ADR-1505: An Intel GPU tester image with a backend-neutral GPU section in the report¶

• Status: Accepted
• Date: 2026-10-03
• Deciders: Lusoris
• Tags: ci, docker, sycl, testing, parity, fork-local

Context¶

The SYCL twins are proven on the project's Arc A380, and since 2026-10-03 on an Arc B580 and an Arc Pro B60 (ADR-1501). The Xe-LP half of T-SYCL-ROW-KERNELS-SG16-OTHER-DEVICES-2026-10-02 needs an integrated GPU the project does not own: an outside tester's Intel Core with its UHD 770, whose operating system may be Linux or Windows with WSL2. The tester must run one command and build nothing, and the result must say, row by row, what it measured. Two more kits (CUDA and HIP, a Windows bundle) follow, so the report needs one GPU section that any backend fills, not a third Metal-style section.

The tester image of ADR-1492 is CPU-only. The development container carries every SDK and is about 70 GB; a hosted runner cannot pull it. The oneAPI release image (ADR-1368) shows how to build a SYCL binary on Debian 13 at the pinned versions, but its runtime stage installs Intel's runtime meta-package, 1.1 GB that includes an OpenCL CPU device, an OpenMP offload library and an LTO plugin, none of which a GPU run loads.

Decision¶

We will publish ghcr.io/vmafx/vmafx:<describe>-tester-sycl (linux/amd64) from a final-sycl target of docker/Dockerfile.tester, built and published by docker-publish-tester.yml in new jobs that follow the CPU image's pattern (dispatch on master with ref, the tester-publish environment, keyless signature, provenance, an attested SPDX SBOM and a -source image) under the licensing rules of ADR-1503: the image is the artifact sycl-image of licensing.json, and its build cannot finish without the licence check. The build stage is the oneAPI release builder (Debian 13, install-intel-oneapi.sh --mode=builder, install-intel-ocloc.sh --components build): the SYCL build with the default AOT list plus SPIR-V, the gpu and sycl suites' executables and shell tests, and the parity gate. The runtime stage holds the Intel SYCL runtime files the compiler's credist.txt lists and the build needs (libsycl, the Unified Runtime loader and Level Zero adapters, the compiler runtime libraries), copied unmodified with their licence texts (tools/rc1-tester/image/sycl-runtime.json), UMF and hwloc, and from the pinned compute-runtime release only the Level Zero GPU driver, IGC core, gmmlib and the Level Zero loader: the offline compiler, the OpenCL ICD and IGC's OpenCL front end are purged, because SPIR-V reaches the GPU through Level Zero and IGC alone (measured on the A380 with a SPIR-V-only probe). The report gains schema version 3 and a gpu section (hw_gpu.py) that a backend fills through one small interface (GpuBackend: discover devices, pin a run to one, parse audit output, name a row map): per device the twins against the CPU of the same image per fixture and per parity-gate feature at the gate's bound for that twin, the gate's cells of the backend held exact, the device tests, the audits and the state rows of that device's family. The SYCL backend finds every Level Zero GPU (at most four), pins each run with ONEAPI_DEVICE_SELECTOR=level_zero:<n>, records whether it reached the GPU through a Linux render node or WSL2's /dev/dxg, and reads the GPU IP version to place the device in a family (ocloc ids): the row map (sycl-rows.json) applies the SG16 row's Xe-LP part only to an Xe-LP or Xe-LPG device. No usable device is no_device with the missing docker run option as the reason, not a failure, as for Metal.

Alternatives considered¶

Consequences¶

• Positive: one command closes the Xe-LP part of the SG16 row on a tester's UHD 770 and measures every SYCL twin on any Intel GPU; the CUDA and HIP kits add a backend module and a row map, not a report format. The A380 run of the documented Linux command passes in about 90 s: 69 device tests, the scratch audit (125 kernels, none in scratch memory), every twin identical to the CPU on four fixtures, the gate's cells at 0 (ciede at its 1e-9 bound, 0 measured).
• Negative: the image is about 1.6 GB unpacked and 0.95 GB compressed, mostly the device test executables, each carrying the AOT images of libvmaf (about 10 MB each). The WSL2 path rests on a probe of the same compute runtime under Docker Desktop on Windows 11 (Research-2128), not on a run of this image there; the report records the path it took so the first WSL2 report shows it.
• Neutral / follow-ups: the licence record gains a dpkg-foreign component kind (vendor packages installed by dpkg, which carry no Debian copyright file and no Debian source) and pinned fetched_texts; a backend plugs in as described in the maintainer notes; the Metal sections keep their schema-2 form and may move into gpu later.

References¶

• req (maintainer brief, 2026-10-03, paraphrased): an Intel GPU tester kit an outsider runs with one command and no build, first for a UHD 770 on Linux or WSL2; it closes the Xe-LP half of the SG16 row, measures every SYCL twin on any Intel GPU, and its report section is generic so the CUDA and HIP kits reuse it.
• ADR-1492, ADR-1493, ADR-1496, ADR-1368, ADR-1468, ADR-1395, ADR-1501, ADR-1503, Research-2128 (the compute runtime under WSL2).
• Intel compute-runtime release 26.35.39758.10 notes (supported platforms, WSL testing) and documentation/WSL.md of intel/compute-runtime.