| 1 |
// Copyright (c) 2019-2022, The rav1e contributors. All rights reserved |
| 2 |
// |
| 3 |
// This source code is subject to the terms of the BSD 2 Clause License and |
| 4 |
// the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
| 5 |
// was not distributed with this source code in the LICENSE file, you can |
| 6 |
// obtain it at www.aomedia.org/license/software. If the Alliance for Open |
| 7 |
// Media Patent License 1.0 was not distributed with this source code in the |
| 8 |
// PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
| 9 |
|
| 10 |
use crate::cpu_features::CpuFeatureLevel; |
| 11 |
use crate::frame::PlaneSlice; |
| 12 |
use crate::lrf::*; |
| 13 |
use crate::util::Pixel; |
| 14 |
#[cfg(target_arch = "x86")] |
| 15 |
use std::arch::x86::*; |
| 16 |
#[cfg(target_arch = "x86_64")] |
| 17 |
use std::arch::x86_64::*; |
| 18 |
use std::mem; |
| 19 |
|
| 20 |
// computes an intermediate (ab) row for stripe_w + 2 columns at row y |
| 21 |
#[inline] |
| 22 |
pub fn sgrproj_box_ab_r1<const BD: usize>( |
| 23 |
af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32], |
| 24 |
iimg_stride: usize, y: usize, stripe_w: usize, s: u32, cpu: CpuFeatureLevel, |
| 25 |
) { |
| 26 |
// only use 8-bit AVX2 assembly when bitdepth equals 8 |
| 27 |
if cpu >= CpuFeatureLevel::AVX2 && BD == 8 { |
| 28 |
// SAFETY: Calls Assembly code. |
| 29 |
return unsafe { |
| 30 |
sgrproj_box_ab_r1_avx2::<BD>( |
| 31 |
af, |
| 32 |
bf, |
| 33 |
iimg, |
| 34 |
iimg_sq, |
| 35 |
iimg_stride, |
| 36 |
y, |
| 37 |
stripe_w, |
| 38 |
s, |
| 39 |
); |
| 40 |
}; |
| 41 |
} |
| 42 |
|
| 43 |
rust::sgrproj_box_ab_r1::<BD>( |
| 44 |
af, |
| 45 |
bf, |
| 46 |
iimg, |
| 47 |
iimg_sq, |
| 48 |
iimg_stride, |
| 49 |
y, |
| 50 |
stripe_w, |
| 51 |
s, |
| 52 |
cpu, |
| 53 |
); |
| 54 |
} |
| 55 |
|
| 56 |
// computes an intermediate (ab) row for stripe_w + 2 columns at row y |
| 57 |
#[inline] |
| 58 |
pub fn sgrproj_box_ab_r2<const BD: usize>( |
| 59 |
af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32], |
| 60 |
iimg_stride: usize, y: usize, stripe_w: usize, s: u32, cpu: CpuFeatureLevel, |
| 61 |
) { |
| 62 |
// only use 8-bit AVX2 assembly when bitdepth equals 8 |
| 63 |
if cpu >= CpuFeatureLevel::AVX2 && BD == 8 { |
| 64 |
// SAFETY: Calls Assembly code. |
| 65 |
return unsafe { |
| 66 |
sgrproj_box_ab_r2_avx2::<BD>( |
| 67 |
af, |
| 68 |
bf, |
| 69 |
iimg, |
| 70 |
iimg_sq, |
| 71 |
iimg_stride, |
| 72 |
y, |
| 73 |
stripe_w, |
| 74 |
s, |
| 75 |
); |
| 76 |
}; |
| 77 |
} |
| 78 |
|
| 79 |
rust::sgrproj_box_ab_r2::<BD>( |
| 80 |
af, |
| 81 |
bf, |
| 82 |
iimg, |
| 83 |
iimg_sq, |
| 84 |
iimg_stride, |
| 85 |
y, |
| 86 |
