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
use super::*;
use strided_view::auxiliary::index_order;

#[test]
fn test_fuse_dims_contiguous() {
    // Two contiguous dimensions: [3, 4] with strides [1, 3] -> fused to [12, 1]
    let dims = [3, 4];
    let strides1 = [1isize, 3];
    let strides2 = [1isize, 3];
    let all_strides: Vec<&[isize]> = vec![&strides1, &strides2];

    let fused = fuse_dims(&dims, &all_strides);
    assert_eq!(fused.as_slice(), vec![12, 1]);
}

#[test]
fn test_fuse_dims_allows_broadcast_operand_across_fused_axes() {
    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 all_strides: Vec<&[isize]> = vec![&out, &lhs, &rhs];

    let fused = fuse_dims(&dims, &all_strides);

    assert_eq!(fused.as_slice(), vec![256, 1, 64, 64]);
}

#[test]
fn test_fuse_dims_non_contiguous() {
    // Non-contiguous: strides don't match
    let dims = [3, 4];
    let strides1 = [1isize, 10]; // Not contiguous (should be 3)
    let all_strides: Vec<&[isize]> = vec![&strides1];

    let fused = fuse_dims(&dims, &all_strides);
    assert_eq!(fused.as_slice(), vec![3, 4]); // No fusion
}

#[test]
fn test_fuse_dims_partial() {
    // 3D: first two fuse, third doesn't
    let dims = [2, 3, 4];
    let strides = [1isize, 2, 100]; // dims[0]*strides[0]=2=strides[1], but 6≠100
    let all_strides: Vec<&[isize]> = vec![&strides];

    let fused = fuse_dims(&dims, &all_strides);
    assert_eq!(fused.as_slice(), vec![6, 1, 4]); // Fused first two
}

#[test]
fn test_fuse_dims_multiple_arrays() {
    // Only fuse if ALL arrays are contiguous
    let dims = [3, 4];
    let strides1 = [1isize, 3]; // Contiguous
    let strides2 = [1isize, 10]; // Not contiguous
    let all_strides: Vec<&[isize]> = vec![&strides1, &strides2];

    let fused = fuse_dims(&dims, &all_strides);
    assert_eq!(fused.as_slice(), vec![3, 4]); // No fusion because strides2 isn't contiguous
}

#[test]
fn test_compute_importance_2_arrays() {
    // Example with 2 arrays, dims [4, 5]
    let dims = [4usize, 5];
    let strides1 = [1isize, 4]; // Column-major output
    let strides2 = [5isize, 1]; // Row-major input
    let all_strides: Vec<&[isize]> = vec![&strides1, &strides2];

    let order1 = index_order(&strides1);
    let order2 = index_order(&strides2);
    let index_orders = vec![order1.into_vec(), order2.into_vec()];

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    // With output weighted 2x, dimension 0 should have higher importance
    // since it has smaller stride in the output array
    assert!(importance[0] > importance[1]);
}

#[test]
fn test_sort_by_importance() {
    let importance = vec![100u64, 50, 200, 10];
    let perm = sort_by_importance(&importance);
    assert_eq!(perm.as_slice(), vec![2, 0, 1, 3]); // Indices sorted by descending importance
}

#[test]
fn test_compute_costs() {
    let strides1 = [1isize, 4, 0];
    let strides2 = [2isize, 1, 0];
    let all_strides: Vec<&[isize]> = vec![&strides1, &strides2];

    let costs = compute_costs(&all_strides);
    // min strides: [1, 1, 0], transformed: [2, 2, 1]
    assert_eq!(costs.as_slice(), vec![2, 2, 1]);
}

#[test]
fn test_compute_importance_with_zero_stride() {
    // Zero stride (broadcast) still gets index_order = 1 for compatibility,
    // but it does not contribute to ordering importance.
    let dims = [4usize, 5];
    let strides1 = [0isize, 1]; // First dim is broadcast
    let all_strides: Vec<&[isize]> = vec![&strides1];

    let order1 = index_order(&strides1);
    // For stride 0: order = 1 (zero strides always get 1)
    // For stride 1: order = 1 (no non-zero stride < 1)
    assert_eq!(order1.as_slice(), &[1, 1]);

    let index_orders = vec![order1.into_vec()];
    let importance = compute_importance(&dims, &all_strides, &index_orders);

    assert_eq!(importance[0], 0);
    assert!(importance[1] > 0);
}

#[test]
fn test_compute_importance_size_one_dim() {
    // Size-1 dimensions get zero importance
    let dims = [4usize, 1, 5];
    let strides1 = [1isize, 4, 4];
    let all_strides: Vec<&[isize]> = vec![&strides1];

    let order1 = index_order(&strides1);
    let index_orders = vec![order1.into_vec()];
    let importance = compute_importance(&dims, &all_strides, &index_orders);

