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
use super::*;
use crate::cost::KernelClass;

#[test]
fn test_kernel_inner_block() {
    let dims = vec![2, 4];
    let strides1 = vec![4isize, 1];
    let strides2 = vec![4isize, 1];
    let strides_list: Vec<&[isize]> = vec![&strides1, &strides2];
    let plan = build_plan(&dims, &strides_list, Some(0), &[8, 8]);

    let mut total_elements = 0usize;
    for_each_inner_block(&dims, &plan, &strides_list, |_offsets, len, _strides| {
        total_elements += len;
        Ok(())
    })
    .unwrap();

    assert_eq!(total_elements, 8);
}

#[test]
fn test_contiguous_layout_row_vs_col() {
    let dims = [3usize, 4];
    let row = [4isize, 1];
    let col = [1isize, 3];
    assert_eq!(
        contiguous_layout(&dims, &row),
        Some(ContiguousLayout::RowMajor)
    );
    assert_eq!(
        contiguous_layout(&dims, &col),
        Some(ContiguousLayout::ColMajor)
    );
    assert!(contiguous_layout(&dims, &row).is_some());
    assert!(contiguous_layout(&dims, &col).is_some());
}

#[test]
fn test_contiguous_layout_ignores_dim1_axes() {
    let dims = [2usize, 1, 3];
    // Middle stride is irrelevant since that axis never varies.
    let strides = [3isize, 999, 1];
    assert_eq!(
        contiguous_layout(&dims, &strides),
        Some(ContiguousLayout::RowMajor)
    );
}

// ---- same_contiguous_layout tests ----

#[test]
fn test_same_contiguous_layout_all_row_major() {
    let dims = [3usize, 4];
    let s1 = [4isize, 1];
    let s2 = [4isize, 1];
    assert_eq!(
        same_contiguous_layout(&dims, &[&s1, &s2]),
        Some(ContiguousLayout::RowMajor)
    );
}

#[test]
fn test_same_contiguous_layout_all_col_major() {
    let dims = [3usize, 4];
    let s1 = [1isize, 3];
    let s2 = [1isize, 3];
    assert_eq!(
        same_contiguous_layout(&dims, &[&s1, &s2]),
        Some(ContiguousLayout::ColMajor)
    );
}

#[test]
fn test_same_contiguous_layout_mixed_layouts() {
    let dims = [3usize, 4];
    let row = [4isize, 1];
    let col = [1isize, 3];
    assert_eq!(same_contiguous_layout(&dims, &[&row, &col]), None);
}

#[test]
fn test_same_contiguous_layout_one_noncontiguous() {
    let dims = [3usize, 4];
    let row = [4isize, 1];
    let bad = [8isize, 2];
    assert_eq!(same_contiguous_layout(&dims, &[&row, &bad]), None);
}

#[test]
fn test_same_contiguous_layout_empty_strides_list() {
    let dims = [3usize, 4];
    let empty: &[&[isize]] = &[];
    assert_eq!(same_contiguous_layout(&dims, empty), None);
}

#[test]
fn test_same_contiguous_layout_single_array() {
    let dims = [3usize, 4];
    let s = [4isize, 1];
    assert_eq!(
        same_contiguous_layout(&dims, &[&s[..]]),
        Some(ContiguousLayout::RowMajor)
    );
}

#[test]
fn test_same_contiguous_layout_many_arrays() {
    let dims = [2usize, 3];
    let s = [3isize, 1];
    assert_eq!(
        same_contiguous_layout(&dims, &[&s[..], &s[..], &s[..], &s[..], &s[..]]),
        Some(ContiguousLayout::RowMajor)
    );
}

#[test]
fn test_same_contiguous_layout_empty_dims() {
    let dims: [usize; 0] = [];
    let s: [isize; 0] = [];
    assert_eq!(
        same_contiguous_layout(&dims, &[&s[..], &s[..]]),
        Some(ContiguousLayout::RowMajor)
    );
}