stripe_w, |
| 87 |
s, |
| 88 |
cpu, |
| 89 |
); |
| 90 |
} |
| 91 |
|
| 92 |
#[inline] |
| 93 |
pub fn sgrproj_box_f_r0<T: Pixel>( |
| 94 |
f: &mut [u32], y: usize, w: usize, cdeffed: &PlaneSlice<T>, |
| 95 |
cpu: CpuFeatureLevel, |
| 96 |
) { |
| 97 |
if cpu >= CpuFeatureLevel::AVX2 { |
| 98 |
// SAFETY: Calls Assembly code. |
| 99 |
return unsafe { |
| 100 |
sgrproj_box_f_r0_avx2(f, y, w, cdeffed); |
| 101 |
}; |
| 102 |
} |
| 103 |
|
| 104 |
rust::sgrproj_box_f_r0(f, y, w, cdeffed, cpu); |
| 105 |
} |
| 106 |
|
| 107 |
#[inline] |
| 108 |
pub fn sgrproj_box_f_r1<T: Pixel>( |
| 109 |
af: &[&[u32]; 3], bf: &[&[u32]; 3], f: &mut [u32], y: usize, w: usize, |
| 110 |
cdeffed: &PlaneSlice<T>, cpu: CpuFeatureLevel, |
| 111 |
) { |
| 112 |
if cpu >= CpuFeatureLevel::AVX2 { |
| 113 |
// SAFETY: Calls Assembly code. |
| 114 |
return unsafe { |
| 115 |
sgrproj_box_f_r1_avx2(af, bf, f, y, w, cdeffed); |
| 116 |
}; |
| 117 |
} |
| 118 |
|
| 119 |
rust::sgrproj_box_f_r1(af, bf, f, y, w, cdeffed, cpu); |
| 120 |
} |
| 121 |
|
| 122 |
#[inline] |
| 123 |
pub fn sgrproj_box_f_r2<T: Pixel>( |
| 124 |
af: &[&[u32]; 2], bf: &[&[u32]; 2], f0: &mut [u32], f1: &mut [u32], |
| 125 |
y: usize, w: usize, cdeffed: &PlaneSlice<T>, cpu: CpuFeatureLevel, |
| 126 |
) { |
| 127 |
if cpu >= CpuFeatureLevel::AVX2 { |
| 128 |
// SAFETY: Calls Assembly code. |
| 129 |
return unsafe { |
| 130 |
sgrproj_box_f_r2_avx2(af, bf, f0, f1, y, w, cdeffed); |
| 131 |
}; |
| 132 |
} |
| 133 |
|
| 134 |
rust::sgrproj_box_f_r2(af, bf, f0, f1, y, w, cdeffed, cpu); |
| 135 |
} |
| 136 |
|
| 137 |
static X_BY_XPLUS1: [u32; 256] = [ |
| 138 |
// Special case: Map 0 -> 1 (corresponding to a value of 1/256) |
| 139 |
// instead of 0. See comments in selfguided_restoration_internal() for why |
| 140 |
1, 128, 171, 192, 205, 213, 219, 224, 228, 230, 233, 235, 236, 238, 239, 240, |
| 141 |
241, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247, 247, 247, 247, 248, |
| 142 |
248, 248, 248, 249, 249, 249, 249, 249, 250, 250, 250, 250, 250, 250, 250, |
| 143 |
251, 251, 251, 251, 251, 251, 251, 251, 251, 251, 252, 252, 252, 252, 252, |
| 144 |
252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 253, 253, 253, |
| 145 |
253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, |
| 146 |
253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 254, 254, 254, 254, |
| 147 |
254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, |
| 148 |
254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, |
| 149 |
254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, |
| 150 |
254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, |
| 151 |
