pixon 0.8.0

SIMD-dispatched color conversion for the FFmpeg AVPixelFormat space: one-call decode of validated source frames to RGB / RGBA / Luma / HSV, fused resampling while converting, colorimetry-honest matrices and chroma siting, no_std-friendly.
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
//! Sinker impl for the packed V410 source format — Ship 12a (Tier 5
//! 10-bit packed YUV 4:4:4). Full output coverage: u8 + native-depth
//! u16 RGB / RGBA + u8 / u16 luma + u8 HSV.
//!
//! V410 packs **one pixel per 32-bit word** (`(V << 20) | (Y << 10) | U`)
//! with 10-bit channels and 2 bits of padding (unlike the MSB-aligned
//! u16 quadruple layout used by Y2xx formats). The packed slice type
//! is `&[u32]`, not `&[u16]`. There is no chroma subsampling — every
//! pixel carries its own independent U / Y / V triplet (4:4:4).
//!
//! Outputs map to the sink's standard channels:
//! - `with_rgb` / `with_rgba` — packed YUV → RGB Q15 pipeline at
//!   `BITS = 10`, downshifted to u8; RGBA alpha is forced to `0xFF`
//!   (V410 has no alpha channel).
//! - `with_rgb_u16` / `with_rgba_u16` — same pipeline at native
//!   10-bit depth, low-bit-packed in `u16` (`[0, 1023]`); RGBA alpha
//!   is `0x3FF` (10-bit max).
//! - `with_luma` — extracts the 10-bit Y values from each V410 word
//!   and downshifts `>> 2` to u8.
//! - `with_luma_u16` — extracts the 10-bit Y values at native depth,
//!   low-bit-packed in `u16` (`[0, 1023]`). Each 10-bit Y is read
//!   directly from bits `[19:10]` of the V410 word (no shift needed
//!   beyond the bit-field extraction), yielding values in `[0, 0x3FF]`.
//! - `with_hsv` — when HSV is the only u8 colour output (no `with_rgb`
//!   / `with_rgba`), the direct `v410_to_hsv_row` kernel computes HSV
//!   straight from the packed YUV with no source-width RGB scratch
//!   (#263). When RGB or RGBA is also attached, HSV derives from the
//!   already-staged u8 RGB row via `rgb_to_hsv_row` (the cheap path).
//!
//! When both u8 RGB and u8 RGBA outputs are requested, the RGBA plane
//! is derived from the just-computed u8 RGB row via
//! [`expand_rgb_to_rgba_row`] (Strategy A) instead of running a
//! second YUV→RGB kernel. The same Strategy A applies on the u16
//! path via [`expand_rgb_u16_to_rgba_u16_row::<10>`]. When only the
//! RGBA variant is wanted, the dedicated `_to_rgba_row` /
//! `_to_rgba_u16_row` kernel writes the output buffer directly
//! without staging RGB.

use super::{
  GeometryOverflow, InsufficientBuffer, MixedSinker, MixedSinkerError, RowIndexOutOfRange,
  RowShapeMismatch, RowSlice, check_dimensions_match, packed_yuv444_triple_filter_resample,
  packed_yuv444_triple_resample, rgb_row_buf_or_scratch, rgba_plane_row_slice,
  rgba_u16_plane_row_slice,
};
// `NativeRouteChanged` is raised only by the native fast tier's route-flip
// guard, and `packed_yuv444_hb_process_native` exists only when the reused
// high-bit planar join is compiled in. Gated to the native tier's feature
// intersection.
#[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
use super::{NativeRouteChanged, packed_yuv444_hb_process_native};
// The RFC #238 insertion-point selector decides the native-vs-row-stage
// splice; consulted only inside the native tier's `cfg`, so its import
// shares that intersection.
#[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
use crate::resample::{AveragingDomain, InsertionContext, InsertionPoint, select_insertion_point};
use crate::{
  PixelSink,
  row::{
    expand_rgb_to_rgba_row, expand_rgb_u16_to_rgba_u16_row, rgb_to_hsv_row, v410_to_hsv_row,
    v410_to_luma_row, v410_to_luma_u16_row, v410_to_rgb_row, v410_to_rgb_u16_row, v410_to_rgba_row,
    v410_to_rgba_u16_row,
  },
  source::{V410, V410Row, V410Sink},
};

