launchbound-bench 2.2.1

Measurement: bench plans, the PTX runner (CUDA driver API via dlopen), and interval statistics
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
//! Plan execution: resumable, checkpointed after every candidate (the box
//! dies — idle guard, dead-man switch, spot reclaim — so the harness is
//! resumable or it is broken, docs/ARCHITECTURE.md), with a CPU heartbeat so a
//! GPU-bound sweep never looks idle to the 30-minute CPU alarm.

use crate::cuda::Device;
use crate::plan::{ArgSpec, BenchPlan, Candidate};
use crate::stats::{Summary, summarize};
use serde::{Deserialize, Serialize};
use std::path::Path;
use std::sync::atomic::{AtomicBool, Ordering};
use std::time::Instant;

/// What measuring one candidate produced — including the ways it can fail.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CandidateResult {
    /// Its canonical `config.v1` ID, joining this to the plan and verdicts.
    pub id: String,
    /// Its dimension assignments.
    pub config: String,
    /// `ok`, `error`, or `timeout` — the last meaning a watchdog fired on
    /// a candidate the gate had refused, which is the refusal being right.
    pub status: String,
    /// Why it failed, when it did.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub error: Option<String>,
    /// Untimed launches performed before measuring.
    pub warmup: u32,
    /// Timed launches requested.
    pub repeats: u32,
    /// Every raw sample, in launch order. Kept so a summary can be
    /// recomputed without re-running the GPU.
    #[serde(default)]
    pub times_ms: Vec<f64>,
    /// The reduction of `times_ms`. `None` when nothing was measured.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub summary: Option<Summary>,
    /// Wall-clock seconds this candidate consumed on the GPU host.
    pub gpu_seconds: f64,
}

/// A `results.v1` document: every measurement, and the machine that took it.
///
/// Checkpointed after each candidate, so an interrupted sweep resumes from
/// what it already measured rather than starting over.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Results {
    /// Schema tag; always `results.v1`.
    pub schema: String,
    /// The kernel measured.
    pub kernel: String,
    /// The entry point launched.
    pub entry: String,
    /// The capability the plan was gated at.
    pub plan_cc: String,
    /// Product name of the device, as the driver reports it.
    pub device_name: String,
    /// The device's actual capability, which need not equal `plan_cc`.
    pub device_cc: String,
    /// Driver version. Part of what makes a timing reproducible, and part
    /// of why results do not port (`docs/LIMITATIONS.md`).
    pub driver_version: String,
    /// One entry per candidate visited, in visiting order.
    pub candidates: Vec<CandidateResult>,
    /// GPU seconds the whole sweep consumed.
    pub total_gpu_seconds: f64,
    /// Strategy that produced the visiting order (`exhaustive` | `random:<seed>`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub strategy: Option<String>,
    /// True when the sweep stopped because the wall budget ran out.
    #[serde(default)]
    pub budget_exhausted: bool,
}

/// How to visit a plan's candidates, and when to stop.
pub struct RunOptions {
    /// Visiting order over plan.candidates indices (a permutation or
    /// prefix); defaults to plan order.
    pub order: Vec<usize>,
    /// Wall-clock budget; the sweep stops (resumably) when it is spent.
    pub budget_secs: Option<f64>,
    /// Recorded in results for provenance.
    pub strategy: Option<String>,
}

impl RunOptions {
    /// Visit every candidate in plan order, with no budget.
    pub fn exhaustive(plan: &BenchPlan) -> Self {
        RunOptions {
            order: (0..plan.candidates.len()).collect(),
            budget_secs: None,
            strategy: Some("exhaustive".into()),
        }
    }
}

