| 1 |
// Modified version of https://github.com/TooManyBees/lab |
| 2 |
|
| 3 |
use lab::Lab; |
| 4 |
|
| 5 |
// κ and ε parameters used in conversion between XYZ and La*b*. See |
| 6 |
// http://www.brucelindbloom.com/LContinuity.html for explanation as to why |
| 7 |
// those are different values than those provided by CIE standard. |
| 8 |
const KAPPA: f32 = 24389.0 / 27.0; |
| 9 |
const EPSILON: f32 = 216.0 / 24389.0; |
| 10 |
|
| 11 |
pub fn rgb_to_lab(rgb: &[f32; 3]) -> Lab { |
| 12 |
xyz_to_lab(rgb_to_xyz(rgb)) |
| 13 |
} |
| 14 |
|
| 15 |
#[allow(clippy::excessive_precision)] |
| 16 |
fn rgb_to_xyz(rgb: &[f32; 3]) -> [f32; 3] { |
| 17 |
let r = rgb_to_xyz_map(rgb[0]); |
| 18 |
let g = rgb_to_xyz_map(rgb[1]); |
| 19 |
let b = rgb_to_xyz_map(rgb[2]); |
| 20 |
|
| 21 |
[ |
| 22 |
r * 0.4124564390896921 + g * 0.357576077643909 + b * 0.18043748326639894, |
| 23 |
r * 0.21267285140562248 + g * 0.715152155287818 + b * 0.07217499330655958, |
| 24 |
r * 0.019333895582329317 + g * 0.119192025881303 + b * 0.9503040785363677, |
| 25 |
] |
| 26 |
} |
| 27 |
|
| 28 |
#[inline] |
| 29 |
fn rgb_to_xyz_map(c: f32) -> f32 { |
| 30 |
if c > 10. / 255. { |
| 31 |
const A: f32 = 0.055; |
| 32 |
const D: f32 = 1.0 / 1.055; |
| 33 |
pow_2_4((c + A) * D) |
| 34 |
} else { |
| 35 |
const D: f32 = 1.0 / 12.92; |
| 36 |
c * D |
| 37 |
} |
| 38 |
} |
| 39 |
|
| 40 |
fn xyz_to_lab(xyz: [f32; 3]) -> Lab { |
| 41 |
let x = xyz_to_lab_map(xyz[0] * (1.0 / 0.95047)); |
| 42 |
let y = xyz_to_lab_map(xyz[1]); |
| 43 |
let z = xyz_to_lab_map(xyz[2] * (1.0 / 1.08883)); |
| 44 |
|
| 45 |
Lab { |
| 46 |
l: (116.0 * y) - 16.0, |
| 47 |
a: 500.0 * (x - y), |
| 48 |
b: 200.0 * (y - z), |
| 49 |
} |
| 50 |
} |
| 51 |
|
| 52 |
#[inline] |
| 53 |
fn xyz_to_lab_map(c: f32) -> f32 { |
| 54 |
if c > EPSILON { |
| 55 |
cbrt_approx(c) |
| 56 |
} else { |
| 57 |
(KAPPA * c + 16.0) * (1.0 / 116.0) |
| 58 |
} |
| 59 |
} |
| 60 |
|
| 61 |
macro_rules! lookup_table_8 { |
| 62 |
(start: $start:expr, closure: $closure:expr) => { |
| 63 |
[ |
| 64 |
$closure($start + 0), |
| 65 |
$closure($start + 1), |
| 66 |
$closure($start + 2), |
| 67 |
$closure($start + 3), |
| 68 |
$closure($start + 4), |
| 69 |
$closure($start + 5), |
| 70 |
$closure($start + 6), |
| 71 |
$closure($start + 7), |
| 72 |
] |
| 73 |
}; |
| 74 |
} |
| 75 |
|
| 76 |
macro_rules! lookup_table_16 { |
| 77 |
(start: $start:expr, closure: $closure:expr) => { |
| 78 |
[ |
| 79 |
$closure($start + 0), |
| 80 |
$closure($start + 1), |
| 81 |
$closure($start + 2), |
| 82 |
$closure($start + 3), |
| 83 |
$closure($start + 4), |
