colconv 0.1.0

SIMD-dispatched color-conversion kernels covering the FFmpeg AVPixelFormat space, with a Sink-based API so consumers pick which derived outputs (RGB / Luma / HSV / custom) they want without paying for the ones they don't.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
//! AVX2 kernels for the Tier 12 (DCP / `Xyz12`) source.
//!
//! Each kernel processes **8 pixels per SIMD iteration** in `__m256`
//! (256-bit, 8 f32 lanes) registers. The matmul, integer narrow, and
//! interleaved store are vectorized; the SMPTE 428-1 inverse OETF and
//! sRGB-shape forward OETF run **scalar per lane** via the same
//! scalar functions the reference kernel uses, preserving the 0-ULP
//! scalar↔SIMD parity contract by construction.
//!
//! Pipeline:
//! 1. Three `__m128i` loads via SSE4.1's `load_endian_u16x8::<BE>`
//!    feed the 3-channel deinterleave shuffle (the same byte-shuffle
//!    pattern the SSE4.1 backend uses, since AVX2 reuses
//!    `_mm_shuffle_epi8` per 128-bit lane). The result is three
//!    `__m128i` channel vectors (8 u16 each).
//! 2. Each `(X8, Y8, Z8)` is right-shifted by 4 (`_mm_srli_epi16::<4>`)
//!    to extract the active 12-bit code from the high-bit-packed `u16`
//!    (FFmpeg `AV_PIX_FMT_XYZ12LE/BE`: code in `[15:4]`, low 4 bits
//!    zero), defensively masked (`_mm_and_si128`), widened to
//!    `__m256i` u32x8 via `_mm256_cvtepu16_epi32`, then cast to
//!    `__m256` via `_mm256_cvtepi32_ps`.
//! 3. Per-lane scalar `smpte428_inverse_oetf` produces linear XYZ as
//!    `__m256`.
//! 4. Vectorized 3×3 matmul uses plain `_mm256_mul_ps + _mm256_add_ps`
//!    (NOT FMA via `_mm256_fmadd_ps`) — single-rounding FMA breaks the
//!    0-ULP parity contract on integer-narrow output paths.
//! 5. Per-lane scalar `oetf_srgb` (only for u8 / u16 / f16 outputs).
//! 6. Clamp `[0, 1]` × `scale` + `+0.5` truncate via
//!    `_mm256_cvttps_epi32` (round-half-up matches scalar's
//!    `(c * scale + 0.5) as int`); saturate to u16 via
//!    `_mm256_packus_epi32` followed by `_mm256_permute4x64_epi64`
//!    to undo per-128-bit-lane reordering.
//!
//! Width remainder (`width % 8`) is handled by the scalar reference
//! kernel.
//!
//! # Numerical contract
//!
//! Bit-identical scalar↔SIMD output across all integer / f16 / f32
//! paths (verified by `tests::xyz12`).

use core::arch::x86_64::*;

use super::super::{
  x86_common::{write_rgb_u8_8, write_rgba_u8_8},
  x86_sse41::endian::load_endian_u16x8,
};
use crate::{
  DcpTargetGamut,
  row::scalar::{
    self,
    xyz12::{oetf_srgb, smpte428_inverse_oetf},
    xyz12_constants::xyz_to_rgb_matrix,
  },
};

const PIXELS_PER_ITER: usize = 8;
const SAMPLE_MASK_U16: u16 = 0x0FFF;

// ---- Internal helpers --------------------------------------------------

/// Deinterleaves 3 `__m128i` registers (24 u16 = 8 packed XYZ pixels)
/// into `(X8, Y8, Z8)` u16x8 channel vectors. Same shuffle pattern as
/// the SSE4.1 backend's `deinterleave_xyz12_8px` — kept inline here
/// rather than imported across module boundaries.
#[inline(always)]
unsafe fn deinterleave_xyz12_8px(
  v0: __m128i,
  v1: __m128i,
  v2: __m128i,
) -> (__m128i, __m128i, __m128i) {
  unsafe {
    let x_v0 = _mm_setr_epi8(0, 1, 6, 7, 12, 13, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
    let x_v1 = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, 2, 3, 8, 9, 14, 15, -1, -1, -1, -1);
    let x_v2 = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 4, 5, 10, 11);
    let x = _mm_or_si128(
      _mm_or_si128(_mm_shuffle_epi8(v0, x_v0), _mm_shuffle_epi8(v1, x_v1)),
      _mm_shuffle_epi8(v2, x_v2),
    );

