strided-basic 0.15.0

Shared typed strided CPU primitives and copy/reduction execution.
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
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
//! The cache-tiled plan and replay walk: every entry equals an independent
//! per-element decode bit for bit, on serial and team tokens; layouts whose
//! tiles would not change the visit order keep the column-major walk; split
//! leaves partition the box; non-injective destinations stay serial.

use super::*;
use crate::{CopyPlan, UpdatePlan};
use num_complex::Complex64;
use threadteam::{ledger, Team};

fn team() -> Team {
    threadteam::Pool::builder()
        .threads(4)
        .dispatch_costs(crate::threading::test_dispatch_costs())
        .acquire_timeout(core::time::Duration::from_secs(10))
        .build()
        .unwrap()
        .into_team(4)
        .expect("an idle unit-test pool")
}

/// One layout with its base offsets and backing lengths.
struct Layout {
    dims: Vec<usize>,
    dst: Vec<isize>,
    src: Vec<isize>,
    dst_offset: isize,
    src_offset: isize,
}

impl Layout {
    fn new(dims: &[usize], dst: &[isize], src: &[isize], src_offset: isize) -> Self {
        Self {
            dims: dims.to_vec(),
            dst: dst.to_vec(),
            src: src.to_vec(),
            dst_offset: lowest(dims, dst),
            src_offset,
        }
    }

    fn total(&self) -> usize {
        self.dims.iter().product()
    }

    /// Destination and source offset of every logical index, column-major
    /// over the given axes. Extent-1 axes contribute nothing whatever their
    /// stride.
    fn offsets(&self) -> Vec<(usize, usize)> {
        (0..self.total())
            .map(|linear| {
                let mut rest = linear;
                let (mut d, mut s) = (self.dst_offset, self.src_offset);
                for axis in 0..self.dims.len() {
                    let coord = rest % self.dims[axis];
                    rest /= self.dims[axis];
                    if coord > 0 {
                        d += coord as isize * self.dst[axis];
                        s += coord as isize * self.src[axis];
                    }
                }
                (d as usize, s as usize)
            })
            .collect()
    }

    fn lens(&self) -> (usize, usize) {
        let offsets = self.offsets();
        let d = offsets.iter().map(|o| o.0).max().unwrap() + 1;
        let s = offsets.iter().map(|o| o.1).max().unwrap() + 1;
        (d, s)
    }

    fn src_view<'a, T>(&'a self, data: &'a [T]) -> RawStridedRef<'a, T> {
        RawStridedRef::new(data, &self.dims, &self.src, self.src_offset).unwrap()
    }

    fn dst_view<'a, T>(&'a self, data: &'a mut [T]) -> RawStridedMut<'a, T> {
        RawStridedMut::new(data, &self.dims, &self.dst, self.dst_offset).unwrap()
    }

    /// The tiles [`PairTraversal::given`] selects, whatever their shape.
    fn tiles<D, S>(&self) -> Option<Tiles> {
        PairTraversal::given(&self.dims, &self.dst, &self.src).tiles::<D, S>()
    }

    /// The tiles of the inline walk only.
    fn inline_tiles<D, S>(&self) -> Option<PairTiles> {
        self.tiles::<D, S>()
            .as_ref()
            .and_then(Tiles::inline)
            .copied()
    }
}

/// Offset that makes the lowest reachable slot 0.
fn lowest(dims: &[usize], strides: &[isize]) -> isize {
    dims.iter()
        .zip(strides)
        .filter(|(&d, &s)| d > 1 && s < 0)
        .map(|(&d, &s)| (d as isize - 1) * -s)
        .sum()
}

/// `dims` with `strides(axis)` reversed over the axes.
fn reversed(dims: &[usize]) -> Layout {
    let mut stride = 1isize;
    let dst: Vec<isize> = dims
        .iter()
        .map(|&d| {
            let s = stride;
            stride *= d as isize;
            s
        })
        .collect();
    let mut stride = 1isize;
    let mut src = vec![0isize; dims.len()];
    for axis in (0..dims.len()).rev() {
        src[axis] = stride;
        stride *= dims[axis] as isize;
    }
    Layout::new(dims, &dst, &src, 0)
}

fn layouts() -> Vec<(&'static str, Layout)> {
    vec![
        // Clipped tiles: blocks [43, 42] do not divide [67, 45].
        (
            "67x45 transpose",
            Layout::new(&[67, 45], &[1, 67], &[45, 1], 0),
        ),
        // A stored [37, 25, 41] source, permuted, reversed along its unit
        // axis and cropped at a nonzero base.
        (
            "rank 3 permuted, negative, cropped",
            Layout::new(&[40, 36, 24], &[1, 40, 1440], &[925, -1, 37], 36 + 37),
        ),
        ("16^4 reversed", reversed(&[16; 4])),
        ("4^8 reversed", reversed(&[4; 8])),
        // Fused rank above the inline limit: the spilled walk.
        ("4^9 reversed", reversed(&[4; 9])),
        // The spilled walk with a negative source axis and a nonzero base: a
        // walker that drops the sign still matches the positive case above.
        ("4^9 reversed source, negative last axis, cropped base", {
            let dims = [4usize; 9];
            let dst: Vec<isize> = (0..9).map(|axis| 1isize << (2 * axis)).collect();
            let mut src: Vec<isize> = (0..9).map(|axis| 1isize << (2 * (8 - axis))).collect();
            src[8] = -1;
            Layout::new(&dims, &dst, &src, 8)
        }),
        (
            "rank 12 reversed",
            reversed(&[2, 2, 2, 2, 2, 2, 2, 2, 4, 4, 4, 4]),
        ),
        // An extent-1 axis that `pair_tiles` drops.
        (
            "[1, 65, 65] transpose",
            Layout::new(&[1, 65, 65], &[3, 1, 65], &[5, 65, 1], 0),
        ),
    ]
}

trait Value: Copy + Send + Sync + PartialEq + core::fmt::Debug {
    fn value(k: usize) -> Self;
    fn bits(self) -> [u64; 2];
}

impl Value for f64 {
    fn value(k: usize) -> Self {
        (k % 97) as f64 * 0.25 - 7.0
    }
    fn bits(self) -> [u64; 2] {
        [self.to_bits(), 0]
    }
}

impl Value for Complex64 {
    fn value(k: usize) -> Self {
        Complex64::new((k % 89) as f64 * 0.5 - 3.0, (k % 13) as f64 - 6.0)
    }
    fn bits(self) -> [u64; 2] {
        [self.re.to_bits(), self.im.to_bits()]
    }
}

fn bits<T: Value>(values: &[T]) -> Vec<[u64; 2]> {
    values.iter().map(|v| v.bits()).collect()
}

