| 1 |
use criterion::{black_box, criterion_group, criterion_main, Bencher, BenchmarkId, Criterion}; |
| 2 |
use half::prelude::*; |
| 3 |
use std::{f32, f64, iter}; |
| 4 |
|
| 5 |
const SIMD_LARGE_BENCH_SLICE_LEN: usize = 1024; |
| 6 |
|
| 7 |
fn bench_f32_to_f16(c: &mut Criterion) { |
| 8 |
let mut group = c.benchmark_group("Convert f16 From f32"); |
| 9 |
for val in &[ |
| 10 |
0., |
| 11 |
-0., |
| 12 |
1., |
| 13 |
f32::MIN, |
| 14 |
f32::MAX, |
| 15 |
f32::MIN_POSITIVE, |
| 16 |
f32::NEG_INFINITY, |
| 17 |
f32::INFINITY, |
| 18 |
f32::NAN, |
| 19 |
f32::consts::E, |
| 20 |
f32::consts::PI, |
| 21 |
] { |
| 22 |
group.bench_with_input(BenchmarkId::new("f16::from_f32", val), val, |b, i| { |
| 23 |
b.iter(|| f16::from_f32(*i)) |
| 24 |
}); |
| 25 |
} |
| 26 |
} |
| 27 |
|
| 28 |
fn bench_f64_to_f16(c: &mut Criterion) { |
| 29 |
let mut group = c.benchmark_group("Convert f16 From f64"); |
| 30 |
for val in &[ |
| 31 |
0., |
| 32 |
-0., |
| 33 |
1., |
| 34 |
f64::MIN, |
| 35 |
f64::MAX, |
| 36 |
f64::MIN_POSITIVE, |
| 37 |
f64::NEG_INFINITY, |
| 38 |
f64::INFINITY, |
| 39 |
f64::NAN, |
| 40 |
f64::consts::E, |
| 41 |
f64::consts::PI, |
| 42 |
] { |
| 43 |
group.bench_with_input(BenchmarkId::new("f16::from_f64", val), val, |b, i| { |
| 44 |
b.iter(|| f16::from_f64(*i)) |
| 45 |
}); |
| 46 |
} |
| 47 |
} |
| 48 |
|
| 49 |
fn bench_f16_to_f32(c: &mut Criterion) { |
| 50 |
let mut group = c.benchmark_group("Convert f16 to f32"); |
| 51 |
for val in &[ |
| 52 |
f16::ZERO, |
| 53 |
f16::NEG_ZERO, |
| 54 |
f16::ONE, |
| 55 |
f16::MIN, |
| 56 |
f16::MAX, |
| 57 |
f16::MIN_POSITIVE, |
| 58 |
f16::NEG_INFINITY, |
| 59 |
f16::INFINITY, |
| 60 |
f16::NAN, |
| 61 |
f16::E, |
| 62 |
f16::PI, |
| 63 |
] { |
| 64 |
group.bench_with_input(BenchmarkId::new("f16::to_f32", val), val, |b, i| { |
| 65 |
b.iter(|| i.to_f32()) |
| 66 |
}); |
| 67 |
} |
| 68 |
} |
| 69 |
|
| 70 |
fn bench_f16_to_f64(c: &mut Criterion) { |
| 71 |
let mut group = c.benchmark_group("Convert f16 to f64"); |
| 72 |
for val in &[ |
| 73 |
f16::ZERO, |
| 74 |
f16::NEG_ZERO, |
| 75 |
f16::ONE, |
| 76 |
f16::MIN, |
| 77 |
f16::MAX, |
| 78 |
f16::MIN_POSITIVE, |
| 79 |
f16::NEG_INFINITY, |
| 80 |
f16::INFINITY, |
| 81 |
f16::NAN, |
| 82 |
f16::E, |
| 83 |
f16::PI, |
| 84 |
] { |
| 85 |
group.bench_with_input(BenchmarkId::new("f16::to_f64", val), val, |b, i| { |
| 86 |
b.iter(|| i.to_f64()) |
| 87 |
}); |
| 88 |
} |
| 89 |
} |
| 90 |
|
| 91 |
criterion_group!( |
| 92 |
f16_sisd, |
| 93 |
bench_f32_to_f16, |
