scenesdetect 0.1.0

Scene/shot cut detection ported from PySceneDetect — Sans-I/O streaming API with SIMD-accelerated detectors for histogram, pHash, threshold, content, and adaptive algorithms.
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
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
//! Perceptual hash (pHash) scene detection via DCT signatures.
//!
//! This module implements [`Detector`](crate::phash::Detector), a port of
//! PySceneDetect's `detect-hash` algorithm. Where
//! [`histogram::Detector`](crate::histogram::Detector) looks at *brightness
//! distribution*, the pHash detector looks at *spatial structure*: a cut
//! fires when the low-frequency DCT signature of the frame changes
//! significantly.
//!
//! # Algorithm
//!
//! For each incoming [`LumaFrame`](crate::frame::LumaFrame):
//!
//! 1. **Resize** the Y plane to `imsize × imsize` (where `imsize = size *
//!    lowpass`) using area-weighted downsampling.
//! 2. **Normalize** to `[0, 1]` by dividing by the max sample.
//! 3. **2D DCT-II** (orthonormal, matching OpenCV's `cv2.dct` scaling) on
//!    the resized image.
//! 4. **Crop** to the top-left `size × size` low-frequency block.
//! 5. **Median threshold:** set bit `i` iff that coefficient is strictly
//!    greater than the block's median.
//!
//! The resulting `size²` bits are the frame's pHash. Between consecutive
//! frames, the normalized Hamming distance
//! `popcount(h1 ^ h2) / (size²)` is compared against `threshold`; a cut is
//! emitted when it is `>=` and at least `min_duration` has elapsed since the
//! previous cut.
//!
//! Default parameters (`size=16`, `lowpass=2`) → resize to `32 × 32`, DCT,
//! then a `16 × 16 = 256`-bit fingerprint per frame. Comparison cost is a
//! handful of `XOR` + `popcount` instructions.
//!
//! # Attribution
//!
//! Based on Neal Krawetz's DCT-based pHash (2011) and Johannes Buchner's
//! `imagehash` library. Directly ported from PySceneDetect's `detect-hash`
//! (BSD 3-Clause).

use core::{f32::consts::PI, time::Duration};
use derive_more::IsVariant;
use thiserror::Error;

use crate::frame::{LumaFrame, Timebase, Timestamp};

#[cfg(feature = "serde")]
use serde::{Deserialize, Serialize};

use std::{vec, vec::Vec};

use super::{ceil_32, cos_32, floor_32, sqrt_32};

/// Configuration for [`Detector`].
#[derive(Debug, Clone)]
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
pub struct Options {
  threshold: f64,
  size: u32,
  lowpass: u32,
  #[cfg_attr(feature = "serde", serde(with = "humantime_serde"))]
  min_duration: Duration,
  initial_cut: bool,
}

impl Default for Options {
  #[cfg_attr(not(tarpaulin), inline(always))]
  fn default() -> Self {
    Self::new()
  }
}

impl Options {
  /// Creates a new [`Options`] with the specified parameters.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn new() -> Self {
    Self {
      threshold: 0.395,
      size: 16,
      lowpass: 2,
      min_duration: Duration::from_secs(1),
      initial_cut: true,
    }
  }

  /// Returns the threshold for scene change detection. Higher values are more sensitive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn threshold(&self) -> f64 {
    self.threshold
  }

  /// Sets the scene change threshold. Higher values are more sensitive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn with_threshold(mut self, threshold: f64) -> Self {
    self.set_threshold(threshold);
    self
  }

  /// Sets the scene change threshold. Higher values are more sensitive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn set_threshold(&mut self, threshold: f64) -> &mut Self {
    self.threshold = threshold;
    self
  }

  /// Returns the hash size. Higher values are more sensitive but more expensive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn size(&self) -> u32 {
    self.size
  }

  /// Sets the hash size. Higher values are more sensitive but more expensive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn with_size(mut self, size: u32) -> Self {
    self.set_size(size);
    self
  }

  /// Sets the hash size. Higher values are more sensitive but more expensive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn set_size(&mut self, size: u32) -> &mut Self {
    self.size = size;
    self
  }

  /// Returns the lowpass filter size used to smooth the image before hashing. Higher values are more sensitive but more expensive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn lowpass(&self) -> u32 {
    self.lowpass
  }

  /// Sets the lowpass filter size. Higher values are more sensitive but more expensive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn with_lowpass(mut self, lowpass: u32) -> Self {
    self.set_lowpass(lowpass);
    self
  }

  /// Sets the lowpass filter size. Higher values are more sensitive but more expensive.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn set_lowpass(&mut self, lowpass: u32) -> &mut Self {
    self.lowpass = lowpass;
    self
  }

  /// Returns the minimum scene duration. Shorter scenes are ignored.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn min_duration(&self) -> Duration {
    self.min_duration
  }

  /// Sets the minimum scene duration. Shorter scenes are ignored.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn with_min_duration(mut self, min_duration: Duration) -> Self {
    self.set_min_duration(min_duration);
    self
  }

  /// Sets the minimum scene duration. Shorter scenes are ignored.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn set_min_duration(&mut self, min_duration: Duration) -> &mut Self {
    self.min_duration = min_duration;
    self
  }