| 84 |
$closure($start + 5), |
| 85 |
$closure($start + 6), |
| 86 |
$closure($start + 7), |
| 87 |
$closure($start + 8), |
| 88 |
$closure($start + 9), |
| 89 |
$closure($start + 10), |
| 90 |
$closure($start + 11), |
| 91 |
$closure($start + 12), |
| 92 |
$closure($start + 13), |
| 93 |
$closure($start + 14), |
| 94 |
$closure($start + 15), |
| 95 |
] |
| 96 |
}; |
| 97 |
} |
| 98 |
|
| 99 |
fn pow_2_4(x: f32) -> f32 { |
| 100 |
// Closely approximate x^2.4. |
| 101 |
// Divide x by its exponent and a truncated version of itself to get it as close to 1 as |
| 102 |
// possible. Calculate the power of 2.4 using the binomial method. Multiply what was divided to |
| 103 |
// the power of 2.4. |
| 104 |
|
| 105 |
// Lookup tables still have to be hardcoded. |
| 106 |
const FRAC_BITS: u32 = 3; |
| 107 |
|
| 108 |
// Cast x into an integer to manipulate its exponent and fractional parts into indexes for |
| 109 |
// lookup tables. |
| 110 |
let bits = x.to_bits(); |
| 111 |
|
| 112 |
// Get the integer log2 from the exponent part of bits |
| 113 |
let log2 = (bits >> 23) as i32 - 0x7f; |
| 114 |
|
| 115 |
// x is always >= (10/255 + A)*D so we only have to deal with a limited range in the exponent. |
| 116 |
// log2 range is [-4, 3] |
| 117 |
// Use a lookup table to offset for dividing by 2^log of x. |
| 118 |
// x^2.4 = (2^log2)^2.4 * (x/(2^log2))^2.4 |
| 119 |
let lookup_entry_exp_pow_2_4 = |
| 120 |
|log2: i32| (f32::from_bits(((log2 + 0x7f) << 23) as u32) as f64).powf(2.4) as f32; |
| 121 |
let lookup_table_exp_pow_2_4 = lookup_table_8!(start: -4, closure: lookup_entry_exp_pow_2_4); |
| 122 |
let exp_pow_2_4 = lookup_table_exp_pow_2_4[(log2 + 4) as usize]; |
| 123 |
|
| 124 |
// Zero the exponent of x or divide by 2^log. |
| 125 |
let x = f32::from_bits((bits & 0x807fffff) | 0x3f800000); |
| 126 |
|
| 127 |
// Use lookup tables to divide by a truncated version of x and get an offset for that division. |
| 128 |
// x^2.4 = a^2.4 * (x/a)^2.4 |
| 129 |
let lookup_entry_inv_truncated = |fraction: i32| { |
| 130 |
let truncated = 1.0 + (fraction as f64 + 0.5) / ((1 << FRAC_BITS) as f64); |
| 131 |
(1.0 / truncated) as f32 |
| 132 |
}; |
| 133 |
let lookup_table_inv_truncated = lookup_table_8!(start: 0, closure: lookup_entry_inv_truncated); |
| 134 |
let lookup_entry_truncated_pow_2_4 = |
| 135 |
|fraction: i32| (lookup_entry_inv_truncated(fraction) as f64).powf(-2.4) as f32; |
| 136 |
let lookup_table_truncated_pow_2_4 = |
| 137 |
lookup_table_8!(start: 0, closure: lookup_entry_truncated_pow_2_4); |
| 138 |
|
| 139 |