254, 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 152 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 153 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 154 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 155 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 156 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 256, |
| 157 |
]; |
| 158 |
|
| 159 |
#[inline] |
| 160 |
#[target_feature(enable = "avx2")] |
| 161 |
unsafe fn sgrproj_box_ab_8_avx2<const BD: usize>( |
| 162 |
r: usize, af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32], |
| 163 |
iimg_stride: usize, x: usize, y: usize, s: u32, |
| 164 |
) { |
| 165 |
let bdm8 = BD - 8; |
| 166 |
let d: usize = r * 2 + 1; |
| 167 |
let n: i32 = (d * d) as i32; |
| 168 |
let one_over_n = if r == 1 { 455 } else { 164 }; |
| 169 |
|
| 170 |
// Using an integral image, compute the sum of a square region |
| 171 |
#[inline] |
| 172 |
#[target_feature(enable = "avx2")] |
| 173 |
unsafe fn get_integral_square_avx2( |
| 174 |
iimg: &[u32], stride: usize, x: usize, y: usize, size: usize, |
| 175 |
) -> __m256i { |
| 176 |
let iimg = iimg.as_ptr().add(y * stride + x); |
| 177 |
// Cancel out overflow in iimg by using wrapping arithmetic |
| 178 |
_mm256_sub_epi32( |
| 179 |
_mm256_add_epi32( |
| 180 |
_mm256_loadu_si256(iimg as *const _), |
| 181 |
_mm256_loadu_si256(iimg.add(size * stride + size) as *const _), |
| 182 |
), |
| 183 |
_mm256_add_epi32( |
| 184 |
_mm256_loadu_si256(iimg.add(size * stride) as *const _), |
| 185 |
_mm256_loadu_si256(iimg.add(size) as *const _), |
| 186 |
), |
| 187 |
) |
| 188 |
} |
| 189 |
|
| 190 |
let sum = get_integral_square_avx2(iimg, iimg_stride, x, y, d); |
| 191 |
let ssq = get_integral_square_avx2(iimg_sq, iimg_stride, x, y, d); |
| 192 |
let scaled_sum = _mm256_srlv_epi32( |
| 193 |
_mm256_add_epi32(sum, _mm256_set1_epi32(1 << bdm8 as i32 >> 1)), |
| 194 |
_mm256_set1_epi32(bdm8 as i32), |
| 195 |
); |
| 196 |
let scaled_ssq = _mm256_srlv_epi32( |
| 197 |
_mm256_add_epi32(ssq, _mm256_set1_epi32(1 << (2 * bdm8) as i32 >> 1)), |
| 198 |
_mm256_set1_epi32(2 * bdm8 as i32), |
| 199 |
); |
| 200 |
let p = _mm256_max_epi32( |
| 201 |
_mm256_setzero_si256(), |
| 202 |
_mm256_sub_epi32( |
| 203 |
_mm256_mullo_epi32(scaled_ssq, _mm256_set1_epi32(n)), |
| 204 |
_mm256_madd_epi16(scaled_sum, scaled_sum), |
| 205 |
), |
| 206 |
); |
| 207 |
let z = _mm256_srli_epi32( |
| 208 |
_mm256_add_epi32( |
| 209 |
_mm256_mullo_epi32(p, _mm256_set1_epi32(s as i32)), |
| 210 |
_mm256_set1_epi32(1 << SGRPROJ_MTABLE_BITS as i32 >> 1), |
| 211 |
), |
| 212 |
SGRPROJ_MTABLE_BITS as i32, |
| 213 |
); |
| 214 |
let a = _mm256_i32gather_epi32( |
