tenferro-cpu 0.12.0

CPU backend, kernels, provider selection, and CPU resource pools for tenferro.
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
use super::*;
use crate::Pool;
use crate::{Tensor, TensorRead};
use tenferro_tensor::{TensorDeviceTransfer, TensorElementwise, TensorStructural};

fn pool(threads: usize) -> Pool {
    Pool::builder().threads(threads).build().expect("test pool")
}

fn vector(values: &[f64]) -> Tensor {
    Tensor::from_vec_col_major(vec![values.len()], values.to_vec()).expect("vector")
}

fn engine_error(error: &crate::Error) -> Option<CpuEngineError> {
    let mut source: Option<&(dyn std::error::Error + 'static)> = Some(error);
    while let Some(current) = source {
        if let Some(found) = current.downcast_ref::<CpuEngineError>() {
            return Some(*found);
        }
        source = current.source();
    }
    None
}

#[test]
fn serial_engine_has_no_pool_and_budget_one() {
    let mut engine = CpuEngine::serial();
    assert_eq!(engine.num_threads(), 1);
    assert_eq!(engine.exec().expect("serial token").budget(), 1);
}

#[test]
fn pooled_engine_budget_is_the_pool_size() {
    let mut engine = CpuEngine::new(pool(3));
    assert_eq!(engine.num_threads(), 3);
    assert_eq!(engine.exec().expect("pooled token").budget(), 3);
}

#[test]
fn operation_from_a_worker_of_a_different_pool_runs_on_the_engine() {
    let engine = std::sync::Mutex::new(CpuEngine::new(pool(2)));
    let a = vector(&[1.0, 2.0]);
    let outer = pool(1);
    // A worker of a different pool may call the engine; a worker of its own
    // pool may too, because the token enters the pool in place (no refusal).
    let ok = std::thread::scope(|scope| {
        scope
            .spawn(|| {
                // `Pool` has no `install`; hand the closure to its only worker
                // through a token fork on a one-slot token.
                let token = threadteam::Exec::new(&outer, 1).unwrap();
                let slot = std::sync::Mutex::new(None);
                token.fork(1, |_, _| {
                    *slot.lock().unwrap() = Some(
                        engine
                            .lock()
                            .unwrap()
                            .add_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&a))
                            .map(|tensor| tensor.as_slice::<f64>().map(<[f64]>::to_vec)),
                    );
                });
                slot.into_inner().unwrap().expect("the job ran")
            })
            .join()
            .expect("foreign pool caller")
    });
    assert_eq!(ok.expect("add").expect("f64"), vec![2.0, 4.0]);
}

#[test]
fn panicking_operation_breaks_the_engine() {
    let mut engine = CpuEngine::serial();
    let unwind = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
        let _ = engine.with_linalg_pool::<()>(|_, buffers| {
            let _scratch = <f64 as crate::buffer_pool::PoolScalar>::pool_acquire_zeroed(buffers, 8);
            panic!("exercise the broken flag");
        });
    }));
    assert!(unwind.is_err());
    assert!(engine.is_broken());
    let a = vector(&[1.0]);
    let error = engine
        .neg_read(TensorRead::from_tensor(&a))
        .expect_err("broken engine refuses work");
    assert_eq!(engine_error(&error), Some(CpuEngineError::Broken));
}

#[test]
fn engines_share_a_plan_cache_only_when_given_one() {
    let shared = CpuPlanCache::new();
    let first = CpuEngine::serial().with_plan_cache(shared.clone());
    let second = CpuEngine::serial().with_plan_cache(shared.clone());
    let private = CpuEngine::serial();
    assert!(first.plan_cache().shares_with(second.plan_cache()));
    assert!(!first.plan_cache().shares_with(private.plan_cache()));
}

#[test]
fn buffer_limit_and_reset_apply_to_this_engine() {
    let mut engine = CpuEngine::serial();
    let a = Tensor::from_vec_col_major(vec![2, 3], vec![1.0_f64; 6]).expect("matrix");
    let transposed = engine
        .transpose_read(TensorRead::from_tensor(&a), &[1, 0])
        .expect("transpose");
    tenferro_tensor::TensorBuffer::reclaim_buffer(&mut engine, transposed);
    assert!(engine.buffer_pool_len() > 0);
    engine.set_buffer_pool_limit_bytes(0);
    assert_eq!(engine.buffer_pool_limit_bytes(), 0);
    engine.reset_buffer_pool();
    assert_eq!(engine.buffer_pool_len(), 0);
    engine.clear_runtime_caches();
    assert_eq!(engine.runtime_cache_stats().clears, 1);
}