    let y_v0 = _mm_setr_epi8(2, 3, 8, 9, 14, 15, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
    let y_v1 = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, 4, 5, 10, 11, -1, -1, -1, -1, -1, -1);
    let y_v2 = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 0, 1, 6, 7, 12, 13);
    let y = _mm_or_si128(
      _mm_or_si128(_mm_shuffle_epi8(v0, y_v0), _mm_shuffle_epi8(v1, y_v1)),
      _mm_shuffle_epi8(v2, y_v2),
    );

    let z_v0 = _mm_setr_epi8(4, 5, 10, 11, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
    let z_v1 = _mm_setr_epi8(-1, -1, -1, -1, 0, 1, 6, 7, 12, 13, -1, -1, -1, -1, -1, -1);
    let z_v2 = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 2, 3, 8, 9, 14, 15);
    let z = _mm_or_si128(
      _mm_or_si128(_mm_shuffle_epi8(v0, z_v0), _mm_shuffle_epi8(v1, z_v1)),
      _mm_shuffle_epi8(v2, z_v2),
    );

    (x, y, z)
  }
}

/// Widens a u16x8 channel vector (already 12-bit-masked) to f32x8.
/// `_mm256_cvtepu16_epi32` zero-extends u16 → i32 (no sign issue
/// because samples are < 4096); `_mm256_cvtepi32_ps` casts to f32.
#[inline(always)]
unsafe fn u16x8_to_f32x8(v: __m128i) -> __m256 {
  unsafe { _mm256_cvtepi32_ps(_mm256_cvtepu16_epi32(v)) }
}

/// Per-lane scalar SMPTE 428-1 inverse OETF on an `__m256` (8 lanes).
#[inline(always)]
unsafe fn smpte428_inv_oetf_scalar8(v: __m256) -> __m256 {
  unsafe {
    let mut buf = [0.0_f32; 8];
    _mm256_storeu_ps(buf.as_mut_ptr(), v);
    for slot in &mut buf {
      *slot = smpte428_inverse_oetf(*slot as u16);
    }
    _mm256_loadu_ps(buf.as_ptr())
  }
}

/// Per-lane scalar sRGB OETF on an `__m256` (8 lanes).
#[inline(always)]
unsafe fn oetf_srgb_scalar8(v: __m256) -> __m256 {
  unsafe {
    let mut buf = [0.0_f32; 8];
    _mm256_storeu_ps(buf.as_mut_ptr(), v);
    for slot in &mut buf {
      *slot = oetf_srgb(*slot);
    }
    _mm256_loadu_ps(buf.as_ptr())
  }
}

/// Vectorized 3×3 matmul on an 8-lane f32 vector. Plain mul + add
/// (NOT FMA) — see `super::super::neon::xyz12::matmul_xyz_to_rgb`
/// docstring for the rounding-schedule rationale.
#[inline(always)]
unsafe fn matmul_xyz_to_rgb_8lane(
  m: &[[f32; 3]; 3],
  x: __m256,
  y: __m256,
  z: __m256,
) -> (__m256, __m256, __m256) {
  unsafe {
    let m00 = _mm256_set1_ps(m[0][0]);
    let m01 = _mm256_set1_ps(m[0][1]);
    let m02 = _mm256_set1_ps(m[0][2]);
    let m10 = _mm256_set1_ps(m[1][0]);
    let m11 = _mm256_set1_ps(m[1][1]);
    let m12 = _mm256_set1_ps(m[1][2]);
    let m20 = _mm256_set1_ps(m[2][0]);
    let m21 = _mm256_set1_ps(m[2][1]);
    let m22 = _mm256_set1_ps(m[2][2]);
    let r = _mm256_add_ps(
      _mm256_add_ps(_mm256_mul_ps(m00, x), _mm256_mul_ps(m01, y)),
      _mm256_mul_ps(m02, z),
    );
    let g = _mm256_add_ps(
      _mm256_add_ps(_mm256_mul_ps(m10, x), _mm256_mul_ps(m11, y)),
      _mm256_mul_ps(m12, z),
    );
    let b = _mm256_add_ps(
      _mm256_add_ps(_mm256_mul_ps(m20, x), _mm256_mul_ps(m21, y)),
      _mm256_mul_ps(m22, z),
    );
    (r, g, b)
  }
}

