xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f32-vclamp.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1#!/bin/sh
2# Copyright 2020 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#################################### Scalar ###################################
8### Generic C micro-kernels
9tools/xngen src/f32-vclamp/scalar.c.in -D BATCH_TILE=1 -D WASM=0 -o src/f32-vclamp/gen/vclamp-scalar-x1.c &
10tools/xngen src/f32-vclamp/scalar.c.in -D BATCH_TILE=2 -D WASM=0 -o src/f32-vclamp/gen/vclamp-scalar-x2.c &
11tools/xngen src/f32-vclamp/scalar.c.in -D BATCH_TILE=4 -D WASM=0 -o src/f32-vclamp/gen/vclamp-scalar-x4.c &
12
13### WAsm-specific micro-kernels
14tools/xngen src/f32-vclamp/scalar.c.in -D BATCH_TILE=1 -D WASM=1 -o src/f32-vclamp/gen/vclamp-wasm-x1.c &
15tools/xngen src/f32-vclamp/scalar.c.in -D BATCH_TILE=2 -D WASM=1 -o src/f32-vclamp/gen/vclamp-wasm-x2.c &
16tools/xngen src/f32-vclamp/scalar.c.in -D BATCH_TILE=4 -D WASM=1 -o src/f32-vclamp/gen/vclamp-wasm-x4.c &
17
18################################## WAsm SIMD ##################################
19tools/xngen src/f32-vclamp/wasmsimd.c.in -D BATCH_TILE=4 -D X86=0 -o src/f32-vclamp/gen/vclamp-wasmsimd-arm-x4.c &
20tools/xngen src/f32-vclamp/wasmsimd.c.in -D BATCH_TILE=8 -D X86=0 -o src/f32-vclamp/gen/vclamp-wasmsimd-arm-x8.c &
21
22tools/xngen src/f32-vclamp/wasmsimd.c.in -D BATCH_TILE=4 -D X86=1 -o src/f32-vclamp/gen/vclamp-wasmsimd-x86-x4.c &
23tools/xngen src/f32-vclamp/wasmsimd.c.in -D BATCH_TILE=8 -D X86=1 -o src/f32-vclamp/gen/vclamp-wasmsimd-x86-x8.c &
24
25################################### ARM NEON ##################################
26tools/xngen src/f32-vclamp/neon.c.in -D BATCH_TILE=4 -o src/f32-vclamp/gen/vclamp-neon-x4.c &
27tools/xngen src/f32-vclamp/neon.c.in -D BATCH_TILE=8 -o src/f32-vclamp/gen/vclamp-neon-x8.c &
28
29################################# x86 128-bit #################################
30tools/xngen src/f32-vclamp/sse.c.in -D BATCH_TILE=4 -o src/f32-vclamp/gen/vclamp-sse-x4.c &
31tools/xngen src/f32-vclamp/sse.c.in -D BATCH_TILE=8 -o src/f32-vclamp/gen/vclamp-sse-x8.c &
32
33################################# x86 256-bit #################################
34tools/xngen src/f32-vclamp/avx.c.in -D BATCH_TILE=8  -o src/f32-vclamp/gen/vclamp-avx-x8.c &
35tools/xngen src/f32-vclamp/avx.c.in -D BATCH_TILE=16 -o src/f32-vclamp/gen/vclamp-avx-x16.c &
36
37################################# x86 512-bit #################################
38tools/xngen src/f32-vclamp/avx512f.c.in -D BATCH_TILE=16 -o src/f32-vclamp/gen/vclamp-avx512f-x16.c &
39tools/xngen src/f32-vclamp/avx512f.c.in -D BATCH_TILE=32 -o src/f32-vclamp/gen/vclamp-avx512f-x32.c &
40
41################################## Unit tests #################################
42tools/generate-vunary-test.py --spec test/f32-vclamp.yaml --output test/f32-vclamp.cc &
43
44wait
45