// Expose only FRAC_BITS of the fraction. |
| 140 |
let fraction = (bits >> (23 - FRAC_BITS) & ((1 << FRAC_BITS) - 1)) as usize; |
| 141 |
let truncated_pow_2_4 = lookup_table_truncated_pow_2_4[fraction]; |
| 142 |
let x = x * lookup_table_inv_truncated[fraction]; |
| 143 |
|
| 144 |
// Binomial series |
| 145 |
// Greater than 12 bits of precision. |
| 146 |
//let est = 7. / 25. - 24. / 25. * x + 42. / 25. * x.powi(2); |
| 147 |
// Plenty of precision. |
| 148 |
let est = 7. / 125. - 36. / 125. * x + 126. / 125. * x.powi(2) + 28. / 125. * x.powi(3); |
| 149 |
|
| 150 |
est * (truncated_pow_2_4 * exp_pow_2_4) |
| 151 |
} |
| 152 |
|
| 153 |
fn cbrt_approx(x: f32) -> f32 { |
| 154 |
// Closely approximate x^(1/3). |
| 155 |
// Divide x by its exponent and a truncated version of itself to get it as close to 1 as |
| 156 |
// possible. Calculate the power of 1/3 using the binomial method. Multiply what was divided to |
| 157 |
// the power of 1/3. |
| 158 |
|
| 159 |
// Lookup tables still have to be hardcoded. |
| 160 |
const FRAC_BITS: u32 = 3; |
| 161 |
|
| 162 |
// Cast x into an integer to manipulate its exponent and fractional parts into indexes for |
| 163 |
// lookup tables. |
| 164 |
let bits = x.to_bits(); |
| 165 |
|
| 166 |
// Get the integer log2 from the exponent part of bits |
| 167 |
let log2 = (bits >> 23) as i32 - 0x7f; |
| 168 |
|
| 169 |
// x is always > EPSILON so we only have to deal with a limited range in the exponent. |
| 170 |
// log2 range is [-7, 8] |
| 171 |
// Use a lookup table to offset for dividing by 2^log of x. |
| 172 |
// x^(1/3) = (2^log2)^(1/3) * (x/(2^log2))^(1/3) |
| 173 |
let lookup_entry_exp_cbrt = |
| 174 |
|log2: i32| (f32::from_bits(((log2 + 0x7f) << 23) as u32) as f64).powf(1. / 3.) as f32; |
| 175 |
let lookup_table_exp_cbrt = lookup_table_16!(start: -7, closure: lookup_entry_exp_cbrt); |
| 176 |
let exp_pow_cbrt = lookup_table_exp_cbrt[(log2 + 7) as usize]; |
| 177 |
|
| 178 |
// Zero the exponent of x or divide by 2^log. |
| 179 |
let x = f32::from_bits((bits & 0x807fffff) | 0x3f800000); |
| 180 |
|
| 181 |
// Use lookup tables to divide by a truncated version of x and get an offset for that division. |
| 182 |
// x^(1/3) = a^(1/3) * (x/a)^(1/3) |
| 183 |
let lookup_entry_inv_truncated = |fraction: i32| { |
| 184 |
let truncated = 1.0 + (fraction as f64 + 0.5) / ((1 << FRAC_BITS) as f64); |
| 185 |
(1.0 / truncated) as f32 |
| 186 |
}; |
| 187 |
let lookup_table_inv_truncated = lookup_table_8!(start: 0, closure: lookup_entry_inv_truncated); |
| 188 |
let lookup_entry_truncated_cbrt = |
| 189 |
