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
use super::*;
use crate::fuse::compute_costs;
use strided_view::auxiliary::index_order;

#[test]
fn test_total_memory_region_contiguous() {
    // Contiguous array: 100 elements * 8 bytes = 800 bytes
    // Should be about 800 bytes (plus cache line rounding)
    let dims = [100usize];
    let strides = [8u128]; // 8 bytes per element
    let byte_strides: Vec<&[u128]> = vec![&strides];

    let region = total_memory_region(&dims, &byte_strides);

    // 99 * 8 = 792 bytes contiguous
    // 792 / 64 + 1 = 13 cache lines
    // 64 * 13 = 832 bytes
    assert_eq!(region, 832);
}

#[test]
fn test_total_memory_region_strided() {
    // Non-contiguous: stride larger than cache line
    let dims = [10usize];
    let strides = [128u128]; // Stride >= cache line
    let byte_strides: Vec<&[u128]> = vec![&strides];

    let region = total_memory_region(&dims, &byte_strides);

    // Each element touches separate cache line blocks
    // 64 * 1 * 10 = 640 bytes
    assert_eq!(region, 640);
}

#[test]
fn test_compute_blocks_small() {
    // Small array that fits in cache
    let dims = [10usize, 10];
    let costs = [2isize, 2];
    let strides = [8u128, 80];
    let orders = [1usize, 2];
    let byte_strides: Vec<&[u128]> = vec![&strides];
    let stride_orders: Vec<&[usize]> = vec![&orders];

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // Should use full dimensions since it fits
    assert_eq!(blocks.as_slice(), vec![10, 10]);
}

#[test]
fn test_compute_blocks_large() {
    // Large array that needs blocking
    let dims = [1000usize, 1000];
    let costs = [2isize, 2];
    let strides = [8u128, 8000]; // 8 bytes per f64
    let orders = [1usize, 2];
    let byte_strides: Vec<&[u128]> = vec![&strides];
    let stride_orders: Vec<&[usize]> = vec![&orders];

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // Should reduce block sizes
    assert!(blocks[0] <= dims[0]);
    assert!(blocks[1] <= dims[1]);
    // But should still be reasonable
    assert!(blocks[0] >= 1);
    assert!(blocks[1] >= 1);
}

#[test]
fn test_last_argmax_weighted() {
    let blocks = [10usize, 20, 5];
    let costs = [1isize, 1, 2];

    let idx = last_argmax_weighted(&blocks, &costs);

    // (10-1)*1=9, (20-1)*1=19, (5-1)*2=8
    // Maximum is 19 at index 1
    assert_eq!(idx, Some(1));
}

#[test]
fn test_last_argmax_weighted_tie() {
    // When tied, returns last index
    let blocks = [10usize, 10];
    let costs = [1isize, 1];

    let idx = last_argmax_weighted(&blocks, &costs);

    // Both have score 9, return last (index 1)
    assert_eq!(idx, Some(1));
}

#[test]
fn test_compute_block_sizes_full_pipeline() {
    // Test the full compute_block_sizes function
    let dims = [100usize, 100];
    let strides = [1isize, 100];
    let strides_list: Vec<&[isize]> = vec![&strides];

    let blocks = compute_block_sizes(&dims, &strides_list, &[8]);

    // Should return valid block sizes
    assert_eq!(blocks.len(), 2);
    assert!(blocks[0] >= 1 && blocks[0] <= 100);
    assert!(blocks[1] >= 1 && blocks[1] <= 100);
}

// ========== Julia-comparison tests ==========

#[test]
fn test_total_memory_region_julia_match_2d() {
    // Julia: dims=(10,10), bytestrides=((8,80),)
    // numcontigeouscachelines: 9*8 + 9*80 = 72 + 720 = 792 (but 80 >= 64!)
    // Actually: stride 8 < 64, so += (10-1)*8 = 72
    //          stride 80 >= 64, so numcachelineblocks *= 10
    // contiguous_lines = 72/64 + 1 = 2
    // memory = 64 * 2 * 10 = 1280
    let dims = [10usize, 10];
    let strides = [8u128, 80];
    let byte_strides: Vec<&[u128]> = vec![&strides];

    let region = total_memory_region(&dims, &byte_strides);
    assert_eq!(region, 1280);
}

#[test]
fn test_total_memory_region_julia_match_multiple_arrays() {
    // Two arrays with different strides
    let dims = [10usize, 10];
    let strides1 = [8u128, 80]; // Column-major-ish
    let strides2 = [80u128, 8]; // Row-major-ish
    let byte_strides: Vec<&[u128]> = vec![&strides1, &strides2];

    let region = total_memory_region(&dims, &byte_strides);