  /// Set the minimum scene length as a number of frames at a given frame rate.
  ///
  /// Convenience for users coming from frame-count APIs (e.g., PySceneDetect's
  /// `min_scene_len`). Internally this converts to [`Self::min_duration`] via
  /// [`Timebase::frames_to_duration`]. On VFR content the duration stays fixed
  /// while frame counts drift — that's the desired behavior.
  ///
  /// `fps` is interpreted as frames per second: 30 fps = `Timebase::new(30, 1)`,
  /// NTSC = `Timebase::new(30000, 1001)`.
  ///
  /// # Panics
  ///
  /// Panics if `fps.num() == 0`.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn with_min_frames(mut self, frames: u32, fps: Timebase) -> Self {
    self.set_min_frames(frames, fps);
    self
  }

  /// In-place form of [`Self::with_min_frames`].
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn set_min_frames(&mut self, frames: u32, fps: Timebase) -> &mut Self {
    self.min_duration = fps.frames_to_duration(frames);
    self
  }

  /// Whether the first detected cut is allowed to fire immediately.
  ///
  /// - `true` (default): the first detected cut fires as soon as the
  ///   normalized Hamming distance exceeds `threshold`.
  /// - `false`: suppresses cuts until the stream has actually run for at
  ///   least [`Self::min_duration`]. Matches PySceneDetect's default.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn initial_cut(&self) -> bool {
    self.initial_cut
  }

  /// Sets whether the first detected cut may fire immediately.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn with_initial_cut(mut self, val: bool) -> Self {
    self.initial_cut = val;
    self
  }

  /// Sets `initial_cut` in place.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn set_initial_cut(&mut self, val: bool) -> &mut Self {
    self.initial_cut = val;
    self
  }
}

/// Error returned by [`Detector::try_new`] when the provided [`Options`] are
/// inconsistent.
#[derive(Debug, Clone, PartialEq, Eq, IsVariant, Error)]
#[non_exhaustive]
pub enum Error {
  /// `options.size() < 2`. The algorithm needs at least a `2 × 2` hash block
  /// to have a meaningful median threshold.
  #[error("phash size ({size}) must be >= 2")]
  SizeTooSmall {
    /// The provided size.
    size: u32,
  },
  /// `options.lowpass() < 1`. The resize multiplier must be at least 1 so
  /// that `imsize = size * lowpass >= size`.
  #[error("phash lowpass ({lowpass}) must be >= 1")]
  LowpassTooSmall {
    /// The provided lowpass multiplier.
    lowpass: u32,
  },
  /// `size * lowpass` or its square would exceed `usize`. Only reachable
  /// with pathological values on 32-bit targets.
  #[error("phash dimensions overflow usize: size ({size}) * lowpass ({lowpass}) squared")]
  DimensionsOverflow {
    /// The provided size.
    size: u32,
    /// The provided lowpass multiplier.
    lowpass: u32,
  },
}

/// Perceptual-hash scene detector. See the
/// [module-level documentation](crate::phash) for the algorithm.
///
/// After construction the detector allocates nothing per frame: the DCT
/// cosine basis matrix is precomputed, and scratch buffers for the resized
/// image, the DCT intermediate/result, the low-frequency block, and a sort
/// scratch for the median are all reused.
#[derive(Debug, Clone)]
pub struct Detector {
  options: Options,
  /// `size * lowpass` — side length of the resized square image.
  imsize: usize,
  /// `options.size` as `usize` — side length of the low-frequency block.
  size: usize,
  /// `options.threshold` cached as f64 for fast comparison.
  threshold: f64,
  /// Precomputed orthonormal DCT-II basis: `dct_cos[k*imsize + n] = α(k) · cos(π(2n+1)k / 2N)`.
  dct_cos: Vec<f32>,
  /// Area-weighted resize weights. Lazily built on the first frame, then
  /// reused across frames of matching dimensions. Rebuilt if the input
  /// resolution changes mid-stream (seeks, adaptive bitrate).
  resize_table: ResizeTable,
  /// Resized (`imsize × imsize`) and normalized (`[0, 1]`) image.
  resized: Vec<f32>,
  /// Row-transformed intermediate for the 2D DCT.
  dct_tmp: Vec<f32>,
  /// Full 2D DCT result.
  dct_result: Vec<f32>,
  /// Flattened `size × size` low-frequency crop (order preserved for bit packing).
  low_freq: Vec<f32>,
  /// Sort scratch for the median — avoids disturbing `low_freq`.
  sort_scratch: Vec<f32>,
  /// Packed bits of the current frame's hash; `len = ceil(size² / 64)`.
  current_hash: Vec<u64>,
  /// Packed bits of the previous frame's hash.
  previous_hash: Vec<u64>,
  has_previous: bool,
  last_cut_ts: Option<Timestamp>,
  last_distance: Option<f64>,
}

impl Detector {
  /// Creates a new detector with the given options, validating them.
  ///
  /// Prefer [`Self::try_new`] at runtime call sites where invalid options
  /// are possible; this constructor is meant for call sites where the
  /// options are statically known-good (tests, fixtures, defaults).
  ///
  /// # Panics
  ///
  /// Panics if the options are invalid — see [`enum@Error`] for the specific
  /// conditions.
  pub fn new(options: Options) -> Self {
    Self::try_new(options).expect("invalid phash Options")
  }