| 215 |
X_BY_XPLUS1.as_ptr() as *const _, |
| 216 |
_mm256_min_epi32(z, _mm256_set1_epi32(255)), |
| 217 |
4, |
| 218 |
); |
| 219 |
let b = _mm256_mullo_epi32( |
| 220 |
_mm256_madd_epi16( |
| 221 |
_mm256_sub_epi32(_mm256_set1_epi32(1 << SGRPROJ_SGR_BITS as i32), a), |
| 222 |
sum, |
| 223 |
), |
| 224 |
_mm256_set1_epi32(one_over_n), |
| 225 |
); |
| 226 |
let b = _mm256_srlv_epi32( |
| 227 |
_mm256_add_epi32( |
| 228 |
b, |
| 229 |
_mm256_set1_epi32(1 << SGRPROJ_RECIP_BITS as i32 >> 1), |
| 230 |
), |
| 231 |
_mm256_set1_epi32(SGRPROJ_RECIP_BITS as i32), |
| 232 |
); |
| 233 |
_mm256_storeu_si256(af.as_mut_ptr().add(x) as *mut _, a); |
| 234 |
_mm256_storeu_si256(bf.as_mut_ptr().add(x) as *mut _, b); |
| 235 |
} |
| 236 |
|
| 237 |
#[target_feature(enable = "avx2")] |
| 238 |
pub(crate) unsafe fn sgrproj_box_ab_r1_avx2<const BD: usize>( |
| 239 |
af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32], |
| 240 |
iimg_stride: usize, y: usize, stripe_w: usize, s: u32, |
| 241 |
) { |
| 242 |
for x in (0..stripe_w + 2).step_by(8) { |
| 243 |
if x + 8 <= stripe_w + 2 { |
| 244 |
sgrproj_box_ab_8_avx2::<BD>( |
| 245 |
1, |
| 246 |
af, |
| 247 |
bf, |
| 248 |
iimg, |
| 249 |
iimg_sq, |
| 250 |
iimg_stride, |
| 251 |
x, |
| 252 |
y, |
| 253 |
s, |
| 254 |
); |
| 255 |
} else { |
| 256 |
// finish using scalar |
| 257 |
rust::sgrproj_box_ab_internal::<BD>( |
| 258 |
1, |
| 259 |
af, |
| 260 |
bf, |
| 261 |
iimg, |
| 262 |
iimg_sq, |
| 263 |
iimg_stride, |
| 264 |
x, |
| 265 |
y, |
| 266 |
stripe_w, |
| 267 |
s, |
| 268 |
); |
| 269 |
} |
| 270 |
} |
| 271 |
|
| 272 |
#[cfg(feature = "check_asm")] |
| 273 |
{ |
| 274 |
let mut af_ref: Vec<u32> = vec![0; stripe_w + 2]; |
| 275 |
let mut bf_ref: Vec<u32> = vec![0; stripe_w + 2]; |
| 276 |
rust::sgrproj_box_ab_internal::<BD>( |
| 277 |
1, |
| 278 |
&mut af_ref, |
| 279 |
&mut bf_ref, |
| 280 |
iimg, |
| 281 |
iimg_sq, |
| 282 |
iimg_stride, |
| 283 |
0, |
| 284 |
y, |
| 285 |
stripe_w, |
| 286 |
s, |
| 287 |
); |
| 288 |
assert_eq!(&af[..stripe_w + 2], &af_ref[..]); |
| 289 |
assert_eq!(&bf[..stripe_w + 2], &bf_ref[..]); |
| 290 |
} |
| 291 |
} |
| 292 |
|
| 293 |
#[target_feature(enable = "avx2")] |
| 294 |
pub(crate) unsafe fn sgrproj_box_ab_r2_avx2<const BD: usize>( |
| 295 |
af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32], |
| 296 |
iimg_stride: usize, y: usize, stripe_w: usize, s: u32, |
| 297 |
) { |
| 298 |
for x in (0..stripe_w + 2).step_by(8) { |
| 299 |
if x + 8 <= stripe_w + 2 { |
| 300 |
sgrproj_box_ab_8_avx2::<BD>( |
| 301 |
2, |
| 302 |
af, |
| 303 |
bf, |
| 304 |
iimg, |
| 305 |
iimg_sq, |
| 306 |
iimg_stride, |
| 307 |
x, |
| 308 |