|fraction: i32| (lookup_entry_inv_truncated(fraction) as f64).powf(-1. / 3.) as f32; |
| 190 |
let lookup_table_truncated_cbrt = |
| 191 |
lookup_table_8!(start: 0, closure: lookup_entry_truncated_cbrt); |
| 192 |
|
| 193 |
// Expose only FRAC_BITS of the fraction. |
| 194 |
let fraction = (bits >> (23 - FRAC_BITS) & ((1 << FRAC_BITS) - 1)) as usize; |
| 195 |
let truncated_pow_cbrt = lookup_table_truncated_cbrt[fraction]; |
| 196 |
let x = x * lookup_table_inv_truncated[fraction]; |
| 197 |
|
| 198 |
// Binomial series |
| 199 |
let est = 40. / 81. + 60. / 81. * x - 24. / 81. * x.powi(2) + 5. / 81. * x.powi(3); |
| 200 |
|
| 201 |
est * (truncated_pow_cbrt * exp_pow_cbrt) |
| 202 |
} |
| 203 |
|
| 204 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 205 |
pub use self::avx2::*; |
| 206 |
|
| 207 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 208 |
mod avx2 { |
| 209 |
use super::*; |
| 210 |
|
| 211 |
#[cfg(target_arch = "x86")] |
| 212 |
use std::arch::x86::*; |
| 213 |
#[cfg(target_arch = "x86_64")] |
| 214 |
use std::arch::x86_64::*; |
| 215 |
|
| 216 |
macro_rules! lookup_table_8_avx2 { |
| 217 |
(start: $start:expr, closure: $closure:expr) => { |
| 218 |
_mm256_setr_ps( |
| 219 |
$closure($start + 0), |
| 220 |
$closure($start + 1), |
| 221 |
$closure($start + 2), |
| 222 |
$closure($start + 3), |
| 223 |
$closure($start + 4), |
| 224 |
$closure($start + 5), |
| 225 |
$closure($start + 6), |
| 226 |
$closure($start + 7), |
| 227 |
) |
| 228 |
}; |
| 229 |
} |
| 230 |
|
| 231 |
macro_rules! lookup_table_16_avx2 { |
| 232 |
(start: $start:expr, closure: $closure:expr) => { |
| 233 |
( |
| 234 |
lookup_table_8_avx2!(start: $start, closure: $closure), |
| 235 |
lookup_table_8_avx2!(start: $start + 8, closure: $closure), |
| 236 |
) |
| 237 |
}; |
| 238 |
} |
| 239 |
|
| 240 |
macro_rules! sum_mult_avx { |
| 241 |
(($init:expr), $(($vec:expr, $mul:expr)),* ) => { |
| 242 |
{ |
| 243 |
let mut sum = _mm256_set1_ps($init); |
| 244 |
$( |
| 245 |
sum = _mm256_add_ps(sum, _mm256_mul_ps($vec, _mm256_set1_ps($mul))); |
| 246 |
)* |
| 247 |
sum |
| 248 |
} |
| 249 |
}; |
| 250 |
( $(($vec:expr, $mul:expr)),* ) => { |
| 251 |
sum_mult_avx!((0.0), $(($vec, $mul)),*) |
| 252 |
}; |
| 253 |
} |
| 254 |
|
| 255 |
#[target_feature(enable = "avx2")] |
| 256 |
pub unsafe fn rgb_to_lab_avx2(rgb: &[__m256; 3]) -> [Lab; 8] { |
| 257 |
xyz_to_lab_avx2(rgb_to_xyz_avx2(rgb)) |
| 258 |
} |
| 259 |
|
| 260 |
#[target_feature(enable = "avx2")] |
| 261 |
#[allow(clippy::excessive_precision)] |
| 262 |
#[allow(clippy::many_single_char_names)] |
| 263 |
