eenn 0.1.0

A hybrid neural-symbolic constraint solver with cognitive reasoning capabilities
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

use std::sync::Arc;
use std::time::Instant;

use eenn::{
    Calibration, ExecutionContext, GreedyPartitioner, InMemoryProfiler, KernelCache, Partitioner,
    Profiler,
};

// Test-only Planner that uses GreedyPartitioner to create ExecutionPlans and
// a stubbed precompile worker that writes quick artifacts into the KernelCache.
#[test]
fn planner_and_precompile_worker_integration() {
    // Create a set of mock ops (reuse types from the library tests via local definitions)

    // Create a kernel cache in tempdir and persist calibration into the cache root
    let dir = tempfile::tempdir().expect("tmpdir");
    let calib = Calibration {
        bandwidth_bytes_per_sec: 500.0 * 1024.0 * 1024.0,
    }; // 500 MB/s
    calib
        .persist_to_cache_root(dir.path())
        .expect("persist calibration");

    // Create cache and then load calibration from the cache root
    let mut cache = KernelCache::new(dir.path()).expect("cache create");
    let loaded = Calibration::load_from_cache_root(dir.path())
        .expect("load calibration")
        .expect("calib present");
    let bytes_per_ms = loaded.bytes_per_ms();
    let lookahead_values = [0usize, 1usize, 2usize, 4usize];
    for &lookahead in &lookahead_values {
        let partitioner = GreedyPartitioner {
            lookahead,
            bytes_per_ms,
            max_ops_in_fusion: 128,
        };
        struct OpMock(&'static str, bool, f64, usize);
        impl eenn::partitioner::Op for OpMock {
            fn name(&self) -> &str {
                self.0
            }
            fn gpu_capable(&self) -> bool {
                self.1
            }
            fn estimate(&self) -> (f64, usize) {
                (self.2, self.3)
            }
        }

        let ops: Vec<Arc<dyn eenn::partitioner::Op>> = vec![
            Arc::new(OpMock("a", true, 2.0, 0)),
            Arc::new(OpMock("b", true, 2.0, 0)),
            Arc::new(OpMock("c", true, 0.5, 1000)),
            Arc::new(OpMock("d", false, 0.0, 0)),
        ];

        // Partition
        let segments = partitioner.partition(&ops);
        assert!(
            !segments.is_empty(),
            "Expected at least one GPU segment for lookahead={}",
            lookahead
        );

        // Set up profiler + execution context
        let profiler = std::sync::Arc::new(InMemoryProfiler::new());
        let ctx = ExecutionContext::with_profiler(profiler.clone());

        // In-test simple cache stats
        let mut misses = 0usize;

        // For each segment, compute a fake fingerprint and check cache; if missing, simulate compile and write artifact
        for seg in &segments {
            let fp = format!("seg-{}-{}", seg.start, seg.end);
            if cache.lookup(&fp).is_none() {
                misses += 1;
                if let Some(p) = &ctx.profiler {
                    p.record_event("cache_miss", 1);
                }
                // simulate compile time (quick tier)
                std::thread::sleep(std::time::Duration::from_millis(1));
                let artifact = format!("compiled:generic:{}", fp).into_bytes();
                cache.write_artifact(&fp, &artifact).expect("write");
            } else if let Some(p) = &ctx.profiler {
                p.record_event("cache_hit", 1);
            }
            let bytes = cache.read_artifact(&fp).expect("read");
            assert!(bytes.starts_with(b"compiled:"));

            // Simulate metrics and hotness scoring
            #[derive(Clone)]
            struct ArtifactMetrics {
                use_count: usize,
                last_used_at: std::time::Instant,
                avg_runtime_ms: f64,
                compile_time_ms: f64,
            }

            // fake metrics
            let mut metrics = ArtifactMetrics {
                use_count: 10,
                last_used_at: Instant::now(),
                avg_runtime_ms: 2.0,
                compile_time_ms: 5.0,
            };

            fn compute_hotness(metrics: &ArtifactMetrics, now: Instant) -> f64 {
                let age_s = (now.duration_since(metrics.last_used_at).as_secs_f64()).max(1.0);
                let time_decay =
                    (-0.1 * (now.duration_since(metrics.last_used_at).as_secs_f64())).exp();
                let usage_velocity = metrics.use_count as f64 / age_s;
                let avg_saved_ms = (metrics.avg_runtime_ms - 0.5).max(0.0); // assume jit overhead 0.5ms
                let compile_roi = if metrics.compile_time_ms > 0.0 {
                    (metrics.use_count as f64 * avg_saved_ms) / metrics.compile_time_ms
                } else {
                    0.0
                };
                usage_velocity * compile_roi * time_decay
            }

            // Evaluate hotness and optionally promote to specialized
            let now = Instant::now();
            let hotness = compute_hotness(&metrics, now);
            // promotion threshold is arbitrary for the test
            if hotness >= 0.1 {
                // simulate heavy specialized compile and atomic replacement
                if let Some(p) = &ctx.profiler {
                    p.record_event("promotion_start", 1);
                }
                std::thread::sleep(std::time::Duration::from_millis(2));
                let specialized = format!("compiled:specialized:{}", fp).into_bytes();
                cache
                    .write_artifact(&fp, &specialized)
                    .expect("write specialized");
                if let Some(p) = &ctx.profiler {
                    p.record_event("promotion_done", 1);
                }
            } else if let Some(p) = &ctx.profiler {
                p.record_event("promotion_skipped", 1);
            }

            // Simulate demotion if artifact cools off: reduce use_count and recompute
            metrics.use_count = 0; // cooled off
            let hotness2 = compute_hotness(&metrics, Instant::now());
            if hotness2 < 0.01 {
                // demote: rewrite generic artifact
                let generic = format!("compiled:generic:{}", fp).into_bytes();
                cache.write_artifact(&fp, &generic).expect("write demoted");
                if let Some(p) = &ctx.profiler {
                    p.record_event("demoted", 1);
                }
            }
        }

        // Validate profiler recorded events for this iteration
        let snap = profiler.snapshot();
        assert_eq!(
            snap.events.get("cache_miss").copied().unwrap_or(0) as usize,
            misses
        );
        // Either we started promotions for some segments or we explicitly skipped them
        let promotion_start = snap.events.get("promotion_start").copied().unwrap_or(0);
        let promotion_skipped = snap.events.get("promotion_skipped").copied().unwrap_or(0);
        assert!(
            promotion_start > 0 || promotion_skipped > 0,
            "expected promotion_start or promotion_skipped to be recorded"
        );
        // Ensure demotion path exercised at least once (we force cooled-off metrics in test)
        let demoted = snap.events.get("demoted").copied().unwrap_or(0);
        assert!(demoted > 0, "expected at least one demotion event");
    }
}