// ---- sequential_contiguous_layout tests ----

#[test]
fn test_sequential_contiguous_layout_small_array() {
    let dims = [3usize, 4];
    let s1 = [4isize, 1];
    let s2 = [4isize, 1];
    assert_eq!(
        sequential_contiguous_layout(KernelClass::Map, &dims, &[&s1, &s2], &Exec::serial())
            .unwrap(),
        Some(ContiguousLayout::RowMajor)
    );
}

#[test]
fn test_sequential_contiguous_layout_noncontiguous() {
    let dims = [3usize, 4];
    let s1 = [4isize, 1];
    let s2 = [8isize, 2];
    assert_eq!(
        sequential_contiguous_layout(KernelClass::Map, &dims, &[&s1, &s2], &Exec::serial())
            .unwrap(),
        None
    );
}

#[test]
fn test_sequential_contiguous_layout_col_major() {
    let dims = [3usize, 4];
    let col = [1isize, 3];
    assert_eq!(
        sequential_contiguous_layout(KernelClass::Map, &dims, &[&col], &Exec::serial()).unwrap(),
        Some(ContiguousLayout::ColMajor)
    );
}

#[test]
fn test_build_plan_fused_compresses() {
    // A contiguous 2x3 column-major array fuses [2,3] -> [6,1] -> compress -> [6]
    let dims = [2usize, 3];
    let strides = [1isize, 2];
    let strides_list: Vec<&[isize]> = vec![&strides];
    let (fused_dims, fused_strides, plan) = build_plan_fused(&dims, &strides_list, Some(0), &[8]);
    // After fusion + compression, should be 1D
    assert_eq!(fused_dims.as_slice(), vec![6]);
    assert_eq!(fused_strides.len(), 1);
    assert_eq!(fused_strides[0].as_slice(), vec![1]);
    assert_eq!(plan.block.len(), 1);
}

#[test]
fn test_build_plan_fused_keeps_broadcast_compact_inner_loop() {
    let dims = [16usize, 16, 64, 64];
    let out = [1isize, 16, 256, 16_384];
    let lhs = [1isize, 16, 0, 256];
    let rhs = [0isize, 0, 1, 64];
    let strides_list: Vec<&[isize]> = vec![&out, &lhs, &rhs];

    let (fused_dims, fused_strides, _) = build_plan_fused(&dims, &strides_list, Some(0), &[8; 3]);

    assert_eq!(fused_dims[0], 256);
    assert_eq!(fused_strides[0][0], 1);
    assert_eq!(fused_strides[1][0], 1);
    assert_eq!(fused_strides[2][0], 0);
}

#[test]
fn test_kernel_nd_iterative_total_elements_match() {
    // Use rank 9 to test kernel_nd_inner_iterative (fallback for rank >= 9)
    let dims = vec![2usize, 2, 2, 2, 2, 2, 2, 2, 3];
    let blocks = vec![2usize, 1, 1, 1, 1, 1, 1, 1, 1];
    let mut stride_val = 1isize;
    let mut sv = Vec::new();
    for &d in &dims {
        sv.push(stride_val);
        stride_val *= d as isize;
    }
    let strides = [sv.clone(), sv];
    let strides: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
    let mut offsets = vec![0isize, 0isize];
    let mut total = 0usize;

    kernel_nd_inner_iterative(
        &dims,
        &blocks,
        &strides,
        &mut offsets,
        &mut |_off, len, _s| {
            total += len;
            Ok(())
        },
    )
    .unwrap();

    assert_eq!(total, dims.iter().product::<usize>());
    assert_eq!(offsets, vec![0isize, 0isize]);
}

/// Helper: run a macro-generated kernel and verify total elements and offset reset.
fn verify_kernel_total_and_offsets(rank: usize) {
    // Build dims: dim[0] = 3 for block variety, rest = 2
    let mut dims = vec![3usize];
    dims.resize(rank, 2);
    let blocks = vec![2usize; rank];