    // Array 1: stride 8 < 64 -> contiguous += 72, stride 80 >= 64 -> blocks *= 10
    //          contiguous_lines = 2, region1 = 64 * 2 * 10 = 1280
    // Array 2: stride 80 >= 64 -> blocks *= 10, stride 8 < 64 -> contiguous += 72
    //          contiguous_lines = 2, region2 = 64 * 2 * 10 = 1280
    // Total = 2560
    assert_eq!(region, 2560);
}

#[test]
fn test_total_memory_region_all_contiguous() {
    // All strides < cache line
    let dims = [10usize, 5];
    let strides = [8u128, 40]; // Both < 64
    let byte_strides: Vec<&[u128]> = vec![&strides];

    let region = total_memory_region(&dims, &byte_strides);

    // contiguous = (10-1)*8 + (5-1)*40 = 72 + 160 = 232
    // contiguous_lines = 232/64 + 1 = 4
    // blocks = 1 (no large strides)
    // region = 64 * 4 * 1 = 256
    assert_eq!(region, 256);
}

#[test]
fn test_total_memory_region_all_large_strides() {
    // All strides >= cache line
    let dims = [5usize, 4];
    let strides = [64u128, 320]; // Both >= 64
    let byte_strides: Vec<&[u128]> = vec![&strides];

    let region = total_memory_region(&dims, &byte_strides);

    // contiguous = 0
    // contiguous_lines = 0/64 + 1 = 1
    // blocks = 5 * 4 = 20
    // region = 64 * 1 * 20 = 1280
    assert_eq!(region, 1280);
}

#[test]
fn test_compute_blocks_first_dim_smallest_stride() {
    // When first dimension has smallest stride for all arrays,
    // Julia keeps d1 and recurses on tail
    let dims = [100usize, 10, 10];
    let costs = [2isize, 2, 2];
    // All arrays have first dim with smallest stride
    let strides1 = [8u128, 800, 8000];
    let orders1 = [1usize, 2, 3]; // First dim has order 1 (smallest)
    let byte_strides: Vec<&[u128]> = vec![&strides1];
    let stride_orders: Vec<&[usize]> = vec![&orders1];

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // First dimension should be kept as-is (100)
    assert_eq!(blocks[0], 100);
}

#[test]
fn test_compute_blocks_min_stride_larger_than_blocksize() {
    // When minimum stride > blocksize AND total memory > blocksize, return all 1s
    // Need:
    // 1. Total memory > blocksize
    // 2. First dim does NOT have smallest stride (to skip the special case)
    // 3. All strides > blocksize
    let dims = [100usize, 100];
    let costs = [2isize, 2];
    // Stride order: dim 1 has smaller stride than dim 0, but both > blocksize
    let strides = [40000000u128, 40000]; // dim 1 is smaller but still > 32768
    let orders = [2usize, 1]; // dim 1 has smaller stride order
    let byte_strides: Vec<&[u128]> = vec![&strides];
    let stride_orders: Vec<&[usize]> = vec![&orders];

    // Verify total memory exceeds blocksize first
    let initial_mem = total_memory_region(&dims, &byte_strides);
    assert!(
        initial_mem > BLOCK_MEMORY_SIZE as u128,
        "Initial memory {} should exceed {}",
        initial_mem,
        BLOCK_MEMORY_SIZE
    );

    // Verify minimum stride > blocksize
    let min_stride = strides.iter().copied().min().unwrap();
    assert!(
        min_stride > BLOCK_MEMORY_SIZE as u128,
        "Min stride {} should exceed {}",
        min_stride,
        BLOCK_MEMORY_SIZE
    );

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // With minimum stride > blocksize, should return all 1s
    assert_eq!(blocks.as_slice(), vec![1, 1]);
}

#[test]
fn test_compute_blocks_4d_array_first_dim_smallest() {
    // 4D array where first dim has smallest stride (column-major case)
    // Julia's algorithm keeps first dimension when it has smallest stride for cache locality
    let dims = [10usize, 10, 10, 10];
    let costs = [2isize, 2, 2, 2];
    // Column-major strides: first dim has smallest stride
    let strides = [8u128, 80, 800, 8000];
    let orders = [1usize, 2, 3, 4];
    let byte_strides: Vec<&[u128]> = vec![&strides];
    let stride_orders: Vec<&[usize]> = vec![&orders];

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // When first dim has smallest stride, Julia keeps full dims via recursion
    // This is by design for cache efficiency
    assert_eq!(blocks.len(), 4);
    for (i, &b) in blocks.iter().enumerate() {
        assert!(b >= 1 && b <= dims[i], "Block {} out of range", i);
    }
}