/// Loads 8 XYZ12 pixels and produces (R, G, B) `__m256` after the
/// inverse-OETF + 3×3 matmul.
#[inline(always)]
unsafe fn load_and_matmul_8px<const BE: bool>(
  p: *const u8,
  m: &[[f32; 3]; 3],
) -> (__m256, __m256, __m256) {
  unsafe {
    let v0 = load_endian_u16x8::<BE>(p);
    let v1 = load_endian_u16x8::<BE>(p.add(16));
    let v2 = load_endian_u16x8::<BE>(p.add(32));
    let (x_u, y_u, z_u) = deinterleave_xyz12_8px(v0, v1, v2);
    // Shift right 4 to extract active 12-bit code (high-bit-packed
    // FFmpeg `AV_PIX_FMT_XYZ12LE/BE`); mask defensively.
    let mask = _mm_set1_epi16(SAMPLE_MASK_U16 as i16);
    let x_shr = _mm_and_si128(_mm_srli_epi16::<4>(x_u), mask);
    let y_shr = _mm_and_si128(_mm_srli_epi16::<4>(y_u), mask);
    let z_shr = _mm_and_si128(_mm_srli_epi16::<4>(z_u), mask);
    let x_lin = smpte428_inv_oetf_scalar8(u16x8_to_f32x8(x_shr));
    let y_lin = smpte428_inv_oetf_scalar8(u16x8_to_f32x8(y_shr));
    let z_lin = smpte428_inv_oetf_scalar8(u16x8_to_f32x8(z_shr));
    matmul_xyz_to_rgb_8lane(m, x_lin, y_lin, z_lin)
  }
}

/// Loads 8 XYZ12 pixels and produces (X, Y, Z) `__m256` linear-XYZ —
/// step 1 only, no matmul.
#[inline(always)]
unsafe fn load_xyz_linear_8px<const BE: bool>(p: *const u8) -> (__m256, __m256, __m256) {
  unsafe {
    let v0 = load_endian_u16x8::<BE>(p);
    let v1 = load_endian_u16x8::<BE>(p.add(16));
    let v2 = load_endian_u16x8::<BE>(p.add(32));
    let (x_u, y_u, z_u) = deinterleave_xyz12_8px(v0, v1, v2);
    let mask = _mm_set1_epi16(SAMPLE_MASK_U16 as i16);
    let x_shr = _mm_and_si128(_mm_srli_epi16::<4>(x_u), mask);
    let y_shr = _mm_and_si128(_mm_srli_epi16::<4>(y_u), mask);
    let z_shr = _mm_and_si128(_mm_srli_epi16::<4>(z_u), mask);
    let x_lin = smpte428_inv_oetf_scalar8(u16x8_to_f32x8(x_shr));
    let y_lin = smpte428_inv_oetf_scalar8(u16x8_to_f32x8(y_shr));
    let z_lin = smpte428_inv_oetf_scalar8(u16x8_to_f32x8(z_shr));
    (x_lin, y_lin, z_lin)
  }
}

/// Vectorized clamp `[0, 1]` × scale + `+0.5` truncate, returning a
/// u32x8 (`__m256i`) ready for saturating narrow.
#[inline(always)]
unsafe fn clamp_scale_to_u32x8(v: __m256, zero: __m256, one: __m256, scale: __m256) -> __m256i {
  unsafe {
    let half = _mm256_set1_ps(0.5);
    let clamped = _mm256_min_ps(_mm256_max_ps(v, zero), one);
    let scaled = _mm256_add_ps(_mm256_mul_ps(clamped, scale), half);
    _mm256_cvttps_epi32(scaled)
  }
}