/// `f(dst_slot, src_slot)` applied per logical index in column-major order.
fn oracle<D: Copy, S: Copy>(layout: &Layout, dst: &mut [D], src: &[S], f: impl Fn(&mut D, S)) {
    for (d, s) in layout.offsets() {
        f(&mut dst[d], src[s]);
    }
}

fn uninit_bits<T: Value>(values: &[MaybeUninit<T>]) -> Vec<[u64; 2]> {
    // SAFETY: test storage starts initialized and only receives initialized values.
    values
        .iter()
        .map(|v| unsafe { v.assume_init() }.bits())
        .collect()
}

/// Every typed entry on `layout` with token `exec()` against [`oracle`].
fn check_entries<'e, T>(layout: &Layout, exec: impl Fn() -> Exec<'e>, label: &str)
where
    T: Value
        + Mul<Output = T>
        + Add<Output = T>
        + ElementOpApply
        + num_traits::Zero
        + num_traits::One,
{
    let (dst_len, src_len) = layout.lens();
    let src: Vec<T> = (0..src_len).map(T::value).collect();
    let init: Vec<T> = (0..dst_len).map(|k| T::value(k + 11)).collect();
    let source = layout.src_view(&src);
    let scale = T::value(3);
    let beta = T::value(40);

    let check = |name: &str, f: &dyn Fn(&mut T, T), run: &dyn Fn(&mut RawStridedMut<'_, T>)| {
        let mut expected = init.clone();
        oracle(layout, &mut expected, &src, f);
        let mut actual = init.clone();
        run(&mut layout.dst_view(&mut actual));
        assert_eq!(bits(&actual), bits(&expected), "{label}: {name}");
    };
    check("replay_copy_raw", &|d, s| *d = s, &|d| {
        replay_copy_raw(d, &source, exec()).unwrap()
    });
    check("replay_map_raw +=", &|d, s| *d = *d + s, &|d| {
        replay_map_raw(d, &source, |d: &mut T, s: T| *d = *d + s, exec()).unwrap()
    });
    let copy = CopyPlan::compile(&layout.dims, &layout.dst, &layout.src).unwrap();
    let update = UpdatePlan::compile(&layout.dims, &layout.dst, &layout.src).unwrap();
    check("CopyPlan::execute", &|d, s| *d = s, &|d| {
        copy.execute(d, &source, exec()).unwrap()
    });
    check("CopyPlan::execute_scale", &|d, s| *d = scale * s, &|d| {
        copy.execute_scale(d, &source, scale, exec()).unwrap()
    });
    check("CopyPlan::execute_conj", &|d, s| *d = s.conj(), &|d| {
        copy.execute_conj(d, &source, exec()).unwrap()
    });
    check(
        "UpdatePlan::execute beta != 0",
        &|d, s| *d = scale * (s * scale) + beta * *d,
        &|d| {
            update
                .execute(d, &source, scale, beta, |x| x * scale, exec())
                .unwrap()
        },
    );

    // Uninitialized destinations: the holes keep their sentinel.
    let sentinel = T::value(1000);
    let check_uninit =
        |name: &str,
         f: &dyn Fn(&mut T, T),
         run: &dyn Fn(&mut RawStridedMut<'_, MaybeUninit<T>>)| {
            let mut expected = vec![sentinel; dst_len];
            oracle(layout, &mut expected, &src, f);
            let mut actual = vec![MaybeUninit::new(sentinel); dst_len];
            run(&mut layout.dst_view(&mut actual));
            assert_eq!(uninit_bits(&actual), bits(&expected), "{label}: {name}");
        };
    check_uninit("replay_copy_raw_uninit", &|d, s| *d = s, &|d| {
        replay_copy_raw_uninit(d, &source, exec()).unwrap()
    });
    check_uninit("replay_map_raw_uninit", &|d, s| *d = s, &|d| {
        replay_map_raw_uninit(d, &source, |s| s, exec()).unwrap()
    });
    check_uninit("CopyPlan::execute_uninit", &|d, s| *d = s, &|d| {
        copy.execute_uninit(d, &source, exec()).unwrap()
    });
    check_uninit("UpdatePlan::execute_uninit", &|d, s| *d = scale * s, &|d| {
        update
            .execute_uninit(d, &source, scale, |x| x, exec())
            .unwrap()
    });
}

/// The `f64`-only rows: beta = 0 over a NaN destination, and a mixed-type
/// uninitialized map.
fn check_f64_rows<'e>(layout: &Layout, exec: impl Fn() -> Exec<'e>, label: &str) {
    let (dst_len, src_len) = layout.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let update = UpdatePlan::compile(&layout.dims, &layout.dst, &layout.src).unwrap();
    let mut expected = vec![f64::NAN; dst_len];
    oracle(layout, &mut expected, &src, |d, s| *d = 2.0 * s);
    let mut actual = vec![f64::NAN; dst_len];
    update
        .execute(
            &mut layout.dst_view(&mut actual),
            &layout.src_view(&src),
            2.0,
            0.0,
            |x| x,
            exec(),
        )
        .unwrap();
    assert_eq!(bits(&actual), bits(&expected), "{label}: beta = 0");

    let ints: Vec<i32> = (0..src_len as i32).map(|k| k * 3 - 50).collect();
    let mut expected = vec![-0.5f64; dst_len];
    oracle(layout, &mut expected, &ints, |d, s| *d = f64::from(s) * 0.5);
    let mut actual = vec![MaybeUninit::new(-0.5f64); dst_len];
    replay_map_raw_uninit(
        &mut layout.dst_view(&mut actual),
        &layout.src_view(&ints),
        |s| f64::from(s) * 0.5,
        exec(),
    )
    .unwrap();
    assert_eq!(uninit_bits(&actual), bits(&expected), "{label}: i32 -> f64");
}

#[test]
fn spilled_layouts_tile_as_their_compressed_rank_would() {
    // [#425]: a fused rank above the inline limit walks untiled, unless the
    // shared model's tiles are applied to the spilled axes too.
    let nine = reversed(&[4; 9]);
    let (dst_len, src_len) = nine.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&nine, &mut expected, &src, |d, s| *d = s);
    let mut actual = vec![0.0f64; dst_len];
    crate::threading::take_tile_boxes();
    replay_copy_raw(
        &mut nine.dst_view(&mut actual),
        &nine.src_view(&src),
        Exec::serial(),
    )
    .unwrap();
    assert_eq!(bits(&actual), bits(&expected));
    let boxes = crate::threading::take_tile_boxes();
    assert!(
        boxes >= 16,
        "the rank-9 copy must visit cache tiles, got {boxes} boxes"
    );
    assert!(matches!(nine.tiles::<f64, f64>(), Some(Tiles::Spilled(_))));

