xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f16-raddstoreexpminusmax.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1#!/bin/sh
2# Copyright 2022 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+FP16ARITH
8tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=32 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x32.c &
9tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=32 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x32-acc2.c &
10tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=32 -D ACCUMULATORS=4 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x32-acc4.c &
11tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=40 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x40.c &
12tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=40 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x40-acc2.c &
13tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=40 -D ACCUMULATORS=5 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x40-acc5.c &
14tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=48 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x48.c &
15tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=48 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x48-acc2.c &
16tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=48 -D ACCUMULATORS=3 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x48-acc3.c &
17tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=64 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x64.c &
18tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=64 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x64-acc2.c &
19tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=64 -D ACCUMULATORS=4 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x64-acc4.c &
20tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=72 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x72.c &
21tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=72 -D ACCUMULATORS=3 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x72-acc3.c &
22tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=80 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x80.c &
23tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=80 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x80-acc2.c &
24tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=80 -D ACCUMULATORS=5 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x80-acc5.c &
25tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x96.c &
26tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x96-acc2.c &
27tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=3 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x96-acc3.c &
28tools/xngen src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=6 -o src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x96-acc6.c &
29
30# x86 AVX2
31tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=32 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x32.c &
32tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=32 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x32-acc2.c &
33tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=32 -D ACCUMULATORS=4 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x32-acc4.c &
34tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=40 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x40.c &
35tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=40 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x40-acc2.c &
36tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=40 -D ACCUMULATORS=5 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x40-acc5.c &
37tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=48 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x48.c &
38tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=48 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x48-acc2.c &
39tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=48 -D ACCUMULATORS=3 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x48-acc3.c &
40tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=64 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x64.c &
41tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=64 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x64-acc2.c &
42tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=64 -D ACCUMULATORS=4 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x64-acc4.c &
43tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=72 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x72.c &
44tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=72 -D ACCUMULATORS=3 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x72-acc3.c &
45tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=80 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x80.c &
46tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=80 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x80-acc2.c &
47tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=80 -D ACCUMULATORS=5 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x80-acc5.c &
48tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=1 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x96.c &
49tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=2 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x96-acc2.c &
50tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=3 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x96-acc3.c &
51tools/xngen src/f16-raddstoreexpminusmax/avx2-rr1-p2.c.in -D BATCH_TILE=96 -D ACCUMULATORS=6 -o src/f16-raddstoreexpminusmax/gen/avx2-rr1-p2-x96-acc6.c &
52
53################################## Unit tests #################################
54tools/generate-raddstoreexpminusmax-test.py --spec test/f16-raddstoreexpminusmax.yaml --output test/f16-raddstoreexpminusmax.cc &
55
56wait
57