  /// Creates a new detector with the given options, returning [`enum@Error`] if
  /// the options are inconsistent.
  ///
  /// Validates:
  /// - `options.size() >= 2` (need a non-trivial hash block)
  /// - `options.lowpass() >= 1` (need at least unit resize)
  /// - `size * lowpass * size * lowpass` fits in `usize` (avoids overflow
  ///   when sizing scratch buffers on 32-bit targets)
  ///
  /// Precomputes the DCT basis and allocates all scratch buffers on success.
  pub fn try_new(options: Options) -> Result<Self, Error> {
    if options.size < 2 {
      return Err(Error::SizeTooSmall { size: options.size });
    }
    if options.lowpass < 1 {
      return Err(Error::LowpassTooSmall {
        lowpass: options.lowpass,
      });
    }

    let size = options.size as usize;
    let lowpass = options.lowpass as usize;
    let imsize = match size.checked_mul(lowpass) {
      Some(v) => v,
      None => {
        return Err(Error::DimensionsOverflow {
          size: options.size,
          lowpass: options.lowpass,
        });
      }
    };
    let total = match imsize.checked_mul(imsize) {
      Some(v) => v,
      None => {
        return Err(Error::DimensionsOverflow {
          size: options.size,
          lowpass: options.lowpass,
        });
      }
    };

    let threshold = options.threshold;
    let bits = size * size;
    let hash_words = bits.div_ceil(64);
    let dct_cos = build_dct_cos(imsize);

    Ok(Self {
      options,
      imsize,
      size,
      threshold,
      dct_cos,
      resize_table: ResizeTable::new(),
      resized: vec![0.0f32; total],
      dct_tmp: vec![0.0f32; total],
      dct_result: vec![0.0f32; total],
      low_freq: vec![0.0f32; bits],
      sort_scratch: vec![0.0f32; bits],
      current_hash: vec![0u64; hash_words],
      previous_hash: vec![0u64; hash_words],
      has_previous: false,
      last_cut_ts: None,
      last_distance: None,
    })
  }

  /// Returns a reference to the options used by this detector.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn options(&self) -> &Options {
    &self.options
  }

  /// Returns the normalized Hamming distance between the last two frames'
  /// hashes, or `None` if fewer than two frames have been processed.
  ///
  /// Range: `[0.0, 1.0]`. `0.0` means identical hashes; `1.0` means every
  /// bit flipped. Useful for logging / diagnostics.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub const fn last_distance(&self) -> Option<f64> {
    self.last_distance
  }

  /// Resets the detector's streaming state so it can be reused on a fresh
  /// stream (e.g., when the next video begins) without rebuilding the DCT
  /// basis or reallocating scratch buffers.
  ///
  /// After `clear()` the next [`Self::process`] call is treated as if it
  /// were the first frame of a new stream: no cut is emitted, and the frame
  /// re-seeds `last_cut_ts`. The previous video's hashes, `last_cut_ts`,
  /// and `last_distance` are all discarded.
  ///
  /// The resize table is kept. It will reuse its weights if the new stream
  /// has the same resolution, or auto-rebuild on the first frame otherwise.
  #[cfg_attr(not(tarpaulin), inline(always))]
  pub fn clear(&mut self) {
    self.has_previous = false;
    self.last_cut_ts = None;
    self.last_distance = None;
  }

  /// Processes the next frame. Returns `Some(ts)` if a cut is detected at
  /// the frame's timestamp, otherwise `None`.
  ///
  /// The first frame establishes the baseline hash and cut-gating reference;
  /// no cut is emitted for it.
  pub fn process(&mut self, frame: LumaFrame<'_>) -> Option<Timestamp> {
    let ts = frame.timestamp();

    if self.last_cut_ts.is_none() {
      self.last_cut_ts = Some(if self.options.initial_cut {
        ts.saturating_sub_duration(self.options.min_duration)
      } else {
        ts
      });
    }

    self.compute_hash(&frame);

    let mut cut: Option<Timestamp> = None;
    if self.has_previous {
      let dist = hamming_distance(&self.previous_hash, &self.current_hash);
      let bits = self.size * self.size;
      let norm = dist as f64 / bits as f64;
      self.last_distance = Some(norm);

      let min_elapsed = self
        .last_cut_ts
        .as_ref()
        .and_then(|last| ts.duration_since(last))
        .is_some_and(|d| d >= self.options.min_duration);

      if norm >= self.threshold && min_elapsed {
        cut = Some(ts);
        self.last_cut_ts = Some(ts);
      }
    }

    core::mem::swap(&mut self.current_hash, &mut self.previous_hash);
    self.has_previous = true;
    cut
  }

  /// Builds the current frame's hash into `self.current_hash`.
  fn compute_hash(&mut self, frame: &LumaFrame<'_>) {
    // 1. Ensure resize table matches the frame dimensions. This rebuilds on
    //    the first frame and on any subsequent dimension change. For a CFR
    //    stream this cost is paid once.
    self
      .resize_table
      .ensure(frame.width(), frame.height(), self.imsize);

