Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -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 | |
| 7 | #################################### Scalar ################################### |
| 8 | tools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=1 -o src/f32-vsqrt/gen/scalar-sqrt-x1.c |
| 9 | tools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=2 -o src/f32-vsqrt/gen/scalar-sqrt-x2.c |
| 10 | tools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/scalar-sqrt-x4.c |
| 11 | |
| 12 | ################################## WAsm SIMD ################################## |
| 13 | tools/xngen src/f32-vsqrt/wasmsimd-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c |
| 14 | tools/xngen src/f32-vsqrt/wasmsimd-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c |
| 15 | |
| 16 | ################################### ARM NEON ################################## |
| 17 | tools/xngen src/f32-vsqrt/neon-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/neon-sqrt-x4.c |
| 18 | tools/xngen src/f32-vsqrt/neon-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/neon-sqrt-x8.c |
| 19 | |
| 20 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c |
| 21 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c |
| 22 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=12 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c |
| 23 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c |
| 24 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=20 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c |
| 25 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c |
| 26 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=28 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c |
| 27 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c |
| 28 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=36 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c |
| 29 | tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c |
| 30 | |
| 31 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c |
| 32 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c |
| 33 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=12 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c |
| 34 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c |
| 35 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=20 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c |
| 36 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c |
| 37 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=28 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c |
| 38 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c |
| 39 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=36 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c |
| 40 | tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c |
| 41 | |
| 42 | ################################# x86 128-bit ################################# |
| 43 | tools/xngen src/f32-vsqrt/sse-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/sse-sqrt-x4.c |
| 44 | tools/xngen src/f32-vsqrt/sse-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/sse-sqrt-x8.c |
| 45 | |
| 46 | ################################# x86 256-bit ################################# |
| 47 | tools/xngen src/f32-vsqrt/avx-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/avx-sqrt-x8.c |
| 48 | tools/xngen src/f32-vsqrt/avx-sqrt.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/avx-sqrt-x16.c |
| 49 | |
| 50 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c |
| 51 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c |
| 52 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c |
| 53 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c |
| 54 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c |
| 55 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=48 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c |
| 56 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=56 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c |
| 57 | tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=64 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c |
| 58 | |
| 59 | ################################# x86 512-bit ################################# |
| 60 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c |
| 61 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c |
| 62 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=48 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c |
| 63 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=64 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c |
| 64 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=80 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c |
| 65 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=96 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c |
| 66 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=112 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c |
| 67 | tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=128 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c |
| 68 | |
| 69 | ################################## Unit tests ################################# |
| 70 | tools/generate-vunary-test.py --spec test/f32-vsqrt.yaml --output test/f32-vsqrt.cc |