The evidence (P1–P4)
In short: every gpuemu default comes from a measured study. The headline results: 9/9 LLM-style bugs caught across 5 GPU classes with 0 false positives on 15 controls (P1); recall lifted from 65% to 82% by calibrated tolerances (P2); 99% recall from adversarial input generation (P3).
The correctness illusion in LLM-generated GPU kernels
A hardware-free fuzz oracle catches 9/9 LLM-style bugs across 5 GPU classes (RTX 3060, A10, L40S, A100 SXM4, H100 NVL) with 0 false positives on 15/15 controls.
Operator-aware mixed-precision tolerance calibration
A p95-of-controls × 1.5 envelope raises kernel-bug recall from 65% to 82% over the field-standard fixed atol/rtol — at zero precision cost.
Test-input generation for tensor programs
A seven-strategy ablation: adversarial value sampling wins at 99% recall; "regular shape only" misses 100% of tail-mask bugs.
Static PTX metrics track structural regressions but miss semantic ones
Structural Δregs / Δinstrs predicts Δperf% consistently across 5 GPU classes; semantic bugs compile to identical PTX and need the correctness oracle.
P3: the recall swing from input strategy
Holding the kernel and oracle fixed, only the input-generation strategy changed. Adversarial sampling beats the field-standard default by +28 percentage points, and beats “regular shapes only” by 35.
| Input strategy | Bug recall |
|---|---|
| Adversarial values (gpuemu default) | 99% |
| Field-standard default oracle | 71% |
| Regular shapes only | 64% |
Replayable, not just reported
Each study ships as a LaTeX manuscript plus run-id records and a kernel corpus you can replay. Because gpuemu’s RNG is bit-identical across Rust and Python, every result reproduces byte-for-byte from its seed. The full manuscripts and run records live in the gpuemu research program.