/// Narrow a u32x8 → u16x8 via `_mm256_packus_epi32` + lane fixup.
///
/// `_mm256_packus_epi32` packs per-128-bit-lane (low-pair from each
/// lane → low half, high-pair → high half), so the lane-split output
/// has channels {0..3, 8..11, 4..7, 12..15} pre-permute. We pack with
/// itself and read the low 128 bits as the natural `[0..7]` channel.
#[inline(always)]
unsafe fn narrow_u32x8_to_u16x8(lo: __m256i) -> __m128i {
  unsafe {
    // Pack with self → 16 u16 lanes (duplicated). Extract the low 128.
    let packed = _mm256_packus_epi32(lo, lo);
    // Lane fixup so the natural `[0..7]` u16 order falls in lane 0.
    let permuted = _mm256_permute4x64_epi64::<0xD8>(packed);
    _mm256_castsi256_si128(permuted)
  }
}

// ---- Per-output kernels ------------------------------------------------

/// XYZ12 → packed u8 RGB (8 px/iter).
///
/// # Safety
///
/// 1. AVX2 must be available (caller obligation).
/// 2. `xyz.len() >= width * 3`; `rgb_out.len() >= width * 3`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgb_row<const BE: bool>(
  xyz: &[u16],
  rgb_out: &mut [u8],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgb_out.len() >= width * 3, "rgb_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);

  unsafe {
    let zero_ps = _mm256_setzero_ps();
    let one_ps = _mm256_set1_ps(1.0);
    let scale = _mm256_set1_ps(255.0);
    let zero_si = _mm_setzero_si128();
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let r_oetf = oetf_srgb_scalar8(r_lin);
      let g_oetf = oetf_srgb_scalar8(g_lin);
      let b_oetf = oetf_srgb_scalar8(b_lin);
      let r_u32 = clamp_scale_to_u32x8(r_oetf, zero_ps, one_ps, scale);
      let g_u32 = clamp_scale_to_u32x8(g_oetf, zero_ps, one_ps, scale);
      let b_u32 = clamp_scale_to_u32x8(b_oetf, zero_ps, one_ps, scale);
      let r_u16 = narrow_u32x8_to_u16x8(r_u32);
      let g_u16 = narrow_u32x8_to_u16x8(g_u32);
      let b_u16 = narrow_u32x8_to_u16x8(b_u32);
      let r_u8 = _mm_packus_epi16(r_u16, zero_si);
      let g_u8 = _mm_packus_epi16(g_u16, zero_si);
      let b_u8 = _mm_packus_epi16(b_u16, zero_si);
      // In-register 8-pixel RGB interleave + 16-byte + 8-byte stores
      // avoid a 3× stack-temp + scalar-scatter loop.
      write_rgb_u8_8(r_u8, g_u8, b_u8, rgb_out.as_mut_ptr().add(x * 3));
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgb_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgb_out[x * 3..width * 3],
        width - x,
        target_gamut,
      );
    }
  }
}

/// XYZ12 → packed u8 RGBA (alpha = `0xFF`).
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `rgba_out.len() >= width * 4`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgba_row<const BE: bool>(
  xyz: &[u16],
  rgba_out: &mut [u8],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgba_out.len() >= width * 4, "rgba_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);