    // Column-major strides
    let mut stride_val = 1isize;
    let mut sv = Vec::new();
    for &d in &dims {
        sv.push(stride_val);
        stride_val *= d as isize;
    }
    let strides = [sv.clone(), sv];
    let strides: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
    let mut offsets = vec![0isize, 0isize];
    let mut total = 0usize;
    let expected: usize = dims.iter().product();

    let result = match rank {
        1 => kernel_1d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        2 => kernel_2d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        3 => kernel_3d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        4 => kernel_4d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        5 => kernel_5d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        6 => kernel_6d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        7 => kernel_7d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        8 => kernel_8d_inner(&dims, &blocks, &strides, &mut offsets, &mut |_o, l, _s| {
            total += l;
            Ok(())
        }),
        _ => panic!("unsupported rank"),
    };
    result.unwrap();
    assert_eq!(total, expected, "rank={rank}: total mismatch");
    assert_eq!(offsets, vec![0, 0], "rank={rank}: offsets not reset");
}

#[test]
fn test_macro_kernels_total_elements_1d() {
    verify_kernel_total_and_offsets(1);
}

#[test]
fn test_macro_kernels_total_elements_2d() {
    verify_kernel_total_and_offsets(2);
}

#[test]
fn test_macro_kernels_total_elements_3d() {
    verify_kernel_total_and_offsets(3);
}

#[test]
fn test_macro_kernels_total_elements_4d() {
    verify_kernel_total_and_offsets(4);
}

#[test]
fn test_macro_kernels_total_elements_5d() {
    verify_kernel_total_and_offsets(5);
}

#[test]
fn test_macro_kernels_total_elements_6d() {
    verify_kernel_total_and_offsets(6);
}

#[test]
fn test_macro_kernels_total_elements_7d() {
    verify_kernel_total_and_offsets(7);
}

#[test]
fn test_macro_kernels_total_elements_8d() {
    verify_kernel_total_and_offsets(8);
}

/// Verify that macro-generated kernels visit every element exactly once
/// by collecting all linear offsets and checking them against the expected set.
fn verify_kernel_visits_all_elements(rank: usize) {
    assert!((2..=8).contains(&rank));
    let mut dims = vec![3usize];
    dims.resize(rank, 2);
    let blocks = vec![2usize; rank];

    // Column-major strides (single array for simplicity)
    let mut stride_val = 1isize;
    let mut sv = Vec::new();
    for &d in &dims {
        sv.push(stride_val);
        stride_val *= d as isize;
    }
    let strides = [sv];
    let strides: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();

    // Collect all linear offsets visited
    let mut visited = std::collections::HashSet::new();
    let mut offsets = vec![0isize];

    let result = match rank {
        2 => kernel_2d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        3 => kernel_3d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        4 => kernel_4d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        5 => kernel_5d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        6 => kernel_6d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        7 => kernel_7d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        8 => kernel_8d_inner(&dims, &blocks, &strides, &mut offsets, &mut |o, len, s| {
            for i in 0..len {
                visited.insert(o[0] + (i as isize) * s[0]);
            }
            Ok(())
        }),
        _ => unreachable!(),
    };
    result.unwrap();

    // Expected: all offsets 0..total
    let total: usize = dims.iter().product();
    let expected: std::collections::HashSet<isize> = (0..total as isize).collect();
    assert_eq!(
        visited, expected,
        "rank={rank}: not all elements visited exactly once"
    );
}

#[test]
fn test_macro_kernel_5d_visits_all_elements() {
    verify_kernel_visits_all_elements(5);
}

#[test]
fn test_macro_kernel_6d_visits_all_elements() {
    verify_kernel_visits_all_elements(6);
}

#[test]
fn test_macro_kernel_7d_visits_all_elements() {
    verify_kernel_visits_all_elements(7);
}

