1024px image generation end-to-end on a consumer RDNA2 GPU, and a NULL-deref segfault root-caused

Contribution

Claimed. A 1024px image-generation end-to-end pipeline runs to completion on a consumer-class RDNA2 GPU under a ROCm/HIP toolchain with the vendor BLAS library fully stubbed out inline (de-BLAS'd), producing a final latent whose cosine similarity to a reference diffusion implementation's golden output is 0.99999999 (acceptance gate >= 0.999, PASS). The run completes cleanly and emits its end-of-pipeline marker reproducibly across two invocations. Separately and primarily, this note root-causes a segmentation fault that occurred after the denoise phase: a missing output-directory argument caused the latent-write file open to return a NULL handle, and an unguarded write call then dereferenced NULL before the decode stage was ever reached. The fix creates the output directory and adds NULL guards at all write sites, with no kernel or math change.

NOT claimed. This is an engineering / correctness-validation result, not a novel algorithm. It does NOT claim throughput competitiveness, generation quality beyond single-latent parity, generalization to other resolutions or image models, or that the cosine parity reflects perceptual or sample quality. The performance numbers are single-box, single-run wall-clock timings, not a benchmarked distribution. The segfault fix restores reachability of the output path; it does not alter or re-validate the numerical pipeline beyond confirming parity is untouched.

Methods

Results

Threats to validity / Limitations

• Sample size. Parity is a single final-latent cosine on one configuration, a correctness smoke test, not a benchmark-scale evaluation. No multi-prompt, multi-seed, or distributional coverage is recorded.
• No variance. Phase and wall-clock figures are single-box, single-run timings (the two reruns confirmed clean reachability, not timing variance). No mean or standard deviation across runs.
• Parity is not quality. A cosine of 0.99999999 to a reference latent demonstrates numerical agreement of the latent, not perceptual or sample-generation quality, and not human-judged fidelity.
• Precision ceiling unrecorded. The numerical precision of the engine path is not recorded in the source notes, so the cosine result cannot be bounded against a known ULP floor.
• Scope bound. The result is specific to 1024px on this consumer RDNA2 GPU with a single image model; other resolutions, models, or GPUs are out of scope.
• Baseline version unpinned. The reference diffusion golden was used as oracle, but its version is not recorded in the source notes; the "validate the oracle" provenance is therefore incomplete.
• The fix is reachability, not math. The segfault fix restores the output path; it adds no new numerical validation beyond asserting parity was untouched.
• Reproduction strength. The reproducibility evidence is limited to two clean reruns of the same configuration, not an independent re-derivation.

Reproducibility / Provenance

• Failure mode (pre-fix): A missing output-directory argument caused the latent-write file open to return a NULL handle; the subsequent unguarded write dereferenced NULL, producing a deterministic segmentation fault on the box immediately after the denoise phase (~188s).
• Fix: Ensure the output directory exists (create-if-missing) in the entry path, and guard the file handle at all write sites. No kernel or math change; parity untouched.
• Build note: Compiling and linking against the appropriate GPU architecture requires explicitly linking the core HIP runtime; the default link invocation otherwise fails with an undefined runtime-registration symbol.
• Dependency footprint: Dynamic-dependency inspection shows only the core HIP runtime and system libraries (no vendor BLAS / MKL / inference-primitives library).
• Source provenance: Specific commit, branch, and file paths are omitted from this disclosure layer.