impl<'a, R, const BE: bool> MixedSinker<'a, V410<BE>, R> {
  /// Attaches a packed **8-bit** RGBA output buffer. Alpha is filled
  /// with constant `0xFF` (V410 has no alpha channel).
  ///
  /// Returns `Err(InsufficientRgbaBuffer)` if
  /// `buf.len() < width x height x 4`, or `Err(GeometryOverflow)` on
  /// 32‑bit targets when the product overflows.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn with_rgba(mut self, buf: &'a mut [u8]) -> Result<Self, MixedSinkerError> {
    self.set_rgba(buf)?;
    Ok(self)
  }
  /// In-place variant of [`with_rgba`](Self::with_rgba).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn set_rgba(&mut self, buf: &'a mut [u8]) -> Result<&mut Self, MixedSinkerError> {
    let expected = self.frame_elems(4)?;
    if buf.len() < expected {
      return Err(MixedSinkerError::InsufficientRgbaBuffer(
        InsufficientBuffer::new(expected, buf.len()),
      ));
    }
    self.rgba = Some(buf);
    Ok(self)
  }

  /// Attaches a packed **`u16`** RGB output buffer. 10-bit
  /// low-bit-packed (`[0, 1023]`); length is measured in `u16`
  /// **elements** (`width x height x 3`).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn with_rgb_u16(mut self, buf: &'a mut [u16]) -> Result<Self, MixedSinkerError> {
    self.set_rgb_u16(buf)?;
    Ok(self)
  }
  /// In-place variant of [`with_rgb_u16`](Self::with_rgb_u16).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn set_rgb_u16(&mut self, buf: &'a mut [u16]) -> Result<&mut Self, MixedSinkerError> {
    let expected = self.frame_elems(3)?;
    if buf.len() < expected {
      return Err(MixedSinkerError::InsufficientRgbU16Buffer(
        InsufficientBuffer::new(expected, buf.len()),
      ));
    }
    self.rgb_u16 = Some(buf);
    Ok(self)
  }

  /// Attaches a packed **`u16`** RGBA output buffer. 10-bit
  /// low-bit-packed (`[0, 1023]`); alpha element is `0x3FF` (10-bit
  /// max). Length is measured in `u16` **elements**
  /// (`width x height x 4`).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn with_rgba_u16(mut self, buf: &'a mut [u16]) -> Result<Self, MixedSinkerError> {
    self.set_rgba_u16(buf)?;
    Ok(self)
  }
  /// In-place variant of [`with_rgba_u16`](Self::with_rgba_u16).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn set_rgba_u16(&mut self, buf: &'a mut [u16]) -> Result<&mut Self, MixedSinkerError> {
    let expected = self.frame_elems(4)?;
    if buf.len() < expected {
      return Err(MixedSinkerError::InsufficientRgbaU16Buffer(
        InsufficientBuffer::new(expected, buf.len()),
      ));
    }
    self.rgba_u16 = Some(buf);
    Ok(self)
  }

  /// Attaches a native-depth **`u16`** luma output buffer. The 10-bit
  /// Y samples are extracted from each V410 word at native depth,
  /// low-bit-packed in `u16` (`[0, 1023]`). Length is measured in
  /// `u16` **elements** (`width x height`).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn with_luma_u16(mut self, buf: &'a mut [u16]) -> Result<Self, MixedSinkerError> {
    self.set_luma_u16(buf)?;
    Ok(self)
  }
  /// In-place variant of [`with_luma_u16`](Self::with_luma_u16).
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn set_luma_u16(&mut self, buf: &'a mut [u16]) -> Result<&mut Self, MixedSinkerError> {
    let expected = self.frame_pixels()?;
    if buf.len() < expected {
      return Err(MixedSinkerError::InsufficientLumaU16Buffer(
        InsufficientBuffer::new(expected, buf.len()),
      ));
    }
    self.luma_u16 = Some(buf);
    Ok(self)
  }
}

impl<const BE: bool, R> V410Sink<BE> for MixedSinker<'_, V410<BE>, R> {}

impl<const BE: bool, R> PixelSink for MixedSinker<'_, V410<BE>, R> {
  type Input<'r> = V410Row<'r>;
  type Error = MixedSinkerError;

