Glossary

The vocabulary of GPU-kernel correctness, defined plainly. If you have read the problem or the P1–P4 evidence and hit an unfamiliar term, it is here.

Adversarial inputs

Values and shapes chosen specifically to expose numerical edge cases (large magnitudes, near-cancellations, non-block-aligned dimensions). In study P3 they reached 99% bug recall, +28 pp over the default oracle.

Bug recall

The fraction of known-buggy kernels an oracle flags as wrong. Per-op calibrated tolerances raise recall from 65% to 82% (P2); adversarial input generation reaches 99% (P3).

Calibrated tolerance

A pass/fail threshold fit per operator and dtype (gpuemu uses a p95-of-controls × 1.5 envelope) rather than one global atol/rtol, so tight ops are not under-tested and loose ops are not falsely failed.

Correctness oracle

The procedure that decides whether a kernel’s output is “correct”. The field-standard oracle is a single torch.allclose on one shape, dtype, and seed; gpuemu’s oracle is a high-precision fp64 reference checked across op-schema-aware adversarial inputs.

False positive

A correct kernel wrongly flagged as buggy. gpuemu recorded 0 false positives on 15/15 control kernels in study P1 — a low false-alarm rate is what makes a correctness gate usable in CI.

fp64 reference oracle

A double-precision CPU implementation of an operator used as ground truth. The kernel under test is compared against it per dtype, which is far more sensitive to small numerical errors than comparing two low-precision results.

Op-schema-aware fuzzing

Input generation that understands each operator’s shape and value constraints, so it can produce boundary, regular, and adversarial cases — including the partial-tile shapes where masked kernels break.

PTX / SASS

PTX is NVIDIA’s virtual ISA; SASS is the machine assembly a GPU actually executes. gpuemu’s static lint reads register pressure, spills, and instruction counts from these artifacts to flag performance regressions — but semantic bugs can compile to identical PTX (P4).

SARIF

Static Analysis Results Interchange Format — the JSON schema GitHub code scanning consumes. gpuemu ci emits SARIF so failing kernels surface as annotations in pull requests.

Silent kernel bug

A GPU kernel that compiles, runs, and passes the standard correctness check yet returns numerically wrong output. Because it never crashes, it ships green and degrades model quality downstream. gpuemu exists to catch this class.

Tail-mask leak

A bug where a kernel forgets to mask the last partial tile (e.g. softmax when the sequence length is not a multiple of the block size). It only fires on non-block-aligned shapes, so one-shape tests miss it entirely.

torch.allclose

PyTorch’s elementwise closeness check (and torch.testing.assert_close). Ubiquitous as a kernel correctness test, but run on one shape it accepted 9/9 LLM-style buggy kernels in a measured 26-op corpus.

Ready to see it in practice? How gpuemu works walks the full validation pipeline, and Get started validates a kernel in five minutes.