#[test]
fn batch_policy_setter_returns_previous() {
    let mut engine = CpuEngine::serial();
    let previous = engine.set_batch_policy(crate::CpuBatchPolicy::new(
        crate::CpuBatchStrategy::Sequential,
    ));
    assert_eq!(previous.strategy, crate::CpuBatchStrategy::Auto);
    assert_eq!(
        engine.batch_policy().strategy,
        crate::CpuBatchStrategy::Sequential
    );
}

#[test]
fn nested_engines_keep_their_pools_independent_and_match_the_serial_result() {
    use std::collections::HashSet;
    use std::sync::Barrier;

    const LARGE: usize = crate::tests::native_participants::LARGE_NATIVE_TEST_LEN;
    let mut serial = CpuEngine::serial();
    let (_, expected) = serial
        .with_linalg_pool(|exec, _| {
            Ok(crate::tests::native_participants::run_native_map_of_len(
                LARGE,
                false,
                exec.token(),
            ))
        })
        .expect("serial map");

    for k in [2_usize, 3] {
        for m in [2_usize, 4] {
            let barrier = std::sync::Arc::new(Barrier::new(k));
            let handles: Vec<_> = (0..k)
                .map(|_| {
                    let barrier = std::sync::Arc::clone(&barrier);
                    std::thread::spawn(move || {
                        let mut engine = CpuEngine::new(pool(m));
                        barrier.wait();
                        let (participants, output) = engine
                            .with_linalg_pool(|exec, _| {
                                Ok(crate::tests::native_participants::run_native_map_of_len(
                                    LARGE,
                                    true,
                                    exec.token(),
                                ))
                            })
                            .expect("nested map");
                        // The ledger counts live tasks per token: never more
                        // than the engine's budget, at any moment.
                        #[cfg(feature = "ledger")]
                        assert!(engine.take_ledger_max_live() <= m);
                        (participants, output)
                    })
                })
                .collect();

            let mut workers = HashSet::new();
            for handle in handles {
                let (participants, output) = handle.join().expect("nested caller");
                assert_eq!(output, expected, "k={k} m={m}");
                assert!(
                    participants.max_active() <= m,
                    "k={k} m={m}: {} workers ran at once",
                    participants.max_active()
                );
                assert!(
                    participants.thread_count() <= m,
                    "k={k} m={m}: {} workers participated",
                    participants.thread_count()
                );
                workers.extend(participants.thread_ids());
            }
            assert!(
                workers.len() <= k * m,
                "k={k} m={m}: {} tenferro-controlled workers across {k} engines",
                workers.len()
            );
        }
    }
}

#[test]
fn a_small_strided_operation_stays_on_the_calling_thread_without_a_pool_entry() {
    const SMALL: usize = 64;
    let mut serial = CpuEngine::serial();
    let (_, expected) = serial
        .with_linalg_pool(|exec, _| {
            Ok(crate::tests::native_participants::run_native_map_of_len(
                SMALL,
                false,
                exec.token(),
            ))
        })
        .expect("serial map");
    // A serial engine mints a serial token: no root, no pool.
    #[cfg(feature = "ledger")]
    assert!(serial.take_ledger_roots().is_empty());

    let mut engine = CpuEngine::new(pool(4));
    let caller = std::thread::current().id();
    let (participants, actual) = engine
        .with_linalg_pool(|exec, _| {
            Ok(crate::tests::native_participants::run_native_map_of_len(
                SMALL,
                false,
                exec.token(),
            ))
        })
        .expect("pooled map");
    assert_eq!(actual, expected);
    assert_eq!(participants.thread_ids(), vec![caller]);
    assert_eq!(participants.max_active(), 1);
    assert_eq!(participants.thread_count(), 1);
    #[cfg(feature = "ledger")]
    {
        let roots = engine.take_ledger_roots();
        assert!(!roots.is_empty());
        for root in roots {
            assert_eq!(threadteam::ledger::installs(root), 0);
        }
    }
    assert_eq!(std::thread::current().id(), caller);
}