y, |
| 309 |
s, |
| 310 |
); |
| 311 |
} else { |
| 312 |
// finish using scalar |
| 313 |
rust::sgrproj_box_ab_internal::<BD>( |
| 314 |
2, |
| 315 |
af, |
| 316 |
bf, |
| 317 |
iimg, |
| 318 |
iimg_sq, |
| 319 |
iimg_stride, |
| 320 |
x, |
| 321 |
y, |
| 322 |
stripe_w, |
| 323 |
s, |
| 324 |
); |
| 325 |
} |
| 326 |
} |
| 327 |
|
| 328 |
#[cfg(feature = "check_asm")] |
| 329 |
{ |
| 330 |
let mut af_ref: Vec<u32> = vec![0; stripe_w + 2]; |
| 331 |
let mut bf_ref: Vec<u32> = vec![0; stripe_w + 2]; |
| 332 |
rust::sgrproj_box_ab_internal::<BD>( |
| 333 |
2, |
| 334 |
&mut af_ref, |
| 335 |
&mut bf_ref, |
| 336 |
iimg, |
| 337 |
iimg_sq, |
| 338 |
iimg_stride, |
| 339 |
0, |
| 340 |
y, |
| 341 |
stripe_w, |
| 342 |
s, |
| 343 |
); |
| 344 |
assert_eq!(&af[..stripe_w + 2], &af_ref[..]); |
| 345 |
assert_eq!(&bf[..stripe_w + 2], &bf_ref[..]); |
| 346 |
} |
| 347 |
} |
| 348 |
|
| 349 |
#[inline] |
| 350 |
#[target_feature(enable = "avx2")] |
| 351 |
unsafe fn sgrproj_box_f_r0_8_avx2<T: Pixel>( |
| 352 |
f: &mut [u32], x: usize, y: usize, cdeffed: &PlaneSlice<T>, |
| 353 |
) { |
| 354 |
_mm256_storeu_si256( |
| 355 |
f.as_mut_ptr().add(x) as *mut _, |
| 356 |
_mm256_slli_epi32( |
| 357 |
if mem::size_of::<T>() == 1 { |
| 358 |
_mm256_cvtepu8_epi32(_mm_loadl_epi64( |
| 359 |
cdeffed.subslice(x, y).as_ptr() as *const _ |
| 360 |
)) |
| 361 |
} else { |
| 362 |
_mm256_cvtepu16_epi32(_mm_loadu_si128( |
| 363 |
cdeffed.subslice(x, y).as_ptr() as *const _ |
| 364 |
)) |
| 365 |
}, |
| 366 |
SGRPROJ_RST_BITS as i32, |
| 367 |
), |
| 368 |
); |
| 369 |
} |
| 370 |
|
| 371 |
#[target_feature(enable = "avx2")] |
| 372 |
pub(crate) unsafe fn sgrproj_box_f_r0_avx2<T: Pixel>( |
| 373 |
f: &mut [u32], y: usize, w: usize, cdeffed: &PlaneSlice<T>, |
| 374 |
) { |
| 375 |
for x in (0..w).step_by(8) { |
| 376 |
if x + 8 <= w { |
| 377 |
sgrproj_box_f_r0_8_avx2(f, x, y, cdeffed); |
| 378 |
} else { |
| 379 |
// finish using scalar |
| 380 |
rust::sgrproj_box_f_r0_internal(f, x, y, w, cdeffed); |
| 381 |
} |
| 382 |
} |
| 383 |
|
| 384 |
#[cfg(feature = "check_asm")] |
| 385 |
{ |
| 386 |
let mut f_ref: Vec<u32> = vec![0; w]; |
| 387 |
rust::sgrproj_box_f_r0_internal(&mut f_ref, 0, y, w, cdeffed); |
| 388 |
assert_eq!(&f[..w], &f_ref[..]); |
| 389 |
} |
| 390 |
} |
| 391 |
|
| 392 |
#[inline] |
| 393 |
#[target_feature(enable = "avx2")] |
| 394 |
unsafe fn sgrproj_box_f_r1_8_avx2<T: Pixel>( |
| 395 |
af: &[&[u32]; 3], bf: &[&[u32]; 3], f: &mut [u32], x: usize, y: usize, |
| 396 |
cdeffed: &PlaneSlice<T>, |
| 397 |
) { |
| 398 |
let three = _mm256_set1_epi32(3); |
| 399 |
let four = _mm256_set1_epi32(4); |
| 400 |
let a0 = af[0].as_ptr(); |