#[test]
fn test_macro_kernel_8d_visits_all_elements() {
    verify_kernel_visits_all_elements(8);
}

// ============================================================================
// L1 set fit of the planned tiles (#287)
// ============================================================================

/// Lines a tile may keep live in one 8-way L1 set.
const LIVE_LINES: u128 = 6;

/// The per-set load of the lines a tile keeps live, computed independently of
/// the planner from the L1d geometry the tiles target: 32 KiB, 8 ways, so a
/// 4 KiB way and 64-byte lines.
///
/// For operand `k`, `c_k` is its first axis whose byte stride is below a line;
/// the lines along the axes before it are revisited on the next step of
/// `c_k`. An axis whose byte stride is a multiple of the 4 KiB way puts all
/// `b` of its lines in one set; other axes spread theirs and are not counted.
/// A tile may keep [`LIVE_LINES`] of them: the 8 ways less a quarter of
/// headroom for pseudo-LRU replacement and the stride prefetcher.
fn l1_set_load(blocks: &[usize], strides_list: &[&[isize]], elem_sizes: &[usize]) -> u128 {
    const LINE: u128 = 64;
    const WAY: u128 = 32 * 1024 / 8;
    strides_list
        .iter()
        .zip(elem_sizes)
        .filter(|&(_, &e)| e > 0)
        .filter_map(|(strides, &e)| {
            let byte = |i: usize| strides[i].unsigned_abs() as u128 * e as u128;
            // `c_k = 0`: the operand streams its lines and keeps none live.
            let c = (0..strides.len())
                .find(|&i| byte(i) < LINE)
                .filter(|&c| c > 0)?;
            Some(
                (0..c)
                    .filter(|&i| byte(i) % WAY == 0)
                    .fold(1u128, |load, i| load * blocks[i] as u128),
            )
        })
        .sum()
}

/// Column-major strides of `dims`.
fn col_major(dims: &[usize]) -> Vec<isize> {
    let mut acc = 1isize;
    dims.iter()
        .map(|&d| {
            let s = acc;
            acc *= d as isize;
            s
        })
        .collect()
}

/// Strides of `dims` stored with axis `storage[0]` fastest, then `storage[1]`, ...
fn stored(dims: &[usize], storage: &[usize]) -> Vec<isize> {
    let mut strides = vec![0isize; dims.len()];
    let mut acc = 1isize;
    for &axis in storage {
        strides[axis] = acc;
        acc *= dims[axis] as isize;
    }
    strides
}

/// `D = A op B` layouts whose transposed or permuted operands walk
/// power-of-two strides along the destination's run: the tiles of #287.
/// A named layout: dims and one stride vector per operand, destination first.
type NamedLayout = (&'static str, Vec<usize>, Vec<Vec<isize>>);

fn transposed_operand_layouts() -> Vec<NamedLayout> {
    let mut cases = Vec::new();
    for (m, n) in [(2048, 2048), (1024, 4096), (4096, 1024)] {
        let d = col_major(&[m, n]);
        let t = stored(&[m, n], &[1, 0]);
        cases.push((
            "B transposed",
            vec![m, n],
            vec![d.clone(), d.clone(), t.clone()],
        ));
        cases.push((
            "A and B transposed",
            vec![m, n],
            vec![d.clone(), t.clone(), t.clone()],
        ));
        cases.push((
            "B transposed, A row broadcast",
            vec![m, n],
            vec![d, vec![0, 1], t],
        ));
    }
    let dims = [128usize, 128, 256];
    let d = col_major(&dims);
    for storage in [[2, 0, 1], [1, 2, 0], [2, 1, 0]] {
        cases.push((
            "B permuted",
            dims.to_vec(),
            vec![d.clone(), d.clone(), stored(&dims, &storage)],
        ));
    }
    cases
}