#[test]
fn test_compute_blocks_4d_needs_reduction() {
    // 4D array where blocking reduction is actually needed
    // Mix strides so first dim doesn't have globally smallest stride
    let dims = [100usize, 100, 100, 100];
    let costs = [2isize, 2, 2, 2];
    // Mixed strides: dim 1 has smallest stride, not dim 0
    let strides = [80u128, 8, 8000, 800];
    let orders = [2usize, 1, 4, 3]; // Reflects actual stride magnitudes
    let byte_strides: Vec<&[u128]> = vec![&strides];
    let stride_orders: Vec<&[usize]> = vec![&orders];

    // Verify total memory exceeds blocksize
    let initial_mem = total_memory_region(&dims, &byte_strides);
    assert!(initial_mem > BLOCK_MEMORY_SIZE as u128);

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // Blocks should be reduced from original dims
    assert_eq!(blocks.len(), 4);
    let total_elements: usize = blocks.iter().product();
    let original_elements: usize = dims.iter().product();
    assert!(
        total_elements < original_elements,
        "Blocks should be smaller than original"
    );
}

#[test]
fn test_compute_blocks_4d_permuted() {
    // 4D array with permuted strides (Issue #5 scenario)
    let dims = [10usize, 10, 10, 10];
    let costs = [2isize, 4, 8, 16]; // Different costs
                                    // Permuted strides: last dim has smallest stride
    let strides = [8000u128, 800, 80, 8];
    let orders = [4usize, 3, 2, 1];
    let byte_strides: Vec<&[u128]> = vec![&strides];
    let stride_orders: Vec<&[usize]> = vec![&orders];

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    assert_eq!(blocks.len(), 4);
    // Verify blocks are valid
    for (i, &b) in blocks.iter().enumerate() {
        assert!(b >= 1 && b <= dims[i]);
    }
}

#[test]
fn test_compute_blocks_mixed_strides_two_arrays() {
    // Two arrays with conflicting stride patterns
    let dims = [100usize, 100];
    let costs = [2isize, 2];
    let strides1 = [8u128, 800]; // Column-major
    let strides2 = [800u128, 8]; // Row-major
    let orders1 = [1usize, 2];
    let orders2 = [2usize, 1];
    let byte_strides: Vec<&[u128]> = vec![&strides1, &strides2];
    let stride_orders: Vec<&[usize]> = vec![&orders1, &orders2];

    let blocks = compute_blocks(
        &dims,
        &costs,
        &byte_strides,
        &stride_orders,
        BLOCK_MEMORY_SIZE,
    );

    // Should not use the first-dim-smallest special case
    // since arrays have different stride orders
    assert_eq!(blocks.len(), 2);
    assert!(blocks[0] >= 1 && blocks[0] <= 100);
    assert!(blocks[1] >= 1 && blocks[1] <= 100);
}

#[test]
fn test_last_argmax_weighted_all_ones() {
    // When all blocks are 1, should return None
    let blocks = [1usize, 1, 1];
    let costs = [1isize, 2, 3];

    let idx = last_argmax_weighted(&blocks, &costs);
    assert_eq!(idx, None);
}

#[test]
fn test_last_argmax_weighted_mixed() {
    // Mix of 1s and larger values
    let blocks = [1usize, 5, 3];
    let costs = [100isize, 1, 1];

    let idx = last_argmax_weighted(&blocks, &costs);

    // (1-1)*100=0, (5-1)*1=4, (3-1)*1=2
    // Max is 4 at index 1
    assert_eq!(idx, Some(1));
}

#[test]
fn test_last_argmax_weighted_cost_matters() {
    // Higher cost can outweigh larger block
    let blocks = [3usize, 10];
    let costs = [10isize, 1];

    let idx = last_argmax_weighted(&blocks, &costs);

    // (3-1)*10=20, (10-1)*1=9
    // Max is 20 at index 0
    assert_eq!(idx, Some(0));
}

#[test]
fn test_compute_blocks_negative_strides() {
    // Negative strides are handled via their magnitude: same blocks as the
    // positive layout.
    let dims = [100usize, 100];
    let positive = compute_block_sizes(&dims, &[&[1, 100], &[100, 1]], &[8, 8]);
    let negative = compute_block_sizes(&dims, &[&[-1, -100], &[-100, 1]], &[8, 8]);
    assert_eq!(negative, positive);
}

#[test]
fn test_compute_block_sizes_4d_column_major() {
    // Full pipeline test for 4D arrays (Issue #5)
    // Column-major layout: first dim has smallest stride
    let dims = [32usize, 32, 32, 32];
    let strides = [1isize, 32, 1024, 32768]; // Column-major
    let strides_list: Vec<&[isize]> = vec![&strides];

    let blocks = compute_block_sizes(&dims, &strides_list, &[8]);

    assert_eq!(blocks.len(), 4);
    // Verify blocks are valid
    for (i, &b) in blocks.iter().enumerate() {
        assert!(b >= 1 && b <= dims[i], "Block {} = {} out of range", i, b);
    }
}