| 401 |
let a1 = af[1].as_ptr(); |
| 402 |
let a2 = af[2].as_ptr(); |
| 403 |
let b0 = bf[0].as_ptr(); |
| 404 |
let b1 = bf[1].as_ptr(); |
| 405 |
let b2 = bf[2].as_ptr(); |
| 406 |
let a = _mm256_add_epi32( |
| 407 |
_mm256_madd_epi16( |
| 408 |
_mm256_add_epi32( |
| 409 |
_mm256_add_epi32( |
| 410 |
_mm256_loadu_si256(a0.add(x) as *const _), |
| 411 |
_mm256_loadu_si256(a0.add(x + 2) as *const _), |
| 412 |
), |
| 413 |
_mm256_add_epi32( |
| 414 |
_mm256_loadu_si256(a2.add(x) as *const _), |
| 415 |
_mm256_loadu_si256(a2.add(x + 2) as *const _), |
| 416 |
), |
| 417 |
), |
| 418 |
three, |
| 419 |
), |
| 420 |
_mm256_madd_epi16( |
| 421 |
_mm256_add_epi32( |
| 422 |
_mm256_add_epi32( |
| 423 |
_mm256_loadu_si256(a1.add(x) as *const _), |
| 424 |
_mm256_loadu_si256(a0.add(x + 1) as *const _), |
| 425 |
), |
| 426 |
_mm256_add_epi32( |
| 427 |
_mm256_add_epi32( |
| 428 |
_mm256_loadu_si256(a1.add(x + 1) as *const _), |
| 429 |
_mm256_loadu_si256(a2.add(x + 1) as *const _), |
| 430 |
), |
| 431 |
_mm256_loadu_si256(a1.add(x + 2) as *const _), |
| 432 |
), |
| 433 |
), |
| 434 |
four, |
| 435 |
), |
| 436 |
); |
| 437 |
let b = _mm256_add_epi32( |
| 438 |
_mm256_mullo_epi32( |
| 439 |
_mm256_add_epi32( |
| 440 |
_mm256_add_epi32( |
| 441 |
_mm256_loadu_si256(b0.add(x) as *const _), |
| 442 |
_mm256_loadu_si256(b0.add(x + 2) as *const _), |
| 443 |
), |
| 444 |
_mm256_add_epi32( |
| 445 |
_mm256_loadu_si256(b2.add(x) as *const _), |
| 446 |
_mm256_loadu_si256(b2.add(x + 2) as *const _), |
| 447 |
), |
| 448 |
), |
| 449 |
three, |
| 450 |
), |
| 451 |
_mm256_mullo_epi32( |
| 452 |
_mm256_add_epi32( |
| 453 |
_mm256_add_epi32( |
| 454 |
_mm256_loadu_si256(b1.add(x) as *const _), |
| 455 |
_mm256_loadu_si256(b0.add(x + 1) as *const _), |
| 456 |
), |
| 457 |
_mm256_add_epi32( |
| 458 |
_mm256_add_epi32( |
| 459 |
_mm256_loadu_si256(b1.add(x + 1) as *const _), |
| 460 |
_mm256_loadu_si256(b2.add(x + 1) as *const _), |
| 461 |
), |
| 462 |
_mm256_loadu_si256(b1.add(x + 2) as *const _), |
| 463 |
), |
| 464 |
), |
| 465 |
four, |
| 466 |
), |
| 467 |
); |
| 468 |
let v = _mm256_add_epi32( |
| 469 |
_mm256_madd_epi16( |
| 470 |
a, |
| 471 |
if mem::size_of::<T>() == 1 { |
| 472 |
_mm256_cvtepu8_epi32(_mm_loadl_epi64( |
| 473 |
cdeffed.subslice(x, y).as_ptr() as *const _ |
| 474 |
)) |
| 475 |
} else { |
| 476 |
_mm256_cvtepu16_epi32(_mm_loadu_si128( |
| 477 |
cdeffed.subslice(x, y).as_ptr() as *const _ |
| 478 |
)) |
| 479 |
}, |
| 480 |
), |
| 481 |
b, |
| 482 |
); |
| 483 |
const SHIFT: i32 = (5 + SGRPROJ_SGR_BITS - SGRPROJ_RST_BITS) as i32; |
| 484 |
_mm256_storeu_si256( |
| 485 |
f.as_mut_ptr().add(x) as *mut _, |