/// Copy-shaped (one destination, one source) layouts whose source is
/// transposed or permuted along the destination's run.
fn transposed_pair_layouts() -> Vec<NamedLayout> {
    let mut cases = Vec::new();
    for (m, n) in [(2048, 2048), (1024, 4096), (4096, 1024)] {
        let d = col_major(&[m, n]);
        let t = stored(&[m, n], &[1, 0]);
        cases.push(("source transposed", vec![m, n], vec![d, t]));
    }
    let dims = [128usize, 128, 256];
    for storage in [[1, 2, 0], [2, 1, 0]] {
        cases.push((
            "source permuted",
            dims.to_vec(),
            vec![col_major(&dims), stored(&dims, &storage)],
        ));
    }
    cases
}

#[test]
fn transposed_pair_tiles_fit_the_l1_sets() {
    for elem in [8usize, 4] {
        for (name, dims, strides) in transposed_pair_layouts() {
            let refs: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
            let sizes = vec![elem; refs.len()];
            let (fused, ordered, plan) = build_plan_fused(&dims, &refs, Some(0), &sizes);
            let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
            let load = l1_set_load(&plan.block, &ordered, &sizes);
            assert!(
                load <= LIVE_LINES,
                "{name} {dims:?} elem {elem}: tiles {:?} over {:?} keep {load} lines live in one 8-way L1 set",
                plan.block.as_slice(),
                fused.as_slice(),
            );
        }
    }
}

/// Walks with more than two operands keep the model's tiles: the #287 gate
/// measured the fit regressing three-operand `mul_into`.
#[test]
fn wider_walks_keep_the_model_tiles() {
    for elem in [8usize, 4] {
        for (name, dims, strides) in transposed_operand_layouts() {
            let refs: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
            let sizes = vec![elem; refs.len()];
            let (fused, ordered, plan) = build_plan_fused(&dims, &refs, Some(0), &sizes);
            let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
            let model = crate::block::compute_block_sizes(&fused, &ordered, &sizes);
            assert_eq!(
                plan.block.as_slice(),
                model.as_slice(),
                "{name} {dims:?} elem {elem}"
            );
        }
    }
}

#[test]
fn transposed_copy_pair_tiles_fit_the_l1_sets() {
    use crate::raw_ops::{pair_tiles, FusedPairLayout, RAW_FUSED_RANK_LIMIT};
    for (m, n) in [(2048usize, 2048usize), (1024, 4096), (4096, 1024)] {
        let mut layout = FusedPairLayout {
            rank: 2,
            dims: [1; RAW_FUSED_RANK_LIMIT],
            dst_strides: [0; RAW_FUSED_RANK_LIMIT],
            src_strides: [0; RAW_FUSED_RANK_LIMIT],
        };
        layout.dims[..2].copy_from_slice(&[m, n]);
        layout.dst_strides[..2].copy_from_slice(&[1, m as isize]);
        layout.src_strides[..2].copy_from_slice(&[n as isize, 1]);
        let tiles = pair_tiles::<f64, f64>(&layout).expect("a transposed copy is tiled");
        let blocks = &tiles.blocks[..2];
        let load = l1_set_load(blocks, &[&[1, m as isize], &[n as isize, 1]], &[8, 8]);
        assert!(
            load <= LIVE_LINES,
            "copy [{m}, {n}] transposed: tiles {blocks:?} keep {load} lines live in one 8-way L1 set"
        );
    }
}