    // One inert axis appended to the rank-8 layout: `given` keeps the axis and
    // spills, while the compressed layout is the rank-8 one, so the shared
    // model must reach the same blocks and the same slots.
    let rank8 = reversed(&[4; 8]);
    let mut padded = Layout::new(
        &[4, 4, 4, 4, 4, 4, 4, 4, 1],
        &[1, 4, 16, 64, 256, 1024, 4096, 16384, 3],
        &[16384, 4096, 1024, 256, 64, 16, 4, 1, 5],
        0,
    );
    padded.dst_offset = rank8.dst_offset;
    padded.src_offset = rank8.src_offset;
    let Some(Tiles::Inline(compressed)) = padded.tiles::<f64, f64>() else {
        panic!("a padded rank-9 layout must tile as its compressed rank-8 one");
    };
    assert_eq!(compressed.layout.rank, 8);
    assert_eq!(compressed.layout.dims[..8], rank8.dims[..]);
    let rank8_tiles = rank8
        .inline_tiles::<f64, f64>()
        .expect("the rank-8 layout tiles");
    assert_eq!(compressed.blocks, rank8_tiles.blocks);

    let (dst_len, src_len) = padded.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&padded, &mut expected, &src, |d, s| *d = s);
    let mut actual = vec![0.0f64; dst_len];
    crate::threading::take_tile_boxes();
    replay_copy_raw(
        &mut padded.dst_view(&mut actual),
        &padded.src_view(&src),
        Exec::serial(),
    )
    .unwrap();
    assert_eq!(bits(&actual), bits(&expected));
    let boxes = crate::threading::take_tile_boxes();
    assert!(
        boxes >= 16,
        "the padded rank-9 copy must visit cache tiles, got {boxes} boxes"
    );

    // The spilled walk with a negative source axis and a nonzero base reaches
    // the tiles too, and only a sign-preserving offset reproduces the oracle.
    let cropped = layouts()
        .into_iter()
        .find(|(label, _)| label.starts_with("4^9 reversed source"))
        .expect("the negative spilled layout");
    assert!(matches!(
        cropped.1.tiles::<f64, f64>(),
        Some(Tiles::Spilled(_))
    ));
    let (dst_len, src_len) = cropped.1.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&cropped.1, &mut expected, &src, |d, s| *d = s);
    let mut actual = vec![0.0f64; dst_len];
    crate::threading::take_tile_boxes();
    replay_copy_raw(
        &mut cropped.1.dst_view(&mut actual),
        &cropped.1.src_view(&src),
        Exec::serial(),
    )
    .unwrap();
    assert_eq!(bits(&actual), bits(&expected), "{}", cropped.0);
    let boxes = crate::threading::take_tile_boxes();
    assert!(
        boxes >= 16,
        "the negative spilled copy must visit tiles, got {boxes}"
    );

    // A layout whose column-major walk is kept stays on it: one box count of
    // zero is the same answer the inline route gives the rank-8 layout.
    let dense = Layout::new(&[4096, 512], &[1, 4096], &[1, 4096], 0);
    assert!(dense.tiles::<f64, f64>().is_none());

    // The split leaves of the spilled route partition the box, as the inline
    // ones do.
    let team = team();
    let nine = reversed(&[4; 9]);
    let (dst_len, src_len) = nine.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut actual = vec![0.0f64; dst_len];
    crate::threading::take_tiled_leaves();
    replay_copy_raw(
        &mut nine.dst_view(&mut actual),
        &nine.src_view(&src),
        team.exec(4),
    )
    .unwrap();
    let leaves = crate::threading::take_tiled_leaves();
    #[cfg(feature = "parallel")]
    {
        assert!(!leaves.is_empty(), "the spilled box must split at width 4");
        assert!(leaf_visits(&nine, &leaves).iter().all(|&v| v == 1));
    }
    #[cfg(not(feature = "parallel"))]
    let _ = leaves;
}

/// [#425] review follow-up: every spilled plan entry, not only the raw replay,
/// must reach the tiles. A mutation that suppresses tiles for
/// `PairTraversal::Spilled` inside `CopyPlan` or `UpdatePlan` has to fail here.
/// The model cuts after axis 0 here and reads a **nine**-entry tail in
/// `slabs_share_no_line`; the fixed eight-entry buffer this replaced panicked
/// on it, and no other fixture in the file produces a tail above eight (#425).
#[test]
fn a_tail_above_eight_entries_is_decided_without_panicking() {
    let dims = [8usize, 256, 2, 2, 2, 2, 2, 2, 2, 2];
    let dst = [
        1isize, 8, 2048, 4096, 8192, 16384, 32768, 65536, 131072, 262144,
    ];
    let src = [1isize, 2048, 1024, 512, 256, 128, 64, 32, 16, 8];
    let layout = Layout::new(&dims, &dst, &src, 0);
    assert!(layout.tiles::<f64, f64>().is_none());
}

#[test]
fn spilled_plan_entries_visit_tiles() {
    let nine = reversed(&[4; 9]);
    let (dst_len, src_len) = nine.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let source = nine.src_view(&src);
    let copy = CopyPlan::compile(&nine.dims, &nine.dst, &nine.src).unwrap();
    let update = UpdatePlan::compile(&nine.dims, &nine.dst, &nine.src).unwrap();

    let run = |name: &str, f: &mut dyn FnMut(&mut RawStridedMut<'_, f64>)| {
        let mut actual = vec![0.0f64; dst_len];
        crate::threading::take_tile_boxes();
        f(&mut nine.dst_view(&mut actual));
        let boxes = crate::threading::take_tile_boxes();
        assert!(
            boxes >= 16,
            "{name}: the spilled plan must visit cache tiles, got {boxes} boxes"
        );
    };
    run("CopyPlan::execute", &mut |d| {
        copy.execute(d, &source, Exec::serial()).unwrap()
    });
    run("CopyPlan::execute_scale", &mut |d| {
        copy.execute_scale(d, &source, 2.0, Exec::serial()).unwrap()
    });
    run("CopyPlan::execute_conj", &mut |d| {
        copy.execute_conj(d, &source, Exec::serial()).unwrap()
    });
    run("UpdatePlan::execute", &mut |d| {
        update
            .execute(d, &source, 2.0, 3.0, |x| x, Exec::serial())
            .unwrap()
    });

    let run_uninit = |name: &str, f: &mut dyn FnMut(&mut RawStridedMut<'_, MaybeUninit<f64>>)| {
        let mut actual = vec![MaybeUninit::new(0.0f64); dst_len];
        crate::threading::take_tile_boxes();
        f(&mut nine.dst_view(&mut actual));
        let boxes = crate::threading::take_tile_boxes();
        assert!(
            boxes >= 16,
            "{name}: the spilled plan must visit cache tiles, got {boxes} boxes"
        );
    };
    run_uninit("CopyPlan::execute_uninit", &mut |d| {
        copy.execute_uninit(d, &source, Exec::serial()).unwrap()
    });
    run_uninit("UpdatePlan::execute_uninit", &mut |d| {
        update
            .execute_uninit(d, &source, 2.0, |x| x, Exec::serial())
            .unwrap()
    });
}