unsafe fn rgb_to_xyz_avx2(rgb: &[__m256; 3]) -> [__m256; 3] { |
| 264 |
let r = rgb_to_xyz_map_avx2(rgb[0]); |
| 265 |
let g = rgb_to_xyz_map_avx2(rgb[1]); |
| 266 |
let b = rgb_to_xyz_map_avx2(rgb[2]); |
| 267 |
|
| 268 |
let x = sum_mult_avx!( |
| 269 |
(r, 0.4124564390896921), |
| 270 |
(g, 0.357576077643909), |
| 271 |
(b, 0.18043748326639894) |
| 272 |
); |
| 273 |
let y = sum_mult_avx!( |
| 274 |
(r, 0.21267285140562248), |
| 275 |
(g, 0.715152155287818), |
| 276 |
(b, 0.07217499330655958) |
| 277 |
); |
| 278 |
let z = sum_mult_avx!( |
| 279 |
(r, 0.019333895582329317), |
| 280 |
(g, 0.119192025881303), |
| 281 |
(b, 0.9503040785363677) |
| 282 |
); |
| 283 |
|
| 284 |
[x, y, z] |
| 285 |
} |
| 286 |
|
| 287 |
#[inline] |
| 288 |
#[target_feature(enable = "avx2")] |
| 289 |
unsafe fn rgb_to_xyz_map_avx2(c: __m256) -> __m256 { |
| 290 |
let low = _mm256_mul_ps(c, _mm256_set1_ps(1.0 / 12.92)); |
| 291 |
let hi = pow_2_4_avx2(_mm256_mul_ps( |
| 292 |
_mm256_add_ps(c, _mm256_set1_ps(0.055)), |
| 293 |
_mm256_set1_ps(1.0 / 1.055), |
| 294 |
)); |
| 295 |
let select = _mm256_cmp_ps(c, _mm256_set1_ps(10. / 255.), _CMP_GT_OS); |
| 296 |
_mm256_blendv_ps(low, hi, select) |
| 297 |
} |
| 298 |
|
| 299 |
#[inline] |
| 300 |
#[target_feature(enable = "avx2")] |
| 301 |
#[allow(clippy::many_single_char_names)] |
| 302 |
unsafe fn xyz_to_lab_avx2(xyz: [__m256; 3]) -> [Lab; 8] { |
| 303 |
let x = xyz_to_lab_map_avx2(_mm256_mul_ps(xyz[0], _mm256_set1_ps(1.0 / 0.95047))); |
| 304 |
let y = xyz_to_lab_map_avx2(xyz[1]); |
| 305 |
let z = xyz_to_lab_map_avx2(_mm256_mul_ps(xyz[2], _mm256_set1_ps(1.0 / 1.08883))); |
| 306 |
|
| 307 |
let l = _mm256_sub_ps( |
| 308 |
_mm256_mul_ps(_mm256_set1_ps(116.0), y), |
| 309 |
_mm256_set1_ps(16.0), |
| 310 |
); |
| 311 |
let a = _mm256_mul_ps(_mm256_sub_ps(x, y), _mm256_set1_ps(500.0)); |
| 312 |
let b = _mm256_mul_ps(_mm256_sub_ps(y, z), _mm256_set1_ps(200.0)); |
| 313 |
|
| 314 |
#[target_feature(enable = "avx2")] |
| 315 |
unsafe fn to_array(reg: __m256) -> [f32; 8] { |
| 316 |
std::mem::transmute(reg) |
| 317 |
} |
| 318 |
let l = to_array(l); |
| 319 |
let a = to_array(a); |
| 320 |
let b = to_array(b); |
| 321 |
|
| 322 |
let mut output = [Lab { |
| 323 |
l: 0., |
| 324 |
a: 0., |
| 325 |
b: 0., |
| 326 |
}; 8]; |
| 327 |
for i in 0..8 { |
| 328 |
output[i] = Lab { |
| 329 |
l: l[i], |
| 330 |
a: a[i], |
| 331 |
b: b[i], |
| 332 |
}; |
| 333 |
} |
| 334 |
output |
| 335 |
} |
| 336 |
|
| 337 |
#[inline] |
| 338 |
#[target_feature(enable = "avx2")] |
| 339 |
unsafe fn xyz_to_lab_map_avx2(c: __m256) -> __m256 { |