  #[cfg_attr(not(tarpaulin), inline(always))]
  fn kernel_matrix(&self) -> crate::KernelMatrix {
    self.kernel_matrix
  }
  fn begin_frame(&mut self, width: u32, height: u32) -> Result<(), Self::Error> {
    check_dimensions_match(self.width, self.height, width, height)?;
    // New frame: restart the row-stage streams (lazily created in
    // `process`) and drop the frozen output set.
    if let Some(stream) = self.rgb_stream.as_mut() {
      stream.reset();
    }
    if let Some(stream) = self.rgb_stream_u16.as_mut() {
      stream.reset();
    }
    if let Some(stream) = self.luma_stream_u16.as_mut() {
      stream.reset();
    }
    if let Some(stream) = self.rgb_filter_stream.as_mut() {
      stream.reset();
    }
    if let Some(stream) = self.rgb_filter_stream_u16.as_mut() {
      stream.reset();
    }
    if let Some(stream) = self.luma_filter_stream_u16.as_mut() {
      stream.reset();
    }
    // New frame: restart the native join and clear the per-frame frozen
    // native/row-stage route so the next frame may pick either tier; a
    // mid-frame flip stays rejected. Gated to the native tier's feature
    // intersection (the planar join the native tier reuses is compiled only
    // under `yuv-planar`).
    #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
    if let Some(native) = self.native_planar_u16.as_mut() {
      native.reset();
    }
    #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
    {
      self.frozen_native_route = None;
    }
    self.resample_outputs = None;
    Ok(())
  }

  fn process(&mut self, row: V410Row<'_>) -> Result<(), Self::Error> {
    const BITS: u32 = 10;
    let w = self.width;
    let h = self.height;
    let idx = row.row();
    let use_simd = self.simd;

    // V410 row = `width` u32 elements (one pixel per word).
    let packed_expected = w;
    if row.packed().len() != packed_expected {
      return Err(MixedSinkerError::RowShapeMismatch(RowShapeMismatch::new(
        RowSlice::V410Packed,
        idx,
        packed_expected,
        row.packed().len(),
      )));
    }
    if idx >= self.height {
      return Err(MixedSinkerError::RowIndexOutOfRange(
        RowIndexOutOfRange::new(idx, self.height),
      ));
    }

    let Self {
      rgb,
      rgb_u16,
      rgba,
      rgba_u16,
      luma,
      luma_u16,
      hsv,
      rgb_scratch,
      rgb_scratch_u16,
      luma_scratch_u16,
      plan,
      rgb_stream,
      rgb_stream_u16,
      luma_stream_u16,
      rgb_filter_stream,
      rgb_filter_stream_u16,
      luma_filter_stream_u16,
      resample_outputs,
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      native,
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      native_planar_u16,
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      packed_444_y_full_u16,
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      packed_444_u_full_u16,
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      packed_444_v_full_u16,
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      frozen_native_route,
      ..
    } = self;