    // 2. Area-weighted downsample, returning `max` in the same pass so we
    //    fold the normalization pre-scan into the resize loop.
    let max = self.resize_table.apply(
      &mut self.resized,
      frame.data(),
      frame.stride() as usize,
      self.imsize,
    );

    // 3. Normalize by max. Second pass over the 1 KiB `resized` buffer.
    let scale = if max == 0.0 { 1.0 } else { 1.0 / max };
    for v in self.resized.iter_mut() {
      *v *= scale;
    }

    // 4. 2D DCT-II (orthonormal, matching cv2.dct).
    dct2(
      &self.dct_cos,
      &self.resized,
      &mut self.dct_tmp,
      &mut self.dct_result,
      self.imsize,
    );

    // 5. Crop top-left size×size block into a flat buffer.
    for y in 0..self.size {
      let src_row = &self.dct_result[y * self.imsize..y * self.imsize + self.size];
      let dst_row = &mut self.low_freq[y * self.size..(y + 1) * self.size];
      dst_row.copy_from_slice(src_row);
    }

    // 6. Median via O(N) quick-select on sort_scratch (preserves `low_freq`).
    self.sort_scratch.clone_from(&self.low_freq);
    let median = median_f32(&mut self.sort_scratch);

    // 7. Pack bits: bit i set iff low_freq[i] > median. Bit 0 = (0,0) = DC term.
    self.current_hash.fill(0);
    for (i, &v) in self.low_freq.iter().enumerate() {
      if v > median {
        self.current_hash[i / 64] |= 1u64 << (i % 64);
      }
    }
  }
}

/// Builds the orthonormal DCT-II basis: `C[k, n] = α(k) · cos(π(2n+1)k / 2N)`,
/// where `α(0) = 1/√N` and `α(k≠0) = √(2/N)`. This matches `cv2.dct`.
fn build_dct_cos(n: usize) -> Vec<f32> {
  let mut c = vec![0.0f32; n * n];
  let alpha0 = sqrt_32(1.0 / n as f32);
  let alpha_k = sqrt_32(2.0 / n as f32);
  for k in 0..n {
    let a = if k == 0 { alpha0 } else { alpha_k };
    for m in 0..n {
      let angle = PI * (2.0 * m as f32 + 1.0) * k as f32 / (2.0 * n as f32);
      c[k * n + m] = a * cos_32(angle);
    }
  }
  c
}

/// Separable 2D DCT-II: `result = C · input · Cᵀ`.
/// `tmp` is a scratch buffer of size `n*n`.
fn dct2(c: &[f32], input: &[f32], tmp: &mut [f32], result: &mut [f32], n: usize) {
  debug_assert_eq!(c.len(), n * n);
  debug_assert_eq!(input.len(), n * n);
  debug_assert_eq!(tmp.len(), n * n);
  debug_assert_eq!(result.len(), n * n);

  // tmp = input · Cᵀ   (row transform; output column j = Σ_k input[m, k] · C[j, k])
  for m in 0..n {
    for j in 0..n {
      let mut s = 0.0f32;
      for k in 0..n {
        s += input[m * n + k] * c[j * n + k];
      }
      tmp[m * n + j] = s;
    }
  }
  // result = C · tmp    (column transform; output[k, j] = Σ_m C[k, m] · tmp[m, j])
  for k in 0..n {
    for j in 0..n {
      let mut s = 0.0f32;
      for m in 0..n {
        s += c[k * n + m] * tmp[m * n + j];
      }
      result[k * n + j] = s;
    }
  }
}

/// Precomputed area-weighted resize weights for a fixed
/// `src_{w,h} → dst_size × dst_size` mapping.
///
/// Factors the 2D area weight as a product of 1D horizontal and vertical
/// overlap fractions. For each destination row / column, we store a
/// contiguous run of `(src_idx, weight)` pairs, indexed via prefix-sum
/// `x_range_starts` / `y_range_starts`. Empty `(src_w = 0, src_h = 0)`
/// is the "not yet built" sentinel — [`Self::ensure`] detects it.
#[derive(Debug, Clone)]
struct ResizeTable {
  src_w: u32,
  src_h: u32,
  inv_area: f32,
  /// Source column indices contributing to each destination column, flattened.
  x_offsets: Vec<u32>,
  x_weights: Vec<f32>,
  /// Prefix sum; `x_range_starts[dst_x]..x_range_starts[dst_x+1]` indexes
  /// the contiguous run of pairs for destination column `dst_x`. Length
  /// `dst_size + 1`.
  x_range_starts: Vec<u32>,
  /// Same, for rows.
  y_offsets: Vec<u32>,
  y_weights: Vec<f32>,
  y_range_starts: Vec<u32>,
}

impl ResizeTable {
  /// Creates an empty (not-yet-built) table.
  fn new() -> Self {
    Self {
      src_w: 0,
      src_h: 0,
      inv_area: 0.0,
      x_offsets: Vec::new(),
      x_weights: Vec::new(),
      x_range_starts: Vec::new(),
      y_offsets: Vec::new(),
      y_weights: Vec::new(),
      y_range_starts: Vec::new(),
    }
  }