#[test]
fn tiled_replays_equal_the_column_major_oracle() {
    let team = team();
    for (label, layout) in layouts() {
        assert!(
            layout.tiles::<f64, f64>().is_some(),
            "{label} must take the tiled walk"
        );
        for parallel in [false, true] {
            let exec = || {
                if parallel {
                    team.exec(4)
                } else {
                    Exec::serial()
                }
            };
            check_entries::<f64>(&layout, exec, label);
            check_entries::<Complex64>(&layout, exec, label);
            check_f64_rows(&layout, exec, label);
        }
    }
    let layout = |prefix: &str| {
        layouts()
            .into_iter()
            .find(|(label, _)| label.starts_with(prefix))
            .unwrap_or_else(|| panic!("no layout named {prefix}"))
            .1
    };
    // One clipped tile: 67 = 43 + 24 and 45 = 42 + 3.
    let tiles = layout("67x45").inline_tiles::<f64, f64>().unwrap();
    assert_eq!(tiles.blocks[..2], [43, 42]);
    // The extent-1 axis is dropped before the model.
    let tiles = layout("[1, 65, 65]").inline_tiles::<f64, f64>().unwrap();
    assert_eq!(tiles.layout.rank, 2);
    assert_eq!(tiles.layout.dims[..2], [65, 65]);
}

/// A spilled layout whose block does not divide every extent: the last tile of
/// each such axis must be clipped to the remaining extent. The exact-division
/// spilled fixtures cannot see that arithmetic — a `min` against the whole box
/// extent instead of the remaining one still matches the oracle on them.
#[test]
fn spilled_tiles_clip_the_last_tile_to_the_remaining_extent() {
    let team = team();
    let layout = reversed(&[5; 9]);
    let tiles = layout
        .tiles::<f64, f64>()
        .expect("5^9 reversed must take the tiled walk");
    let Tiles::Spilled(tiles) = tiles else {
        panic!("5^9 reversed must reach the spilled walk, not the inline one");
    };
    assert!(
        tiles
            .blocks()
            .iter()
            .zip(tiles.dims())
            .any(|(&block, &dim)| dim % block != 0),
        "the fixture must leave a clipped last tile: dims {:?} blocks {:?}",
        tiles.dims(),
        tiles.blocks()
    );
    for parallel in [false, true] {
        let exec = || {
            if parallel {
                team.exec(4)
            } else {
                Exec::serial()
            }
        };
        check_entries::<f64>(&layout, exec, "5^9 reversed clipped");
    }
}

#[test]
fn tiled_dynamic_slice_window_equals_the_oracle() {
    use crate::DynamicSlicePlan;
    // A transposing window of a [220, 200] operand at runtime start (7, 11).
    let operand_dims = [220usize, 200];
    let operand_strides = [200isize, 1];
    let sizes = [197usize, 185];
    let dest_strides = [1isize, 197];
    assert!(pair_tiles::<MaybeUninit<f64>, f64>(
        &fuse_pair_layout(&sizes, &dest_strides, &operand_strides).unwrap()
    )
    .is_some());
    let plan = DynamicSlicePlan::compile(
        &operand_dims,
        &operand_strides,
        &[2],
        &[1],
        &sizes,
        &dest_strides,
        &sizes,
    )
    .unwrap();
    let data: Vec<f64> = (0..44_000).map(f64::value).collect();
    let operand = RawStridedRef::new(&data, &operand_dims, &operand_strides, 0).unwrap();
    let starts_data = [7i64, 11];
    let starts = RawStridedRef::new(&starts_data, &[2], &[1], 0).unwrap();
    let mut expected = vec![0.0f64; 197 * 185];
    for j in 0..185 {
        for i in 0..197 {
            expected[i + 197 * j] = data[(i + 7) * 200 + (j + 11)];
        }
    }
    let team = team();
    crate::threading::take_tiled_leaves();
    for exec in [Exec::serial(), team.exec(4)] {
        let mut out = vec![0.0f64; 197 * 185];
        plan.execute(
            &mut RawStridedMut::new(&mut out, &sizes, &dest_strides, 0).unwrap(),
            &operand,
            &starts,
            exec,
        )
        .unwrap();
        assert_eq!(bits(&out), bits(&expected));
    }
    let leaves = crate::threading::take_tiled_leaves();
    #[cfg(feature = "parallel")]
    {
        let window = Layout::new(&sizes, &dest_strides, &operand_strides, 7 * 200 + 11);
        assert!(!leaves.is_empty());
        assert!(leaf_visits(&window, &leaves).iter().all(|&v| v == 1));
    }
    #[cfg(not(feature = "parallel"))]
    let _ = leaves;
}

#[test]
fn tiled_dynamic_update_slice_window_equals_the_oracle() {
    use crate::DynamicUpdateSlicePlan;
    // A transposing [197, 185] update into a dense [220, 200] operand at runtime
    // start (7, 11); the operand copy is contiguous and does not tile.
    let dims = [220usize, 200];
    let strides = [200isize, 1];
    let sizes = [197usize, 185];
    let update_strides = [1isize, 197];
    assert!(pair_tiles::<MaybeUninit<f64>, f64>(
        &fuse_pair_layout(&sizes, &strides, &update_strides).unwrap()
    )
    .is_some());
    let plan = DynamicUpdateSlicePlan::compile(
        &dims,
        &strides,
        &[2],
        &[1],
        &sizes,
        &update_strides,
        &dims,
        &strides,
    )
    .unwrap();
    let data: Vec<f64> = (0..44_000).map(f64::value).collect();
    let operand = RawStridedRef::new(&data, &dims, &strides, 0).unwrap();
    let update_data: Vec<f64> = (44_000..44_000 + 197 * 185).map(f64::value).collect();
    let update = RawStridedRef::new(&update_data, &sizes, &update_strides, 0).unwrap();
    let starts_data = [7i64, 11];
    let starts = RawStridedRef::new(&starts_data, &[2], &[1], 0).unwrap();
    let mut expected = data.clone();
    for j in 0..185 {
        for i in 0..197 {
            expected[(i + 7) * 200 + (j + 11)] = update_data[i + 197 * j];
        }
    }
    let team = team();
    crate::threading::take_tiled_leaves();
    for exec in [Exec::serial(), team.exec(4)] {
        let mut out = vec![0.0f64; 44_000];
        plan.execute(
            &mut RawStridedMut::new(&mut out, &dims, &strides, 0).unwrap(),
            &operand,
            &update,
            &starts,
            exec,
        )
        .unwrap();
        assert_eq!(bits(&out), bits(&expected));
    }
    let leaves = crate::threading::take_tiled_leaves();
    #[cfg(feature = "parallel")]
    {
        // The fused window walks its axes in destination-stride order.
        let window = Layout {
            dst_offset: 7 * 200 + 11,
            ..Layout::new(&[185, 197], &[1, 200], &[197, 1], 0)
        };
        assert!(!leaves.is_empty());
        assert!(leaf_visits(&window, &leaves).iter().all(|&v| v == 1));
    }
    #[cfg(not(feature = "parallel"))]
    let _ = leaves;
}