    // Dimension 1 has size 1 -> importance = 0
    assert_eq!(importance[1], 0);
    // Other dimensions should have non-zero importance
    assert!(importance[0] > 0);
    assert!(importance[2] > 0);
}

#[test]
fn test_compute_importance_output_weight() {
    // Output (first array) is strongly weighted
    // With same strides, dimension with smaller stride in output wins
    let dims = [4usize, 5];
    let out_strides = [1isize, 4]; // Column-major output
    let in_strides = [5isize, 1]; // Row-major input
    let all_strides: Vec<&[isize]> = vec![&out_strides, &in_strides];

    let order_out = index_order(&out_strides); // [1, 2]
    let order_in = index_order(&in_strides); // [2, 1]
    let index_orders = vec![order_out.into_vec(), order_in.into_vec()];

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    // Output weighting makes dimension 0 (smaller stride in output) win.
    assert!(importance[0] > importance[1]);
}

#[test]
fn test_compute_costs_owned_vecs() {
    // Ported from threading.rs: verify compute_costs works with Vec<Vec<isize>>
    let strides_list: Vec<Vec<isize>> = vec![vec![1, 0, 3], vec![2, 0, 4]];
    let costs = compute_costs(&strides_list);
    assert_eq!(costs.as_slice(), vec![2, 1, 6]);
}

#[test]
fn test_compute_costs_with_zero() {
    // Zero strides become cost 1, non-zero become 2*abs
    let strides1 = [0isize, 2, -3];
    let strides2 = [1isize, 0, 2];
    let all_strides: Vec<&[isize]> = vec![&strides1, &strides2];

    let costs = compute_costs(&all_strides);
    // min abs: [0, 0, 2]
    // transform: [1, 1, 4]
    assert_eq!(costs.as_slice(), vec![1, 1, 4]);
}

// ---- compress_dims tests ----

/// Compare the compressed per-operand stride lists against expected slices.
fn assert_strides(cs: &[crate::inline::InlineVec<isize>], expected: &[&[isize]]) {
    assert_eq!(cs.len(), expected.len());
    for (actual, expected) in cs.iter().zip(expected) {
        assert_eq!(actual.as_slice(), *expected);
    }
}

#[test]
fn test_compress_dims_removes_fused() {
    let dims = vec![12usize, 1];
    let strides = vec![vec![1isize, 3]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![12]);
    assert_strides(&cs, &[&[1]]);
}

#[test]
fn test_compress_dims_removes_multiple() {
    let dims = vec![6usize, 1, 4];
    let strides = vec![vec![1isize, 2, 100]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![6, 4]);
    assert_strides(&cs, &[&[1, 100]]);
}

#[test]
fn test_compress_dims_no_removal() {
    let dims = vec![3usize, 4];
    let strides = vec![vec![1isize, 3]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![3, 4]);
    assert_strides(&cs, &[&[1, 3]]);
}

#[test]
fn test_compress_dims_all_ones() {
    let dims = vec![1usize, 1, 1];
    let strides = vec![vec![1isize, 1, 1]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![1]);
    assert_strides(&cs, &[&[1]]);
}

#[test]
fn test_compress_dims_multi_arrays() {
    let dims = vec![6usize, 1, 4];
    let strides = vec![vec![1isize, 6, 6], vec![4isize, 24, 1]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![6, 4]);
    assert_strides(&cs, &[&[1, 6], &[4, 1]]);
}

#[test]
fn test_compress_dims_single_dim() {
    let dims = vec![5usize];
    let strides = vec![vec![1isize]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![5]);
    assert_strides(&cs, &[&[1]]);
}

#[test]
fn test_compress_dims_single_dim_one() {
    let dims = vec![1usize];
    let strides = vec![vec![1isize]];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert_eq!(cd.as_slice(), vec![1]);
    assert_strides(&cs, &[&[1]]);
}

#[test]
fn test_compress_dims_empty() {
    let dims: Vec<usize> = vec![];
    let strides: Vec<Vec<isize>> = vec![];
    let (cd, cs) = compress_dims(&dims, &strides);
    assert!(cd.is_empty());
    assert!(cs.is_empty());
}