/// Layouts without a large-stride run under a reused operand, a destination
/// with zero-stride (reduced) axes, and the transposed set: the planner's tiles
/// never exceed the model's, equal them where no line is kept live, and keep
/// every destination zero-stride axis at the model's block.
#[test]
fn planned_tiles_never_exceed_the_model_and_keep_the_reduced_axes() {
    let mut cases = transposed_operand_layouts();
    cases.extend(transposed_pair_layouts());
    for (m, n) in [(2048usize, 2048usize), (2047, 2047), (1000, 4100)] {
        let d = col_major(&[m, n]);
        let t = stored(&[m, n], &[1, 0]);
        // No operand keeps a line live: the model's tiles exactly.
        cases.push((
            "column-major",
            vec![m, n],
            vec![d.clone(), d.clone(), d.clone()],
        ));
        cases.push((
            "row broadcast",
            vec![m, n],
            vec![d.clone(), d.clone(), vec![0, 1]],
        ));
        cases.push((
            "column broadcast",
            vec![m, n],
            vec![d.clone(), d.clone(), vec![1, 0]],
        ));
        // A reduction into a vector: the destination's zero-stride axis.
        cases.push(("sum over columns", vec![m, n], vec![vec![1, 0], t.clone()]));
        cases.push(("sum over rows", vec![m, n], vec![vec![0, 1], d.clone()]));
        cases.push((
            "sum over rows of a transpose",
            vec![m, n],
            vec![vec![0, 1], t],
        ));
    }
    for elem in [8usize, 4] {
        for (name, dims, strides) in &cases {
            let refs: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
            let sizes = vec![elem; refs.len()];
            let (fused, ordered, plan) = build_plan_fused(dims, &refs, Some(0), &sizes);
            let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
            let model = crate::block::compute_block_sizes(&fused, &ordered, &sizes);
            for axis in 0..fused.len() {
                assert!(
                    (1..=model[axis]).contains(&plan.block[axis]),
                    "{name} {dims:?} elem {elem}: tiles {:?} outside 1..=model {:?}",
                    plan.block.as_slice(),
                    model.as_slice(),
                );
                if ordered[0][axis] == 0 {
                    assert_eq!(
                        plan.block[axis], model[axis],
                        "{name} {dims:?} elem {elem}: reduced axis {axis} moved from the model"
                    );
                }
            }
            if l1_set_load(&model, &ordered, &sizes) <= LIVE_LINES {
                assert_eq!(
                    plan.block.as_slice(),
                    model.as_slice(),
                    "{name} {dims:?} elem {elem}: tiles within the sets must stay the model's"
                );
            }
            // A fold has no destination; its tiles order the accumulator and
            // stay the model's.
            let (fold_dims, fold_strides, fold_plan) = build_plan_fused(dims, &refs, None, &sizes);
            let fold_strides: Vec<&[isize]> = fold_strides.iter().map(|s| s.as_slice()).collect();
            assert_eq!(
                fold_plan.block.as_slice(),
                crate::block::compute_block_sizes(&fold_dims, &fold_strides, &sizes).as_slice(),
                "{name} {dims:?} elem {elem}: a fold's tiles moved from the model"
            );
        }
    }
}

/// `dot` folds into one accumulator in tile order, so its plan keeps the
/// model's tiles even where a destination-writing plan of the same layout is
/// fitted to the L1 sets (#287): its result bits stay those of the model.
#[test]
fn a_fold_keeps_the_model_tiles_where_a_destination_is_fitted() {
    for (m, n) in [(2048usize, 2048usize), (1024, 4096)] {
        let a = col_major(&[m, n]);
        let b = stored(&[m, n], &[1, 0]);
        let refs: [&[isize]; 2] = [&a, &b];
        let total = m * n;
        let (fused, ordered, fold) = fold_plan(&[m, n], &refs, 0, &[8, 8], total);
        let ordered: Vec<&[isize]> = ordered.iter().map(|s| s.as_slice()).collect();
        let model = crate::block::compute_block_sizes(&fused, &ordered, &[8, 8]);
        assert_eq!(fold.block.as_slice(), model.as_slice(), "[{m}, {n}]");
        let (_, _, fitted) = pointwise_plan(&[m, n], &refs, Some(0), &[8, 8], total);
        assert_ne!(
            fitted.block, fold.block,
            "[{m}, {n}]: the layout must be one the fit moves"
        );
    }
}