| 94 |
bench_f64_to_f16, |
| 95 |
bench_f16_to_f32, |
| 96 |
bench_f16_to_f64 |
| 97 |
); |
| 98 |
|
| 99 |
fn bench_slice_f32_to_f16(c: &mut Criterion) { |
| 100 |
let mut constant_buffer = [f16::ZERO; 11]; |
| 101 |
let constants = [ |
| 102 |
0., |
| 103 |
-0., |
| 104 |
1., |
| 105 |
f32::MIN, |
| 106 |
f32::MAX, |
| 107 |
f32::MIN_POSITIVE, |
| 108 |
f32::NEG_INFINITY, |
| 109 |
f32::INFINITY, |
| 110 |
f32::NAN, |
| 111 |
f32::consts::E, |
| 112 |
f32::consts::PI, |
| 113 |
]; |
| 114 |
c.bench_function( |
| 115 |
"HalfFloatSliceExt::convert_from_f32_slice/constants", |
| 116 |
|b: &mut Bencher<'_>| { |
| 117 |
b.iter(|| black_box(&mut constant_buffer).convert_from_f32_slice(black_box(&constants))) |
| 118 |
}, |
| 119 |
); |
| 120 |
|
| 121 |
let large: Vec<_> = iter::repeat(0) |
| 122 |
.enumerate() |
| 123 |
.map(|(i, _)| i as f32) |
| 124 |
.take(SIMD_LARGE_BENCH_SLICE_LEN) |
| 125 |
.collect(); |
| 126 |
let mut large_buffer = [f16::ZERO; SIMD_LARGE_BENCH_SLICE_LEN]; |
| 127 |
c.bench_function( |
| 128 |
"HalfFloatSliceExt::convert_from_f32_slice/large", |
| 129 |
|b: &mut Bencher<'_>| { |
| 130 |
b.iter(|| black_box(&mut large_buffer).convert_from_f32_slice(black_box(&large))) |
| 131 |
}, |
| 132 |
); |
| 133 |
} |
| 134 |
|
| 135 |
fn bench_slice_f64_to_f16(c: &mut Criterion) { |
| 136 |
let mut constant_buffer = [f16::ZERO; 11]; |
| 137 |
let constants = [ |
| 138 |
0., |
| 139 |
-0., |
| 140 |
1., |
| 141 |
f64::MIN, |
| 142 |
f64::MAX, |
| 143 |
f64::MIN_POSITIVE, |
| 144 |
f64::NEG_INFINITY, |
| 145 |
f64::INFINITY, |
| 146 |
f64::NAN, |
| 147 |
f64::consts::E, |
| 148 |
f64::consts::PI, |
| 149 |
]; |
| 150 |
c.bench_function( |
| 151 |
"HalfFloatSliceExt::convert_from_f64_slice/constants", |
| 152 |
|b: &mut Bencher<'_>| { |
| 153 |
b.iter(|| black_box(&mut constant_buffer).convert_from_f64_slice(black_box(&constants))) |
| 154 |
}, |
| 155 |
); |
| 156 |
|
| 157 |
let large: Vec<_> = iter::repeat(0) |
| 158 |
.enumerate() |
| 159 |
.map(|(i, _)| i as f64) |
| 160 |
.take(SIMD_LARGE_BENCH_SLICE_LEN) |
| 161 |
.collect(); |
| 162 |
let mut large_buffer = [f16::ZERO; SIMD_LARGE_BENCH_SLICE_LEN]; |
| 163 |
c.bench_function( |
| 164 |
"HalfFloatSliceExt::convert_from_f64_slice/large", |
| 165 |
|b: &mut Bencher<'_>| { |
| 166 |
b.iter(|| black_box(&mut large_buffer).convert_from_f64_slice(black_box(&large))) |
| 167 |
}, |
| 168 |
); |
| 169 |
} |
| 170 |
|
| 171 |
fn bench_slice_f16_to_f32(c: &mut Criterion) { |
| 172 |
