How Hard Is It to Test the GPU Driver?
summary
Builds and hardware run through GitLab CI to hundreds of embedded boards using LAVA and Tron-CI; the Vulkan conformance suite alone is about VK-CTS ~2.8M tests so pipelines use fractions, parallel runs across ~10 devices and nightly full runs to catch what fractions miss. A merge board enforces serial merge+test, retries hide transient GPU/device/net flakiness, and teams provision many duplicate boards because continuous stress drops a 15‑board farm to ~8–12 usable units over years.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:53GPU driver testing stack
- 3:21Mesa’s GitLab CI pipeline
- 5:49Testing with LAVA and TronCI
- 8:38Merge challenges
- 10:24Handling CI flakiness
- 12:07Nightly testing
- 14:40Device connection options
- 16:08Device quantity importance
- 17:34Open vs proprietary CI
- 18:57Simplifying CI maintenance
- 19:43Lowering contributor barriers
- 21:52Q&A
- 24:26Q&A OpenCL and CUDA testing
- 25:53CI pain points
- 28:49Closing
related talks
35mManaging CI Across Multiple Farms for GPU Testing
38mGrowing a Lab for Automated Upstream Testing: Challenges and Lessons Learned
41mThe magical fantasy land of Linux kernel testing
37mKernel CI – How Far Can It Go?
29mTesting, a Journey from Testing Kernels to Testing Debian and Yocto
41mKeeping up with LTS: Linux Kernel Functional Testing (LKFT) on Devices
35mTest Harness: Continuous Testing on Hardware
41mBeyond unit tests