/// Execute `plan`, appending to `results_path` (resume: candidates already
/// present are skipped). Writes the results file after every candidate.
pub fn run_plan(
    plan: &BenchPlan,
    plan_dir: &Path,
    results_path: &Path,
    options: &RunOptions,
    progress: &mut dyn FnMut(&str),
) -> Result<Results, String> {
    let device = Device::open()?;
    progress(&format!(
        "device: {} (cc {}, driver {})",
        device.name, device.cc, device.driver_version
    ));
    if device.cc != plan.cc {
        progress(&format!(
            "WARNING: plan was gated at cc {}, device is cc {} — verdicts do not transfer \
             across parts (docs/SAFETY.md); results will be labelled with the device cc",
            plan.cc, device.cc
        ));
    }

    // A checkpoint that cannot be read is not "start over": resuming would
    // silently discard measurements that cost GPU time, and the file is the
    // only record of them. Say what is wrong and stop.
    let existing = Results::load(results_path)?;
    let mut results = match existing {
        Some(existing) => {
            progress(&format!(
                "resuming: {} candidates already measured",
                existing.candidates.len()
            ));
            existing
        }
        None => Results {
            schema: "results.v1".into(),
            kernel: plan.kernel.clone(),
            entry: plan.entry.clone(),
            plan_cc: plan.cc.clone(),
            device_name: device.name.clone(),
            device_cc: device.cc.clone(),
            driver_version: device.driver_version.clone(),
            candidates: Vec::new(),
            total_gpu_seconds: 0.0,
            strategy: options.strategy.clone(),
            budget_exhausted: false,
        },
    };
    results.budget_exhausted = false;

    let heartbeat_stop = start_heartbeat();
    let sweep_started = Instant::now();

    for &index in &options.order {
        let Some(candidate) = plan.candidates.get(index) else {
            return Err(format!("order index {index} out of range"));
        };
        if results.candidates.iter().any(|c| c.id == candidate.id) {
            continue;
        }
        if let Some(budget) = options.budget_secs
            && sweep_started.elapsed().as_secs_f64() >= budget
        {
            results.budget_exhausted = true;
            progress(&format!(
                "budget exhausted after {:.1}s: {} of {} candidates measured (resumable)",
                sweep_started.elapsed().as_secs_f64(),
                results.candidates.len(),
                plan.candidates.len()
            ));
            break;
        }
        // A gate-refused candidate may genuinely hang (that is why it was
        // refused). Checkpoint a `timeout` record BEFORE launching, so a
        // watchdog abort (or a wedged GPU context) leaves a resumable
        // truth on disk; overwrite it with the real outcome if we survive.
        let watchdog = if candidate.unsafe_candidate {
            results.candidates.push(CandidateResult {
                id: candidate.id.clone(),
                config: candidate.config.clone(),
                status: "timeout".into(),
                error: Some(format!(
                    "unsafe candidate did not complete within {UNSAFE_TIMEOUT_SECS}s;                      presumed hung (this is the failure mode the gate predicts)"
                )),
                warmup: candidate.warmup,
                repeats: candidate.repeats,
                times_ms: Vec::new(),
                summary: None,
                gpu_seconds: unsafe_timeout_secs() as f64,
            });
            results.checkpoint(results_path)?;
            progress(&format!(
                "{} UNSAFE candidate: watchdog armed at {}s",
                candidate.id,
                unsafe_timeout_secs()
            ));
            Some(arm_watchdog(unsafe_timeout_secs()))
        } else {
            None
        };
        let started = Instant::now();
        let outcome = run_candidate(&device, plan, plan_dir, candidate);
        if let Some(armed) = watchdog {
            armed.store(true, Ordering::Relaxed); // disarm
            // Replace the pre-checkpointed timeout record with the truth.
            results.candidates.retain(|c| c.id != candidate.id);
        }
        let gpu_seconds = started.elapsed().as_secs_f64();
        let result = match outcome {
            Ok(times_ms) => {
                let summary = summarize(&times_ms);
                progress(&format!(
                    "{} {}: median {} over {} repeats ({:.1}s)",
                    candidate.id,
                    candidate.config,
                    summary
                        .as_ref()
                        .map(|s| format!(
                            "{:.4} ms [{:.4}, {:.4}]",
                            s.median_ms, s.ci95_lo_ms, s.ci95_hi_ms
                        ))
                        .unwrap_or_else(|| "n/a".into()),
                    candidate.repeats,
                    gpu_seconds,
                ));
                CandidateResult {
                    id: candidate.id.clone(),
                    config: candidate.config.clone(),
                    status: "ok".into(),
                    error: None,
                    warmup: candidate.warmup,
                    repeats: candidate.repeats,
                    times_ms,
                    summary,
                    gpu_seconds,
                }
            }
            Err(e) => {
                progress(&format!("{} ERROR: {e}", candidate.id));
                CandidateResult {
                    id: candidate.id.clone(),
                    config: candidate.config.clone(),
                    status: "error".into(),
                    error: Some(e),
                    warmup: candidate.warmup,
                    repeats: candidate.repeats,
                    times_ms: Vec::new(),
                    summary: None,
                    gpu_seconds,
                }
            }
        };
        results.candidates.push(result);
        results.total_gpu_seconds = sweep_started.elapsed().as_secs_f64();
        results.checkpoint(results_path)?;
    }

