Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 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 |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 9 | tools/xngen src/f32-vrelu/scalar.c.in -D BATCH_TILE=1 -o src/f32-vrelu/gen/vrelu-scalar-x1.c & |
| 10 | tools/xngen src/f32-vrelu/scalar.c.in -D BATCH_TILE=2 -o src/f32-vrelu/gen/vrelu-scalar-x2.c & |
| 11 | tools/xngen src/f32-vrelu/scalar.c.in -D BATCH_TILE=4 -o src/f32-vrelu/gen/vrelu-scalar-x4.c & |
| 12 | tools/xngen src/f32-vrelu/scalar.c.in -D BATCH_TILE=8 -o src/f32-vrelu/gen/vrelu-scalar-x8.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 13 | |
| 14 | ### WAsm-specific micro-kernels |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 15 | tools/xngen src/f32-vrelu/wasm.c.in -D BATCH_TILE=1 -o src/f32-vrelu/gen/vrelu-wasm-x1.c & |
| 16 | tools/xngen src/f32-vrelu/wasm.c.in -D BATCH_TILE=2 -o src/f32-vrelu/gen/vrelu-wasm-x2.c & |
| 17 | tools/xngen src/f32-vrelu/wasm.c.in -D BATCH_TILE=4 -o src/f32-vrelu/gen/vrelu-wasm-x4.c & |
| 18 | tools/xngen src/f32-vrelu/wasm.c.in -D BATCH_TILE=8 -o src/f32-vrelu/gen/vrelu-wasm-x8.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 19 | |
| 20 | ################################## WAsm SIMD ################################## |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 21 | tools/xngen src/f32-vrelu/wasmsimd.c.in -D BATCH_TILE=4 -o src/f32-vrelu/gen/vrelu-wasmsimd-x4.c & |
| 22 | tools/xngen src/f32-vrelu/wasmsimd.c.in -D BATCH_TILE=8 -o src/f32-vrelu/gen/vrelu-wasmsimd-x8.c & |
| 23 | tools/xngen src/f32-vrelu/wasmsimd.c.in -D BATCH_TILE=16 -o src/f32-vrelu/gen/vrelu-wasmsimd-x16.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 24 | |
| 25 | ################################### ARM NEON ################################## |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 26 | tools/xngen src/f32-vrelu/neon.c.in -D BATCH_TILE=4 -o src/f32-vrelu/gen/vrelu-neon-x4.c & |
| 27 | tools/xngen src/f32-vrelu/neon.c.in -D BATCH_TILE=8 -o src/f32-vrelu/gen/vrelu-neon-x8.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 28 | |
| 29 | ################################# x86 128-bit ################################# |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 30 | tools/xngen src/f32-vrelu/sse.c.in -D BATCH_TILE=4 -o src/f32-vrelu/gen/vrelu-sse-x4.c & |
| 31 | tools/xngen src/f32-vrelu/sse.c.in -D BATCH_TILE=8 -o src/f32-vrelu/gen/vrelu-sse-x8.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 32 | |
| 33 | ################################# x86 256-bit ################################# |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 34 | tools/xngen src/f32-vrelu/avx.c.in -D BATCH_TILE=8 -o src/f32-vrelu/gen/vrelu-avx-x8.c & |
| 35 | tools/xngen src/f32-vrelu/avx.c.in -D BATCH_TILE=16 -o src/f32-vrelu/gen/vrelu-avx-x16.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 36 | |
| 37 | ################################# x86 512-bit ################################# |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 38 | tools/xngen src/f32-vrelu/avx512f.c.in -D BATCH_TILE=16 -o src/f32-vrelu/gen/vrelu-avx512f-x16.c & |
| 39 | tools/xngen src/f32-vrelu/avx512f.c.in -D BATCH_TILE=32 -o src/f32-vrelu/gen/vrelu-avx512f-x32.c & |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 40 | |
| 41 | ################################## Unit tests ################################# |
Frank Barchard | aae722a | 2021-08-30 12:37:30 -0700 | [diff] [blame] | 42 | tools/generate-vunary-test.py --spec test/f32-vrelu.yaml --output test/f32-vrelu.cc & |
| 43 | |
| 44 | wait |