xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f32-vhswish.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1#!/bin/sh
2# Copyright 2019 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-vhswish/scalar.c.in -D BATCH_TILE=1 -D WASM=0 -o src/f32-vhswish/gen/vhswish-scalar-x1.c &
10tools/xngen src/f32-vhswish/scalar.c.in -D BATCH_TILE=2 -D WASM=0 -o src/f32-vhswish/gen/vhswish-scalar-x2.c &
11tools/xngen src/f32-vhswish/scalar.c.in -D BATCH_TILE=4 -D WASM=0 -o src/f32-vhswish/gen/vhswish-scalar-x4.c &
12
13### WAsm-specific micro-kernels
14tools/xngen src/f32-vhswish/scalar.c.in -D BATCH_TILE=1 -D WASM=1 -o src/f32-vhswish/gen/vhswish-wasm-x1.c &
15tools/xngen src/f32-vhswish/scalar.c.in -D BATCH_TILE=2 -D WASM=1 -o src/f32-vhswish/gen/vhswish-wasm-x2.c &
16tools/xngen src/f32-vhswish/scalar.c.in -D BATCH_TILE=4 -D WASM=1 -o src/f32-vhswish/gen/vhswish-wasm-x4.c &
17
18################################## WAsm SIMD ##################################
19tools/xngen src/f32-vhswish/wasmsimd.c.in -D BATCH_TILE=4  -o src/f32-vhswish/gen/vhswish-wasmsimd-x4.c &
20tools/xngen src/f32-vhswish/wasmsimd.c.in -D BATCH_TILE=8  -o src/f32-vhswish/gen/vhswish-wasmsimd-x8.c &
21tools/xngen src/f32-vhswish/wasmsimd.c.in -D BATCH_TILE=16 -o src/f32-vhswish/gen/vhswish-wasmsimd-x16.c &
22
23################################### ARM NEON ##################################
24tools/xngen src/f32-vhswish/neon.c.in -D BATCH_TILE=4  -o src/f32-vhswish/gen/vhswish-neon-x4.c &
25tools/xngen src/f32-vhswish/neon.c.in -D BATCH_TILE=8  -o src/f32-vhswish/gen/vhswish-neon-x8.c &
26tools/xngen src/f32-vhswish/neon.c.in -D BATCH_TILE=16 -o src/f32-vhswish/gen/vhswish-neon-x16.c &
27
28################################# x86 128-bit #################################
29tools/xngen src/f32-vhswish/sse.c.in -D BATCH_TILE=4 -o src/f32-vhswish/gen/vhswish-sse-x4.c &
30tools/xngen src/f32-vhswish/sse.c.in -D BATCH_TILE=8 -o src/f32-vhswish/gen/vhswish-sse-x8.c &
31
32################################# x86 256-bit #################################
33tools/xngen src/f32-vhswish/avx.c.in -D BATCH_TILE=8 -D FMA=0 -o src/f32-vhswish/gen/vhswish-avx-x8.c &
34tools/xngen src/f32-vhswish/avx.c.in -D BATCH_TILE=16 -D FMA=0 -o src/f32-vhswish/gen/vhswish-avx-x16.c &
35
36tools/xngen src/f32-vhswish/avx.c.in -D BATCH_TILE=8 -D FMA=3 -o src/f32-vhswish/gen/vhswish-fma3-x8.c &
37tools/xngen src/f32-vhswish/avx.c.in -D BATCH_TILE=16 -D FMA=3 -o src/f32-vhswish/gen/vhswish-fma3-x16.c &
38
39################################# x86 512-bit #################################
40tools/xngen src/f32-vhswish/avx512f.c.in -D BATCH_TILE=16 -o src/f32-vhswish/gen/vhswish-avx512f-x16.c &
41tools/xngen src/f32-vhswish/avx512f.c.in -D BATCH_TILE=32 -o src/f32-vhswish/gen/vhswish-avx512f-x32.c &
42
43################################## Unit tests #################################
44tools/generate-vunary-test.py --spec test/f32-vhswish.yaml --output test/f32-vhswish.cc &
45
46wait
47