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
//! Fused-downscale coverage for `Grayf16` — routed through a single 1-channel
//! `AreaStream<f32>` that widens the source `f16` luma plane to `f32` and bins
//! it at f32 precision (Grayf16 *is* an f16 luma plane). The wire row widens to
//! a host-native f32 luma plane first (the same kernel the direct `luma_f32`
//! path uses), then every attached output derives from each finalized binned
//! f32 luma row using the `grayf32` kernels — so every resampled output equals
//! the direct `Grayf32` sink run over a frame that already holds the binned f32
//! luma plane (the binned domain is f32, not f16). The half-float twin of
//! `resample_grayf32`.

use crate::{
  KernelMatrix, PixelSink,
  frame::{Grayf16Frame, Grayf32Frame},
  resample::{AreaResampler, ResampleError},
  sinker::{MixedSinker, MixedSinkerError},
  source::{Grayf16, Grayf16Row, Grayf32, grayf16_to, grayf16_to_endian},
};
use half::f16;

const SRC: usize = 8;
const OUT: usize = 4;
const FR: bool = true;
const M: KernelMatrix = KernelMatrix::Bt709;

/// Re-encode a host-native f16 slice as LE-encoded byte storage (the `grayf16le`
/// plane contract), recovered via `u16::from_le`.
fn as_le_f16(host: &[f16]) -> Vec<f16> {
  host
    .iter()
    .map(|v| f16::from_bits(u16::from_ne_bytes(v.to_bits().to_le_bytes())))
    .collect()
}

/// Re-encode a host-native f16 slice as BE-encoded byte storage (`grayf16be`),
/// recovered via `u16::from_be`.
fn as_be_f16(host: &[f16]) -> Vec<f16> {
  host
    .iter()
    .map(|v| f16::from_bits(u16::from_ne_bytes(v.to_bits().to_be_bytes())))
    .collect()
}

/// Re-encode a host-native f32 slice as LE-encoded byte storage (the `grayf32le`
/// plane contract) for the `Grayf32` reference sink over the binned luma.
fn as_le_f32(host: &[f32]) -> Vec<f32> {
  host
    .iter()
    .map(|&v| f32::from_bits(v.to_bits().to_le()))
    .collect()
}

/// Interior f16 luma ramp mixing in-range, HDR (> 1.0), and negative values so
/// the area mean sees real variation per 2x2 block.
fn ramp() -> Vec<f16> {
  let mut y = vec![f16::ZERO; SRC * SRC];
  for (i, p) in y.iter_mut().enumerate() {
    *p = f16::from_f32(match i % 5 {
      0 => i as f32 / 64.0,
      1 => 1.0 + i as f32 / 8.0,
      2 => -(i as f32) / 100.0,
      3 => 0.5,
      _ => 2.5,
    });
  }
  y
}

/// Exact 2x2-block area mean of the *widened* f16 plane (each f16 → f32) to the
/// `OUT` grid. The engine accumulates in `f64` and finalizes `(acc / 4) as f32`
/// — bit-identical to the stream for the uniform 2:1 box.
fn block_mean_2x2(plane: &[f16]) -> Vec<f32> {
  let mut out = vec![0.0f32; OUT * OUT];
  for oy in 0..OUT {
    for ox in 0..OUT {
      let mut s = 0.0f64;
      for dy in 0..2 {
        for dx in 0..2 {
          s += plane[(oy * 2 + dy) * SRC + ox * 2 + dx].to_f32() as f64;
        }
      }
      out[oy * OUT + ox] = (s / 4.0) as f32;
    }
  }
  out
}

#[test]
#[cfg_attr(
  miri,
  ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
)]
fn grayf16_downscale_luma_f32_is_exact_area_mean() {
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let src = Grayf16Frame::new(&pix, SRC as u32, SRC as u32, SRC as u32);

  let mut luma_f32 = vec![0.0f32; OUT * OUT];
  {
    let mut sink =
      MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
        .unwrap()
        .with_luma_f32(&mut luma_f32)
        .unwrap();
    grayf16_to(&src, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  assert_eq!(
    luma_f32,
    block_mean_2x2(&plane),
    "luma_f32 must be the exact 2x2 f32 block mean of the widened f16 plane"
  );
}