#[test]
fn test_compute_block_sizes_4d_permuted_strides() {
    // Row-major strides [32768, 1024, 32, 1] in the reversed iteration order
    // [3, 2, 1, 0]: the caller passes the metadata already in that order, so
    // the strides strictly increase and every axis is kept whole.
    let dims = [32usize, 32, 32, 32];
    let strides = [1isize, 32, 1024, 32768];
    let strides_list: Vec<&[isize]> = vec![&strides];

    let blocks = compute_block_sizes(&dims, &strides_list, &[8]);

    assert_eq!(blocks.as_slice(), [32, 32, 32, 32]);
}

// ========== Exact arithmetic contract ==========

#[path = "oracle.rs"]
mod oracle;

use crate::kernel::{build_plan, build_plan_fused};
use num_complex::Complex64 as C64;
use oracle::{legacy_blocks, model_blocks, model_blocks_sized, LegacyCosts};

fn refs(strides: &[Vec<isize>]) -> Vec<&[isize]> {
    strides.iter().map(|s| s.as_slice()).collect()
}

#[test]
fn compute_block_sizes_equals_the_exact_model_on_the_sweep() {
    let (mut fits, mut overflows) = (0, 0);
    for (dims, strides, elem_size) in oracle::sweep() {
        let strides = refs(&strides);
        let model = model_blocks(&dims, &strides, elem_size);
        let blocks = compute_block_sizes(&dims, &strides, &vec![elem_size; strides.len()]);
        assert_eq!(
            blocks.as_slice(),
            model,
            "{dims:?} {strides:?} e={elem_size}"
        );
        if elem_size == 0 {
            assert_eq!(model, dims, "a zero element size keeps every axis");
        }
        match legacy_blocks(&dims, &strides, elem_size, LegacyCosts::Saturating) {
            Some(legacy) => {
                assert_eq!(legacy, model, "{dims:?} {strides:?} e={elem_size}");
                fits += 1;
            }
            None => overflows += 1,
        }
    }
    assert!(
        fits > 0 && overflows > 0,
        "{fits} fitting, {overflows} overflowing"
    );
    // Rank 0 returns before any conversion of the element size.
    assert!(compute_block_sizes(&[], &[&[]], &[usize::MAX]).is_empty());
    assert_eq!(
        legacy_blocks(&[], &[&[]], usize::MAX, LegacyCosts::Saturating),
        Some(vec![])
    );
}

/// Per-operand element sizes for the sweep: a deterministic spread over real
/// and degenerate sizes, so most cases mix sizes and some put a zero on one
/// operand.
fn sweep_elem_sizes(operands: usize, case: usize, seed: usize) -> Vec<usize> {
    const SIZES: [usize; 6] = [0, 1, 2, 4, 8, 16];
    (0..operands)
        .map(|k| SIZES[(case + 2 * k + seed) % SIZES.len()])
        .collect()
}

#[test]
fn compute_block_sizes_equals_the_exact_model_per_operand_on_the_sweep() {
    // The same sweep, with each operand planned at its own element size.
    let mut mixed = 0;
    for (case, (dims, strides, seed)) in oracle::sweep().into_iter().enumerate() {
        let strides = refs(&strides);
        let elem_sizes = sweep_elem_sizes(strides.len(), case, seed);
        let model = model_blocks_sized(&dims, &strides, &elem_sizes);
        let blocks = compute_block_sizes(&dims, &strides, &elem_sizes);
        assert_eq!(
            blocks.as_slice(),
            model,
            "{dims:?} {strides:?} e={elem_sizes:?}"
        );
        if elem_sizes.iter().any(|&e| e != elem_sizes[0]) {
            mixed += 1;
        }
    }
    assert!(mixed > 0, "{mixed} mixed-size cases");
}

#[test]
fn a_mixed_size_pair_blocks_at_each_operands_own_footprint() {
    // f64 <- u8 and c64 <- f64 over a transposed source, so no axis fuses and
    // both operands stay under their real byte strides. A single shared size
    // (the larger one, as before #426) overstates the small operand and picks
    // a different tile.
    let dims = [128usize, 128];
    let dst = [1isize, 128];
    let src = [128, 1];
    let strides: [&[isize]; 2] = [&dst, &src];
    // (dst size, src size, the model's blocks, the pre-#426 blocks)
    let cases: [(usize, usize, [usize; 2], [usize; 2]); 2] =
        [(8, 1, [57, 56], [64, 32]), (16, 8, [40, 32], [32, 32])];
    for (dst_size, src_size, blocks, old_blocks) in cases {
        let elem_sizes = [dst_size, src_size];
        let model = model_blocks_sized(&dims, &strides, &elem_sizes);
        assert_eq!(
            compute_block_sizes(&dims, &strides, &elem_sizes).as_slice(),
            model
        );
        let uniform = dst_size.max(src_size);
        let old = model_blocks_sized(&dims, &strides, &[uniform, uniform]);
        assert_eq!(
            legacy_blocks(&dims, &strides, uniform, LegacyCosts::Saturating),
            Some(old.clone())
        );
        assert_eq!(model, blocks, "dst={dst_size} src={src_size}");
        assert_eq!(
            old, old_blocks,
            "dst={dst_size} src={src_size}: the shared size {uniform} answer"
        );
        assert_ne!(
            model, old,
            "dst={dst_size} src={src_size}: a shared size {uniform} must change the tile"
        );
    }
}