    // Non-identity plan: feed the shared three-stream tail. `BE` from the
    // marker selects the source decode wire for every conversion. The span
    // kind picks the engine — area binning or signed-coefficient filter
    // (both convert the YUV to RGB with the same closures and resample in
    // RGB space, so filter colour equals the RGB filter of the converted
    // pixels and matches area up to the kernel). Freeze + sequence-check
    // before staging, so a no-output sink stays a no-op and an
    // out-of-sequence row is rejected without allocating.
    if let Some(plan) = plan.as_ref() {
      let matrix = row.matrix();
      let full_range = row.full_range();
      let packed = row.packed();
      // A `Filter` plan routes to the filter resampler (the native fast tier is
      // area-only); branched FIRST, before the area native-route machinery.
      if let crate::resample::SpanKind::Filter = plan.kind() {
        return packed_yuv444_triple_filter_resample::<BITS>(
          rgb_filter_stream,
          rgb_filter_stream_u16,
          luma_filter_stream_u16,
          resample_outputs,
          rgb,
          rgba,
          rgb_u16,
          rgba_u16,
          luma,
          luma_u16,
          hsv,
          rgb_scratch,
          rgb_scratch_u16,
          luma_scratch_u16,
          w,
          plan,
          idx,
          use_simd,
          matrix,
          full_range,
          |scratch| v410_to_rgb_row(packed, scratch, w, matrix, full_range, use_simd, BE),
          |scratch| v410_to_rgb_u16_row(packed, scratch, w, matrix, full_range, use_simd, BE),
          |scratch| v410_to_luma_u16_row(packed, scratch, w, use_simd, BE),
        );
      }
      // Area plan. When the native tier is enabled (and the planar join it
      // reuses is compiled in), bit-extract each V410 word into separate
      // host-native LOGICAL u16 Y / U / V planes (4:4:4, full-width chroma) and
      // bin those at output resolution, converting once per output row.
      // Otherwise (or under `with_native(false)`) feed the row-stage tail. The
      // reused planar join emits the native-depth `luma_u16` (the clamped binned
      // Y), so attaching `luma_u16` no longer forces row-stage; the route
      // depends only on `with_native`. The output set is frozen on the first
      // resampled row, so the route stays stable across a frame and the mid-frame
      // flip guard catches a `set_native` toggle.
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      {
        let need_output = luma.is_some()
          || luma_u16.is_some()
          || rgb.is_some()
          || rgba.is_some()
          || rgb_u16.is_some()
          || rgba_u16.is_some()
          || hsv.is_some();
        // The RFC #238 splice stage. A filter plan already returned above, so
        // `area_plan` is true and the selector reproduces the former `*native`
        // boolean bit-for-bit (`cfg!` is true wherever this block compiles).
        let take_native = matches!(
          select_insertion_point(
            AveragingDomain::Encoded,
            InsertionContext {
              native_eligible: cfg!(all(feature = "yuv-444-packed", feature = "yuv-planar")),
              with_native: *native,
              area_plan: true,
            },
          ),
          InsertionPoint::NativeCodes
        );
        // Reject a mid-frame native/row-stage route flip BEFORE either tier's
        // dispatch (the #186 CHECK-before / SET-after template).
        if need_output
          && let Some(frozen) = *frozen_native_route
          && frozen != take_native
        {
          return Err(MixedSinkerError::NativeRouteChanged(
            NativeRouteChanged::new(idx),
          ));
        }
        if take_native {
          // Dispatch first; freeze the route to native ONLY after the call
          // returns Ok on an output-bearing row. V410 word layout: bits[9:0] U,
          // bits[19:10] Y, bits[29:20] V, bits[31:30] padding — `BE` selects the
          // u32 wire decode before the bit-field extract.
          packed_yuv444_hb_process_native::<BITS>(
            plan,
            native_planar_u16,
            packed_444_y_full_u16,
            packed_444_u_full_u16,
            packed_444_v_full_u16,
            resample_outputs,
            rgb,
            rgba,
            rgb_u16,
            rgba_u16,
            luma,
            luma_u16,
            hsv,
            rgb_scratch,
            rgb_scratch_u16,
            |y_dst| {
              for (dst, &raw) in y_dst.iter_mut().zip(packed) {
                let word = if BE {
                  u32::from_be(raw)
                } else {
                  u32::from_le(raw)
                };
                *dst = ((word >> 10) & 0x3FF) as u16;
              }
            },
            |u_dst, v_dst| {
              for ((u, v), &raw) in u_dst.iter_mut().zip(v_dst.iter_mut()).zip(packed) {
                let word = if BE {
                  u32::from_be(raw)
                } else {
                  u32::from_le(raw)
                };
                *u = (word & 0x3FF) as u16;
                *v = ((word >> 20) & 0x3FF) as u16;
              }
            },
            matrix,
            full_range,
            idx,
            w,
            h,
            use_simd,
          )?;
          if frozen_native_route.is_none() && need_output {
            *frozen_native_route = Some(true);
          }
          return Ok(());
        }
      }
      // Row-stage area tail (the only route under a `yuv-444-packed`-solo build
      // where the planar join is absent). Same CHECK-before / SET-after split.
      packed_yuv444_triple_resample::<BITS>(
        rgb_stream,
        rgb_stream_u16,
        luma_stream_u16,
        resample_outputs,
        rgb,
        rgba,
        rgb_u16,
        rgba_u16,
        luma,
        luma_u16,
        hsv,
        rgb_scratch,
        rgb_scratch_u16,
        luma_scratch_u16,
        w,
        plan,
        idx,
        use_simd,
        matrix,
        full_range,
        |scratch| v410_to_rgb_row(packed, scratch, w, matrix, full_range, use_simd, BE),
        |scratch| v410_to_rgb_u16_row(packed, scratch, w, matrix, full_range, use_simd, BE),
        |scratch| v410_to_luma_u16_row(packed, scratch, w, use_simd, BE),
      )?;
      #[cfg(all(feature = "yuv-444-packed", feature = "yuv-planar"))]
      {
        let need_output = luma.is_some()
          || luma_u16.is_some()
          || rgb.is_some()
          || rgba.is_some()
          || rgb_u16.is_some()
          || rgba_u16.is_some()
          || hsv.is_some();
        if frozen_native_route.is_none() && need_output {
          *frozen_native_route = Some(false);
        }
      }
      return Ok(());
    }

