xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f16-f32-vcvt.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1#!/bin/sh
2# Copyright 2021 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 ###################################
8tools/xngen src/f16-f32-vcvt/neon-int16.c.in -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-neon-int16-x8.c &
9tools/xngen src/f16-f32-vcvt/neon-int16.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-neon-int16-x16.c &
10tools/xngen src/f16-f32-vcvt/neon-int16.c.in -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-neon-int16-x24.c &
11tools/xngen src/f16-f32-vcvt/neon-int16.c.in -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-neon-int16-x32.c &
12
13tools/xngen src/f16-f32-vcvt/neon-int32.c.in -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-neon-int32-x8.c &
14tools/xngen src/f16-f32-vcvt/neon-int32.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-neon-int32-x16.c &
15tools/xngen src/f16-f32-vcvt/neon-int32.c.in -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-neon-int32-x24.c &
16tools/xngen src/f16-f32-vcvt/neon-int32.c.in -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-neon-int32-x32.c &
17
18tools/xngen src/f16-f32-vcvt/neonfp16.c.in -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-neonfp16-x8.c &
19tools/xngen src/f16-f32-vcvt/neonfp16.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-neonfp16-x16.c &
20
21################################# x86 128-bit #################################
22tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-sse2-int16-x8.c &
23tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-sse2-int16-x16.c &
24tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-sse2-int16-x24.c &
25tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-sse2-int16-x32.c &
26
27tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-sse2-int32-x8.c &
28tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-sse2-int32-x16.c &
29tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-sse2-int32-x24.c &
30tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=2 -D AVX=0 -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-sse2-int32-x32.c &
31
32tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-sse41-int16-x8.c &
33tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-sse41-int16-x16.c &
34tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-sse41-int16-x24.c &
35tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-sse41-int16-x32.c &
36
37tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-sse41-int32-x8.c &
38tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-sse41-int32-x16.c &
39tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-sse41-int32-x24.c &
40tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=0 -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-sse41-int32-x32.c &
41
42tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-avx-int16-x8.c &
43tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-avx-int16-x16.c &
44tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-avx-int16-x24.c &
45tools/xngen src/f16-f32-vcvt/sse-int16.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-avx-int16-x32.c &
46
47tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-avx-int32-x8.c &
48tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-avx-int32-x16.c &
49tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-avx-int32-x24.c &
50tools/xngen src/f16-f32-vcvt/sse-int32.c.in -D SSE=4 -D AVX=1 -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-avx-int32-x32.c &
51
52################################# x86 256-bit #################################
53tools/xngen src/f16-f32-vcvt/f16c.c.in -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-f16c-x8.c &
54tools/xngen src/f16-f32-vcvt/f16c.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-f16c-x16.c &
55
56################################# x86 512-bit #################################
57tools/xngen src/f16-f32-vcvt/avx512skx.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-avx512skx-x16.c &
58tools/xngen src/f16-f32-vcvt/avx512skx.c.in -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-avx512skx-x32.c &
59
60################################## WAsm SIMD ##################################
61tools/xngen src/f16-f32-vcvt/wasmsimd-int16.c.in -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int16-x8.c &
62tools/xngen src/f16-f32-vcvt/wasmsimd-int16.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int16-x16.c &
63tools/xngen src/f16-f32-vcvt/wasmsimd-int16.c.in -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int16-x24.c &
64tools/xngen src/f16-f32-vcvt/wasmsimd-int16.c.in -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int16-x32.c &
65
66tools/xngen src/f16-f32-vcvt/wasmsimd-int32.c.in -D BATCH_TILE=8  -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int32-x8.c &
67tools/xngen src/f16-f32-vcvt/wasmsimd-int32.c.in -D BATCH_TILE=16 -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int32-x16.c &
68tools/xngen src/f16-f32-vcvt/wasmsimd-int32.c.in -D BATCH_TILE=24 -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int32-x24.c &
69tools/xngen src/f16-f32-vcvt/wasmsimd-int32.c.in -D BATCH_TILE=32 -o src/f16-f32-vcvt/gen/vcvt-wasmsimd-int32-x32.c &
70
71#################################### Scalar ###################################
72tools/xngen src/f16-f32-vcvt/scalar.c.in -D BATCH_TILE=1 -o src/f16-f32-vcvt/gen/vcvt-scalar-x1.c &
73tools/xngen src/f16-f32-vcvt/scalar.c.in -D BATCH_TILE=2 -o src/f16-f32-vcvt/gen/vcvt-scalar-x2.c &
74tools/xngen src/f16-f32-vcvt/scalar.c.in -D BATCH_TILE=3 -o src/f16-f32-vcvt/gen/vcvt-scalar-x3.c &
75tools/xngen src/f16-f32-vcvt/scalar.c.in -D BATCH_TILE=4 -o src/f16-f32-vcvt/gen/vcvt-scalar-x4.c &
76
77################################## Unit tests #################################
78tools/generate-vcvt-test.py --spec test/f16-f32-vcvt.yaml --output test/f16-f32-vcvt.cc &
79
80wait
81