Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -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 |
| 9 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDNE -D BATCH_TILE=1 -o src/f32-vrnd/gen/vrndne-scalar-libm-x1.c |
| 10 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDNE -D BATCH_TILE=2 -o src/f32-vrnd/gen/vrndne-scalar-libm-x2.c |
| 11 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDNE -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndne-scalar-libm-x4.c |
| 12 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDZ -D BATCH_TILE=1 -o src/f32-vrnd/gen/vrndz-scalar-libm-x1.c |
| 13 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDZ -D BATCH_TILE=2 -o src/f32-vrnd/gen/vrndz-scalar-libm-x2.c |
| 14 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDZ -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-scalar-libm-x4.c |
| 15 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDU -D BATCH_TILE=1 -o src/f32-vrnd/gen/vrndu-scalar-libm-x1.c |
| 16 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDU -D BATCH_TILE=2 -o src/f32-vrnd/gen/vrndu-scalar-libm-x2.c |
| 17 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDU -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-scalar-libm-x4.c |
| 18 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDD -D BATCH_TILE=1 -o src/f32-vrnd/gen/vrndd-scalar-libm-x1.c |
| 19 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDD -D BATCH_TILE=2 -o src/f32-vrnd/gen/vrndd-scalar-libm-x2.c |
| 20 | tools/xngen src/f32-vrnd/scalar-libm.c.in -D OP=RNDD -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-scalar-libm-x4.c |
| 21 | |
Marat Dukhan | 72b399a | 2020-06-28 22:41:38 -0700 | [diff] [blame] | 22 | ################################## WAsm SIMD ################################## |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 23 | tools/xngen src/f32-vrnd/vrndne-wasmsimd-addsub.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c |
| 24 | tools/xngen src/f32-vrnd/vrndne-wasmsimd-addsub.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c |
| 25 | tools/xngen src/f32-vrnd/vrndz-wasmsimd-addsub.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c |
| 26 | tools/xngen src/f32-vrnd/vrndz-wasmsimd-addsub.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c |
| 27 | tools/xngen src/f32-vrnd/vrndu-wasmsimd-addsub.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c |
| 28 | tools/xngen src/f32-vrnd/vrndu-wasmsimd-addsub.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c |
| 29 | tools/xngen src/f32-vrnd/vrndd-wasmsimd-addsub.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c |
| 30 | tools/xngen src/f32-vrnd/vrndd-wasmsimd-addsub.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c |
| 31 | |
| 32 | tools/xngen src/f32-vrnd/vrndz-wasmsimd-cvt.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c |
| 33 | tools/xngen src/f32-vrnd/vrndz-wasmsimd-cvt.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c |
| 34 | tools/xngen src/f32-vrnd/vrndu-wasmsimd-cvt.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c |
| 35 | tools/xngen src/f32-vrnd/vrndu-wasmsimd-cvt.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c |
| 36 | tools/xngen src/f32-vrnd/vrndd-wasmsimd-cvt.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c |
| 37 | tools/xngen src/f32-vrnd/vrndd-wasmsimd-cvt.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c |
Marat Dukhan | 72b399a | 2020-06-28 22:41:38 -0700 | [diff] [blame] | 38 | |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 39 | ################################### ARM NEON ################################## |
| 40 | tools/xngen src/f32-vrnd/vrndne-neon.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndne-neon-x4.c |
| 41 | tools/xngen src/f32-vrnd/vrndne-neon.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndne-neon-x8.c |
| 42 | tools/xngen src/f32-vrnd/vrndz-neon.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-neon-x4.c |
| 43 | tools/xngen src/f32-vrnd/vrndz-neon.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-neon-x8.c |
| 44 | tools/xngen src/f32-vrnd/vrndu-neon.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-neon-x4.c |
| 45 | tools/xngen src/f32-vrnd/vrndu-neon.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-neon-x8.c |
| 46 | tools/xngen src/f32-vrnd/vrndd-neon.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-neon-x4.c |
| 47 | tools/xngen src/f32-vrnd/vrndd-neon.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-neon-x8.c |
| 48 | |
| 49 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDNE -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndne-neonv8-x4.c |
| 50 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDNE -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndne-neonv8-x8.c |
| 51 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDZ -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-neonv8-x4.c |
| 52 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDZ -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-neonv8-x8.c |
| 53 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDU -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-neonv8-x4.c |