let mut constant_buffer = [0f32; 11]; |
| 173 |
let constants = [ |
| 174 |
f16::ZERO, |
| 175 |
f16::NEG_ZERO, |
| 176 |
f16::ONE, |
| 177 |
f16::MIN, |
| 178 |
f16::MAX, |
| 179 |
f16::MIN_POSITIVE, |
| 180 |
f16::NEG_INFINITY, |
| 181 |
f16::INFINITY, |
| 182 |
f16::NAN, |
| 183 |
f16::E, |
| 184 |
f16::PI, |
| 185 |
]; |
| 186 |
c.bench_function( |
| 187 |
"HalfFloatSliceExt::convert_to_f32_slice/constants", |
| 188 |
|b: &mut Bencher<'_>| { |
| 189 |
b.iter(|| black_box(&constants).convert_to_f32_slice(black_box(&mut constant_buffer))) |
| 190 |
}, |
| 191 |
); |
| 192 |
|
| 193 |
let large: Vec<_> = iter::repeat(0) |
| 194 |
.enumerate() |
| 195 |
.map(|(i, _)| f16::from_f32(i as f32)) |
| 196 |
.take(SIMD_LARGE_BENCH_SLICE_LEN) |
| 197 |
.collect(); |
| 198 |
let mut large_buffer = [0f32; SIMD_LARGE_BENCH_SLICE_LEN]; |
| 199 |
c.bench_function( |
| 200 |
"HalfFloatSliceExt::convert_to_f32_slice/large", |
| 201 |
|b: &mut Bencher<'_>| { |
| 202 |
b.iter(|| black_box(&large).convert_to_f32_slice(black_box(&mut large_buffer))) |
| 203 |
}, |
| 204 |
); |
| 205 |
} |
| 206 |
|
| 207 |
fn bench_slice_f16_to_f64(c: &mut Criterion) { |
| 208 |
let mut constant_buffer = [0f64; 11]; |
| 209 |
let constants = [ |
| 210 |
f16::ZERO, |
| 211 |
f16::NEG_ZERO, |
| 212 |
f16::ONE, |
| 213 |
f16::MIN, |
| 214 |
f16::MAX, |
| 215 |
f16::MIN_POSITIVE, |
| 216 |
f16::NEG_INFINITY, |
| 217 |
f16::INFINITY, |
| 218 |
f16::NAN, |
| 219 |
f16::E, |
| 220 |
f16::PI, |
| 221 |
]; |
| 222 |
c.bench_function( |
| 223 |
"HalfFloatSliceExt::convert_to_f64_slice/constants", |
| 224 |
|b: &mut Bencher<'_>| { |
| 225 |
b.iter(|| black_box(&constants).convert_to_f64_slice(black_box(&mut constant_buffer))) |
| 226 |
}, |
| 227 |
); |
| 228 |
|
| 229 |
let large: Vec<_> = iter::repeat(0) |
| 230 |
.enumerate() |
| 231 |
.map(|(i, _)| f16::from_f64(i as f64)) |
| 232 |
.take(SIMD_LARGE_BENCH_SLICE_LEN) |
| 233 |
.collect(); |
| 234 |
let mut large_buffer = [0f64; SIMD_LARGE_BENCH_SLICE_LEN]; |
| 235 |
c.bench_function( |
| 236 |
"HalfFloatSliceExt::convert_to_f64_slice/large", |
| 237 |
|b: &mut Bencher<'_>| { |
| 238 |
b.iter(|| black_box(&large).convert_to_f64_slice(black_box(&mut large_buffer))) |
| 239 |
}, |
| 240 |
); |
| 241 |
} |
| 242 |
|
| 243 |
criterion_group!( |
| 244 |
f16_simd, |
| 245 |
bench_slice_f32_to_f16, |
| 246 |
bench_slice_f64_to_f16, |
| 247 |
bench_slice_f16_to_f32, |
| 248 |
bench_slice_f16_to_f64 |
| 249 |
); |
| 250 |
|
| 251 |
fn bench_f32_to_bf16(c: &mut Criterion) { |