| 340 |
let low = _mm256_mul_ps( |
| 341 |
_mm256_add_ps( |
| 342 |
_mm256_mul_ps(c, _mm256_set1_ps(KAPPA)), |
| 343 |
_mm256_set1_ps(16.0), |
| 344 |
), |
| 345 |
_mm256_set1_ps(1.0 / 116.0), |
| 346 |
); |
| 347 |
let hi = cbrt_approx_avx2(c); |
| 348 |
let select = _mm256_cmp_ps(c, _mm256_set1_ps(EPSILON), _CMP_GT_OS); |
| 349 |
_mm256_blendv_ps(low, hi, select) |
| 350 |
} |
| 351 |
|
| 352 |
#[target_feature(enable = "avx2")] |
| 353 |
unsafe fn pow_2_4_avx2(x: __m256) -> __m256 { |
| 354 |
// See non-avx2 version |
| 355 |
|
| 356 |
const FRAC_BITS: u32 = 3; |
| 357 |
|
| 358 |
let bits = _mm256_castps_si256(x); |
| 359 |
|
| 360 |
let log2_index = |
| 361 |
_mm256_add_epi32(_mm256_srli_epi32(bits, 23), _mm256_set1_epi32(-0x7f + 4)); |
| 362 |
|
| 363 |
let lookup_entry_exp_pow_2_4 = |
| 364 |
|log2: i32| (f32::from_bits(((log2 + 0x7f) << 23) as u32) as f64).powf(2.4) as f32; |
| 365 |
let lookup_table_exp_pow_2_4 = |
| 366 |
lookup_table_8_avx2!(start: -4, closure: lookup_entry_exp_pow_2_4); |
| 367 |
|
| 368 |
let exp_pow_2_4 = _mm256_permutevar8x32_ps(lookup_table_exp_pow_2_4, log2_index); |
| 369 |
|
| 370 |
let x = _mm256_or_ps( |
| 371 |
_mm256_and_ps( |
| 372 |
x, |
| 373 |
_mm256_castsi256_ps(_mm256_set1_epi32(0x807fffffu32 as i32)), |
| 374 |
), |
| 375 |
_mm256_castsi256_ps(_mm256_set1_epi32(0x3f800000)), |
| 376 |
); |
| 377 |
|
| 378 |
let lookup_entry_inv_truncated = |fraction: i32| { |
| 379 |
let truncated = 1.0 + (fraction as f64 + 0.5) / ((1 << FRAC_BITS) as f64); |
| 380 |
(1.0 / truncated) as f32 |
| 381 |
}; |
| 382 |
let lookup_table_inv_truncated = |
| 383 |
lookup_table_8_avx2!(start: 0, closure: lookup_entry_inv_truncated); |
| 384 |
let lookup_entry_truncated_pow_2_4 = |
| 385 |
|fraction: i32| (lookup_entry_inv_truncated(fraction) as f64).powf(-2.4) as f32; |
| 386 |
let lookup_table_truncated_pow_2_4 = |
| 387 |
lookup_table_8_avx2!(start: 0, closure: lookup_entry_truncated_pow_2_4); |
| 388 |
|
| 389 |
// No reason to mask the higher bits |
| 390 |
let fraction = _mm256_srli_epi32(bits, 23 - FRAC_BITS as i32); |
| 391 |
let truncated_pow_2_4 = _mm256_permutevar8x32_ps(lookup_table_truncated_pow_2_4, fraction); |
| 392 |
let x = _mm256_mul_ps( |
| 393 |
x, |
| 394 |
_mm256_permutevar8x32_ps(lookup_table_inv_truncated, fraction), |
| 395 |
); |
| 396 |
|
| 397 |
let x2 = _mm256_mul_ps(x, x); |
| 398 |
let x3 = _mm256_mul_ps(x2, x); |
| 399 |
let est = sum_mult_avx!( |
| 400 |
(7.0 / 125.0), |
| 401 |
(x, -36. / 125.), |
| 402 |
(x2, 126. / 125.), |
| 403 |
(x3, 28. / 125.) |
| 404 |
); |