/// The pointwise planner's oracle (#287): the model's blocks, then the L1 set
/// fit for the destination at operand 0, unless the model kept the full
/// dimensions (an exit or a decline, which the fit never overrides).
fn planned_blocks_sized(dims: &[usize], strides: &[&[isize]], elem_sizes: &[usize]) -> Vec<usize> {
    let mut blocks = model_blocks_sized(dims, strides, elem_sizes);
    if blocks != dims {
        super::fit_l1_sets(dims, strides, elem_sizes, 0, &mut blocks);
    }
    blocks
}

#[test]
fn the_pointwise_plan_uses_each_operands_element_size() {
    // Every typed pointwise entry passes its operands' sizes here, and this is
    // where they reach the model. The plan is checked against the oracle on
    // the axes the planner actually returns (ordered, fused, compressed).
    let dims = [128usize, 128];
    let dst = [1isize, 128];
    let src = [128, 1];
    let strides: [&[isize]; 2] = [&dst, &src];
    let total = 128 * 128;
    let cases: [[usize; 2]; 2] = [
        [std::mem::size_of::<f64>(), std::mem::size_of::<u8>()],
        [std::mem::size_of::<C64>(), std::mem::size_of::<f64>()],
    ];
    for elem_sizes in cases {
        let (fused_dims, ordered, plan) =
            crate::kernel::pointwise_plan(&dims, &strides, Some(0), &elem_sizes, total);
        let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
        assert_eq!(
            plan.block.as_slice(),
            planned_blocks_sized(&fused_dims, &ordered, &elem_sizes),
            "{elem_sizes:?}"
        );
        // The pre-#426 wiring: one element size for every operand.
        let uniform = elem_sizes.iter().copied().max().unwrap();
        let (_, _, shared) =
            crate::kernel::pointwise_plan(&dims, &strides, Some(0), &[uniform; 2], total);
        assert_ne!(plan.block, shared.block, "{elem_sizes:?}");
    }
}

/// [#428]: the view path answers the layout of #422 the way plan and replay
/// walk it. `pointwise_plan` is what `copy_into` and `map_into` call, and its
/// block vector is the whole routing decision for their walk.
#[test]
fn the_view_path_keeps_the_flat_walk_where_the_tiles_share_no_cache_line() {
    // F3 (`raw_ops::tests::tiled_tests`): every tile keeps axis 0 whole, so a
    // tile is a 32 KiB slab, and the tail axes only permute slabs whose
    // strides are line multiples and nested. No visit order reuses a line.
    let dims = [4096usize, 17, 33];
    let dst = [1isize, 4096, 69_632];
    let src = [1, 135_168, 4096];
    let strides: [&[isize]; 2] = [&dst, &src];
    let elem_sizes = [std::mem::size_of::<f64>(); 2];
    let total = 4096 * 17 * 33;
    // The model itself still offers tiles; the decline is the caller's.
    assert_eq!(
        model_blocks_sized(&dims, &strides, &elem_sizes),
        [4096, 16, 16]
    );
    let (fused_dims, ordered, plan) =
        crate::kernel::pointwise_plan(&dims, &strides, Some(0), &elem_sizes, total);
    let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
    assert_eq!(fused_dims.as_slice(), dims);
    assert_eq!(ordered.as_slice(), strides.as_slice());
    assert_eq!(
        plan.block.as_slice(),
        fused_dims.as_slice(),
        "the view path must keep the flat walk where no visit order reuses a line"
    );

    // A broadcast source whose slabs are one cache line wide and re-read, so
    // its slab intervals are not disjoint: the view path keeps the tiles.
    let dims = [64usize, 64, 64];
    let dst = [1isize, 64, 4096];
    let src = [0isize, 1, 1];
    let strides: [&[isize]; 2] = [&dst, &src];
    let (fused_dims, ordered, plan) =
        crate::kernel::pointwise_plan(&dims, &strides, Some(0), &elem_sizes, 64 * 64 * 64);
    let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
    assert_eq!(
        model_blocks_sized(&fused_dims, &ordered, &elem_sizes),
        [64, 31, 16]
    );
    assert_eq!(plan.block.as_slice(), [64, 31, 16]);