| 486 |
_mm256_srli_epi32( |
| 487 |
_mm256_add_epi32(v, _mm256_set1_epi32(1 << SHIFT >> 1)), |
| 488 |
SHIFT, |
| 489 |
), |
| 490 |
); |
| 491 |
} |
| 492 |
|
| 493 |
#[target_feature(enable = "avx2")] |
| 494 |
pub(crate) unsafe fn sgrproj_box_f_r1_avx2<T: Pixel>( |
| 495 |
af: &[&[u32]; 3], bf: &[&[u32]; 3], f: &mut [u32], y: usize, w: usize, |
| 496 |
cdeffed: &PlaneSlice<T>, |
| 497 |
) { |
| 498 |
for x in (0..w).step_by(8) { |
| 499 |
if x + 8 <= w { |
| 500 |
sgrproj_box_f_r1_8_avx2(af, bf, f, x, y, cdeffed); |
| 501 |
} else { |
| 502 |
// finish using scalar |
| 503 |
rust::sgrproj_box_f_r1_internal(af, bf, f, x, y, w, cdeffed); |
| 504 |
} |
| 505 |
} |
| 506 |
|
| 507 |
#[cfg(feature = "check_asm")] |
| 508 |
{ |
| 509 |
let mut f_ref: Vec<u32> = vec![0; w]; |
| 510 |
rust::sgrproj_box_f_r1_internal(af, bf, &mut f_ref, 0, y, w, cdeffed); |
| 511 |
assert_eq!(&f[..w], &f_ref[..]); |
| 512 |
} |
| 513 |
} |
| 514 |
|
| 515 |
#[inline] |
| 516 |
#[target_feature(enable = "avx2")] |
| 517 |
unsafe fn sgrproj_box_f_r2_8_avx2<T: Pixel>( |
| 518 |
af: &[&[u32]; 2], bf: &[&[u32]; 2], f0: &mut [u32], f1: &mut [u32], |
| 519 |
x: usize, y: usize, cdeffed: &PlaneSlice<T>, |
| 520 |
) { |
| 521 |
let five = _mm256_set1_epi32(5); |
| 522 |
let six = _mm256_set1_epi32(6); |
| 523 |
let a0 = af[0].as_ptr(); |
| 524 |
let a2 = af[1].as_ptr(); |
| 525 |
let b0 = bf[0].as_ptr(); |
| 526 |
let b2 = bf[1].as_ptr(); |
| 527 |
let a = _mm256_add_epi32( |
| 528 |
_mm256_madd_epi16( |
| 529 |
_mm256_add_epi32( |
| 530 |
_mm256_loadu_si256(a0.add(x) as *const _), |
| 531 |
_mm256_loadu_si256(a0.add(x + 2) as *const _), |
| 532 |
), |
| 533 |
five, |
| 534 |
), |
| 535 |
_mm256_madd_epi16(_mm256_loadu_si256(a0.add(x + 1) as *const _), six), |
| 536 |
); |
| 537 |
let b = _mm256_add_epi32( |
| 538 |
_mm256_mullo_epi32( |
| 539 |
_mm256_add_epi32( |
| 540 |
_mm256_loadu_si256(b0.add(x) as *const _), |
| 541 |
_mm256_loadu_si256(b0.add(x + 2) as *const _), |
| 542 |
), |
| 543 |
five, |
| 544 |
), |
| 545 |
_mm256_mullo_epi32(_mm256_loadu_si256(b0.add(x + 1) as *const _), six), |
| 546 |
); |
| 547 |
let ao = _mm256_add_epi32( |
| 548 |
_mm256_madd_epi16( |
| 549 |
_mm256_add_epi32( |
| 550 |
_mm256_loadu_si256(a2.add(x) as *const _), |
| 551 |
_mm256_loadu_si256(a2.add(x + 2) as *const _), |
| 552 |
), |
| 553 |
five, |
| 554 |
), |
| 555 |
_mm256_madd_epi16(_mm256_loadu_si256(a2.add(x + 1) as *const _), six), |
| 556 |
); |
| 557 |
let bo = _mm256_add_epi32( |
| 558 |
_mm256_mullo_epi32( |
| 559 |
_mm256_add_epi32( |
| 560 |
_mm256_loadu_si256(b2.add(x) as *const _), |
| 561 |
_mm256_loadu_si256(b2.add(x + 2) as *const _), |