  /// Ensures the table matches the given dimensions, rebuilding if needed.
  ///
  /// Fast path when dimensions are unchanged: single comparison, no work.
  fn ensure(&mut self, src_w: u32, src_h: u32, dst_size: usize) {
    if self.src_w == src_w && self.src_h == src_h {
      return;
    }
    self.rebuild(src_w, src_h, dst_size);
  }

  /// Rebuilds the table for the given dimensions. Reuses existing `Vec`
  /// capacity via `clear` — no heap churn after the first resolution.
  fn rebuild(&mut self, src_w: u32, src_h: u32, dst_size: usize) {
    debug_assert!(src_w > 0 && src_h > 0, "source dimensions must be non-zero");
    debug_assert!(dst_size > 0);

    self.x_offsets.clear();
    self.x_weights.clear();
    self.x_range_starts.clear();
    self.y_offsets.clear();
    self.y_weights.clear();
    self.y_range_starts.clear();

    let scale_x = src_w as f32 / dst_size as f32;
    let scale_y = src_h as f32 / dst_size as f32;

    build_axis(
      &mut self.x_offsets,
      &mut self.x_weights,
      &mut self.x_range_starts,
      src_w,
      dst_size,
      scale_x,
    );
    build_axis(
      &mut self.y_offsets,
      &mut self.y_weights,
      &mut self.y_range_starts,
      src_h,
      dst_size,
      scale_y,
    );

    self.inv_area = 1.0 / (scale_x * scale_y);
    self.src_w = src_w;
    self.src_h = src_h;
  }

  /// Applies the table to an 8-bit source plane, writing f32 values into
  /// `dst` and returning the max value seen — so the normalization pre-scan
  /// is folded into this single pass.
  fn apply(&self, dst: &mut [f32], src: &[u8], src_stride: usize, dst_size: usize) -> f32 {
    debug_assert_eq!(dst.len(), dst_size * dst_size);
    debug_assert_eq!(self.x_range_starts.len(), dst_size + 1);
    debug_assert_eq!(self.y_range_starts.len(), dst_size + 1);

    let mut max = 0.0f32;

    for dst_y in 0..dst_size {
      let y_start = self.y_range_starts[dst_y] as usize;
      let y_end = self.y_range_starts[dst_y + 1] as usize;

      for dst_x in 0..dst_size {
        let x_start = self.x_range_starts[dst_x] as usize;
        let x_end = self.x_range_starts[dst_x + 1] as usize;

        let mut sum = 0.0f32;
        for yi in y_start..y_end {
          let sy = self.y_offsets[yi] as usize;
          let wy = self.y_weights[yi];
          let row_off = sy * src_stride;
          let mut row_sum = 0.0f32;
          for xi in x_start..x_end {
            let sx = self.x_offsets[xi] as usize;
            row_sum += (src[row_off + sx] as f32) * self.x_weights[xi];
          }
          sum += row_sum * wy;
        }

        let v = sum * self.inv_area;
        dst[dst_y * dst_size + dst_x] = v;
        if v > max {
          max = v;
        }
      }
    }

    max
  }
}

/// Populates one axis (horizontal or vertical) of a resize table. Pushes
/// `(src_idx, weight)` pairs to `offsets`/`weights` and `range_starts`
/// entries such that `range_starts[dst]..range_starts[dst+1]` is the run of
/// pairs for destination index `dst`. The final `range_starts.len()` is
/// `dst_size + 1` (prefix-sum style — last entry is the total length).
fn build_axis(
  offsets: &mut Vec<u32>,
  weights: &mut Vec<f32>,
  range_starts: &mut Vec<u32>,
  src_size: u32,
  dst_size: usize,
  scale: f32,
) {
  for dst in 0..dst_size {
    range_starts.push(offsets.len() as u32);
    let a = dst as f32 * scale;
    let b = (dst + 1) as f32 * scale;
    let s_start = floor_32(a) as u32;
    let s_end = (ceil_32(b) as u32).min(src_size);
    for s in s_start..s_end {
      let w = ((s + 1) as f32).min(b) - (s as f32).max(a);
      if w > 0.0 {
        offsets.push(s);
        weights.push(w);
      }
    }
  }
  range_starts.push(offsets.len() as u32);
}

/// Median of a slice in O(N) via quick-select. Destroys the input order.
///
/// For odd `n`, returns the (`n/2`)th order statistic directly. For even
/// `n`, returns the average of the (`n/2 − 1`)th and (`n/2`)th — matching
/// `numpy.median` and therefore PySceneDetect.
fn median_f32(buf: &mut [f32]) -> f32 {
  let n = buf.len();
  debug_assert!(n > 0);
  if n == 1 {
    return buf[0];
  }
  let mid = n / 2;
  let (left, pivot, _right) = buf.select_nth_unstable_by(mid, |a, b| a.total_cmp(b));
  let m2 = *pivot;
  if n % 2 == 1 {
    m2
  } else {
    // Even length: also need the (mid − 1)th order statistic, which is the
    // max of the left partition produced by the select above.
    let m1 = left.iter().copied().fold(f32::NEG_INFINITY, f32::max);
    (m1 + m2) / 2.0
  }
}

