xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f16-raddstoreexpminusmax.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker#!/bin/sh
2*4bdc9457SAndroid Build Coastguard Worker# Copyright 2022 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# ARM NEON+FP16ARITH
8*4bdc9457SAndroid Build Coastguard Workertools/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 &
9*4bdc9457SAndroid Build Coastguard Workertools/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 &
10*4bdc9457SAndroid Build Coastguard Workertools/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 &
11*4bdc9457SAndroid Build Coastguard Workertools/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 &
12*4bdc9457SAndroid Build Coastguard Workertools/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 &
13*4bdc9457SAndroid Build Coastguard Workertools/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 &
14*4bdc9457SAndroid Build Coastguard Workertools/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 &
15*4bdc9457SAndroid Build Coastguard Workertools/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 &
16*4bdc9457SAndroid Build Coastguard Workertools/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 &
17*4bdc9457SAndroid Build Coastguard Workertools/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 &
18*4bdc9457SAndroid Build Coastguard Workertools/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 &
19*4bdc9457SAndroid Build Coastguard Workertools/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 &
20*4bdc9457SAndroid Build Coastguard Workertools/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 &
21*4bdc9457SAndroid Build Coastguard Workertools/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 &
22*4bdc9457SAndroid Build Coastguard Workertools/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 &
23*4bdc9457SAndroid Build Coastguard Workertools/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 &
24*4bdc9457SAndroid Build Coastguard Workertools/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 &
25*4bdc9457SAndroid Build Coastguard Workertools/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 &
26*4bdc9457SAndroid Build Coastguard Workertools/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 &
27*4bdc9457SAndroid Build Coastguard Workertools/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 &
28*4bdc9457SAndroid Build Coastguard Workertools/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*4bdc9457SAndroid Build Coastguard Worker
30*4bdc9457SAndroid Build Coastguard Worker# x86 AVX2
31*4bdc9457SAndroid Build Coastguard Workertools/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 &
32*4bdc9457SAndroid Build Coastguard Workertools/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 &
33*4bdc9457SAndroid Build Coastguard Workertools/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 &
34*4bdc9457SAndroid Build Coastguard Workertools/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 &
35*4bdc9457SAndroid Build Coastguard Workertools/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 &
36*4bdc9457SAndroid Build Coastguard Workertools/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 &
37*4bdc9457SAndroid Build Coastguard Workertools/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 &
38*4bdc9457SAndroid Build Coastguard Workertools/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 &
39*4bdc9457SAndroid Build Coastguard Workertools/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 &
40*4bdc9457SAndroid Build Coastguard Workertools/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 &
41*4bdc9457SAndroid Build Coastguard Workertools/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 &
42*4bdc9457SAndroid Build Coastguard Workertools/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 &
43*4bdc9457SAndroid Build Coastguard Workertools/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 &
44*4bdc9457SAndroid Build Coastguard Workertools/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 &
45*4bdc9457SAndroid Build Coastguard Workertools/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 &
46*4bdc9457SAndroid Build Coastguard Workertools/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 &
47*4bdc9457SAndroid Build Coastguard Workertools/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 &
48*4bdc9457SAndroid Build Coastguard Workertools/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 &
49*4bdc9457SAndroid Build Coastguard Workertools/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 &
50*4bdc9457SAndroid Build Coastguard Workertools/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 &
51*4bdc9457SAndroid Build Coastguard Workertools/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*4bdc9457SAndroid Build Coastguard Worker
53*4bdc9457SAndroid Build Coastguard Worker################################## Unit tests #################################
54*4bdc9457SAndroid Build Coastguard Workertools/generate-raddstoreexpminusmax-test.py --spec test/f16-raddstoreexpminusmax.yaml --output test/f16-raddstoreexpminusmax.cc &
55*4bdc9457SAndroid Build Coastguard Worker
56*4bdc9457SAndroid Build Coastguard Workerwait
57