#[test]
fn layouts_whose_tiles_keep_the_order_take_the_single_pass() {
    let none = |dims: &[usize], dst: &[isize], src: &[isize]| {
        PairTraversal::given(dims, dst, src)
            .inline()
            .and_then(pair_tiles::<f64, f64>)
            .is_none()
    };
    // Strictly increasing strides in both operands.
    assert!(none(&[3840, 512], &[1, 3840], &[1, 4096]));
    assert!(none(&[4096, 512], &[1, 4096], &[-1, 4096]));
    // Zero-length and single-element disagreeing layouts.
    assert!(none(&[0, 64], &[64, 1], &[1, 64]));
    assert!(none(&[1, 1], &[64, 1], &[1, 64]));
    // Every byte stride above the budget: all blocks are 1.
    assert!(none(&[64, 64], &[5000, 320_000], &[320_000, 5000]));
    // One run: rank one, or one moving axis among extent-1 axes.
    assert!(none(&[4096], &[1], &[-3]));
    assert!(none(&[1, 4096, 1], &[7, 1, 9], &[5, -3, 11]));

    // On a team token the empty and single-element cases dispatch nothing.
    let team = team();
    let src = [2.5f64; 1];
    let mut dst = [0.0f64; 1];
    crate::threading::take_dispatched_chunks();
    replay_copy_raw(
        &mut RawStridedMut::new(&mut dst, &[1, 1], &[64, 1], 0).unwrap(),
        &RawStridedRef::new(&src, &[1, 1], &[1, 64], 0).unwrap(),
        team.exec(4),
    )
    .unwrap();
    assert_eq!(dst, [2.5]);
    let empty: [f64; 0] = [];
    let mut out: [f64; 0] = [];
    replay_copy_raw(
        &mut RawStridedMut::new(&mut out, &[0, 64], &[64, 1], 0).unwrap(),
        &RawStridedRef::new(&empty, &[0, 64], &[1, 64], 0).unwrap(),
        team.exec(4),
    )
    .unwrap();
    assert_eq!(crate::threading::take_dispatched_chunks(), 0);
}

/// Whether `pair_tiles::<D, S>` selects tiles for the given axes.
fn tiles<D, S>(dims: &[usize], dst: &[isize], src: &[isize]) -> bool {
    PairTraversal::given(dims, dst, src)
        .tiles::<D, S>()
        .is_some()
}

#[test]
fn slabs_that_share_no_cache_line_keep_the_flat_walk() {
    // F3: both operands stream whole 32 KiB runs along axis 0 and the tail
    // axes only permute those runs; the model still offers [4096, 16, 16].
    let f3 = Layout::new(&[4096, 17, 33], &[1, 4096, 69_632], &[1, 135_168, 4096], 0);
    let model = crate::block::reduced_block_sizes(&f3.dims, &[&f3.dst, &f3.src], &[8, 8]).unwrap();
    assert_eq!(model.as_slice(), [4096, 16, 16]);
    // The view path reads the same exit (#428): its planner is the model's
    // caller, and the answer it gets is the full dimensions.
    assert_eq!(
        crate::block::compute_block_sizes(&f3.dims, &[&f3.dst, &f3.src], &[8, 8]).as_slice(),
        f3.dims
    );
    assert!(f3.tiles::<f64, f64>().is_none());
    // A zero-sized source moves no line; the destination alone decides.
    assert!(f3.tiles::<f64, ()>().is_none());
    // Sparse prefixes whose bounding intervals are separated and line-sized,
    // the second with a whole-layout span above the model's 64 * N region.
    assert!(!tiles::<f64, f64>(
        &[8, 17, 33],
        &[81, 568, 9656],
        &[81, 18744, 568]
    ));
    assert!(!tiles::<f64, f64>(
        &[8, 17, 33],
        &[9, 72, 1224],
        &[9, 2376, 72]
    ));

    // Not declined: equal source tail strides re-read the same slab, so
    // per-axis separation alone is not enough.
    assert!(tiles::<f64, f64>(
        &[4096, 17, 16],
        &[1, 65_536, 4096],
        &[1, 4096, 4096]
    ));
    // Not declined: the u8 source's slabs are 8 elements of 1 byte, below a
    // line, so the pair still pays for tiling. The former single-size model
    // gave both operands the 8-byte size of the u64 destination and this
    // fixture then looked line-sized on both sides (#426).
    let mixed = Layout::new(&[8, 512, 512], &[1, 8, 4096], &[1, 4096, 8], 0);
    assert!(mixed.tiles::<u64, u8>().is_some());
    assert!(mixed.tiles::<f64, f64>().is_none());

    let team = team();
    let (dst_len, src_len) = f3.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&f3, &mut expected, &src, |d, s| *d = s);
    let mut actual = vec![0.0f64; dst_len];
    crate::threading::take_tiled_leaves();
    crate::threading::take_dispatched_chunks();
    replay_copy_raw(
        &mut f3.dst_view(&mut actual),
        &f3.src_view(&src),
        team.exec(4),
    )
    .unwrap();
    assert_eq!(bits(&actual), bits(&expected));
    assert!(crate::threading::take_tiled_leaves().is_empty());
    let chunks = crate::threading::take_dispatched_chunks();
    #[cfg(feature = "parallel")]
    assert!(chunks >= 2, "F3 runs the flat split ({chunks} chunks)");
    #[cfg(not(feature = "parallel"))]
    let _ = chunks;

    let (dst_len, src_len) = mixed.lens();
    let bytes: Vec<u8> = (0..src_len).map(|k| (k * 7) as u8).collect();
    let mut expected = vec![0u64; dst_len];
    oracle(&mixed, &mut expected, &bytes, |d, s| *d = u64::from(s) * 3);
    let mut actual = vec![0u64; dst_len];
    replay_map_raw(
        &mut mixed.dst_view(&mut actual),
        &mixed.src_view(&bytes),
        |d: &mut u64, s: u8| *d = u64::from(s) * 3,
        team.exec(4),
    )
    .unwrap();
    assert_eq!(actual, expected);
}