#[test]
fn test_compute_importance_high_rank_levels_fit_in_u64() {
    // tensor4all/strided-rs#450: `g * (n - order)` reached 93 bits at rank 32
    // with 3 operands, so `1u64 << shift` panicked in debug and wrapped in
    // release. Weights must stay exact enough to keep the level order.
    let mut dims = vec![1usize; 32];
    dims[0] = 2048;
    let mut strides0 = vec![2048isize; 32];
    strides0[0] = 1;
    let strides = [strides0.clone(), strides0.clone(), strides0];
    let all_strides: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
    let index_orders: Vec<Vec<usize>> = strides.iter().map(|s| index_order(s).into_vec()).collect();

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    assert!(importance[0] > 0);
    // Singleton axes carry no ordering information.
    assert!(importance[1..].iter().all(|&w| w == 0));
}

#[test]
fn test_compute_importance_high_rank_keeps_level_order() {
    // Column-major strides on 32 live axes with 3 operands: 98 bits naive (the
    // highest term is 2^97), so a
    // wrapping shift scrambles the level order instead of panicking (#450).
    let n = 32;
    let dims = vec![2usize; n];
    let strides: Vec<isize> = (0..n).map(|i| 1isize << i).collect();
    let all_strides: Vec<&[isize]> = (0..3).map(|_| strides.as_slice()).collect();
    let index_orders: Vec<Vec<usize>> = (0..3).map(|_| index_order(&strides).into_vec()).collect();

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    // Importance drops with the stride order, so it must be strictly decreasing.
    for i in 1..n {
        assert!(
            importance[i - 1] > importance[i],
            "axis {} ({}) is not above axis {} ({})",
            i - 1,
            importance[i - 1],
            i,
            importance[i]
        );
    }
}

#[test]
fn test_compute_importance_keeps_the_original_width_when_it_fits() {
    // Rank 32, two operands, striding only in the input array: the original
    // arithmetic fits (2^62), so the per-level width must stay `g` and the
    // weight must stay exact. Narrowing here would reorder the plan (review of
    // #450: a width of 1 bit gives 2^31).
    let n = 32;
    let mut dims = vec![1usize; n];
    dims[0] = 2048;
    let out_strides = vec![0isize; n];
    let mut in_strides = vec![2048isize; n];
    in_strides[0] = 2;
    let all_strides: [&[isize]; 2] = [&out_strides, &in_strides];
    let index_orders: Vec<Vec<usize>> = [&out_strides[..], &in_strides[..]]
        .iter()
        .map(|s| index_order(s).into_vec())
        .collect();

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    assert_eq!(importance[0], 1u64 << 62);
    assert!(importance[1..].iter().all(|&w| w == 0));
}

#[test]
fn test_compute_importance_rank_64_shares_the_top_weight() {
    // Past rank ~61 even one bit per level overflows, so the top levels share a
    // weight: the plan still runs, and the two live axes come out equal.
    let n = 64;
    let mut dims = vec![1usize; n];
    dims[0] = 1024;
    dims[1] = 2;
    let mut strides = vec![4096isize; n];
    strides[0] = 4;
    strides[1] = 1;
    let all_strides: Vec<&[isize]> = (0..3).map(|_| strides.as_slice()).collect();
    let index_orders: Vec<Vec<usize>> = (0..3).map(|_| index_order(&strides).into_vec()).collect();

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    // The widest cap that fits is 61: 4 * 2^61 from the output plus 2 * 2^61.
    assert_eq!(importance[0], 6u64 << 61);
    assert_eq!(importance[0], importance[1]);
    assert!(importance[2..].iter().all(|&w| w == 0));
}

#[test]
fn test_compute_importance_rank_65_broadcast_output_keeps_live_axes_distinct() {
    // Review of #450: the fallback searched level caps only up to 61, so a
    // shape whose live axes are on the input side (zero-strided output) lost a
    // distinction that does fit: cap 63 gives 2^63 and 2^62 here, cap 61 gives
    // 2^61 and 2^61. The cap must be chosen per shape, not capped at 61.
    let n = 65;
    let mut dims = vec![1usize; n];
    dims[0] = 2;
    dims[2] = 2;
    let out_strides = vec![0isize; n];
    let in_strides: Vec<isize> = (0..n).map(|i| i as isize + 1).collect();
    let all_strides: [&[isize]; 2] = [&out_strides, &in_strides];
    let index_orders: Vec<Vec<usize>> = [&out_strides[..], &in_strides[..]]
        .iter()
        .map(|s| index_order(s).into_vec())
        .collect();

