xref: /aosp_15_r20/external/XNNPACK/scripts/generate-bf16-gemm.sh (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1#!/bin/sh
2# Copyright 2022 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 ##################################
8### LD128 micro-kernels
9tools/xngen src/bf16-gemm/c2-neonbf16-bfdot-lane-ld128.c.in -D MR=1 -D NR=8 -o src/bf16-gemm/gen/1x8c2-minmax-neonbf16-bfdot-lane-ld128.c &
10tools/xngen src/bf16-gemm/c2-neonbf16-bfdot-lane-ld128.c.in -D MR=4 -D NR=8 -o src/bf16-gemm/gen/4x8c2-minmax-neonbf16-bfdot-lane-ld128.c &
11tools/xngen src/bf16-gemm/c2-neonbf16-bfdot-lane-ld128.c.in -D MR=5 -D NR=8 -o src/bf16-gemm/gen/5x8c2-minmax-neonbf16-bfdot-lane-ld128.c &
12tools/xngen src/bf16-gemm/c2-neonbf16-bfdot-lane-ld128.c.in -D MR=6 -D NR=8 -o src/bf16-gemm/gen/6x8c2-minmax-neonbf16-bfdot-lane-ld128.c &
13
14tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=1 -D NR=4 -D EXTOPT=SHLAND -o src/bf16-gemm/gen/1x4c8-minmax-neonfma-shland.c &
15tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=2 -D NR=4 -D EXTOPT=SHLAND -o src/bf16-gemm/gen/2x4c8-minmax-neonfma-shland.c &
16tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=3 -D NR=4 -D EXTOPT=SHLAND -o src/bf16-gemm/gen/3x4c8-minmax-neonfma-shland.c &
17tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=4 -D NR=4 -D EXTOPT=SHLAND -o src/bf16-gemm/gen/4x4c8-minmax-neonfma-shland.c &
18tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=5 -D NR=4 -D EXTOPT=SHLAND -o src/bf16-gemm/gen/5x4c8-minmax-neonfma-shland.c &
19
20tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=1 -D NR=4 -D EXTOPT=ZIP -o src/bf16-gemm/gen/1x4c8-minmax-neonfma-zip.c &
21tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=2 -D NR=4 -D EXTOPT=ZIP -o src/bf16-gemm/gen/2x4c8-minmax-neonfma-zip.c &
22tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=3 -D NR=4 -D EXTOPT=ZIP -o src/bf16-gemm/gen/3x4c8-minmax-neonfma-zip.c &
23tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=4 -D NR=4 -D EXTOPT=ZIP -o src/bf16-gemm/gen/4x4c8-minmax-neonfma-zip.c &
24tools/xngen src/bf16-gemm/c8-neon.c.in -D MR=5 -D NR=4 -D EXTOPT=ZIP -o src/bf16-gemm/gen/5x4c8-minmax-neonfma-zip.c &
25
26tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=1 -D NR=4 -D BFOPT=BFDOT -o src/bf16-gemm/gen/1x4c8-minmax-neonbf16-bfdot.c &
27tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=2 -D NR=4 -D BFOPT=BFDOT -o src/bf16-gemm/gen/2x4c8-minmax-neonbf16-bfdot.c &
28tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=3 -D NR=4 -D BFOPT=BFDOT -o src/bf16-gemm/gen/3x4c8-minmax-neonbf16-bfdot.c &
29tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=4 -D NR=4 -D BFOPT=BFDOT -o src/bf16-gemm/gen/4x4c8-minmax-neonbf16-bfdot.c &
30tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=5 -D NR=4 -D BFOPT=BFDOT -o src/bf16-gemm/gen/5x4c8-minmax-neonbf16-bfdot.c &
31
32tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=1 -D NR=4 -D BFOPT=BFMLAL -o src/bf16-gemm/gen/1x4c8-minmax-neonbf16-bfmlal.c &
33tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=2 -D NR=4 -D BFOPT=BFMLAL -o src/bf16-gemm/gen/2x4c8-minmax-neonbf16-bfmlal.c &
34tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=3 -D NR=4 -D BFOPT=BFMLAL -o src/bf16-gemm/gen/3x4c8-minmax-neonbf16-bfmlal.c &
35tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=4 -D NR=4 -D BFOPT=BFMLAL -o src/bf16-gemm/gen/4x4c8-minmax-neonbf16-bfmlal.c &
36tools/xngen src/bf16-gemm/c8-neonbf16.c.in -D MR=5 -D NR=4 -D BFOPT=BFMLAL -o src/bf16-gemm/gen/5x4c8-minmax-neonbf16-bfmlal.c &
37
38################################## Unit tests #################################
39tools/generate-gemm-test.py --spec test/bf16-gemm-minmax.yaml --output test/bf16-gemm-minmax.cc &
40
41wait
42