/// Hamming distance between two equal-length bit strings stored as `u64` words.
#[cfg_attr(not(tarpaulin), inline(always))]
fn hamming_distance(a: &[u64], b: &[u64]) -> u32 {
  debug_assert_eq!(a.len(), b.len());
  a.iter()
    .zip(b.iter())
    .map(|(x, y)| (x ^ y).count_ones())
    .sum()
}

#[cfg(all(test, feature = "std"))]
mod tests {
  use super::*;
  use crate::frame::Timebase;
  use core::num::NonZeroU32;
  use std::{vec, vec::Vec};

  const fn nz32(n: u32) -> NonZeroU32 {
    match NonZeroU32::new(n) {
      Some(v) => v,
      None => panic!("zero"),
    }
  }

  fn make_frame<'a>(data: &'a [u8], w: u32, h: u32, pts: i64) -> LumaFrame<'a> {
    let tb = Timebase::new(1, nz32(1000));
    LumaFrame::new(data, w, h, w, Timestamp::new(pts, tb))
  }

  #[test]
  fn with_min_frames_matches_python_default() {
    // PySceneDetect's default is 15 frames; at 30 fps that's 500 ms.
    let fps = Timebase::new(30, nz32(1));
    let opts = Options::default().with_min_frames(15, fps);
    assert_eq!(opts.min_duration(), Duration::from_millis(500));
  }

  #[test]
  fn with_min_frames_ntsc() {
    let fps = Timebase::new(30_000, nz32(1001));
    let opts = Options::default().with_min_frames(15, fps);
    assert_eq!(opts.min_duration(), Duration::from_nanos(500_500_000));
  }

  #[test]
  fn try_new_success() {
    let det = Detector::try_new(Options::default()).expect("defaults are valid");
    assert_eq!(det.options().size(), 16);
    assert_eq!(det.options().lowpass(), 2);
  }

  #[test]
  fn try_new_rejects_size_too_small() {
    let opts = Options::default().with_size(1);
    let err = Detector::try_new(opts).expect_err("should fail");
    assert_eq!(err, Error::SizeTooSmall { size: 1 });

    let opts = Options::default().with_size(0);
    let err = Detector::try_new(opts).expect_err("should fail");
    assert_eq!(err, Error::SizeTooSmall { size: 0 });
  }

  #[test]
  fn try_new_rejects_lowpass_zero() {
    let opts = Options::default().with_lowpass(0);
    let err = Detector::try_new(opts).expect_err("should fail");
    assert_eq!(err, Error::LowpassTooSmall { lowpass: 0 });
  }

  #[test]
  #[should_panic(expected = "invalid phash Options")]
  fn new_panics_on_invalid() {
    let _ = Detector::new(Options::default().with_size(1));
  }

  #[test]
  fn error_display() {
    let e = Error::SizeTooSmall { size: 1 };
    assert_eq!(format!("{e}"), "phash size (1) must be >= 2");
    let e = Error::LowpassTooSmall { lowpass: 0 };
    assert_eq!(format!("{e}"), "phash lowpass (0) must be >= 1");
  }

  #[test]
  fn hamming_distance_basic() {
    assert_eq!(hamming_distance(&[0, 0], &[0, 0]), 0);
    assert_eq!(hamming_distance(&[0xFF, 0], &[0, 0]), 8);
    assert_eq!(hamming_distance(&[!0u64, !0u64], &[0, 0]), 128);
    assert_eq!(hamming_distance(&[0b1010_1010], &[0b0101_0101]), 8);
  }

  #[test]
  fn build_dct_cos_is_orthonormal() {
    // C · Cᵀ should be the identity for the orthonormal DCT basis.
    let n = 8;
    let c = build_dct_cos(n);
    for i in 0..n {
      for j in 0..n {
        let mut s = 0.0f32;
        for k in 0..n {
          s += c[i * n + k] * c[j * n + k];
        }
        let expected = if i == j { 1.0 } else { 0.0 };
        assert!(
          (s - expected).abs() < 1e-5,
          "C·Cᵀ at ({i},{j}) = {s}, want {expected}",
        );
      }
    }
  }

  #[test]
  fn dct_dc_of_constant_input() {
    // DCT of a constant signal: all energy in the DC bin (0, 0).
    let n = 8;
    let c = build_dct_cos(n);
    let input = vec![1.0f32; n * n];
    let mut tmp = vec![0.0f32; n * n];
    let mut result = vec![0.0f32; n * n];
    dct2(&c, &input, &mut tmp, &mut result, n);
    // DC = α(0)² · n · n · 1 = (1/√n)² · n · n = n  (for each dim)
    // 2D DC = n · α(0)² · n = n for 1D, squared for 2D = n
    // Actually: for orthonormal 2D DCT of constant 1: Y[0,0] = n (since α(0) = 1/√n
    // and summing n values gives n/√n = √n per dim, then 2D = n).
    assert!((result[0] - n as f32).abs() < 1e-4, "DC = {}", result[0]);
    // All other coefficients ≈ 0.
    (1..n * n).for_each(|k| {
      assert!(result[k].abs() < 1e-4, "AC [{k}] = {}", result[k]);
    });
  }