    // The parity rows of `benches/view_decline_428.rs` tile and are not
    // declined, so the decline cannot move them: the view path still returns
    // the model's blocks, fitted to the L1 sets, script for script.
    let parity = [
        ([1024usize, 1024], [1024isize, 1], [1isize, 1024]),
        ([512usize, 512], [1isize, 512], [0isize, 1]),
    ];
    for (dims, dst, src) in parity {
        let strides: [&[isize]; 2] = [&dst, &src];
        let (fused_dims, ordered, plan) = crate::kernel::pointwise_plan(
            &dims,
            &strides,
            Some(0),
            &elem_sizes,
            dims.iter().product(),
        );
        let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
        assert_ne!(plan.block.as_slice(), fused_dims.as_slice(), "must tile");
        // The decline cannot move them; the L1 set fit (#287) may.
        assert_eq!(
            plan.block.as_slice(),
            planned_blocks_sized(&fused_dims, &ordered, &elem_sizes),
            "{dims:?}"
        );
    }
}

#[test]
fn a_zero_sized_destination_leaves_the_reduction_to_the_source() {
    // The other orientation of the ZST contract: the destination has no
    // storage (a `map_into` into a `Zst`), the source is f64. The destination
    // contributes one cache line and order 1 everywhere; the model is still
    // the oracle's.
    let dims = [2usize, 8192];
    let dst = [1isize, 2];
    let src = [2, 1 << 14];
    let strides: [&[isize]; 2] = [&dst, &src];
    let elem_sizes = [0usize, 8];
    let model = model_blocks_sized(&dims, &strides, &elem_sizes);
    assert_eq!(
        compute_block_sizes(&dims, &strides, &elem_sizes).as_slice(),
        model
    );
    assert_eq!(model, [2, 511]);
    // A uniform 8 plans the zero-sized destination by its element strides,
    // which are strictly increasing, so the exit keeps the full dimensions.
    assert_eq!(
        model_blocks_sized(&dims, &strides, &[8, 8]),
        [2, 8192],
        "the pre-#426 answer for the same input"
    );
}

#[test]
fn a_zip_reduce_region_beyond_usize_is_exact() {
    // R1-A: two f64 operands over 1024^6, whose region 2^66 per operand
    // wrapped to 0 ("fits") before the exact contract.
    let dims = [1024usize; 6];
    let a = [8isize, 16, 8, 8, 8, 8];
    let b = [16isize, 8, 8, 8, 8, 8];
    let (dims, strides, plan) = build_plan_fused(&dims, &[&a, &b], None, &[8, 8]);
    assert_eq!(dims.as_slice(), [1024; 6]);
    assert_eq!(strides[0].as_slice(), [8, 8, 8, 8, 8, 16]);
    assert_eq!(strides[1].as_slice(), [8, 8, 8, 8, 16, 8]);
    assert_eq!(plan.block.as_slice(), [1024, 1024, 1024, 1024, 16, 16]);
    let strides: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
    assert_eq!(model_blocks(&dims, &strides, 8), plan.block.as_slice());
    assert_eq!(
        legacy_blocks(&dims, &strides, 8, LegacyCosts::Saturating),
        None
    );
}

/// Zero-sized source element for the mixed-size `map_into` fixture.
#[derive(Clone, Copy, Debug, Default, PartialEq)]
struct Zst;

impl crate::ElementOpApply for Zst {}

static ZSTS: [Zst; 1 << 62] = [Zst; 1 << 62];

#[test]
fn a_zero_sized_source_stops_the_increasing_strides_exit() {
    // R1-B, restated for per-operand element sizes (#426): the f64 destination
    // is planned at 8 bytes and the zero-sized source at 0, so its byte
    // strides are exactly 0 rather than 2^61 * 8 (which a uniform size made
    // overflow u64). The source's constant stride order 1 is then the global
    // minimum on every axis, so the destination's increasing order stops
    // matching it: the increasing-strides exit must not fire and the model
    // reduces axis 1.
    let dims = [2usize, 8192];
    let dst_strides = [1isize, 2];
    let src_strides = [1isize << 61, 1];
    let strides = [&dst_strides[..], &src_strides[..]];
    let (_, _, plan) = build_plan_fused(&dims, &strides, Some(0), &[8, 0]);
    assert_eq!(model_blocks_sized(&dims, &strides, &[8, 0]), [2, 2044]);
    assert_eq!(plan.block.as_slice(), [2, 2044]);
    // The planner before #426 sized both operands at 8 and overflowed here.
    assert_eq!(
        legacy_blocks(&dims, &strides, 8, LegacyCosts::Saturating),
        None
    );

    let src = crate::StridedView::<Zst>::new(&ZSTS[..], &dims, &src_strides, 0).unwrap();
    let mut data = vec![0.0f64; 2 * 8192];
    let mut dst = crate::StridedViewMut::new(&mut data, &dims, &dst_strides, 0).unwrap();
    crate::map_into(&mut dst, &src, |_: Zst| 1.5f64, crate::Exec::serial()).unwrap();
    assert!(data.iter().all(|&x| x == 1.5));
}