#[test]
fn a_small_direct_faer_call_stays_on_the_calling_thread_without_a_pool_entry() {
    let mut engine = CpuEngine::new(pool(4));
    let caller = std::thread::current().id();
    let (same_thread, sequential, width) = engine
        .with_linalg_pool(|exec, _| {
            let width = exec
                .batch_decision(&cpueinsum::BatchQuery::new(1, 8.0))
                .item_threads;
            let observed = exec
                .token()
                .limit(std::num::NonZeroUsize::new(width).unwrap())
                .external(|granted| {
                    (
                        std::thread::current().id() == caller,
                        granted.get() == 1,
                        granted.get(),
                    )
                });
            Ok((observed.0, observed.1, width))
        })
        .expect("direct faer decision");
    assert!(same_thread);
    assert!(sequential);
    assert_eq!(width, 1);
    #[cfg(feature = "ledger")]
    assert_eq!(engine.take_ledger_installs(), 0);
}

#[test]
fn elementwise_families_match_serial_on_pooled_engines_at_one_two_and_four_threads() {
    let a = Tensor::from_vec_col_major(
        vec![4096],
        (0..4096)
            .map(|index| index as f64 * 0.25 - 7.0)
            .collect::<Vec<_>>(),
    )
    .expect("a");
    let b = Tensor::from_vec_col_major(
        vec![4096],
        (0..4096)
            .map(|index| (index % 13) as f64 + 1.5)
            .collect::<Vec<_>>(),
    )
    .expect("b");

    let mut serial = CpuEngine::serial();
    let serial_add = serial
        .add_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&b))
        .expect("serial add");
    let serial_mul = serial
        .mul_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&b))
        .expect("serial mul");

    for threads in [1, 2, 4] {
        let mut engine = CpuEngine::new(pool(threads));
        let add = engine
            .add_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&b))
            .expect("pooled add");
        let mul = engine
            .mul_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&b))
            .expect("pooled mul");
        assert_eq!(
            add.as_slice::<f64>().unwrap(),
            serial_add.as_slice::<f64>().unwrap()
        );
        assert_eq!(
            mul.as_slice::<f64>().unwrap(),
            serial_mul.as_slice::<f64>().unwrap()
        );
    }
}

#[test]
fn fast_paths_refuse_a_broken_engine_before_touching_storage() {
    let mut engine = CpuEngine::serial();
    let unwind = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
        let _ = engine.with_linalg_pool::<()>(|_, _| panic!("break the engine"));
    }));
    assert!(unwind.is_err());
    assert!(engine.is_broken());

    let input = vector(&[1.0, 2.0]);
    let read = TensorRead::from_tensor(&input);
    let upload = engine
        .upload_host_tensor(read.clone())
        .expect_err("a broken engine refuses upload");
    let download = engine
        .download_to_host(read.clone())
        .expect_err("a broken engine refuses download");
    let reshape = engine
        .reshape_read(read, &[2])
        .expect_err("a broken engine refuses the small reshape fast path");
    for error in [upload, download, reshape] {
        assert_eq!(engine_error(&error), Some(CpuEngineError::Broken));
    }
}

fn carries_team(engine: &mut CpuEngine) -> bool {
    engine
        .with_linalg_pool(|exec, _| Ok(exec.token().carries_team()))
        .expect("token")
}

#[test]
fn a_multi_thread_engine_mints_team_tokens_and_smaller_engines_do_not() {
    for (mut engine, team) in [
        (CpuEngine::new(pool(3)), true),
        (CpuEngine::new(pool(1)), false),
        (CpuEngine::serial(), false),
    ] {
        assert_eq!(engine.has_team(), team);
        assert_eq!(carries_team(&mut engine), team);
    }
}