#[test]
#[cfg_attr(
  miri,
  ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
)]
fn grayf16_all_outputs_match_direct_grayf32_over_binned_luma() {
  // Every attached output must be exactly what the direct `Grayf32` sink
  // produces over the (exact) binned f32 luma plane — the Grayf16 emit applies
  // the identical grayf32 kernels per finalized binned f32 row.
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let src = Grayf16Frame::new(&pix, SRC as u32, SRC as u32, SRC as u32);

  let mut rgb = vec![0u8; OUT * OUT * 3];
  let mut rgba = vec![0u8; OUT * OUT * 4];
  let mut rgb_u16 = vec![0u16; OUT * OUT * 3];
  let mut rgba_u16 = vec![0u16; OUT * OUT * 4];
  let mut luma = vec![0u8; OUT * OUT];
  let mut luma_u16 = vec![0u16; OUT * OUT];
  let mut rgb_f32 = vec![0.0f32; OUT * OUT * 3];
  let mut luma_f32 = vec![0.0f32; OUT * OUT];
  let mut h = vec![0u8; OUT * OUT];
  let mut s_ = vec![0u8; OUT * OUT];
  let mut v_ = vec![0u8; OUT * OUT];
  {
    let mut sink =
      MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
        .unwrap()
        .with_rgb(&mut rgb)
        .unwrap()
        .with_rgba(&mut rgba)
        .unwrap()
        .with_rgb_u16(&mut rgb_u16)
        .unwrap()
        .with_rgba_u16(&mut rgba_u16)
        .unwrap()
        .with_luma(&mut luma)
        .unwrap()
        .with_luma_u16(&mut luma_u16)
        .unwrap()
        .with_rgb_f32(&mut rgb_f32)
        .unwrap()
        .with_luma_f32(&mut luma_f32)
        .unwrap()
        .with_hsv(&mut h, &mut s_, &mut v_)
        .unwrap();
    grayf16_to(&src, FR, sink.set_kernel_matrix(M)).unwrap();
  }

  // Reference: the direct Grayf32 sink over the exact binned f32 luma plane.
  let binned = block_mean_2x2(&plane);
  let binned_pix = as_le_f32(&binned);
  let mut ref_rgb = vec![0u8; OUT * OUT * 3];
  let mut ref_rgba = vec![0u8; OUT * OUT * 4];
  let mut ref_rgb_u16 = vec![0u16; OUT * OUT * 3];
  let mut ref_rgba_u16 = vec![0u16; OUT * OUT * 4];
  let mut ref_luma = vec![0u8; OUT * OUT];
  let mut ref_luma_u16 = vec![0u16; OUT * OUT];
  let mut ref_rgb_f32 = vec![0.0f32; OUT * OUT * 3];
  let mut ref_luma_f32 = vec![0.0f32; OUT * OUT];
  let mut ref_h = vec![0u8; OUT * OUT];
  let mut ref_s = vec![0u8; OUT * OUT];
  let mut ref_v = vec![0u8; OUT * OUT];
  {
    let binned_frame = Grayf32Frame::new(&binned_pix, OUT as u32, OUT as u32, OUT as u32);
    let mut sink = MixedSinker::<Grayf32>::new(OUT, OUT)
      .with_rgb(&mut ref_rgb)
      .unwrap()
      .with_rgba(&mut ref_rgba)
      .unwrap()
      .with_rgb_u16(&mut ref_rgb_u16)
      .unwrap()
      .with_rgba_u16(&mut ref_rgba_u16)
      .unwrap()
      .with_luma(&mut ref_luma)
      .unwrap()
      .with_luma_u16(&mut ref_luma_u16)
      .unwrap()
      .with_rgb_f32(&mut ref_rgb_f32)
      .unwrap()
      .with_luma_f32(&mut ref_luma_f32)
      .unwrap()
      .with_hsv(&mut ref_h, &mut ref_s, &mut ref_v)
      .unwrap();
    crate::source::grayf32_to(&binned_frame, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  assert_eq!(luma_f32, ref_luma_f32, "luma_f32");
  assert_eq!(rgb_f32, ref_rgb_f32, "rgb_f32");
  assert_eq!(luma, ref_luma, "luma");
  assert_eq!(luma_u16, ref_luma_u16, "luma_u16");
  assert_eq!(rgb, ref_rgb, "rgb");
  assert_eq!(rgba, ref_rgba, "rgba");
  assert_eq!(rgb_u16, ref_rgb_u16, "rgb_u16");
  assert_eq!(rgba_u16, ref_rgba_u16, "rgba_u16");
  assert_eq!(h, ref_h, "h");
  assert_eq!(s_, ref_s, "s");
  assert_eq!(v_, ref_v, "v");
}