/// Every logical index of `layout` covered by the recorded `leaves`, counted
/// independently of the replay: each origin is decoded from its destination
/// offset through the layout's dense destination strides and checked against
/// its source offset.
#[cfg(feature = "parallel")]
fn leaf_visits(layout: &Layout, leaves: &[crate::threading::TiledLeaf]) -> Vec<u8> {
    let rank = layout.dims.len();
    let mut order: Vec<usize> = (0..rank).collect();
    order.sort_by_key(|&axis| layout.dst[axis]);
    let mut visits = vec![0u8; layout.total()];
    for (dst_offset, src_offset, extents) in leaves {
        assert_eq!(extents.len(), rank);
        let mut rest = dst_offset - layout.dst_offset;
        let mut origin = vec![0usize; rank];
        for &axis in order.iter().rev() {
            origin[axis] = (rest / layout.dst[axis]) as usize;
            rest %= layout.dst[axis];
        }
        assert_eq!(rest, 0);
        let src_at: isize = layout.src_offset
            + origin
                .iter()
                .zip(&layout.src)
                .map(|(&c, &s)| c as isize * s)
                .sum::<isize>();
        assert_eq!(*src_offset, src_at, "source offset of leaf at {origin:?}");
        let count: usize = extents.iter().product();
        for linear in 0..count {
            let mut rest = linear;
            let mut index = 0usize;
            let mut pitch = 1usize;
            for axis in 0..rank {
                let coord = origin[axis] + rest % extents[axis];
                rest /= extents[axis];
                assert!(coord < layout.dims[axis], "leaf beyond the box");
                index += coord * pitch;
                pitch *= layout.dims[axis];
            }
            visits[index] += 1;
        }
    }
    visits
}

#[test]
fn split_leaves_partition_the_box() {
    // A dense permutation that tiles with blocks [35, 17, 2]; at width 4 the
    // split cuts axis 2 into origins 0, 8, 16, 24.
    let dims = [64usize, 17, 33];
    let layout = Layout::new(&dims, &[1, 64, 1088], &[561, 1, 17], 0);
    let tiles = layout.inline_tiles::<f64, f64>().unwrap();
    assert_eq!(tiles.blocks[..3], [35, 17, 2]);
    let (dst_len, src_len) = layout.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&layout, &mut expected, &src, |d, s| *d = s);
    let mut actual = vec![0.0f64; dst_len];
    let team = team();
    crate::threading::take_tiled_leaves();
    replay_copy_raw(
        &mut layout.dst_view(&mut actual),
        &layout.src_view(&src),
        team.exec(4),
    )
    .unwrap();
    assert_eq!(bits(&actual), bits(&expected));
    let leaves = crate::threading::take_tiled_leaves();
    #[cfg(feature = "parallel")]
    {
        assert!(leaf_visits(&layout, &leaves).iter().all(|&v| v == 1));
        assert_eq!(leaves.len(), 4);
        // Each cut keeps at most ceil(d / 2) / d of its parent.
        let n = layout.total() as f64;
        for (_, _, extents) in &leaves {
            let mut bound = n / 4.0;
            for (axis, &extent) in extents.iter().enumerate() {
                let mut d = dims[axis];
                while d > extent {
                    bound *= 2.0 * d.div_ceil(2) as f64 / d as f64;
                    d = d.div_ceil(2);
                }
            }
            assert!(
                extents.iter().product::<usize>() as f64 <= bound,
                "{leaves:?}"
            );
        }
    }
    #[cfg(not(feature = "parallel"))]
    let _ = leaves;
}

#[test]
fn tiled_replay_keeps_the_callers_budget() {
    // With D(2) = estimate / 100 and D(4) = D(6) = 10 * estimate the flat
    // width rule grants 2, while every node of the tree pays a fork at
    // width 2: a tiled replay must split into 4 leaves, as the plan does.
    use threadteam::{DispatchCosts, DispatchKind, HostClass, WidthCosts};
    let layout = Layout::new(&[128, 128], &[1, 128], &[128, 1], 0);
    assert!(layout.tiles::<f64, f64>().is_some());
    let estimate = KernelClass::Copy.estimate(layout.total());
    let costs = WidthCosts::new([estimate / 100, estimate * 10, estimate * 10]);
    let pool = threadteam::Pool::builder()
        .threads(4)
        .dispatch_costs(
            DispatchCosts::for_class(HostClass::Generic)
                .with(DispatchKind::Team, costs)
                .with(DispatchKind::InPool, costs)
                .with(DispatchKind::Outside, costs),
        )
        .acquire_timeout(core::time::Duration::from_secs(10))
        .build()
        .unwrap();
    let team = pool.into_team(4).expect("an idle unit-test pool");
    #[cfg(feature = "parallel")]
    assert_eq!(
        crate::cost::parallel_width(&team.exec(4), KernelClass::Copy, layout.total()).get(),
        2
    );
    let (dst_len, src_len) = layout.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&layout, &mut expected, &src, |d, s| *d = s);
    let copy = CopyPlan::compile(&layout.dims, &layout.dst, &layout.src).unwrap();
    let update = UpdatePlan::compile(&layout.dims, &layout.dst, &layout.src).unwrap();
    for entry in ["replay_map_raw", "CopyPlan::execute", "UpdatePlan::execute"] {
        let mut actual = vec![0.0f64; dst_len];
        crate::threading::take_tiled_leaves();
        if entry == "replay_map_raw" {
            replay_map_raw(
                &mut layout.dst_view(&mut actual),
                &layout.src_view(&src),
                |d: &mut f64, s| *d = s,
                team.exec(4),
            )
            .unwrap();
        } else if entry == "CopyPlan::execute" {
            copy.execute(
                &mut layout.dst_view(&mut actual),
                &layout.src_view(&src),
                team.exec(4),
            )
            .unwrap();
        } else {
            update
                .execute(
                    &mut layout.dst_view(&mut actual),
                    &layout.src_view(&src),
                    1.0,
                    0.0,
                    |x| x,
                    team.exec(4),
                )
                .unwrap();
        }
        assert_eq!(bits(&actual), bits(&expected), "{entry}");
        let leaves = crate::threading::take_tiled_leaves();
        #[cfg(feature = "parallel")]
        {
            let mut extents: Vec<Vec<usize>> = leaves.iter().map(|l| l.2.clone()).collect();
            extents.sort();
            assert_eq!(extents, vec![vec![64, 64]; 4], "{entry}");
            assert!(leaf_visits(&layout, &leaves).iter().all(|&v| v == 1));
        }
        #[cfg(not(feature = "parallel"))]
        let _ = leaves;
    }
}

