xref: /aosp_15_r20/external/XNNPACK/scripts/generate-f32-vmulcaddc.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-vmulcaddc/scalar.c.in -D CHANNEL_TILE=1 -D ROW_TILE=2 -D WASM=0 -o src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c &
10tools/xngen src/f32-vmulcaddc/scalar.c.in -D CHANNEL_TILE=2 -D ROW_TILE=2 -D WASM=0 -o src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c &
11tools/xngen src/f32-vmulcaddc/scalar.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D WASM=0 -o src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c &
12
13### WAsm-specific micro-kernels
14tools/xngen src/f32-vmulcaddc/scalar.c.in -D CHANNEL_TILE=1 -D ROW_TILE=2 -D WASM=1 -o src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c &
15tools/xngen src/f32-vmulcaddc/scalar.c.in -D CHANNEL_TILE=2 -D ROW_TILE=2 -D WASM=1 -o src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c &
16tools/xngen src/f32-vmulcaddc/scalar.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D WASM=1 -o src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c &
17
18################################## WAsm SIMD ##################################
19tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D ARCH=ARM     -D FMA=0 -o src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c &
20tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -D ARCH=ARM     -D FMA=0 -o src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c &
21
22tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D ARCH=X86     -D FMA=0 -o src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c &
23tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -D ARCH=X86     -D FMA=0 -o src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c &
24
25tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D ARCH=RELAXED -D FMA=0 -o src/f32-vmulcaddc/gen/c4-minmax-wasmrelaxedsimd-2x.c &
26tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -D ARCH=RELAXED -D FMA=0 -o src/f32-vmulcaddc/gen/c8-minmax-wasmrelaxedsimd-2x.c &
27
28tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D ARCH=RELAXED -D FMA=1 -o src/f32-vmulcaddc/gen/c4-minmax-wasmrelaxedsimd-fma-2x.c &
29tools/xngen src/f32-vmulcaddc/wasmsimd.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -D ARCH=RELAXED -D FMA=1 -o src/f32-vmulcaddc/gen/c8-minmax-wasmrelaxedsimd-fma-2x.c &
30
31################################### ARM NEON ##################################
32tools/xngen src/f32-vmulcaddc/neon.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D FMA=0 -o src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c &
33tools/xngen src/f32-vmulcaddc/neon.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -D FMA=0 -o src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c &
34
35tools/xngen src/f32-vmulcaddc/neon.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -D FMA=1 -o src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c &
36tools/xngen src/f32-vmulcaddc/neon.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -D FMA=1 -o src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c &
37
38################################### x86 SSE ###################################
39tools/xngen src/f32-vmulcaddc/sse.c.in -D CHANNEL_TILE=4 -D ROW_TILE=2 -o src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c &
40tools/xngen src/f32-vmulcaddc/sse.c.in -D CHANNEL_TILE=8 -D ROW_TILE=2 -o src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c &
41
42################################## Unit tests #################################
43tools/generate-vmulcaddc-test.py --spec test/f32-vmulcaddc-minmax.yaml --output test/f32-vmulcaddc-minmax.cc &
44
45wait
46