| 252 |
let mut group = c.benchmark_group("Convert bf16 From f32"); |
| 253 |
for val in &[ |
| 254 |
0., |
| 255 |
-0., |
| 256 |
1., |
| 257 |
f32::MIN, |
| 258 |
f32::MAX, |
| 259 |
f32::MIN_POSITIVE, |
| 260 |
f32::NEG_INFINITY, |
| 261 |
f32::INFINITY, |
| 262 |
f32::NAN, |
| 263 |
f32::consts::E, |
| 264 |
f32::consts::PI, |
| 265 |
] { |
| 266 |
group.bench_with_input(BenchmarkId::new("bf16::from_f32", val), val, |b, i| { |
| 267 |
b.iter(|| bf16::from_f32(*i)) |
| 268 |
}); |
| 269 |
} |
| 270 |
} |
| 271 |
|
| 272 |
fn bench_f64_to_bf16(c: &mut Criterion) { |
| 273 |
let mut group = c.benchmark_group("Convert bf16 From f64"); |
| 274 |
for val in &[ |
| 275 |
0., |
| 276 |
-0., |
| 277 |
1., |
| 278 |
f64::MIN, |
| 279 |
f64::MAX, |
| 280 |
f64::MIN_POSITIVE, |
| 281 |
f64::NEG_INFINITY, |
| 282 |
f64::INFINITY, |
| 283 |
f64::NAN, |
| 284 |
f64::consts::E, |
| 285 |
f64::consts::PI, |
| 286 |
] { |
| 287 |
group.bench_with_input(BenchmarkId::new("bf16::from_f64", val), val, |b, i| { |
| 288 |
b.iter(|| bf16::from_f64(*i)) |
| 289 |
}); |
| 290 |
} |
| 291 |
} |
| 292 |
|
| 293 |
fn bench_bf16_to_f32(c: &mut Criterion) { |
| 294 |
let mut group = c.benchmark_group("Convert bf16 to f32"); |
| 295 |
for val in &[ |
| 296 |
bf16::ZERO, |
| 297 |
bf16::NEG_ZERO, |
| 298 |
bf16::ONE, |
| 299 |
bf16::MIN, |
| 300 |
bf16::MAX, |
| 301 |
bf16::MIN_POSITIVE, |
| 302 |
bf16::NEG_INFINITY, |
| 303 |
bf16::INFINITY, |
| 304 |
bf16::NAN, |
| 305 |
bf16::E, |
| 306 |
bf16::PI, |
| 307 |
] { |
| 308 |
group.bench_with_input(BenchmarkId::new("bf16::to_f32", val), val, |b, i| { |
| 309 |
b.iter(|| i.to_f32()) |
| 310 |
}); |
| 311 |
} |
| 312 |
} |
| 313 |
|
| 314 |
fn bench_bf16_to_f64(c: &mut Criterion) { |
| 315 |
let mut group = c.benchmark_group("Convert bf16 to f64"); |
| 316 |
for val in &[ |
| 317 |
bf16::ZERO, |
| 318 |
bf16::NEG_ZERO, |
| 319 |
bf16::ONE, |
| 320 |
bf16::MIN, |
| 321 |
bf16::MAX, |
| 322 |
bf16::MIN_POSITIVE, |
| 323 |
bf16::NEG_INFINITY, |
| 324 |
bf16::INFINITY, |
| 325 |
bf16::NAN, |
| 326 |
bf16::E, |
| 327 |
bf16::PI, |
| 328 |
] { |
| 329 |
group.bench_with_input(BenchmarkId::new("bf16::to_f64", val), val, |b, i| { |
| 330 |
b.iter(|| i.to_f64()) |
| 331 |
}); |
| 332 |
} |
| 333 |
} |
| 334 |
|
| 335 |
criterion_group!( |
| 336 |
bf16_sisd, |
| 337 |
bench_f32_to_bf16, |
| 338 |
bench_f64_to_bf16, |
| 339 |
bench_bf16_to_f32, |
| 340 |
bench_bf16_to_f64 |
| 341 |
); |
| 342 |
|
| 343 |
criterion_main!(f16_sisd, bf16_sisd, f16_simd); |