    heartbeat_stop.store(true, Ordering::Relaxed);
    results.total_gpu_seconds = sweep_started.elapsed().as_secs_f64();
    results.checkpoint(results_path)?;
    Ok(results)
}

fn run_candidate(
    device: &Device,
    plan: &BenchPlan,
    plan_dir: &Path,
    candidate: &Candidate,
) -> Result<Vec<f64>, String> {
    let ptx_path = plan_dir.join(&candidate.ptx);
    let ptx = std::fs::read_to_string(&ptx_path)
        .map_err(|e| format!("reading {}: {e}", ptx_path.display()))?;
    let module = device.load_module(&ptx, &plan.entry)?;

    // Materialize buffers and the param pointer table, in ArgSpec order.
    // Each ArgSpec is exactly one .param slot.
    let mut buffers = Vec::new(); // (arg index, Buffer)
    for (i, arg) in candidate.args.iter().enumerate() {
        match arg {
            ArgSpec::InF32 { len } => {
                let host: Vec<f32> = deterministic_f32(*len);
                let buf = device.alloc(host.len() * 4)?;
                device.copy_in(&buf, cast_bytes(&host))?;
                buffers.push((i, buf));
            }
            ArgSpec::InU32 { len, modulo } => {
                let host: Vec<u32> = deterministic_u32(*len, *modulo);
                let buf = device.alloc(host.len() * 4)?;
                device.copy_in(&buf, cast_bytes(&host))?;
                buffers.push((i, buf));
            }
            ArgSpec::OutF32 { len } | ArgSpec::OutU32 { len } => {
                let zero = vec![0u8; (*len as usize) * 4];
                let buf = device.alloc(zero.len())?;
                device.copy_in(&buf, &zero)?;
                buffers.push((i, buf));
            }
            ArgSpec::LenOf { .. } | ArgSpec::U32 { .. } | ArgSpec::U64 { .. } => {}
        }
    }