  unsafe {
    let zero_ps = _mm256_setzero_ps();
    let one_ps = _mm256_set1_ps(1.0);
    let scale = _mm256_set1_ps(255.0);
    let zero_si = _mm_setzero_si128();
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let r_oetf = oetf_srgb_scalar8(r_lin);
      let g_oetf = oetf_srgb_scalar8(g_lin);
      let b_oetf = oetf_srgb_scalar8(b_lin);
      let r_u32 = clamp_scale_to_u32x8(r_oetf, zero_ps, one_ps, scale);
      let g_u32 = clamp_scale_to_u32x8(g_oetf, zero_ps, one_ps, scale);
      let b_u32 = clamp_scale_to_u32x8(b_oetf, zero_ps, one_ps, scale);
      let r_u16 = narrow_u32x8_to_u16x8(r_u32);
      let g_u16 = narrow_u32x8_to_u16x8(g_u32);
      let b_u16 = narrow_u32x8_to_u16x8(b_u32);
      let r_u8 = _mm_packus_epi16(r_u16, zero_si);
      let g_u8 = _mm_packus_epi16(g_u16, zero_si);
      let b_u8 = _mm_packus_epi16(b_u16, zero_si);
      // Alpha = 0xFF in every lane (only the low 8 bytes are read by
      // `write_rgba_u8_8`'s `unpacklo_epi8`).
      let a_u8 = _mm_set1_epi8(-1_i8);
      // In-register 8-pixel RGBA interleave via two unpack stages +
      // two 16-byte stores avoids a 3× stack-temp + scalar-scatter loop.
      write_rgba_u8_8(r_u8, g_u8, b_u8, a_u8, rgba_out.as_mut_ptr().add(x * 4));
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgba_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgba_out[x * 4..width * 4],
        width - x,
        target_gamut,
      );
    }
  }
}

/// XYZ12 → packed u16 RGB (full-range scaling).
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `rgb_out.len() >= width * 3`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgb_u16_row<const BE: bool>(
  xyz: &[u16],
  rgb_out: &mut [u16],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgb_out.len() >= width * 3, "rgb_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);

  unsafe {
    let zero_ps = _mm256_setzero_ps();
    let one_ps = _mm256_set1_ps(1.0);
    let scale = _mm256_set1_ps(65535.0);
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let r_oetf = oetf_srgb_scalar8(r_lin);
      let g_oetf = oetf_srgb_scalar8(g_lin);
      let b_oetf = oetf_srgb_scalar8(b_lin);
      let r_u32 = clamp_scale_to_u32x8(r_oetf, zero_ps, one_ps, scale);
      let g_u32 = clamp_scale_to_u32x8(g_oetf, zero_ps, one_ps, scale);
      let b_u32 = clamp_scale_to_u32x8(b_oetf, zero_ps, one_ps, scale);
      let r_u16 = narrow_u32x8_to_u16x8(r_u32);
      let g_u16 = narrow_u32x8_to_u16x8(g_u32);
      let b_u16 = narrow_u32x8_to_u16x8(b_u32);
      let mut tmp_r = [0u16; 8];
      let mut tmp_g = [0u16; 8];
      let mut tmp_b = [0u16; 8];
      _mm_storeu_si128(tmp_r.as_mut_ptr() as *mut __m128i, r_u16);
      _mm_storeu_si128(tmp_g.as_mut_ptr() as *mut __m128i, g_u16);
      _mm_storeu_si128(tmp_b.as_mut_ptr() as *mut __m128i, b_u16);
      let dst = rgb_out.as_mut_ptr().add(x * 3);
      for i in 0..PIXELS_PER_ITER {
        *dst.add(i * 3) = tmp_r[i];
        *dst.add(i * 3 + 1) = tmp_g[i];
        *dst.add(i * 3 + 2) = tmp_b[i];
      }
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgb_u16_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgb_out[x * 3..width * 3],
        width - x,
        target_gamut,
      );
    }
  }
}

/// XYZ12 → packed u16 RGBA (alpha = `0xFFFF`).
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `rgba_out.len() >= width * 4`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgba_u16_row<const BE: bool>(
  xyz: &[u16],
  rgba_out: &mut [u16],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgba_out.len() >= width * 4, "rgba_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);

