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
//! Checksum recording and divergence detection for BrickProfiler.
//!
//! CORRECTNESS-011: Per-kernel checksum capture for CPU/GPU divergence detection.
//! Extracted from mod.rs to keep file sizes manageable.
use super::BrickProfiler;
use crate::brick::profiler::checksum::{fnv1a_f32_checksum, DivergenceInfo, KernelChecksum};
impl BrickProfiler {
// =======================================================================
// CORRECTNESS-011: Per-kernel checksum capture for divergence detection
// =======================================================================
/// Record a kernel trace with output checksum for divergence detection.
///
/// This enables automated CPU/GPU divergence detection by capturing
/// output checksums alongside timing data. When GPU produces wrong output,
/// this identifies WHICH kernel diverged without hours of manual debugging.
///
/// Five-Whys Root Cause: Hours of manual "let me check X in Y" debugging
/// -> No automated tool identified which kernel diverged
/// -> BrickProfiler only captured timing, not checksums
/// -> Missing feature: per-kernel checksum capture
///
/// # Arguments
/// - `name`: Brick/kernel name
/// - `layer_idx`: Layer index (0-N for transformer layers)
/// - `position`: Position in sequence
/// - `output`: Output tensor data (first 64 floats checksummed)
///
/// # Example
/// ```rust,ignore
/// // After RoPE kernel
/// profiler.record_checksum("RopeNeox", layer_idx, position, &q_rotated);
/// ```
pub fn record_checksum(&mut self, name: &str, layer_idx: usize, position: u32, output: &[f32]) {
if !self.enabled {
return;
}
let checksum = fnv1a_f32_checksum(output);
let trace = KernelChecksum { name: name.to_string(), layer_idx, position, checksum };
self.kernel_checksums.push(trace);
}
/// Get all kernel checksums for divergence comparison.
#[must_use]
pub fn get_checksums(&self) -> &[KernelChecksum] {
&self.kernel_checksums
}
/// Compare checksums with a reference profiler (e.g., CPU baseline).
///
/// Returns None if all checksums match, or the first divergent kernel.
#[must_use]
pub fn find_divergence(&self, reference: &BrickProfiler) -> Option<DivergenceInfo> {
use std::collections::HashMap;
// Index reference checksums by (name, layer_idx, position)
let ref_index: HashMap<(&str, usize, u32), u64> = reference
.kernel_checksums
.iter()
.map(|t| ((t.name.as_str(), t.layer_idx, t.position), t.checksum))
.collect();
// Check each of our checksums against reference
for trace in &self.kernel_checksums {
let key = (trace.name.as_str(), trace.layer_idx, trace.position);
if let Some(&expected) = ref_index.get(&key) {
if trace.checksum != expected {
return Some(DivergenceInfo {
kernel_name: trace.name.clone(),
layer_idx: trace.layer_idx,
position: trace.position,
expected_checksum: expected,
actual_checksum: trace.checksum,
});
}
}
}
None
}
/// Reset checksum tracking (call before new forward pass).
pub fn reset_checksums(&mut self) {
self.kernel_checksums.clear();
}
}