#[test]
fn a_reentrant_token_request_runs_serially_instead_of_panicking() {
    let resources = crate::contraction::ExecutionResources::new(Some(pool(3)));
    let policy = crate::CpuBatchPolicy::default();
    let mut outer = resources.exec(policy);
    assert!(outer.token().carries_team());
    assert_eq!(outer.budget(), 3);
    // Same thread, token alive: serial token, no panic, no wait.
    let mut inner = resources.exec(policy);
    assert!(!inner.token().carries_team());
    assert_eq!(inner.budget(), 1);
    drop(inner);
    drop(outer);
    // The team is free again.
    assert!(resources.exec(policy).token().carries_team());
}

#[test]
fn a_second_thread_waits_for_the_leader_instead_of_taking_a_second_team() {
    let resources = crate::contraction::ExecutionResources::new(Some(pool(3)));
    let policy = crate::CpuBatchPolicy::default();
    let (tx, rx) = std::sync::mpsc::channel();
    std::thread::scope(|scope| {
        let mut leader = resources.exec(policy);
        assert!(leader.token().carries_team());
        scope.spawn(|| {
            let mut exec = resources.exec(policy);
            tx.send(exec.token().carries_team()).unwrap();
        });
        // The second caller blocks while this thread leads.
        assert!(rx
            .recv_timeout(std::time::Duration::from_millis(200))
            .is_err());
        drop(leader);
        // Once the leader is gone it gets the team, not a fallback.
        assert!(rx.recv_timeout(std::time::Duration::from_secs(30)).unwrap());
    });
}

#[test]
fn two_engines_on_two_threads_each_lead_their_own_team_at_once() {
    // The per-lane pattern: engines are built on one thread and led from
    // scoped threads that are not workers of any engine's pool.
    let engines = [CpuEngine::new(pool(2)), CpuEngine::new(pool(2))];
    let barrier = std::sync::Barrier::new(2);
    let a = vector(&(0..4096).map(f64::from).collect::<Vec<_>>());
    let results: Vec<_> = std::thread::scope(|scope| {
        let handles: Vec<_> = engines
            .into_iter()
            .map(|mut engine| {
                let (barrier, a) = (&barrier, &a);
                scope.spawn(move || {
                    // Both teams are led at the same moment.
                    let team = engine
                        .with_linalg_pool(|exec, _| {
                            let team = exec.token().carries_team();
                            barrier.wait();
                            Ok(team)
                        })
                        .expect("token");
                    let sum = engine
                        .add_read(TensorRead::from_tensor(a), TensorRead::from_tensor(a))
                        .expect("add");
                    (team, sum.as_slice::<f64>().expect("f64")[4095])
                })
            })
            .collect();
        handles.into_iter().map(|h| h.join().unwrap()).collect()
    });
    assert_eq!(results, vec![(true, 8190.0), (true, 8190.0)]);
}

#[test]
fn with_exec_lends_the_full_budget_and_breaks_the_engine_on_panic() {
    let mut engine = CpuEngine::new(pool(2));
    let caller = std::thread::current().id();
    let (budget, same_thread) = engine
        .with_exec(|exec| (exec.budget(), std::thread::current().id() == caller))
        .expect("healthy engine");
    assert_eq!(budget, engine.num_threads());
    assert!(same_thread);

    let unwind = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
        engine.with_exec(|_| panic!("break the engine"))
    }));
    assert!(unwind.is_err());
    let error = engine.with_exec(|_| ()).expect_err("broken engine refuses");
    assert_eq!(engine_error(&error), Some(CpuEngineError::Broken));
}

#[test]
fn engine_wide_cache_controls_cover_the_extension_store() {
    use tenferro_runtime::{ExecutionEngine, ExtensionCacheKey};

    let mut engine = CpuEngine::serial();
    let before = engine.runtime_cache_stats();
    let key = ExtensionCacheKey::new("tenferro-cpu.test.engine-controls.v1", "entries", 0);
    let mut caches = engine.take_extension_caches();
    caches.put(key, 7_u64, 96);
    engine.restore_extension_caches(caches);

    let stored = engine.runtime_cache_stats();
    assert_eq!(stored.entries, before.entries + 1);
    assert_eq!(stored.retained_bytes, before.retained_bytes + 96);
    assert_eq!(
        (stored.hits, stored.misses, stored.evictions),
        (before.hits, before.misses, before.evictions),
        "extension store events stay out of the plan counters"
    );

    engine.clear_runtime_caches();
    let cleared = engine.runtime_cache_stats();
    assert_eq!((cleared.entries, cleared.retained_bytes), (0, 0));
    let mut caches = engine.take_extension_caches();
    assert!(caches.get::<u64>(&key).is_none());
    engine.restore_extension_caches(caches);
}

