xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f16-dwconv.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
8*4bdc9457SAndroid Build Coastguard Worker################################### ARM NEON ##################################
9*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=3 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x3-minmax-neonfp16arith.c &
10*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=3 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x3-minmax-neonfp16arith-acc2.c &
11*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=3 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x3-minmax-neonfp16arith.c &
12*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=3 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x3-minmax-neonfp16arith-acc2.c &
13*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=3 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x3-minmax-neonfp16arith.c &
14*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=3 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x3-minmax-neonfp16arith-acc2.c &
15*4bdc9457SAndroid Build Coastguard Worker
16*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c &
17*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c &
18*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c &
19*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c &
20*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x4-minmax-neonfp16arith.c &
21*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x4-minmax-neonfp16arith-acc2.c &
22*4bdc9457SAndroid Build Coastguard Worker
23*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c &
24*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c &
25*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c &
26*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c &
27*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x9-minmax-neonfp16arith.c &
28*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x9-minmax-neonfp16arith-acc2.c &
29*4bdc9457SAndroid Build Coastguard Worker
30*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c &
31*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c &
32*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c &
33*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c &
34*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x25-minmax-neonfp16arith.c &
35*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x25-minmax-neonfp16arith-acc2.c &
36*4bdc9457SAndroid Build Coastguard Worker
37*4bdc9457SAndroid Build Coastguard Worker################################### x86 FMA3 ##################################
38*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=3 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x3-minmax-fma3.c &
39*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=3 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x3-minmax-fma3-acc2.c &
40*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=3 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x3-minmax-fma3.c &
41*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=3 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x3-minmax-fma3-acc2.c &
42*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=3 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x3-minmax-fma3.c &
43*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=3 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x3-minmax-fma3-acc2.c &
44*4bdc9457SAndroid Build Coastguard Worker
45*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x4-minmax-fma3.c &
46*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x4-minmax-fma3-acc2.c &
47*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x4-minmax-fma3.c &
48*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x4-minmax-fma3-acc2.c &
49*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x4-minmax-fma3.c &
50*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x4-minmax-fma3-acc2.c &
51*4bdc9457SAndroid Build Coastguard Worker
52*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x9-minmax-fma3.c &
53*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x9-minmax-fma3-acc2.c &
54*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x9-minmax-fma3.c &
55*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x9-minmax-fma3-acc2.c &
56*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x9-minmax-fma3.c &
57*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x9-minmax-fma3-acc2.c &
58*4bdc9457SAndroid Build Coastguard Worker
59*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x25-minmax-fma3.c &
60*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x25-minmax-fma3-acc2.c &
61*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x25-minmax-fma3.c &
62*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x25-minmax-fma3-acc2.c &
63*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x25-minmax-fma3.c &
64*4bdc9457SAndroid Build Coastguard Workertools/xngen src/f16-dwconv/up-fma3.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x25-minmax-fma3-acc2.c &
65*4bdc9457SAndroid Build Coastguard Worker
66*4bdc9457SAndroid Build Coastguard Worker################################## Unit tests #################################
67*4bdc9457SAndroid Build Coastguard Workertools/generate-dwconv-test.py --spec test/f16-dwconv-minmax.yaml --output test/f16-dwconv-minmax.cc &
68*4bdc9457SAndroid Build Coastguard Worker
69*4bdc9457SAndroid Build Coastguard Workerwait
70