| 54 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDU -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-neonv8-x8.c |
| 55 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDD -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-neonv8-x4.c |
| 56 | tools/xngen src/f32-vrnd/neonv8.c.in -D OP=RNDD -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-neonv8-x8.c |
| 57 | |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 58 | ################################# x86 128-bit ################################# |
| 59 | tools/xngen src/f32-vrnd/vrndne-sse2.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndne-sse2-x4.c |
| 60 | tools/xngen src/f32-vrnd/vrndne-sse2.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndne-sse2-x8.c |
| 61 | tools/xngen src/f32-vrnd/vrndz-sse2.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-sse2-x4.c |
| 62 | tools/xngen src/f32-vrnd/vrndz-sse2.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-sse2-x8.c |
| 63 | tools/xngen src/f32-vrnd/vrndu-sse2.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-sse2-x4.c |
| 64 | tools/xngen src/f32-vrnd/vrndu-sse2.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-sse2-x8.c |
| 65 | tools/xngen src/f32-vrnd/vrndd-sse2.c.in -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-sse2-x4.c |
| 66 | tools/xngen src/f32-vrnd/vrndd-sse2.c.in -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-sse2-x8.c |
| 67 | |
| 68 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDNE -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndne-sse41-x4.c |
| 69 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDNE -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndne-sse41-x8.c |
| 70 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDZ -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndz-sse41-x4.c |
| 71 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDZ -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-sse41-x8.c |
| 72 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDU -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndu-sse41-x4.c |
| 73 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDU -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-sse41-x8.c |
| 74 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDD -D BATCH_TILE=4 -o src/f32-vrnd/gen/vrndd-sse41-x4.c |
| 75 | tools/xngen src/f32-vrnd/sse41.c.in -D OP=RNDD -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-sse41-x8.c |
| 76 | |
| 77 | ################################# x86 256-bit ################################# |
| 78 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDNE -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndne-avx-x8.c |
| 79 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDNE -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndne-avx-x16.c |
| 80 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDZ -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndz-avx-x8.c |
| 81 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDZ -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndz-avx-x16.c |
| 82 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDU -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndu-avx-x8.c |
| 83 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDU -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndu-avx-x16.c |
| 84 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDD -D BATCH_TILE=8 -o src/f32-vrnd/gen/vrndd-avx-x8.c |
| 85 | tools/xngen src/f32-vrnd/avx.c.in -D OP=RNDD -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndd-avx-x16.c |
| 86 | |
| 87 | ################################# x86 512-bit ################################# |
| 88 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDNE -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndne-avx512f-x16.c |
| 89 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDNE -D BATCH_TILE=32 -o src/f32-vrnd/gen/vrndne-avx512f-x32.c |
| 90 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDZ -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndz-avx512f-x16.c |
| 91 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDZ -D BATCH_TILE=32 -o src/f32-vrnd/gen/vrndz-avx512f-x32.c |
| 92 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDU -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndu-avx512f-x16.c |
| 93 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDU -D BATCH_TILE=32 -o src/f32-vrnd/gen/vrndu-avx512f-x32.c |
| 94 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDD -D BATCH_TILE=16 -o src/f32-vrnd/gen/vrndd-avx512f-x16.c |
| 95 | tools/xngen src/f32-vrnd/avx512f.c.in -D OP=RNDD -D BATCH_TILE=32 -o src/f32-vrnd/gen/vrndd-avx512f-x32.c |
| 96 | |
| 97 | ################################## Unit tests ################################# |
| 98 | tools/generate-vunary-test.py --spec test/f32-vrndne.yaml --output test/f32-vrndne.cc |
| 99 | tools/generate-vunary-test.py --spec test/f32-vrndz.yaml --output test/f32-vrndz.cc |
| 100 | tools/generate-vunary-test.py --spec test/f32-vrndu.yaml --output test/f32-vrndu.cc |
| 101 | tools/generate-vunary-test.py --spec test/f32-vrndd.yaml --output test/f32-vrndd.cc |