xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f32-vsqrt.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker#!/bin/sh
2*4bdc9457SAndroid Build Coastguard Worker# Copyright 2020 Google LLC
3*4bdc9457SAndroid Build Coastguard Worker#
4*4bdc9457SAndroid Build Coastguard Worker# This source code is licensed under the BSD-style license found in the
5*4bdc9457SAndroid Build Coastguard Worker# LICENSE file in the root directory of this source tree.
6*4bdc9457SAndroid Build Coastguard Worker
7*4bdc9457SAndroid Build Coastguard Worker#################################### Scalar ###################################
8*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=1 -o src/f32-vsqrt/gen/scalar-sqrt-x1.c &
9*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=2 -o src/f32-vsqrt/gen/scalar-sqrt-x2.c &
10*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/scalar-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/scalar-sqrt-x4.c &
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker################################## WAsm SIMD ##################################
13*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/wasmsimd-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c &
14*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/wasmsimd-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c &
15*4bdc9457SAndroid Build Coastguard Worker
16*4bdc9457SAndroid Build Coastguard Worker################################### ARM NEON ##################################
17*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neon-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/neon-sqrt-x4.c &
18*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neon-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/neon-sqrt-x8.c &
19*4bdc9457SAndroid Build Coastguard Worker
20*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=4  -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c &
21*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c &
22*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=12 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c &
23*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c &
24*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=20 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c &
25*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c &
26*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=28 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c &
27*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c &
28*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=36 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c &
29*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr2fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c &
30*4bdc9457SAndroid Build Coastguard Worker
31*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=4  -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c &
32*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c &
33*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=12 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c &
34*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c &
35*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=20 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c &
36*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c &
37*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=28 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c &
38*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c &
39*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=36 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c &
40*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c &
41*4bdc9457SAndroid Build Coastguard Worker
42*4bdc9457SAndroid Build Coastguard Worker################################# x86 128-bit #################################
43*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/sse-sqrt.c.in -D BATCH_TILE=4 -o src/f32-vsqrt/gen/sse-sqrt-x4.c &
44*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/sse-sqrt.c.in -D BATCH_TILE=8 -o src/f32-vsqrt/gen/sse-sqrt-x8.c &
45*4bdc9457SAndroid Build Coastguard Worker
46*4bdc9457SAndroid Build Coastguard Worker################################# x86 256-bit #################################
47*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx-sqrt.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/avx-sqrt-x8.c &
48*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx-sqrt.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/avx-sqrt-x16.c &
49*4bdc9457SAndroid Build Coastguard Worker
50*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=8  -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c &
51*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=16 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c &
52*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=24 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c &
53*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=32 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c &
54*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=40 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c &
55*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=48 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c &
56*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=56 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c &
57*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/fma3-nr1fma1adj.c.in -D BATCH_TILE=64 -o src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c &
58*4bdc9457SAndroid Build Coastguard Worker
59*4bdc9457SAndroid Build Coastguard Worker################################# x86 512-bit #################################
60*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=16  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c &
61*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=32  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c &
62*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=48  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c &
63*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=64  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c &
64*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=80  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c &
65*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=96  -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c &
66*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=112 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c &
67*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f32-vsqrt/avx512f-nr1fma1adj.c.in -D BATCH_TILE=128 -o src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c &
68*4bdc9457SAndroid Build Coastguard Worker
69*4bdc9457SAndroid Build Coastguard Worker################################## Unit tests #################################
70*4bdc9457SAndroid Build Coastguard Workertools/generate-vunary-test.py --spec test/f32-vsqrt.yaml --output test/f32-vsqrt.cc &
71*4bdc9457SAndroid Build Coastguard Worker
72*4bdc9457SAndroid Build Coastguard Workerwait
73