Native DirectX 12 Experimental Benchmark For Linux — v0.1.5

Native DirectX 12 Experimental Benchmark For Linux — v0.1.5

v0.1.5 packages the r16 maintenance changes for SM86 and SM89 with NVIDIA Open drivers 615.78.08 and 620.70.06. It retains the RTX 30/40 device tables from v0.1.4. SM120 and RTX 50 support are not included.

This is a separate release identity. Existing tags and assets, including v0.1.4, are preserved. Release availability is determined by the v0.1.5 release page.

What changed

• Source-build PATH handling: explicit tool directories containing the PATH separator are refused before any tool probe. Executables selected for version checks are invoked by absolute path. The repair is present in the public builder and both companion snapshots. It changes the source companion only; the runtime archive and starter are unchanged.

• Kernel-event handling: a workqueue CPU-time warning of the form workqueue: <function> hogged CPU for >... no longer stops a run solely because the function contains drm. NVIDIA-driver events, integrity/lockup events, critical kernel priorities and other warning forms keep their stop rules. The observed warning remains in the evidence.

• Native-library preload check: a matching path or inode is refused, including an alias whose btrfs device field differs from stat(). An inode collision on another filesystem is also refused conservatively. The post-load check retains the v0.1.4 btrfs fix.

• Interpreter admission: the native binding keeps its strict GIL-enabled interpreter checks and CPU controls.

• Separate Vulkan reference: a health rejection saves the child outcome and observed kernel records before propagating the stop. Known pre-submit failures close the observation descriptor. A service that ended without its completion record is distinguished from an active-service timeout; its outcome remains UNCERTAIN, with no automatic stop or retry.

• Installer rollback: if replacing a package and restoring the previous one both fail, the installer identifies the preserved folder and reports an incomplete installation. The CodeRabbit repair and three regression tests from PR #7 are retained; this release changes the version and artifact pins.

• Package manual: the runnable archive contains the operational README and readable export history. Export checks compare the packaged desktop and Vulkan tools with both source snapshots.

Offline tool-version probes and build commands now share one execution function and an absolute unshare selected from the system-only search path. Probe or isolation failure stops the build without a host fallback. A missing trusted launcher also writes a FAIL:UNSHARE_MISSING build receipt.

The four runtime inventories use existing compiled inputs. Native ELF files, shaders and the external starter are unchanged; this is not a new compiler or shader build.

Install and share

Follow How to run v0.1.5. Use the v0.1.5 installer with the matching runtime archive and starter. The installer writes a receipt which the launcher checks before naming the release in charts and share-results-v0.1.5.tar.gz.

An older package that does not match this manifest is moved aside before replacement. If rollback fails, preserve the folder named by the installer and resolve the reported file-operation error before retrying. No automatic GPU recovery or retry is added.

Artifacts

File	SHA-256	Bytes
ndx12-native-benchmark-v21-615.78.08-620.70.06.tar.gz	eed38c254c8e8cc8310c8fc84fbca6081696f114aadcb63cc584cb2828b448a5	2506060
ndx12_start.py	8cd0ca349f4535741ae8a673b3aa10de4a0b5f953c283543a453a850337aced5	10958
ndx12-benchmark-v0.1.5-source.tar.gz	fc77498deb4f198a201746dda4e3c5b28bf3341eddb263f1e375f27d97a52d13	4207915
ndx12-install.py (repository file)	e5ab43e95dde9d347ed66e4daed0811005bb00a37cdc0cc44b56372790fa3f4d	21597

The runtime's MANIFEST.sha256 has SHA-256 c6bd8d3575a63304783cb412bafa38d31f2de4eeedfc8d9f9e276fbd0e48c56d and covers 933 files. The source companion manifest covers 673 files. Hashes identify bytes; they do not prove GPU execution.

Validation scope

Saved final-r16 CPU evidence covers four focused runs of 89 test methods (SM86/SM89, normal and optimized Python), eight methods against the actual packaged Vulkan tool, 36 installer tests in each Python mode, and 34 named installation cases across 35 user-namespace runs. The Vulkan process/service controls use CPU models; their filesystem and pipe/parser operations are real.

The eight earlier full suites of 591 tests each belong to predecessor source identities. They were not rerun for the final tool changes. The current release preparation ran 39 installer tests in each Python mode, normal and -O. The set contains 31 existing unit tests, five filesystem/starter rollback controls and the three public rollback regressions from PR #7. These tests loaded the new release installer and checked its pins against the retained package; predecessor installer hashes were not treated as validation of the new file. The separator suite has two regression methods (three failure cases across the separator inputs); they fail on the original public builder and both original companion builders. Its three controls pass on those predecessors: a legal tool directory with spaces, exclusion of the current directory from search, and refusal of a relative tool directory. The offline suite adds 11 methods for probe/build routing, trusted launcher identity, refusal handling and failure receipts. All 16 integrated methods pass normally and with -O on the public builder and both updated source snapshots. The final companion test/module paths are exercised after extraction. Namespace launchers in these CPU tests are routing doubles; successful real namespace creation and a full dependency build were not run.

CodeRabbit CLI completed reviews of the selected final r16 files, with exact-byte coverage and retained triage. The final five-file review reported two findings, dismissed using controls for the pinned build recipe and starter; this was not a zero-finding whole-package review. GitHub review of the release PR covers the public PR files, not the private runtime/source archives.

GPU, real-host installation, live btrfs reproduction and SM120: NOT_RUN for r16. Results from other revisions or experimental candidates do not qualify this archive. No new performance, GPU-correctness or device-compatibility result is claimed. An end-to-end upgrade from a real installed v0.1.4 host was not run.

A pre-existing reporting limitation remains in direct command-line Vulkan auto mode: some non-executed probe states may be described incorrectly in the summary. The interactive launcher uses explicit on/off mode.

Sources and licenses

The source companion provides the SM86/SM89 snapshots, tests, shader sources and pinned build recipes for this runtime package, plus the later source-only dependency-builder repair. Only the builder, its separator and offline tests, affected source locks, companion README and companion manifest change in this source-only successor. It omits the two prebuilt development Vulkan-reference executables; their C++ source and build command remain. The companion does not contain the separately distributed top-level release installer or its GitHub regression-test file; those have public repository history.

Project-owned material remains GPL-3.0-or-later. Retained third-party terms and notices are unchanged; there is no clean-room or blanket original-authorship claim. LICENSE and NOTICE are unchanged in both archives. See Source and licenses and the companion README for provenance, build prerequisites and historical rebuild limits.

One inherited source-build limitation remains open outside this repair: a failed dependency fetch can leave an older receipt marked PASS. Treat the current command failure as a failure and explicitly verify the cache before building; an older receipt alone is not evidence of successful acquisition. This does not affect the installed benchmark launcher.