| 562 |
), |
| 563 |
five, |
| 564 |
), |
| 565 |
_mm256_mullo_epi32(_mm256_loadu_si256(b2.add(x + 1) as *const _), six), |
| 566 |
); |
| 567 |
let v = _mm256_add_epi32( |
| 568 |
_mm256_madd_epi16( |
| 569 |
_mm256_add_epi32(a, ao), |
| 570 |
if mem::size_of::<T>() == 1 { |
| 571 |
_mm256_cvtepu8_epi32(_mm_loadl_epi64( |
| 572 |
cdeffed.subslice(x, y).as_ptr() as *const _ |
| 573 |
)) |
| 574 |
} else { |
| 575 |
_mm256_cvtepu16_epi32(_mm_loadu_si128( |
| 576 |
cdeffed.subslice(x, y).as_ptr() as *const _ |
| 577 |
)) |
| 578 |
}, |
| 579 |
), |
| 580 |
_mm256_add_epi32(b, bo), |
| 581 |
); |
| 582 |
let vo = _mm256_add_epi32( |
| 583 |
_mm256_madd_epi16( |
| 584 |
ao, |
| 585 |
if mem::size_of::<T>() == 1 { |
| 586 |
_mm256_cvtepu8_epi32(_mm_loadl_epi64( |
| 587 |
cdeffed.subslice(x, y + 1).as_ptr() as *const _, |
| 588 |
)) |
| 589 |
} else { |
| 590 |
_mm256_cvtepu16_epi32(_mm_loadu_si128( |
| 591 |
cdeffed.subslice(x, y + 1).as_ptr() as *const _, |
| 592 |
)) |
| 593 |
}, |
| 594 |
), |
| 595 |
bo, |
| 596 |
); |
| 597 |
const SHIFT: i32 = (5 + SGRPROJ_SGR_BITS - SGRPROJ_RST_BITS) as i32; |
| 598 |
_mm256_storeu_si256( |
| 599 |
f0.as_mut_ptr().add(x) as *mut _, |
| 600 |
_mm256_srli_epi32( |
| 601 |
_mm256_add_epi32(v, _mm256_set1_epi32(1 << SHIFT >> 1)), |
| 602 |
SHIFT, |
| 603 |
), |
| 604 |
); |
| 605 |
const SHIFTO: i32 = (4 + SGRPROJ_SGR_BITS - SGRPROJ_RST_BITS) as i32; |
| 606 |
_mm256_storeu_si256( |
| 607 |
f1.as_mut_ptr().add(x) as *mut _, |
| 608 |
_mm256_srli_epi32( |
| 609 |
_mm256_add_epi32(vo, _mm256_set1_epi32(1 << SHIFTO >> 1)), |
| 610 |
SHIFTO, |
| 611 |
), |
| 612 |
); |
| 613 |
} |
| 614 |
|
| 615 |
#[target_feature(enable = "avx2")] |
| 616 |
pub(crate) unsafe fn sgrproj_box_f_r2_avx2<T: Pixel>( |
| 617 |
af: &[&[u32]; 2], bf: &[&[u32]; 2], f0: &mut [u32], f1: &mut [u32], |
| 618 |
y: usize, w: usize, cdeffed: &PlaneSlice<T>, |
| 619 |
) { |
| 620 |
for x in (0..w).step_by(8) { |
| 621 |
if x + 8 <= w { |
| 622 |
sgrproj_box_f_r2_8_avx2(af, bf, f0, f1, x, y, cdeffed); |
| 623 |
} else { |
| 624 |
// finish using scalar |
| 625 |
rust::sgrproj_box_f_r2_internal(af, bf, f0, f1, x, y, w, cdeffed); |
| 626 |
} |
| 627 |
} |
| 628 |
|
| 629 |
#[cfg(feature = "check_asm")] |
| 630 |
{ |
| 631 |
let mut f0_ref: Vec<u32> = vec![0; w]; |
| 632 |
let mut f1_ref: Vec<u32> = vec![0; w]; |
| 633 |
rust::sgrproj_box_f_r2_internal( |
| 634 |
af, |
| 635 |
bf, |
| 636 |
&mut f0_ref, |
| 637 |
&mut f1_ref, |
| 638 |
0, |
| 639 |
y, |
| 640 |
w, |
| 641 |
cdeffed, |
| 642 |
); |
| 643 |
assert_eq!(&f0[..w], &f0_ref[..]); |
| 644 |
assert_eq!(&f1[..w], &f1_ref[..]); |
| 645 |
} |
| 646 |
} |