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    assert_eq!(
        importance[0],
        1u64 << 63,
        "axis 0 (stride 1) takes the top level"
    );
    assert_eq!(
        importance[2],
        1u64 << 62,
        "axis 2 (stride 3) keeps the level below it"
    );
    assert!(importance
        .iter()
        .enumerate()
        .all(|(i, &w)| i == 0 || i == 2 || w == 0));
}

#[test]
fn test_compute_importance_rank_65_two_inputs_reject_cap_63() {
    // Review of #450: the previous test has one input, so every term fits at
    // cap 63. A weight is the *sum* over operands, and two inputs on the same
    // live axis each bring 2^63, which overflows: the fallback must reject cap
    // 63 and settle on 62, where 2^62 + 2^62 is 2^63. Saturating the sum
    // instead would accept 63 and return u64::MAX.
    let n = 65;
    let mut dims = vec![1usize; n];
    dims[0] = 2;
    let out_strides = vec![0isize; n];
    let in_strides: Vec<isize> = (0..n).map(|i| i as isize + 1).collect();
    let all_strides: [&[isize]; 3] = [&out_strides, &in_strides, &in_strides];
    let index_orders: Vec<Vec<usize>> = [&out_strides[..], &in_strides[..], &in_strides[..]]
        .iter()
        .map(|s| index_order(s).into_vec())
        .collect();

    let importance = compute_importance(&dims, &all_strides, &index_orders);

    assert_eq!(
        importance[0],
        1u64 << 63,
        "two inputs at cap 62: 2^62 + 2^62"
    );
    assert!(importance[1..].iter().all(|&w| w == 0));
}

#[test]
fn test_compute_importance_matches_exact_reference() {
    // The original formula evaluated without overflow (u128). Wherever it fits,
    // the 64-bit weights must be identical, so the plan order is unchanged.
    //
    // The grid, for every rank `n = 1..=4` and operand count `m = 1..=4`:
    //   * every independent stride permutation of every operand (the two
    //     budget notes below say where `n = 4, m >= 3` drops to every pair),
    //   * every zero-stride mask (all `2^n` subsets) on the output, on each
    //     input in turn, and on every operand at once,
    //   * every size-1 variant with at most two singleton axes, including the
    //     non-adjacent `[2, 1, 2]` shape.
    fn reference(dims: &[usize], strides: &[&[isize]], orders: &[Vec<usize>]) -> Vec<u128> {
        let (n, m) = (dims.len(), strides.len());
        if n == 0 || m == 0 {
            return Vec::new();
        }
        let g = 64 - (m as u64 + 1).leading_zeros();
        let mut weights = vec![0u128; n];
        for i in 0..n {
            if strides[0][i] != 0 {
                let shift = g * (n - orders[0][i]) as u32;
                weights[i] = (1u128 << (g + 1)) * (1u128 << shift);
            }
        }
        for k in 1..m {
            for i in 0..n {
                if strides[k][i] != 0 {
                    weights[i] += 1u128 << (g * (n - orders[k][i]) as u32);
                }
            }
        }
        for i in 0..n {
            if dims[i] <= 1 {
                weights[i] = 0;
            }
        }
        weights
    }

    fn permutations(items: &[isize]) -> Vec<Vec<isize>> {
        fn rec(items: &mut [isize], pos: usize, out: &mut Vec<Vec<isize>>) {
            if pos == items.len() {
                out.push(items.to_vec());
                return;
            }
            for i in pos..items.len() {
                items.swap(pos, i);
                rec(items, pos + 1, out);
                items.swap(pos, i);
            }
        }
        let mut items = items.to_vec();
        let mut out = Vec::new();
        rec(&mut items, 0, &mut out);
        out
    }

