12.c |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 119 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=16 -D BLEND=0 -o src/f32-sigmoid/gen/sse2-p5-div-x16.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 120 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=20 -D BLEND=0 -o src/f32-sigmoid/gen/sse2-p5-div-x20.c |
| 121 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=24 -D BLEND=0 -o src/f32-sigmoid/gen/sse2-p5-div-x24.c |
| 122 | |
| 123 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=4 -D BLEND=1 -o src/f32-sigmoid/gen/sse41-p5-div-x4.c |
| 124 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=8 -D BLEND=1 -o src/f32-sigmoid/gen/sse41-p5-div-x8.c |
| 125 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=12 -D BLEND=1 -o src/f32-sigmoid/gen/sse41-p5-div-x12.c |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 126 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=16 -D BLEND=1 -o src/f32-sigmoid/gen/sse41-p5-div-x16.c |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 127 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=20 -D BLEND=1 -o src/f32-sigmoid/gen/sse41-p5-div-x20.c |
| 128 | tools/xngen src/f32-sigmoid/sse-p5-div.c.in -D BATCH_TILE=24 -D BLEND=1 -o src/f32-sigmoid/gen/sse41-p5-div-x24.c |
| 129 | |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 130 | ################################### x86 AVX ################################### |
| 131 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c |
| 132 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c |
| 133 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c |
| 134 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=32 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c |
| 135 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=40 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c |
| 136 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=48 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c |
| 137 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=56 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c |
| 138 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=64 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c |
| 139 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=72 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c |
| 140 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=80 -D RR_STEPS=1 -D DIV_ALGO=div -o src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c |
| 141 | |
| 142 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c |
| 143 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c |
| 144 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c |
| 145 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=32 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c |
| 146 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=40 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c |
| 147 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=48 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c |
| 148 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=56 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c |
| 149 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=64 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c |
| 150 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=72 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c |
| 151 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=80 -D RR_STEPS=1 -D DIV_ALGO=nr1fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c |
| 152 | |
| 153 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=8 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c |
| 154 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=16 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c |
| 155 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=24 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c |
| 156 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=32 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c |
| 157 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=40 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c |
| 158 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=48 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c |
| 159 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=56 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c |
| 160 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=64 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c |
| 161 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=72 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c |
| 162 | tools/xngen src/f32-sigmoid/avx2-p5.c.in -D BATCH_TILE=80 -D RR_STEPS=1 -D DIV_ALGO=nr2fma -o src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c |
| 163 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 164 | #################################### PSIMD #################################### |
| 165 | tools/xngen src/f32-sigmoid/psimd-p5-div.c.in -D BATCH_TILE=4 -D BLEND=0 -o src/f32-sigmoid/gen/psimd-p5-div-x4.c |
| 166 | tools/xngen src/f32-sigmoid/psimd-p5-div.c.in -D BATCH_TILE=8 -D BLEND=0 -o src/f32-sigmoid/gen/psimd-p5-div-x8.c |
| 167 | tools/xngen src/f32-sigmoid/psimd-p5-div.c.in -D BATCH_TILE=12 -D BLEND=0 -o src/f32-sigmoid/gen/psimd-p5-div-x12.c |
| 168 | tools/xngen src/f32-sigmoid/psimd-p5-div.c.in -D BATCH_TILE=16 -D BLEND=0 -o src/f32-sigmoid/gen/psimd-p5-div-x16.c |
| 169 | tools/xngen src/f32-sigmoid/psimd-p5-div.c.in -D BATCH_TILE=20 -D BLEND=0 -o src/f32-sigmoid/gen/psimd-p5-div-x20.c |
| 170 | tools/xngen src/f32-sigmoid/psimd-p5-div.c.in -D BATCH_TILE=24 -D BLEND=0 -o src/f32-sigmoid/gen/psimd-p5-div-x24.c |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 171 | |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 172 | ################################### Scalar #################################### |
| 173 | tools/xngen src/f32-sigmoid/scalar-lut2048-p1-div.c.in -D BATCH_TILE=1 -o src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c |
| 174 | tools/xngen src/f32-sigmoid/scalar-lut2048-p1-div.c.in -D BATCH_TILE=2 -o src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c |
| 175 | tools/xngen src/f32-sigmoid/scalar-lut2048-p1-div.c.in -D BATCH_TILE=4 -o src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c |
| 176 | |
| 177 | tools/xngen src/f32-sigmoid/scalar-lut64-p2-div.c.in -D BATCH_TILE=1 -o src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c |
| 178 | tools/xngen src/f32-sigmoid/scalar-lut64-p2-div.c.in -D BATCH_TILE=2 -o src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c |
| 179 | tools/xngen src/f32-sigmoid/scalar-lut64-p2-div.c.in -D BATCH_TILE=4 -o src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c |
| 180 | |
| 181 | tools/xngen src/f32-sigmoid/scalar-p5-div.c.in -D BATCH_TILE=1 -o src/f32-sigmoid/gen/scalar-p5-div-x1.c |
| 182 | tools/xngen src/f32-sigmoid/scalar-p5-div.c.in -D BATCH_TILE=2 -o src/f32-sigmoid/gen/scalar-p5-div-x2.c |
| 183 | tools/xngen src/f32-sigmoid/scalar-p5-div.c.in -D BATCH_TILE=4 -o src/f32-sigmoid/gen/scalar-p5-div-x4.c |
| 184 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 185 | ################################## Unit tests ################################# |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 186 | tools/generate-vunary-test.py --spec test/f32-sigmoid.yaml --output test/f32-sigmoid.cc |