    let one_plane_start = idx * w;
    let one_plane_end = one_plane_start + w;
    let packed = row.packed();

    // Resolve the output set up front so the atomicity preflight below runs
    // before any output row is written.
    let want_rgb = rgb.is_some();
    let want_rgba = rgba.is_some();
    let want_hsv = hsv.is_some();

    // Atomicity preflight (#308, cf. the crate's #180 resample fix and the
    // high-bit semi-planar sibling): reserve the only growable row scratch this
    // identity row can touch — the u8 RGB row buffer — BEFORE any output row is
    // written (the luma / luma_u16 planes below, then the u16 RGB / RGBA
    // fan-out), so an allocator refusal returns a typed `AllocationFailed`
    // leaving the output frame untouched rather than partially mutated. The
    // luma / luma_u16 and u16 RGB / RGBA outputs write straight into their
    // caller buffers and never grow a scratch. `rgb_row_buf_or_scratch`'s
    // allocating (rgb = None) arm is reached exactly when a colour decode needs
    // an RGB row but no caller RGB buffer is borrowable — for this
    // convert-once-then-derive path that is `want_hsv && want_rgba && !want_rgb`
    // (HSV-only routes through the direct `v410_to_hsv_row` kernel, which needs
    // no RGB scratch). The later decode reuses the already-sized buffer, so the
    // default path is byte-identical; only the failure-path ordering changes.
    if want_hsv && want_rgba && !want_rgb {
      rgb_row_buf_or_scratch(
        rgb.as_deref_mut(),
        rgb_scratch,
        one_plane_start,
        one_plane_end,
        w,
        h,
      )?;
    }

    // Luma u8 — extract 8-bit Y bytes from the V410 plane via the
    // dedicated kernel (downshifts 10-bit Y >> 2 to u8).
    if let Some(buf) = luma.as_deref_mut() {
      v410_to_luma_row(
        packed,
        &mut buf[one_plane_start..one_plane_end],
        w,
        use_simd,
        BE,
      );
    }
    // Luma u16 — extract 10-bit Y values at native depth (low-bit-packed
    // in u16, range [0, 0x3FF]).
    if let Some(buf) = luma_u16.as_deref_mut() {
      v410_to_luma_u16_row(
        packed,
        &mut buf[one_plane_start..one_plane_end],
        w,
        use_simd,
        BE,
      );
    }

    // ===== u16 RGB / RGBA path (Strategy A) =====
    let want_rgb_u16 = rgb_u16.is_some();
    let want_rgba_u16 = rgba_u16.is_some();