#[test]
fn a_zero_element_size_keeps_every_axis_for_any_operand_count() {
    // R3-1: 513 operands, region 513 * 64 = 32832 above the budget, but every
    // byte stride is 0, so every stride order is 1 and both axes are kept.
    let dims = [2usize, 2];
    let mut strides = vec![vec![1isize, 2]; 512];
    strides.push(vec![2, 1]);
    let strides = refs(&strides);
    let zeros = vec![0usize; strides.len()];
    assert_eq!(
        compute_block_sizes(&dims, &strides, &zeros).as_slice(),
        [2, 2]
    );
    assert_eq!(model_blocks(&dims, &strides, 0), [2, 2]);
    assert_eq!(
        legacy_blocks(&dims, &strides, 0, LegacyCosts::Saturating),
        Some(vec![2, 2])
    );
    // Stride orders taken from the element strides instead would shrink both.
    let orders: Vec<InlineVec<usize>> = strides.iter().map(|s| index_order(s)).collect();
    let orders: Vec<&[usize]> = orders.iter().map(|o| o.as_slice()).collect();
    let zeros = [0u128; 2];
    let byte_strides = vec![&zeros[..]; 513];
    let costs = compute_costs(&strides);
    assert_eq!(
        compute_blocks(&dims, &costs, &byte_strides, &orders, BLOCK_MEMORY_SIZE).as_slice(),
        [1, 1]
    );
}

#[test]
fn a_halving_score_beyond_isize_selects_the_exact_maximum() {
    // R3-3: costs [2^62, 2]; the score (3 - 1) * 2^62 = 2^63 wrapped to
    // -2^63, which made the old core shrink axis 1 to [3, 3616].
    let dims = [3usize, 8192];
    let a = [1isize << 61, 1];
    let b = [1isize << 61, 2];
    assert_eq!(compute_costs(&[&a, &b]).as_slice(), [1 << 62, 2]);
    // Regions 73728 -> 49152 (halved) -> 24576 (decremented).
    assert_eq!(
        compute_block_sizes(&dims, &[&a, &b], &[1, 1]).as_slice(),
        [1, 8192]
    );
    assert_eq!(model_blocks(&dims, &[&a, &b], 1), [1, 8192]);
    assert_eq!(
        legacy_blocks(&dims, &[&a, &b], 1, LegacyCosts::Saturating),
        None
    );
}

#[test]
fn tied_strides_run_the_full_core() {
    // F1: the source ties (or is broadcast on) the destination's leading axes,
    // so the strictly increasing exit does not apply.
    let dims = [64usize, 64, 64];
    let dst = [1isize, 64, 4096];
    for src in [[0isize, 1, 1], [1, 1, 1]] {
        assert_eq!(
            compute_block_sizes(&dims, &[&dst, &src], &[8, 8]).as_slice(),
            [64, 31, 16]
        );
        assert_eq!(model_blocks(&dims, &[&dst, &src], 8), [64, 31, 16]);
    }
}

#[test]
fn strictly_increasing_strides_keep_every_axis_in_the_model() {
    // crop r2 and reverse r2: the model without exits keeps every axis, as
    // the strictly increasing exit does.
    let cases: [(&[usize], [isize; 2], [isize; 2]); 2] = [
        (&[3840, 4096], [1, 3840], [1, 4096]),
        (&[4096, 4096], [1, 4096], [-1, 4096]),
    ];
    for (dims, dst, src) in cases {
        assert_eq!(model_blocks(dims, &[&dst, &src], 8), dims);
        assert_eq!(
            compute_block_sizes(dims, &[&dst, &src], &[8, 8]).as_slice(),
            dims
        );
    }
}

#[test]
fn build_plan_forwards_ordered_dims_and_strides() {
    // R3-5: [3, 5] row-major iterates in order [1, 0], so the block core
    // sees dims [5, 3] and strides [1, 5].
    let plan = build_plan(&[3, 5], &[&[5, 1]], Some(0), &[8]);
    assert_eq!(plan.order.as_slice(), [1, 0]);
    assert_eq!(plan.block.as_slice(), [5, 3]);
    // [3, 1000]: forwarded strides [1, 1000] fit (region 24000); unreordered
    // strides [1000, 1] would not (64000) and give blocks [512, 3].
    let plan = build_plan(&[3, 1000], &[&[1000, 1]], Some(0), &[8]);
    assert_eq!(plan.order.as_slice(), [1, 0]);
    assert_eq!(plan.block.as_slice(), [1000, 3]);
}