#[test]
#[cfg_attr(
  miri,
  ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
)]
fn grayf16_le_be_resample_outputs_identical() {
  // The binned f32 luma is host-native, so the derive kernels run with
  // `HOST_NATIVE_BE`. LE and BE wire encodings of the same logical plane must
  // resample to identical outputs.
  let plane = ramp();
  let pix_le = as_le_f16(&plane);
  let pix_be = as_be_f16(&plane);

  let mut le_luma_f32 = vec![0.0f32; OUT * OUT];
  let mut le_rgb_u16 = vec![0u16; OUT * OUT * 3];
  let mut le_rgba = vec![0u8; OUT * OUT * 4];
  {
    let frame = Grayf16Frame::new(&pix_le, SRC as u32, SRC as u32, SRC as u32);
    let mut sink =
      MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
        .unwrap()
        .with_luma_f32(&mut le_luma_f32)
        .unwrap()
        .with_rgb_u16(&mut le_rgb_u16)
        .unwrap()
        .with_rgba(&mut le_rgba)
        .unwrap();
    grayf16_to(&frame, FR, sink.set_kernel_matrix(M)).unwrap();
  }

  let mut be_luma_f32 = vec![0.0f32; OUT * OUT];
  let mut be_rgb_u16 = vec![0u16; OUT * OUT * 3];
  let mut be_rgba = vec![0u8; OUT * OUT * 4];
  {
    let frame = Grayf16Frame::<true>::new(&pix_be, SRC as u32, SRC as u32, SRC as u32);
    let mut sink = MixedSinker::<Grayf16<true>, AreaResampler>::with_resampler(
      SRC,
      SRC,
      AreaResampler::to(OUT, OUT),
    )
    .unwrap()
    .with_luma_f32(&mut be_luma_f32)
    .unwrap()
    .with_rgb_u16(&mut be_rgb_u16)
    .unwrap()
    .with_rgba(&mut be_rgba)
    .unwrap();
    grayf16_to_endian::<_, true>(&frame, FR, sink.set_kernel_matrix(M)).unwrap();
  }

  assert_eq!(le_luma_f32, be_luma_f32, "luma_f32 LE/BE diverge");
  assert_eq!(le_rgb_u16, be_rgb_u16, "rgb_u16 LE/BE diverge");
  assert_eq!(le_rgba, be_rgba, "rgba LE/BE diverge");
  assert_eq!(
    le_luma_f32,
    block_mean_2x2(&plane),
    "luma_f32 not area mean"
  );
}

#[test]
#[cfg_attr(
  miri,
  ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
)]
fn grayf16_standalone_rgba_matches_direct_over_binned_luma() {
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let src = Grayf16Frame::new(&pix, SRC as u32, SRC as u32, SRC as u32);

  let mut rgba = vec![0u8; OUT * OUT * 4];
  {
    let mut sink =
      MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
        .unwrap()
        .with_rgba(&mut rgba)
        .unwrap();
    grayf16_to(&src, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  let binned_pix = as_le_f32(&block_mean_2x2(&plane));
  let mut ref_rgba = vec![0u8; OUT * OUT * 4];
  {
    let binned = Grayf32Frame::new(&binned_pix, OUT as u32, OUT as u32, OUT as u32);
    let mut sink = MixedSinker::<Grayf32>::new(OUT, OUT)
      .with_rgba(&mut ref_rgba)
      .unwrap();
    crate::source::grayf32_to(&binned, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  assert_eq!(rgba, ref_rgba, "standalone rgba");
}

#[test]
#[cfg_attr(
  miri,
  ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
)]
fn grayf16_identity_plan_matches_new_sink() {
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let src = Grayf16Frame::new(&pix, SRC as u32, SRC as u32, SRC as u32);

  let mut direct = vec![0.0f32; SRC * SRC * 3];
  {
    let mut sink = MixedSinker::<Grayf16>::new(SRC, SRC)
      .with_rgb_f32(&mut direct)
      .unwrap();
    grayf16_to(&src, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  let mut via_area = vec![0.0f32; SRC * SRC * 3];
  {
    let mut sink =
      MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(SRC, SRC))
        .unwrap()
        .with_rgb_f32(&mut via_area)
        .unwrap();
    grayf16_to(&src, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  assert_eq!(direct, via_area, "identity plan must match the direct sink");
}