| 405 |
|
| 406 |
_mm256_mul_ps(est, _mm256_mul_ps(truncated_pow_2_4, exp_pow_2_4)) |
| 407 |
} |
| 408 |
|
| 409 |
#[target_feature(enable = "avx2")] |
| 410 |
unsafe fn cbrt_approx_avx2(x: __m256) -> __m256 { |
| 411 |
// See non-avx2 version |
| 412 |
|
| 413 |
const FRAC_BITS: u32 = 3; |
| 414 |
|
| 415 |
let bits = _mm256_castps_si256(x); |
| 416 |
|
| 417 |
let log2_index = |
| 418 |
_mm256_add_epi32(_mm256_srli_epi32(bits, 23), _mm256_set1_epi32(-0x7f + 7)); |
| 419 |
|
| 420 |
let lookup_entry_exp_cbrt = |
| 421 |
|log2: i32| (f32::from_bits(((log2 + 0x7f) << 23) as u32) as f64).powf(1. / 3.) as f32; |
| 422 |
let lookup_table_exp_cbrt = |
| 423 |
lookup_table_16_avx2!(start: -7, closure: lookup_entry_exp_cbrt); |
| 424 |
|
| 425 |
let exp_cbrt = _mm256_blendv_ps( |
| 426 |
_mm256_permutevar8x32_ps(lookup_table_exp_cbrt.0, log2_index), |
| 427 |
_mm256_permutevar8x32_ps(lookup_table_exp_cbrt.1, log2_index), |
| 428 |
// Check if log is greater than 7 |
| 429 |
_mm256_castsi256_ps(_mm256_slli_epi32(log2_index, 28)), |
| 430 |
); |
| 431 |
|
| 432 |
let x = _mm256_or_ps( |
| 433 |
_mm256_and_ps( |
| 434 |
x, |
| 435 |
_mm256_castsi256_ps(_mm256_set1_epi32(0x807fffffu32 as i32)), |
| 436 |
), |
| 437 |
_mm256_castsi256_ps(_mm256_set1_epi32(0x3f800000)), |
| 438 |
); |
| 439 |
let lookup_entry_inv_truncated = |fraction: i32| { |
| 440 |
let truncated = 1.0 + (fraction as f64 + 0.5) / ((1 << FRAC_BITS) as f64); |
| 441 |
(1.0 / truncated) as f32 |
| 442 |
}; |
| 443 |
let lookup_table_inv_truncated = |
| 444 |
lookup_table_8_avx2!(start: 0, closure: lookup_entry_inv_truncated); |
| 445 |
let lookup_entry_truncated_cbrt = |
| 446 |
|fraction: i32| (lookup_entry_inv_truncated(fraction) as f64).powf(-1. / 3.) as f32; |
| 447 |
let lookup_table_truncated_cbrt = |
| 448 |
lookup_table_8_avx2!(start: 0, closure: lookup_entry_truncated_cbrt); |
| 449 |
|
| 450 |
// No reason to mask the higher bits |
| 451 |
let fraction = _mm256_srli_epi32(bits, 23 - FRAC_BITS as i32); |
| 452 |
let truncated_cbrt = _mm256_permutevar8x32_ps(lookup_table_truncated_cbrt, fraction); |
| 453 |
let x = _mm256_mul_ps( |
| 454 |
x, |
| 455 |
_mm256_permutevar8x32_ps(lookup_table_inv_truncated, fraction), |
| 456 |
); |
| 457 |
|
| 458 |
let x2 = _mm256_mul_ps(x, x); |
| 459 |
let x3 = _mm256_mul_ps(x2, x); |
| 460 |
let est = sum_mult_avx!( |
| 461 |
(40. / 81.0), |
| 462 |
(x, 60. / 81.), |
| 463 |
(x2, -24. / 81.), |
| 464 |
(x3, 5. / 81.) |
| 465 |
); |
| 466 |
|
| 467 |
_mm256_mul_ps(est, _mm256_mul_ps(truncated_cbrt, exp_cbrt)) |
| 468 |
} |
| 469 |
} |