  #[test]
  fn resize_area_identity() {
    // 4x4 → 4x4 is a no-op.
    let src = [
      10u8, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160,
    ];
    let mut dst = vec![0.0f32; 16];
    let mut table = ResizeTable::new();
    table.ensure(4, 4, 4);
    let max = table.apply(&mut dst, &src, 4, 4);
    for i in 0..16 {
      assert!((dst[i] - src[i] as f32).abs() < 1e-5);
    }
    assert!((max - 160.0).abs() < 1e-5);
  }

  #[test]
  fn resize_area_halve() {
    // 4x4 → 2x2 with a known input — each dest pixel is the average of a 2x2 source block.
    let src = [
      10u8, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160,
    ];
    let mut dst = vec![0.0f32; 4];
    let mut table = ResizeTable::new();
    table.ensure(4, 4, 2);
    let max = table.apply(&mut dst, &src, 4, 2);
    assert!((dst[0] - (10.0 + 20.0 + 50.0 + 60.0) / 4.0).abs() < 1e-4);
    assert!((dst[1] - (30.0 + 40.0 + 70.0 + 80.0) / 4.0).abs() < 1e-4);
    assert!((dst[2] - (90.0 + 100.0 + 130.0 + 140.0) / 4.0).abs() < 1e-4);
    assert!((dst[3] - (110.0 + 120.0 + 150.0 + 160.0) / 4.0).abs() < 1e-4);
    // apply() returns the max — equals the largest destination pixel.
    assert!((max - 135.0).abs() < 1e-4);
  }

  #[test]
  fn resize_table_rebuild_on_dim_change() {
    let mut table = ResizeTable::new();
    // First build.
    table.ensure(1920, 1080, 32);
    let counts_first = (table.x_offsets.len(), table.y_offsets.len());
    // Same dims — fast no-op.
    table.ensure(1920, 1080, 32);
    assert_eq!(table.x_offsets.len(), counts_first.0);
    // Changed dims — rebuild. Weight counts differ for different src size.
    table.ensure(1280, 720, 32);
    assert_ne!(table.x_offsets.len(), counts_first.0);
    assert_eq!(table.src_w, 1280);
    assert_eq!(table.src_h, 720);
  }

  #[test]
  fn median_odd_and_even() {
    // Odd length: returns the middle element.
    let mut v = [5.0f32, 1.0, 3.0, 2.0, 4.0];
    assert_eq!(median_f32(&mut v), 3.0);
    // Even length: returns average of the two middle elements.
    let mut v = [5.0f32, 1.0, 3.0, 2.0, 4.0, 6.0];
    assert_eq!(median_f32(&mut v), (3.0 + 4.0) / 2.0);
  }

  #[test]
  fn identical_frames_produce_no_cut() {
    let mut det = Detector::new(Options::default());
    // A frame with spatial variation (not flat — we want a meaningful DCT).
    let mut buf = vec![0u8; 128 * 96];
    for (i, b) in buf.iter_mut().enumerate() {
      *b = ((i * 7) % 256) as u8;
    }
    assert!(det.process(make_frame(&buf, 128, 96, 0)).is_none());
    assert!(det.process(make_frame(&buf, 128, 96, 2000)).is_none());
    assert!(det.process(make_frame(&buf, 128, 96, 4000)).is_none());
    assert_eq!(det.last_distance(), Some(0.0));
  }

  /// Returns (top/bottom-half, left/right-half) test frames — orthogonal
  /// low-frequency structures that land clearly inside the 16×16 low-freq
  /// DCT block, so the hashes differ reliably.
  fn ortho_halves_frames() -> (Vec<u8>, Vec<u8>) {
    let mut top_bottom = vec![0u8; 128 * 96];
    for y in 0..96 {
      for x in 0..128 {
        top_bottom[y * 128 + x] = if y < 48 { 220 } else { 30 };
      }
    }
    let mut left_right = vec![0u8; 128 * 96];
    for y in 0..96 {
      for x in 0..128 {
        left_right[y * 128 + x] = if x < 64 { 220 } else { 30 };
      }
    }
    (top_bottom, left_right)
  }

  #[test]
  fn very_different_frames_produce_cut() {
    // Use min_duration=0 so the gate can't mask the cut.
    let opts = Options::default().with_min_duration(Duration::from_millis(0));
    let mut det = Detector::new(opts);

    let (a, b) = ortho_halves_frames();

    assert!(det.process(make_frame(&a, 128, 96, 0)).is_none());
    let cut = det.process(make_frame(&b, 128, 96, 33));
    assert!(
      cut.is_some(),
      "expected cut between top/bottom and left/right halves"
    );
    assert!(
      det.last_distance().unwrap() >= Options::default().threshold(),
      "distance {} should meet default threshold 0.395",
      det.last_distance().unwrap(),
    );
  }

  #[test]
  fn min_duration_suppresses_rapid_cuts() {
    // Python-compat mode: no early cuts allowed.
    let opts = Options::default()
      .with_min_duration(Duration::from_secs(1))
      .with_initial_cut(false);
    let mut det = Detector::new(opts);

    let (a, b) = ortho_halves_frames();

    let mut cuts = 0u32;
    for i in 0..30i64 {
      let frame_data = if i % 2 == 0 { &a } else { &b };
      let ts = i * 33;
      if det.process(make_frame(frame_data, 128, 96, ts)).is_some() {
        cuts += 1;
      }
    }
    assert_eq!(cuts, 0, "min_duration should suppress all cuts within 1s");
  }