    /// One operand stride tuple per combination: the operands at `crossed`
    /// range over every permutation of the power-of-two strides independently,
    /// the rest keep the column-major (identity) permutation.
    fn stride_tuples(perms: &[Vec<isize>], m: usize, crossed: &[usize]) -> Vec<Vec<Vec<isize>>> {
        let identity = perms[0].clone();
        let mut tuples = Vec::new();
        let mut pick = vec![0usize; crossed.len()];
        loop {
            let mut tuple = vec![identity.clone(); m];
            for (slot, &operand) in crossed.iter().enumerate() {
                tuple[operand] = perms[pick[slot]].clone();
            }
            tuples.push(tuple);
            let mut carry = true;
            for p in pick.iter_mut() {
                if !carry {
                    break;
                }
                *p += 1;
                if *p == perms.len() {
                    *p = 0;
                } else {
                    carry = false;
                }
            }
            if carry {
                break;
            }
        }
        tuples
    }

    fn check(dims: &[usize], strides: &[Vec<isize>]) {
        let all_strides: Vec<&[isize]> = strides.iter().map(|s| s.as_slice()).collect();
        let orders: Vec<Vec<usize>> = strides.iter().map(|s| index_order(s).into_vec()).collect();

        let got: Vec<u128> = compute_importance(dims, &all_strides, &orders)
            .as_slice()
            .iter()
            .map(|&w| w as u128)
            .collect();
        let want = reference(dims, &all_strides, &orders);
        assert_eq!(
            got.as_slice(),
            want.as_slice(),
            "dims {:?} strides {:?}",
            dims,
            strides
        );
    }

    let mut broadcast_cases = 0usize;

    for n in 1..=4usize {
        let powers: Vec<isize> = (0..n).map(|i| 1isize << i).collect();
        let perms = permutations(&powers);

        // Every size-1 variant with at most two singleton axes: none, each
        // single axis, and every axis pair (including the non-adjacent
        // `[2, 1, 2]` shape).
        let mut dim_grids: Vec<Vec<usize>> = vec![vec![2usize; n]];
        for i in 0..n {
            let mut dims = vec![2usize; n];
            dims[i] = 1;
            dim_grids.push(dims);
        }
        for i in 0..n {
            for j in (i + 1)..n {
                let mut dims = vec![2usize; n];
                dims[i] = 1;
                dims[j] = 1;
                dim_grids.push(dims);
            }
        }

        for m in 1..=4usize {
            // Stride permutations for the size-1 sweep: every operand
            // independent while that stays within the test budget (n <= 3 any
            // m, n = 4 up to m = 2). The exhaustive `24^3 + 24^4 = 345 600`
            // grid at n = 4 is what that budget excludes -- no timing is
            // claimed for it here -- so `n = 4, m >= 3` crosses every pair of
            // operands instead (`24^2` per pair, the rest column-major).
            let independent: Vec<Vec<usize>> = if perms.len().pow(m as u32) <= 2000 {
                vec![(0..m).collect()]
            } else {
                (0..m)
                    .flat_map(|i| ((i + 1)..m).map(move |j| vec![i, j]))
                    .collect()
            };
            let mut tuples: Vec<Vec<Vec<isize>>> = Vec::new();
            for crossed in &independent {
                tuples.extend(stride_tuples(&perms, m, crossed));
            }

            // Every stride permutation against every size-1 variant.
            for dims in &dim_grids {
                for strides in &tuples {
                    check(dims, strides);
                }
            }

            // Every zero-stride mask (a broadcast axis) on every operand
            // position: the output, each input in turn, and all operands at
            // once. This sweep multiplies by `1 + (2^n - 1) * (m + 1)`, so it
            // keeps full permutation independence only up to `(n!)^m <= 600`
            // and permutes one operand at a time above that. Each case gets one
            // dims shape, taken from the grid in rotation, so the sweep covers
            // every shape but no single case sees more than one.
            let broadcast_tuples: Vec<Vec<Vec<isize>>> = if perms.len().pow(m as u32) <= 600 {
                tuples.clone()
            } else {
                let mut t = Vec::new();
                for operand in 0..m {
                    t.extend(stride_tuples(&perms, m, &[operand]));
                }
                t
            };
            for strides in &broadcast_tuples {
                for zero_mask in 0..(1u32 << n) {
                    // Every mask on every position; a zero mask is the same
                    // tuple for every position, so it is checked once.
                    let receivers = if zero_mask == 0 { 1 } else { m + 1 };
                    for role in 0..receivers {
                        let mut strides = strides.clone();
                        for (k, operand) in strides.iter_mut().enumerate() {
                            if role == m || role == k {
                                for (i, s) in operand.iter_mut().enumerate() {
                                    if zero_mask & (1 << i) != 0 {
                                        *s = 0;
                                    }
                                }
                            }
                        }
                        let dims = &dim_grids[broadcast_cases % dim_grids.len()];
                        broadcast_cases += 1;
                        check(dims, &strides);
                    }
                }
            }
        }
    }
}