    // Scalar storage must outlive the launch; the params table points into
    // these vectors and the buffers' device pointers.
    let mut ptr_slots: Vec<u64> = Vec::new();
    let mut u32_slots: Vec<u32> = Vec::new();
    let mut u64_slots: Vec<u64> = Vec::new();
    #[derive(Clone, Copy)]
    enum Slot {
        Ptr(usize),
        U32(usize),
        U64(usize),
    }
    let mut slots = Vec::with_capacity(candidate.args.len());
    for (i, arg) in candidate.args.iter().enumerate() {
        match arg {
            ArgSpec::InF32 { .. }
            | ArgSpec::InU32 { .. }
            | ArgSpec::OutF32 { .. }
            | ArgSpec::OutU32 { .. } => {
                // The loop above materializes a buffer for exactly these
                // four ArgSpec kinds, so the lookup succeeds — an invariant
                // held one loop away from the code that needs it. This
                // function returns a `Result`; a wrong answer here is a bug
                // report, not a crash mid-benchmark.
                let buf = &buffers
                    .iter()
                    .find(|(idx, _)| *idx == i)
                    .ok_or_else(|| format!("internal: argument {i} has no materialized buffer"))?
                    .1;
                ptr_slots.push(buf.ptr);
                slots.push(Slot::Ptr(ptr_slots.len() - 1));
            }
            ArgSpec::LenOf { of } => {
                let len = match candidate.args.get(*of) {
                    Some(ArgSpec::InF32 { len })
                    | Some(ArgSpec::OutF32 { len })
                    | Some(ArgSpec::OutU32 { len })
                    | Some(ArgSpec::InU32 { len, .. }) => *len,
                    other => return Err(format!("len_of {of} points at {other:?}")),
                };
                u64_slots.push(len);
                slots.push(Slot::U64(u64_slots.len() - 1));
            }
            ArgSpec::U32 { value } => {
                u32_slots.push(*value as u32);
                slots.push(Slot::U32(u32_slots.len() - 1));
            }
            ArgSpec::U64 { value } => {
                u64_slots.push(*value);
                slots.push(Slot::U64(u64_slots.len() - 1));
            }
        }
    }
    let mut params: Vec<*mut std::ffi::c_void> = slots
        .iter()
        .map(|slot| match slot {
            Slot::Ptr(k) => std::ptr::from_mut(&mut ptr_slots[*k]).cast(),
            Slot::U32(k) => std::ptr::from_mut(&mut u32_slots[*k]).cast(),
            Slot::U64(k) => std::ptr::from_mut(&mut u64_slots[*k]).cast(),
        })
        .collect();

    for _ in 0..candidate.warmup {
        device.timed_launch(&module, candidate.grid, candidate.block, &mut params)?;
    }
    device.synchronize()?;

    let mut times = Vec::with_capacity(candidate.repeats as usize);
    for _ in 0..candidate.repeats {
        times.push(device.timed_launch(&module, candidate.grid, candidate.block, &mut params)?);
    }
    device.synchronize()?;
    Ok(times)
}

fn cast_bytes<T>(data: &[T]) -> &[u8] {
    unsafe { std::slice::from_raw_parts(data.as_ptr().cast(), std::mem::size_of_val(data)) }
}

/// Deterministic xorshift-seeded data: reproducible across runs and hosts.
fn deterministic_f32(len: u64) -> Vec<f32> {
    let mut state = 0x9e3779b97f4a7c15u64;
    (0..len)
        .map(|_| {
            state ^= state << 13;
            state ^= state >> 7;
            state ^= state << 17;
            ((state >> 40) as f32) / ((1u64 << 24) as f32)
        })
        .collect()
}

fn deterministic_u32(len: u64, modulo: u64) -> Vec<u32> {
    let modulo = modulo.max(1);
    let mut state = 0x2545f4914f6cdd1du64;
    (0..len)
        .map(|_| {
            state ^= state << 13;
            state ^= state >> 7;
            state ^= state << 17;
            (state % modulo) as u32
        })
        .collect()
}