  unsafe {
    let zero_ps = _mm256_setzero_ps();
    let one_ps = _mm256_set1_ps(1.0);
    let scale = _mm256_set1_ps(65535.0);
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let r_oetf = oetf_srgb_scalar8(r_lin);
      let g_oetf = oetf_srgb_scalar8(g_lin);
      let b_oetf = oetf_srgb_scalar8(b_lin);
      let r_u32 = clamp_scale_to_u32x8(r_oetf, zero_ps, one_ps, scale);
      let g_u32 = clamp_scale_to_u32x8(g_oetf, zero_ps, one_ps, scale);
      let b_u32 = clamp_scale_to_u32x8(b_oetf, zero_ps, one_ps, scale);
      let r_u16 = narrow_u32x8_to_u16x8(r_u32);
      let g_u16 = narrow_u32x8_to_u16x8(g_u32);
      let b_u16 = narrow_u32x8_to_u16x8(b_u32);
      let mut tmp_r = [0u16; 8];
      let mut tmp_g = [0u16; 8];
      let mut tmp_b = [0u16; 8];
      _mm_storeu_si128(tmp_r.as_mut_ptr() as *mut __m128i, r_u16);
      _mm_storeu_si128(tmp_g.as_mut_ptr() as *mut __m128i, g_u16);
      _mm_storeu_si128(tmp_b.as_mut_ptr() as *mut __m128i, b_u16);
      let dst = rgba_out.as_mut_ptr().add(x * 4);
      for i in 0..PIXELS_PER_ITER {
        *dst.add(i * 4) = tmp_r[i];
        *dst.add(i * 4 + 1) = tmp_g[i];
        *dst.add(i * 4 + 2) = tmp_b[i];
        *dst.add(i * 4 + 3) = 0xFFFF;
      }
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgba_u16_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgba_out[x * 4..width * 4],
        width - x,
        target_gamut,
      );
    }
  }
}

/// XYZ12 → packed linear RGB f32 (lossless: matrix only, no OETF).
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `rgb_out.len() >= width * 3`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgb_f32_row<const BE: bool>(
  xyz: &[u16],
  rgb_out: &mut [f32],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgb_out.len() >= width * 3, "rgb_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);

  unsafe {
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let mut rb = [0.0_f32; 8];
      let mut gb = [0.0_f32; 8];
      let mut bb = [0.0_f32; 8];
      _mm256_storeu_ps(rb.as_mut_ptr(), r_lin);
      _mm256_storeu_ps(gb.as_mut_ptr(), g_lin);
      _mm256_storeu_ps(bb.as_mut_ptr(), b_lin);
      let dst = rgb_out.as_mut_ptr().add(x * 3);
      for i in 0..PIXELS_PER_ITER {
        *dst.add(i * 3) = rb[i];
        *dst.add(i * 3 + 1) = gb[i];
        *dst.add(i * 3 + 2) = bb[i];
      }
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgb_f32_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgb_out[x * 3..width * 3],
        width - x,
        target_gamut,
      );
    }
  }
}

/// XYZ12 → packed linear XYZ f32 (step 1 only).
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `xyz_out.len() >= width * 3`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_xyz_f32_row<const BE: bool>(
  xyz: &[u16],
  xyz_out: &mut [f32],
  width: usize,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(xyz_out.len() >= width * 3, "xyz_out row too short");

  unsafe {
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (xv, yv, zv) = load_xyz_linear_8px::<BE>(p);
      let mut xb = [0.0_f32; 8];
      let mut yb = [0.0_f32; 8];
      let mut zb = [0.0_f32; 8];
      _mm256_storeu_ps(xb.as_mut_ptr(), xv);
      _mm256_storeu_ps(yb.as_mut_ptr(), yv);
      _mm256_storeu_ps(zb.as_mut_ptr(), zv);
      let dst = xyz_out.as_mut_ptr().add(x * 3);
      for i in 0..PIXELS_PER_ITER {
        *dst.add(i * 3) = xb[i];
        *dst.add(i * 3 + 1) = yb[i];
        *dst.add(i * 3 + 2) = zb[i];
      }
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_xyz_f32_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut xyz_out[x * 3..width * 3],
        width - x,
      );
    }
  }
}