#[cfg_attr(
  miri,
  ignore = "half::f16 uses inline assembly on aarch64 unsupported by Miri"
)]
#[test]
fn grayf16_resample_no_outputs_is_a_no_op() {
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let src = Grayf16Frame::new(&pix, SRC as u32, SRC as u32, SRC as u32);
  let mut sink =
    MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
      .unwrap();
  grayf16_to(&src, FR, sink.set_kernel_matrix(M)).unwrap();
  assert!(
    !sink.luma_stream_f32_allocated(),
    "no-output sink allocated an f32 luma stream"
  );
}

#[cfg_attr(
  miri,
  ignore = "half::f16 uses inline assembly on aarch64 unsupported by Miri"
)]
#[test]
fn grayf16_out_of_sequence_first_row_rejected_before_allocation() {
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let row3 = &pix[3 * SRC..4 * SRC];

  let mut luma_f32 = vec![0.0f32; OUT * OUT];
  let mut sink =
    MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
      .unwrap()
      .with_luma_f32(&mut luma_f32)
      .unwrap();
  sink.begin_frame(SRC as u32, SRC as u32).unwrap();
  let err = sink
    .process(Grayf16Row::for_tests(row3, 3, M, FR))
    .unwrap_err();
  assert!(
    matches!(
      err,
      MixedSinkerError::Resample(ResampleError::OutOfSequenceRow(_))
    ),
    "expected OutOfSequenceRow, got {err:?}"
  );
  assert!(
    !sink.luma_stream_f32_allocated(),
    "stream allocated for a rejected row"
  );
  assert_eq!(
    sink.luma_scratch_f32_capacity(),
    0,
    "source-luma staging grown for a rejected row"
  );
  assert!(
    luma_f32.iter().all(|&b| b == 0.0),
    "rejected row mutated output"
  );
}

#[cfg_attr(
  miri,
  ignore = "half::f16 uses inline assembly on aarch64 unsupported by Miri"
)]
#[test]
fn grayf16_resample_reuses_luma_stream_across_frames() {
  let y1 = ramp();
  let mut y2 = y1.clone();
  for p in y2.iter_mut() {
    *p = f16::from_f32(3.0 - p.to_f32());
  }
  let pix1 = as_le_f16(&y1);
  let pix2 = as_le_f16(&y2);
  let mut luma_f32 = vec![0.0f32; OUT * OUT];
  {
    let mut sink =
      MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
        .unwrap()
        .with_luma_f32(&mut luma_f32)
        .unwrap();
    grayf16_to(
      &Grayf16Frame::new(&pix1, SRC as u32, SRC as u32, SRC as u32),
      FR,
      sink.set_kernel_matrix(M),
    )
    .unwrap();
    grayf16_to(
      &Grayf16Frame::new(&pix2, SRC as u32, SRC as u32, SRC as u32),
      FR,
      sink.set_kernel_matrix(M),
    )
    .unwrap();
  }
  assert_eq!(
    luma_f32,
    block_mean_2x2(&y2),
    "frame 2 luma_f32 must area-downscale frame 2's widened luma"
  );
}

#[cfg_attr(
  miri,
  ignore = "half::f16 uses inline assembly on aarch64 unsupported by Miri"
)]
#[test]
fn grayf16_resample_rejects_mid_frame_output_change() {
  let plane = ramp();
  let pix = as_le_f16(&plane);
  let mut rgb_f32 = vec![0.0f32; OUT * OUT * 3];
  let mut luma_f32 = vec![0.0f32; OUT * OUT];
  let mut sink =
    MixedSinker::<Grayf16, AreaResampler>::with_resampler(SRC, SRC, AreaResampler::to(OUT, OUT))
      .unwrap()
      .with_rgb_f32(&mut rgb_f32)
      .unwrap();
  sink.begin_frame(SRC as u32, SRC as u32).unwrap();
  sink
    .process(Grayf16Row::for_tests(&pix[..SRC], 0, M, FR))
    .unwrap();
  sink.set_luma_f32(&mut luma_f32).unwrap();
  let err = sink
    .process(Grayf16Row::for_tests(&pix[SRC..2 * SRC], 1, M, FR))
    .unwrap_err();
  assert!(
    matches!(err, MixedSinkerError::ResampleOutputsChanged(_)),
    "expected ResampleOutputsChanged, got {err:?}"
  );
  assert!(
    luma_f32.iter().all(|&b| b == 0.0),
    "rejected row mutated the new output"
  );
}