/// Column-major strides of `dims` laid out in the axis order `perm`.
fn permuted_strides(dims: &[usize], perm: &[usize]) -> Vec<isize> {
    let mut strides = vec![0isize; dims.len()];
    let mut stride = 1isize;
    for &axis in perm {
        strides[axis] = stride;
        stride *= dims[axis] as isize;
    }
    strides
}

#[test]
fn flat_scratch_matches_the_model_across_its_spill_boundaries() {
    // T9: ranks around INLINE_RANK, operand counts around OPERAND_CAP, and
    // flat tables of 16, 18, 40, 45 and 72 entries; each operand walks the
    // axes in a different order, so every case reaches the core.
    let cases: [(usize, usize); 5] = [(8, 2), (9, 2), (8, 5), (9, 5), (12, 6)];
    for (rank, operands) in cases {
        let dims: Vec<usize> = (0..rank).map(|i| 2 + i % 3).collect();
        let strides: Vec<Vec<isize>> = (0..operands)
            .map(|k| {
                // Rotate by k, reversed for odd k: a distinct order per operand.
                let mut perm: Vec<usize> = (0..rank).map(|i| (i + k) % rank).collect();
                if k % 2 == 1 {
                    perm.reverse();
                }
                permuted_strides(&dims, &perm)
            })
            .collect();
        let strides = refs(&strides);
        assert!(!strides.iter().all(|s| strictly_increasing(s)));
        let blocks = compute_block_sizes(&dims, &strides, &vec![8usize; operands]);
        assert_ne!(
            blocks.as_slice(),
            dims.as_slice(),
            "rank {rank}, {operands} operands"
        );
        assert_eq!(
            blocks.as_slice(),
            model_blocks(&dims, &strides, 8),
            "rank {rank}, {operands} operands"
        );
    }

    // Keep-first recursion: axes 0 and 1 lead every operand's stride order,
    // so the core keeps them at offsets 0 and 1 and reduces the suffix at
    // offset 2, where the operands disagree; the suffix rows are read at the
    // original pitch.
    let dims = [8usize, 5, 6, 7, 4, 3];
    let strides: Vec<Vec<isize>> = [
        [0usize, 1, 2, 3, 4, 5],
        [0, 1, 5, 4, 3, 2],
        [0, 1, 3, 5, 2, 4],
    ]
    .iter()
    .map(|perm| permuted_strides(&dims, perm))
    .collect();
    let strides = refs(&strides);
    // The model's answer, which this test is about: every operand's slabs
    // share no line here, so `compute_block_sizes` declines it (#428) and the
    // scratch is read through `reduced_block_sizes`.
    let blocks = reduced_block_sizes(&dims, &strides, &[8, 8, 8]).unwrap();
    assert_eq!(blocks[..2], dims[..2]);
    assert_ne!(blocks.as_slice(), dims.as_slice());
    assert_eq!(blocks.as_slice(), model_blocks(&dims, &strides, 8));
    assert_eq!(blocks.as_slice(), [8, 5, 5, 5, 2, 3]);
    assert_eq!(
        compute_block_sizes(&dims, &strides, &[8, 8, 8]).as_slice(),
        dims
    );
}

/// The slab decision is shared with callers that pass uncompressed axes, such
/// as `bench_hooks::planned_blocks`: the F3 layout with 100 stride-0 extent-1
/// axes appended has a 102-entry tail, past the bounded branch, and keeps the
/// helper's answer. The stride-0 singletons sort first and `0 < span`, so the
/// slabs are not separated and the tiles are kept.
#[test]
fn a_padded_tail_beyond_the_bounded_branch_keeps_the_slab_answer() {
    let mut dims = vec![4096usize, 17, 33];
    let mut dst = vec![1isize, 4096, 69_632];
    let mut src = vec![1isize, 135_168, 4096];
    let unpadded = reduced_block_sizes(&dims, &[&dst, &src], &[8, 8]).unwrap();
    dims.extend([1usize; 100]);
    dst.extend([0isize; 100]);
    src.extend([0isize; 100]);
    let blocks = reduced_block_sizes(&dims, &[&dst, &src], &[8, 8]).unwrap();
    let mut expected = unpadded.to_vec();
    expected.extend([1usize; 100]);
    assert_eq!(blocks.as_slice(), expected);
    assert!(dims.len() - 1 > crate::raw_ops::MAX_MOVING_AXES);
    assert!(!declines_tiles(&dims, &[&dst, &src], &[8, 8], &blocks));
}