#[test]
fn non_injective_destination_replays_serially_in_column_major_order() {
    // Axis 2 writes every slot twice; the source disagrees with the
    // destination, so an injective destination of this shape would tile.
    let dims = [64usize, 64, 2];
    let dst = [1isize, 64, 0];
    let src = [64isize, 1, 4096];
    let layout = Layout::new(&dims, &dst, &src, 0);
    assert!(layout.tiles::<i64, i64>().is_some());
    let (dst_len, src_len) = layout.lens();
    let values: Vec<i64> = (0..src_len as i64).map(|k| k * 7 - 1000).collect();
    let team = team();
    for parallel in [false, true] {
        let exec = || {
            let exec = if parallel {
                team.exec(4)
            } else {
                Exec::serial()
            };
            let root = exec.ledger_root();
            (exec, root)
        };
        let assert_serial = |root: Option<u64>| {
            if let Some(root) = root {
                assert_eq!(ledger::max_live(root), 0, "must not fan out");
            }
        };
        let mut expected = vec![3i64; dst_len];
        oracle(&layout, &mut expected, &values, |d, s| *d = s);
        let mut actual = vec![3i64; dst_len];
        let (e, root) = exec();
        replay_copy_raw(
            &mut layout.dst_view(&mut actual),
            &layout.src_view(&values),
            e,
        )
        .unwrap();
        assert_eq!(actual, expected, "last writer");
        assert_serial(root);
        let mut expected = vec![3i64; dst_len];
        oracle(&layout, &mut expected, &values, |d, s| *d = *d * 3 + s);
        let mut actual = vec![3i64; dst_len];
        let (e, root) = exec();
        replay_map_raw(
            &mut layout.dst_view(&mut actual),
            &layout.src_view(&values),
            |d: &mut i64, s| *d = *d * 3 + s,
            e,
        )
        .unwrap();
        assert_eq!(actual, expected, "order-sensitive accumulate");
        assert_serial(root);
    }
}

#[test]
fn inert_singleton_strides_stay_out_of_the_arithmetic() {
    // A stride of `isize::MAX` on an extent-1 axis is valid and unused.
    let m = isize::MAX;
    let layout = Layout::new(&[1, 65, 65], &[m, 1, 65], &[m, 65, 1], 0);
    let tiles = layout.inline_tiles::<f64, f64>().unwrap();
    assert_eq!(tiles.blocks[..2], [56, 33]);
    let (dst_len, src_len) = layout.lens();
    let src: Vec<f64> = (0..src_len).map(f64::value).collect();
    let mut expected = vec![0.0f64; dst_len];
    oracle(&layout, &mut expected, &src, |d, s| *d = s);
    let team = team();
    for parallel in [false, true] {
        let exec = || {
            if parallel {
                team.exec(4)
            } else {
                Exec::serial()
            }
        };
        let mut actual = vec![0.0f64; dst_len];
        replay_copy_raw(
            &mut layout.dst_view(&mut actual),
            &layout.src_view(&src),
            exec(),
        )
        .unwrap();
        assert_eq!(bits(&actual), bits(&expected));
        let mut actual = vec![1.0f64; dst_len];
        let mut accumulated = vec![1.0f64; dst_len];
        oracle(&layout, &mut accumulated, &src, |d, s| *d += s);
        replay_map_raw(
            &mut layout.dst_view(&mut actual),
            &layout.src_view(&src),
            |d: &mut f64, s| *d += s,
            exec(),
        )
        .unwrap();
        assert_eq!(bits(&actual), bits(&accumulated));
    }
}

/// Isolated arithmetic self-check of the tile odometer, not route coverage:
/// hand-built tiles over zero-sized layouts whose last slot lies within one
/// stride of `isize::MAX` visit every index once without overflowing.
#[test]
fn tile_odometer_stays_within_the_reachable_offsets() {
    let m = isize::MAX;
    // (dims, destination strides, source strides, both bases, blocks)
    type Case<'a> = (&'a [usize], &'a [isize], &'a [isize], isize, &'a [usize]);
    let cases: [Case<'_>; 3] = [
        (&[2, 2], &[1, m - 1], &[m - 1, 1], 0, &[1, 1]),
        (&[2, 2], &[-1, -(m - 1)], &[-(m - 1), -1], m, &[1, 2]),
        (&[2, 2, 2], &[1, 2, m - 3], &[m - 3, 2, 1], 0, &[1, 2, 1]),
    ];
    for (dims, dst, src, base, blocks) in cases {
        let axes: Vec<FusedAxis> = dims
            .iter()
            .zip(dst)
            .zip(src)
            .map(|((&dim, &dst), &src)| FusedAxis { dim, dst, src })
            .collect();
        let layout = FusedPairLayout::from_axes(&axes);
        let mut tile_blocks = [1usize; RAW_FUSED_RANK_LIMIT];
        tile_blocks[..blocks.len()].copy_from_slice(blocks);
        let visited = core::cell::Cell::new(0usize);
        let mut unit = ();
        // SAFETY: zero-sized elements make every offset of the dangling
        // pointers a valid access.
        unsafe {
            apply_tiled_box(
                &mut unit as *mut (),
                base,
                &() as *const (),
                base,
                &layout,
                &tile_blocks,
                &mut |_: &mut (), _: ()| visited.set(visited.get() + 1),
            );
        }
        assert_eq!(visited.get(), dims.iter().product::<usize>(), "{dims:?}");
    }
}

/// The destination slots of `tiles` in the order the tiled walk is specified
/// to visit them, built from the specification alone: the tile grid in
/// column-major order (axis 0 fastest), tiles of extents `blocks` anchored at
/// the origin and clipped to the box, and each tile in column-major order.
/// Independent of `apply_tile_range` and `apply_tiled_box`.
fn specified_tile_order(tiles: &PairTiles, dst_offset: isize) -> Vec<usize> {
    let layout = &tiles.layout;
    let rank = layout.rank;
    let dims = &layout.dims[..rank];
    let blocks = &tiles.blocks[..rank];
    let grid: Vec<usize> = dims
        .iter()
        .zip(blocks)
        .map(|(&d, &b)| d.div_ceil(b))
        .collect();
    let mut order = Vec::new();
    for tile in 0..grid.iter().product::<usize>() {
        let mut rest = tile;
        let origin: Vec<usize> = (0..rank)
            .map(|axis| {
                let coord = rest % grid[axis];
                rest /= grid[axis];
                coord * blocks[axis]
            })
            .collect();
        let extent: Vec<usize> = (0..rank)
            .map(|axis| blocks[axis].min(dims[axis] - origin[axis]))
            .collect();
        for element in 0..extent.iter().product::<usize>() {
            let mut rest = element;
            let mut slot = dst_offset;
            for axis in 0..rank {
                let coord = origin[axis] + rest % extent[axis];
                rest /= extent[axis];
                slot += coord as isize * layout.dst_strides[axis];
            }
            order.push(slot as usize);
        }
    }
    order
}