    if want_rgba_u16 && !want_rgb_u16 {
      // Standalone u16 RGBA fast path — write directly into the
      // caller's buffer; no staging.
      let rgba_u16_buf = rgba_u16.as_deref_mut().unwrap();
      let rgba_u16_row =
        rgba_u16_plane_row_slice(rgba_u16_buf, one_plane_start, one_plane_end, w, h)?;
      v410_to_rgba_u16_row(
        packed,
        rgba_u16_row,
        w,
        row.matrix(),
        row.full_range(),
        use_simd,
        BE,
      );
    } else if want_rgb_u16 {
      let rgb_u16_buf = rgb_u16.as_deref_mut().unwrap();
      let rgb_plane_end =
        one_plane_end
          .checked_mul(3)
          .ok_or(MixedSinkerError::GeometryOverflow(GeometryOverflow::new(
            w, h, 3,
          )))?;
      let rgb_plane_start = one_plane_start * 3;
      let rgb_u16_row = &mut rgb_u16_buf[rgb_plane_start..rgb_plane_end];
      v410_to_rgb_u16_row(
        packed,
        rgb_u16_row,
        w,
        row.matrix(),
        row.full_range(),
        use_simd,
        BE,
      );
      if want_rgba_u16 {
        // Strategy A u16 fan-out — derive RGBA from the just-computed
        // RGB row instead of running a second YUV→RGB kernel.
        let rgba_u16_buf = rgba_u16.as_deref_mut().unwrap();
        let rgba_u16_row =
          rgba_u16_plane_row_slice(rgba_u16_buf, one_plane_start, one_plane_end, w, h)?;
        expand_rgb_u16_to_rgba_u16_row::<BITS>(rgb_u16_row, rgba_u16_row, w);
      }
    }

    // ===== u8 RGB / RGBA / HSV path (Strategy A) =====
    // HSV-without-RGB-or-RGBA goes through the direct `v410_to_hsv_row`
    // kernel (no source-width RGB scratch). When RGB or RGBA is *also*
    // attached the RGB kernel runs anyway, so HSV derives off that buffer
    // for free (the cheap path) and `need_u8_rgb_kernel` keeps it alive.
    // (Resample row-stage HSV stays correct via the convert-once path in
    // the plan branch above.)
    // `want_rgb` / `want_rgba` / `want_hsv` were resolved up front for the
    // atomicity preflight (#308).
    let want_hsv_direct = want_hsv && !want_rgb && !want_rgba;
    let need_u8_rgb_kernel = want_rgb || (want_hsv && want_rgba);

    if want_hsv_direct {
      let hsv = hsv.as_mut().expect("want_hsv_direct implies hsv attached");
      let (h, s, v) = hsv.hsv();
      v410_to_hsv_row(
        packed,
        &mut h[one_plane_start..one_plane_end],
        &mut s[one_plane_start..one_plane_end],
        &mut v[one_plane_start..one_plane_end],
        w,
        row.matrix(),
        row.full_range(),
        use_simd,
        BE,
      );
      return Ok(());
    }

    // Standalone u8 RGBA fast path — no RGB / HSV requested. Run the
    // dedicated RGBA kernel directly into the output buffer; avoids
    // both the scratch allocation and the RGB→RGBA expand pass.
    if want_rgba && !need_u8_rgb_kernel {
      let rgba_buf = rgba.as_deref_mut().unwrap();
      let rgba_row = rgba_plane_row_slice(rgba_buf, one_plane_start, one_plane_end, w, h)?;
      v410_to_rgba_row(
        packed,
        rgba_row,
        w,
        row.matrix(),
        row.full_range(),
        use_simd,
        BE,
      );
      return Ok(());
    }

    if !need_u8_rgb_kernel {
      return Ok(());
    }

    let rgb_row = rgb_row_buf_or_scratch(
      rgb.as_deref_mut(),
      rgb_scratch,
      one_plane_start,
      one_plane_end,
      w,
      h,
    )?;
    v410_to_rgb_row(
      packed,
      rgb_row,
      w,
      row.matrix(),
      row.full_range(),
      use_simd,
      BE,
    );

    if let Some(hsv) = hsv.as_mut() {
      let (h, s, v) = hsv.hsv();
      rgb_to_hsv_row(
        rgb_row,
        &mut h[one_plane_start..one_plane_end],
        &mut s[one_plane_start..one_plane_end],
        &mut v[one_plane_start..one_plane_end],
        w,
        use_simd,
      );
    }

    // Strategy A u8 fan-out — derive RGBA from the just-computed RGB
    // row instead of running a second YUV→RGB kernel.
    if let Some(buf) = rgba.as_deref_mut() {
      let rgba_row = rgba_plane_row_slice(buf, one_plane_start, one_plane_end, w, h)?;
      expand_rgb_to_rgba_row(rgb_row, rgba_row, w);
    }

    Ok(())
  }
}