xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f32-vsqrt.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 ###################################
8tools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=1 -o src/f32-vsqrt/gen/scalar-sqrt-x1.c &
9tools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=2 -o src/f32-vsqrt/gen/scalar-sqrt-x2.c &
10tools/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 ##################################
13tools/xngen src/f32-vsqrt/wasmsimd-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c &
14tools/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 ##################################
17tools/xngen src/f32-vsqrt/neon-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/neon-sqrt-x4.c &
18tools/xngen src/f32-vsqrt/neon-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/neon-sqrt-x8.c &
19
20tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=4  -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c &
21tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c &
22tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=12 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c &
23tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c &
24tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=20 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c &
25tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c &
26tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=28 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c &
27tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c &
28tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=36 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c &
29tools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c &
30
31tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=4  -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c &
32tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c &
33tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=12 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c &
34tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c &
35tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=20 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c &
36tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c &
37tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=28 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c &
38tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c &
39tools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=36 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c &
40tools/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 #################################
43tools/xngen src/f32-vsqrt/sse-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/sse-sqrt-x4.c &
44tools/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 #################################
47tools/xngen src/f32-vsqrt/avx-sqrt.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/avx-sqrt-x8.c &
48tools/xngen src/f32-vsqrt/avx-sqrt.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/avx-sqrt-x16.c &
49
50tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c &
51tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c &
52tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c &
53tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c &
54tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c &
55tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=48 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c &
56tools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=56 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c &
57tools/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 #################################
60tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=16  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c &
61tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=32  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c &
62tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=48  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c &
63tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=64  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c &
64tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=80  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c &
65tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=96  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c &
66tools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=112 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c &
67tools/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 #################################
70tools/generate-vunary-test.py --spec test/f32-vsqrt.yaml --output test/f32-vsqrt.cc &
71
72wait
73