  #[test]
  #[cfg_attr(miri, ignore)] // 128×96 phash is extremely slow under Miri (~650s)
  fn clear_resets_stream_state() {
    let opts = Options::default().with_min_duration(Duration::from_millis(0));
    let mut det = Detector::new(opts);

    let (a, b) = ortho_halves_frames();

    // Video 1: prime, then cut.
    assert!(det.process(make_frame(&a, 128, 96, 0)).is_none());
    let cut1 = det.process(make_frame(&b, 128, 96, 33));
    assert!(cut1.is_some());
    assert!(det.last_distance().is_some());

    det.clear();

    // First frame of video 2: no cut, state re-seeded.
    assert!(det.process(make_frame(&a, 128, 96, 1_000_000)).is_none());
    assert!(
      det.last_distance().is_none(),
      "last_distance should be cleared"
    );

    // Second frame of video 2: normal cut detection resumes.
    let cut2 = det.process(make_frame(&b, 128, 96, 1_000_033));
    assert!(cut2.is_some());
  }

  #[test]
  fn clear_preserves_resize_table_when_dims_match() {
    let opts = Options::default().with_min_duration(Duration::from_millis(0));
    let mut det = Detector::new(opts);

    let (a, _) = ortho_halves_frames();
    // First frame builds the resize table for 128×96.
    det.process(make_frame(&a, 128, 96, 0));
    assert_eq!(det.resize_table.src_w, 128);
    assert_eq!(det.resize_table.src_h, 96);
    let x_offsets_len = det.resize_table.x_offsets.len();

    det.clear();
    // Table is preserved across clear — same dims on next video won't rebuild.
    assert_eq!(det.resize_table.src_w, 128);
    assert_eq!(det.resize_table.src_h, 96);
    assert_eq!(det.resize_table.x_offsets.len(), x_offsets_len);
  }

  #[test]
  fn hash_bit_packing_matches_layout() {
    // A small sanity check that bit 0 corresponds to position (0,0) and
    // higher bits walk across rows.
    let mut det = Detector::new(Options::default());
    let size = det.size;
    // Craft a known low_freq pattern: alternating above/below median.
    for i in 0..(size * size) {
      det.low_freq[i] = if i % 2 == 0 { -1.0 } else { 1.0 };
    }
    // Invoke bit-packing logic by mimicking the tail of compute_hash.
    det.sort_scratch.clone_from(&det.low_freq);
    det.sort_scratch.sort_unstable_by(|a, b| a.total_cmp(b));
    let n = det.sort_scratch.len();
    let median = (det.sort_scratch[n / 2 - 1] + det.sort_scratch[n / 2]) / 2.0;
    det.current_hash.fill(0);
    for (i, &v) in det.low_freq.iter().enumerate() {
      if v > median {
        det.current_hash[i / 64] |= 1u64 << (i % 64);
      }
    }
    // Every odd index should be set.
    let set: u32 = det.current_hash.iter().map(|w| w.count_ones()).sum();
    assert_eq!(set as usize, size * size / 2);
  }

  #[test]
  fn options_accessors_builders_setters_roundtrip() {
    let fps30 = Timebase::new(30, nz32(1));

    let opts = Options::default()
      .with_threshold(0.5)
      .with_size(32)
      .with_lowpass(4)
      .with_min_duration(core::time::Duration::from_millis(333))
      .with_initial_cut(false);
    assert_eq!(opts.threshold(), 0.5);
    assert_eq!(opts.size(), 32);
    assert_eq!(opts.lowpass(), 4);
    assert_eq!(opts.min_duration(), core::time::Duration::from_millis(333));
    assert!(!opts.initial_cut());

    let opts_frames = Options::default().with_min_frames(15, fps30);
    assert_eq!(
      opts_frames.min_duration(),
      core::time::Duration::from_millis(500)
    );

    // In-place setters, chainable.
    let mut opts = Options::default();
    opts
      .set_threshold(0.1)
      .set_size(8)
      .set_lowpass(2)
      .set_min_duration(core::time::Duration::from_secs(1))
      .set_initial_cut(true);
    assert_eq!(opts.threshold(), 0.1);
    assert_eq!(opts.size(), 8);
    assert_eq!(opts.lowpass(), 2);
    assert!(opts.initial_cut());

    opts.set_min_frames(30, fps30);
    assert_eq!(opts.min_duration(), core::time::Duration::from_secs(1));
  }

  #[test]
  fn try_new_rejects_imsize_squared_overflow() {
    // imsize = size * lowpass = 100_000 * 100_000 = 1e10 fits in usize on
    // 64-bit. imsize^2 = 1e20 > usize::MAX (≈1.8e19) → DimensionsOverflow.
    let opts = Options::default().with_size(100_000).with_lowpass(100_000);
    let err = Detector::try_new(opts).expect_err("imsize*imsize should overflow");
    assert_eq!(
      err,
      Error::DimensionsOverflow {
        size: 100_000,
        lowpass: 100_000,
      },
    );
  }

  #[test]
  fn median_f32_singleton() {
    let mut buf = [42.0f32];
    assert_eq!(super::median_f32(&mut buf), 42.0);
  }
}