#[test]
fn dispatch_costs_price_every_width_as_the_engines_token_does() {
    use std::num::NonZeroUsize;
    use threadteam::DispatchKind;
    for mut engine in [CpuEngine::new(pool(3)), CpuEngine::serial()] {
        let costs = *engine.dispatch_costs();
        let team = engine.has_team();
        let budget = engine.num_threads();
        let mut exec = engine.exec().expect("token");
        for width in (1..=budget).filter_map(NonZeroUsize::new) {
            let expected = if team && width.get() >= 2 {
                costs.cost(DispatchKind::Team, width)
            } else {
                std::time::Duration::MAX
            };
            assert_eq!(exec.token().dispatch_cost(width), expected, "width {width}");
        }
    }
}

#[test]
fn with_guard_has_the_entry_and_unwind_contract_of_with_exec() {
    for (mut engine, panicking_entry) in [
        (CpuEngine::new(pool(2)), "with_guard"),
        (CpuEngine::new(pool(2)), "with_exec"),
        (CpuEngine::serial(), "with_guard"),
    ] {
        let policy = crate::CpuBatchPolicy::new(crate::CpuBatchStrategy::Outer);
        engine.set_batch_policy(policy);
        let caller = std::thread::current().id();
        let (seen, same_thread) = engine
            .with_guard(|ctx| (ctx.batch_policy(), std::thread::current().id() == caller))
            .expect("healthy engine");
        assert_eq!(seen, policy);
        assert!(same_thread);

        let unwind = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            if panicking_entry == "with_guard" {
                let _ = engine.with_guard(|_| panic!("break the engine"));
            } else {
                let _ = engine.with_exec(|_| panic!("break the engine"));
            }
        }));
        assert!(unwind.is_err());
        assert!(engine.is_broken(), "{panicking_entry}");
        let guard = engine
            .with_guard(|_| ())
            .expect_err("broken engine refuses");
        let exec = engine.with_exec(|_| ()).expect_err("broken engine refuses");
        assert_eq!(engine_error(&guard), Some(CpuEngineError::Broken));
        assert_eq!(engine_error(&exec), Some(CpuEngineError::Broken));
    }
}

#[cfg(feature = "ledger")]
#[test]
fn with_guard_mints_no_token_where_with_exec_mints_one() {
    let mut engine = CpuEngine::new(pool(3));
    assert!(engine.has_team());
    engine.with_guard(|_| ()).expect("healthy engine");
    assert!(engine.take_ledger_roots().is_empty());
    engine.with_exec(|_| ()).expect("healthy engine");
    assert_eq!(engine.take_ledger_roots().len(), 1);
}

#[test]
fn with_guard_bounds_the_workspace_it_lends_on_return_and_unwind() {
    fn grow(ctx: &SerialCtx<'_>) {
        let req = cpueinsum::WorkspaceReq {
            a_bytes: 4096,
            ..Default::default()
        };
        ctx.workspace().with_worker(&req, &mut |_, _, _| {});
    }
    for (limit, retained) in [(0, 0), (usize::MAX, 4096)] {
        let mut engine = CpuEngine::serial();
        engine.set_buffer_pool_limit_bytes(limit);
        let before = engine.runtime_cache_stats().retained_bytes;
        engine.with_guard(|ctx| grow(&ctx)).expect("healthy engine");
        let after = engine.runtime_cache_stats().retained_bytes;
        assert_eq!(after - before, retained, "return, limit {limit}");

        let mut engine = CpuEngine::serial();
        engine.set_buffer_pool_limit_bytes(limit);
        let before = engine.runtime_cache_stats().retained_bytes;
        let unwind = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            let _ = engine.with_guard(|ctx| {
                grow(&ctx);
                panic!("break the engine")
            });
        }));
        assert!(unwind.is_err());
        let after = engine.runtime_cache_stats().retained_bytes;
        assert_eq!(after - before, retained, "unwind, limit {limit}");
    }
}