impl Results {
    /// Read `results.v1` from a run directory.
    ///
    /// `Ok(None)` means one thing only: **the file is not there**, so the
    /// measurement box has not run yet. Everything else is an error.
    ///
    /// This used to be `read_to_string(path).ok()?` then `from_str().ok()`,
    /// so a truncated file, an empty one, `null`, `[]`, a `results.v2` from
    /// a newer runner and a *directory* named `results.json` all collapsed
    /// into that same `None` — and `report` rendered "nothing measured yet",
    /// exit 0, nothing on stderr, with a JSON report that validated. The run
    /// directory is the hand-off between two machines, and the two
    /// conditions call for opposite actions: wait, or go and look. Nothing
    /// told them apart.
    ///
    /// `verdicts.v1` two lines away in the report builder already checked
    /// its schema tag by name; this is the same check, so a future
    /// `results.v2` is refused by name rather than read as nothing.
    ///
    /// # Errors
    ///
    /// Any I/O error that is not `NotFound`, any parse failure, and any
    /// document whose `schema` is not `results.v1` — each naming the path.
    pub fn load(path: &Path) -> Result<Option<Self>, String> {
        let text = match std::fs::read_to_string(path) {
            Ok(text) => text,
            Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok(None),
            Err(e) => return Err(format!("{}: {e}", path.display())),
        };
        // The tag first, so a wrong-schema document is named as one rather
        // than as whatever field happens to be missing from it.
        let value: serde_json::Value = serde_json::from_str(&text)
            .map_err(|e| format!("{}: not JSON: {e}", path.display()))?;
        let declared = value.get("schema").and_then(|s| s.as_str());
        match declared {
            Some("results.v1") => {}
            Some(other) => {
                return Err(format!(
                    "{}: unsupported results schema `{other}` (expected `results.v1`)",
                    path.display()
                ));
            }
            None => {
                return Err(format!(
                    "{}: not a results.v1 document (no `schema` field)",
                    path.display()
                ));
            }
        }
        serde_json::from_str(&text)
            .map(Some)
            .map_err(|e| format!("{}: not a results.v1 document: {e}", path.display()))
    }

    /// Atomic checkpoint: write to a temp file, then rename.
    pub fn checkpoint(&self, path: &Path) -> Result<(), String> {
        let tmp = path.with_extension("json.tmp");
        let json =
            serde_json::to_string_pretty(self).map_err(|e| format!("serializing results: {e}"))?;
        std::fs::write(&tmp, json).map_err(|e| e.to_string())?;
        std::fs::rename(&tmp, path).map_err(|e| e.to_string())
    }
}

const UNSAFE_TIMEOUT_SECS: u64 = 10;

fn unsafe_timeout_secs() -> u64 {
    std::env::var("LAUNCHBOUND_UNSAFE_TIMEOUT_SECS")
        .ok()
        .and_then(|v| v.parse().ok())
        .unwrap_or(UNSAFE_TIMEOUT_SECS)
}

/// Watchdog for unsafe candidates: if not disarmed within the deadline the
/// process exits (a hung kernel cannot be cancelled from user code). The
/// pre-checkpointed `timeout` record makes the rerun skip it.
fn arm_watchdog(deadline_secs: u64) -> std::sync::Arc<AtomicBool> {
    let disarmed = std::sync::Arc::new(AtomicBool::new(false));
    let flag = disarmed.clone();
    std::thread::spawn(move || {
        let start = Instant::now();
        while start.elapsed().as_secs() < deadline_secs {
            if flag.load(Ordering::Relaxed) {
                return;
            }
            std::thread::sleep(std::time::Duration::from_millis(100));
        }
        if !flag.load(Ordering::Relaxed) {
            eprintln!(
                "watchdog: unsafe candidate exceeded {deadline_secs}s; exiting so the                  checkpointed timeout record stands (exit 3, rerun to continue)"
            );
            std::process::exit(3);
        }
    });
    disarmed
}

/// A CPU heartbeat: the box's idle alarm terminates on CPU <5% for 30 min,
/// and a GPU-bound loop can look idle. Burn a configurable duty cycle on
/// one core (LAUNCHBOUND_HEARTBEAT_PCT, default 40) until stopped.
fn start_heartbeat() -> &'static AtomicBool {
    static STOP: AtomicBool = AtomicBool::new(false);
    STOP.store(false, Ordering::Relaxed);
    let duty: u64 = std::env::var("LAUNCHBOUND_HEARTBEAT_PCT")
        .ok()
        .and_then(|v| v.parse().ok())
        .unwrap_or(40)
        .clamp(1, 100);
    std::thread::spawn(move || {
        let mut sink = 0u64;
        while !STOP.load(Ordering::Relaxed) {
            let spin = Instant::now();
            while spin.elapsed().as_millis() < duty as u128 {
                sink = sink.wrapping_mul(6364136223846793005).wrapping_add(1);
            }
            std::hint::black_box(sink);
            std::thread::sleep(std::time::Duration::from_millis(100 - duty.min(99)));
        }
    });
    &STOP
}