/// [#427]: any cut of the tile grid into two ranges, walked first range
/// first, visits exactly the slots of the whole-box walk in the same order:
/// the visit sequence (destination slot of every element, in visit order) of
/// `0..cut` followed by `cut..count` equals that of [`apply_tiled_box`]. And
/// that sequence is the specified one ([`specified_tile_order`], built without
/// either walk), so the order is checked against an independent reference, not
/// only between two walks that share `apply_tile_range`.
///
/// [#427]: https://github.com/tensor4all/strided-rs/issues/427
#[test]
fn tile_ranges_partition_the_tiled_box_walk() {
    use core::cell::RefCell;

    let mut tiled = 0;
    for (label, layout) in layouts() {
        let Some(tiles) = layout.inline_tiles::<f64, f64>() else {
            continue;
        };
        tiled += 1;
        let count = tile_count(&tiles.layout, &tiles.blocks);
        assert!(count >= 2, "{label}: a tiled layout has several tiles");
        let (dst_len, src_len) = layout.lens();
        let src: Vec<f64> = (0..src_len).map(f64::value).collect();
        let mut expected = vec![0.0f64; dst_len];
        oracle(&layout, &mut expected, &src, |d, s| *d = s);

        // The whole-box walk, recording the destination slot of each visit.
        let mut whole = vec![0.0f64; dst_len];
        let whole_base = whole.as_mut_ptr() as usize;
        let reference = RefCell::new(Vec::with_capacity(layout.total()));
        // SAFETY: `lens` sizes both buffers for every slot of the layout, the
        // tiles are `pair_tiles` of it, and the destination is injective.
        unsafe {
            apply_tiled_box(
                whole.as_mut_ptr(),
                layout.dst_offset,
                src.as_ptr(),
                layout.src_offset,
                &tiles.layout,
                &tiles.blocks,
                &mut |d: &mut f64, s: f64| {
                    let slot = (d as *mut f64 as usize - whole_base) / size_of::<f64>();
                    reference.borrow_mut().push(slot);
                    *d = s;
                },
            );
        }
        let reference = reference.into_inner();
        assert_eq!(bits(&whole), bits(&expected), "{label}: whole-box walk");
        assert!(
            reference == specified_tile_order(&tiles, layout.dst_offset),
            "{label}: the whole-box walk must visit the slots in the specified tile order"
        );
        assert_eq!(
            reference.len(),
            layout.total(),
            "{label}: one visit per element"
        );

        let cuts = [0, 1, count / 3, count / 2, count - 1, count];
        for &cut in &cuts {
            let mut actual = vec![0.0f64; dst_len];
            let base = actual.as_mut_ptr() as usize;
            let dst_ptr = actual.as_mut_ptr();
            let visits = RefCell::new(Vec::with_capacity(layout.total()));
            crate::threading::take_tile_boxes();
            for (first, len) in [(0, cut), (cut, count - cut)] {
                // SAFETY: `lens` sizes both buffers for every slot of the
                // layout, the tiles are `pair_tiles` of it, the destination
                // is injective, and `first + len` is at most `count`.
                unsafe {
                    apply_tile_range(
                        dst_ptr,
                        layout.dst_offset,
                        src.as_ptr(),
                        layout.src_offset,
                        &tiles.layout,
                        &tiles.blocks,
                        first,
                        len,
                        &mut |d: &mut f64, s: f64| {
                            let slot = (d as *mut f64 as usize - base) / size_of::<f64>();
                            visits.borrow_mut().push(slot);
                            *d = s;
                        },
                    );
                }
            }
            assert_eq!(bits(&actual), bits(&expected), "{label}: cut at {cut}");
            assert_eq!(
                crate::threading::take_tile_boxes(),
                count,
                "{label}: cut at {cut} must walk every tile once"
            );
            assert!(
                visits.into_inner() == reference,
                "{label}: cut at {cut} must visit the slots in the whole-box order"
            );
        }
    }
    assert!(tiled >= 2, "only {tiled} layouts tiled inline");
}

/// `may_tile` is the element-size-free part of the tiling decision: a
/// traversal it rules out never tiles for any element types, and every
/// traversal that tiles for one is admitted.
#[test]
fn may_tile_admits_every_traversal_that_tiles() {
    let mut cases: Vec<(String, PairTraversal)> = layouts()
        .into_iter()
        .map(|(label, layout)| {
            (
                label.to_string(),
                PairTraversal::given(&layout.dims, &layout.dst, &layout.src),
            )
        })
        .collect();
    for (label, dims, dst, src) in [
        (
            "contiguous run",
            vec![4096usize, 512],
            vec![1isize, 4096],
            vec![1isize, 4096],
        ),
        (
            "one moving axis",
            vec![1, 4096, 1],
            vec![7, 1, 9],
            vec![3, -1, 2],
        ),
        ("one element", vec![1, 1], vec![5, 3], vec![2, 8]),
        ("empty", vec![0, 64], vec![1, 0], vec![64, 1]),
        ("transpose", vec![512, 512], vec![1, 512], vec![512, 1]),
    ] {
        cases.push((label.to_string(), fuse_pair_traversal(&dims, &dst, &src)));
    }
    let mut ruled_out = 0;
    for (label, traversal) in &cases {
        let tiles = [
            traversal.tiles::<f64, f64>().is_some(),
            traversal.tiles::<u8, f64>().is_some(),
            traversal.tiles::<f64, u8>().is_some(),
            traversal.tiles::<Complex64, f64>().is_some(),
            traversal.tiles::<u8, u8>().is_some(),
        ];
        if !traversal.may_tile() {
            ruled_out += 1;
            assert_eq!(tiles, [false; 5], "{label}: ruled out but tiles");
        }
    }
    assert!(ruled_out >= 3, "only {ruled_out} traversals were ruled out");
    assert!(cases
        .iter()
        .any(|(_, t)| t.may_tile() && t.tiles::<f64, f64>().is_some()));
}

/// The wide fixtures of the allocation test (`alloc_matrix.rs`,
/// `check_replay_any_rank_is_allocation_free`) reach the branches it means to
/// cover: (c) the spilled tiled walk of `4^9` with a reversed source, and (d)
/// a 21-axis layout the model's early exits do not answer, so it builds its
/// 42-entry stride-order table, above the 40-entry inline scratch, and still
/// tiles.
#[test]
fn wide_allocation_fixtures_reach_the_spilled_model() {
    let wide = reversed(&[4; 9]);
    assert!(matches!(wide.tiles::<f64, f64>(), Some(Tiles::Spilled(_))));

    let dims = vec![2usize; 21];
    let dst: Vec<isize> = (0..21).map(|k| 1isize << k).collect();
    let src: Vec<isize> = (0..21).rev().map(|k| 1isize << k).collect();
    let axes = dims
        .iter()
        .zip(&dst)
        .zip(&src)
        .map(|((&dim, &dst), &src)| (dim, dst, src));
    assert!(!crate::block::exits_before_metadata(axes, 8, 8));
    assert!(2 * dims.len() > crate::inline::OPERAND_CAP * strided_view::INLINE_RANK);
    let deep = Layout::new(&dims, &dst, &src, 0);
    assert!(matches!(deep.tiles::<f64, f64>(), Some(Tiles::Spilled(_))));
}