/// XYZ12 → packed f16 RGB.
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `rgb_out.len() >= width * 3`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgb_f16_row<const BE: bool>(
  xyz: &[u16],
  rgb_out: &mut [half::f16],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgb_out.len() >= width * 3, "rgb_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);

  unsafe {
    let zero_ps = _mm256_setzero_ps();
    let one_ps = _mm256_set1_ps(1.0);
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let r_oetf = oetf_srgb_scalar8(r_lin);
      let g_oetf = oetf_srgb_scalar8(g_lin);
      let b_oetf = oetf_srgb_scalar8(b_lin);
      let r_clamp = _mm256_min_ps(_mm256_max_ps(r_oetf, zero_ps), one_ps);
      let g_clamp = _mm256_min_ps(_mm256_max_ps(g_oetf, zero_ps), one_ps);
      let b_clamp = _mm256_min_ps(_mm256_max_ps(b_oetf, zero_ps), one_ps);
      let mut rb = [0.0_f32; 8];
      let mut gb = [0.0_f32; 8];
      let mut bb = [0.0_f32; 8];
      _mm256_storeu_ps(rb.as_mut_ptr(), r_clamp);
      _mm256_storeu_ps(gb.as_mut_ptr(), g_clamp);
      _mm256_storeu_ps(bb.as_mut_ptr(), b_clamp);
      for i in 0..PIXELS_PER_ITER {
        let oi = (x + i) * 3;
        rgb_out[oi] = half::f16::from_f32(rb[i]);
        rgb_out[oi + 1] = half::f16::from_f32(gb[i]);
        rgb_out[oi + 2] = half::f16::from_f32(bb[i]);
      }
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgb_f16_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgb_out[x * 3..width * 3],
        width - x,
        target_gamut,
      );
    }
  }
}

/// XYZ12 → packed f16 RGBA (alpha = `1.0`).
///
/// # Safety
///
/// 1. AVX2 must be available.
/// 2. `xyz.len() >= width * 3`; `rgba_out.len() >= width * 4`.
#[inline]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn xyz12_to_rgba_f16_row<const BE: bool>(
  xyz: &[u16],
  rgba_out: &mut [half::f16],
  width: usize,
  target_gamut: DcpTargetGamut,
) {
  debug_assert!(xyz.len() >= width * 3, "xyz row too short");
  debug_assert!(rgba_out.len() >= width * 4, "rgba_out row too short");
  let m = xyz_to_rgb_matrix(target_gamut);
  let one_f16 = half::f16::from_f32(1.0);

  unsafe {
    let zero_ps = _mm256_setzero_ps();
    let one_ps = _mm256_set1_ps(1.0);
    let mut x = 0usize;
    while x + PIXELS_PER_ITER <= width {
      let p = xyz.as_ptr().add(x * 3) as *const u8;
      let (r_lin, g_lin, b_lin) = load_and_matmul_8px::<BE>(p, &m);
      let r_oetf = oetf_srgb_scalar8(r_lin);
      let g_oetf = oetf_srgb_scalar8(g_lin);
      let b_oetf = oetf_srgb_scalar8(b_lin);
      let r_clamp = _mm256_min_ps(_mm256_max_ps(r_oetf, zero_ps), one_ps);
      let g_clamp = _mm256_min_ps(_mm256_max_ps(g_oetf, zero_ps), one_ps);
      let b_clamp = _mm256_min_ps(_mm256_max_ps(b_oetf, zero_ps), one_ps);
      let mut rb = [0.0_f32; 8];
      let mut gb = [0.0_f32; 8];
      let mut bb = [0.0_f32; 8];
      _mm256_storeu_ps(rb.as_mut_ptr(), r_clamp);
      _mm256_storeu_ps(gb.as_mut_ptr(), g_clamp);
      _mm256_storeu_ps(bb.as_mut_ptr(), b_clamp);
      for i in 0..PIXELS_PER_ITER {
        let oi = (x + i) * 4;
        rgba_out[oi] = half::f16::from_f32(rb[i]);
        rgba_out[oi + 1] = half::f16::from_f32(gb[i]);
        rgba_out[oi + 2] = half::f16::from_f32(bb[i]);
        rgba_out[oi + 3] = one_f16;
      }
      x += PIXELS_PER_ITER;
    }
    if x < width {
      scalar::xyz12::xyz12_to_rgba_f16_row::<BE>(
        &xyz[x * 3..width * 3],
        &mut rgba_out[x * 4..width * 4],
        width - x,
        target_gamut,
      );
    }
  }
}