Skip to content

Linux GPU renderer qualification

Results

Repeated post-ready SIGSEGV emulator exits on Linux hosts (#12453) were seen with -gpu swangle -feature -Vulkan on emulator 37.2.12.0 and a medium_phone Android 15/API 35 Google APIs x86_64 image. The cause is still open: no native backtrace was captured, and the crash has not been reproduced.

The manual Linux Qualification workflow (.github/workflows/linux-qualification.yml, scripts under scripts/ci/linux-qualification/) runs each GPU mode under three workloads: AutoMobile tools, a raw adb control, and a heartbeat-only control. Run 38902077979 (main f1fd091b81, two cycles per cell, emulator 37.2.12.0, android-35;google_apis;x86_64, GitHub-hosted ubuntu-latest, no GPU) gave:

GPU mode AutoMobile Raw adb control Heartbeat only
swiftshader_indirect 2/2 clean 2/2 clean 2/2 clean
swangle 2/2 clean 2/2 clean 2/2 clean
angle_indirect 2/2 clean 2/2 clean 2/2 clean
host fails at renderer init fails at renderer init fails at renderer init

No software or indirect GPU cell crashed after readiness. -gpu host stops before readiness with Failed to get EGL display / Could not start renderer because the runner has no GPU render node. That is a startup failure, not the post-ready signature. Two clean cycles per cell is weak evidence: it neither clears nor implicates any software renderer.

The emulator’s -verbose output shows what each flag actually selected on 37.2.12.0:

Flag passed Renderer the emulator used
-gpu swiftshader_indirect mode swiftshader: SwiftShader GLES (gles_swiftshader, GLES 3.0), SwiftShader Vulkan
-gpu swangle -feature -Vulkan ANGLE GLES (gles_swangle, GLES 3.1) on SwiftShader; host Vulkan feature off
-gpu angle_indirect fell back to mode lavapipe: ANGLE GLES on Mesa llvmpipe Vulkan

The angle_indirect row therefore measured ANGLE on Lavapipe, not a distinct ANGLE-indirect path.

The emulator documentation deprecates swiftshader_indirect, swangle_indirect and guest (36.4.9) and angle_indirect (34.2.14). The current software modes are software, swiftshader, lavapipe and swangle. CI (.github/actions/android-emulator/action.yml) passes -gpu swiftshader, the mode the deprecated swiftshader_indirect already mapped to.

Linux configuration guidance until the cause is known:

  • On headless hosts without a GPU, use a current software mode: -gpu swangle or -gpu swiftshader. Both were clean in this matrix. Pass it with AUTOMOBILE_EMULATOR_ARGS='["-gpu","swangle"]'.
  • Use -gpu host only where /dev/dri/renderD* and a working EGL/Vulkan driver exist.
  • -feature -Vulkan is the one non-default part of the crashing hosts’ configuration. It was clean here, but it is unqualified beyond these runs; omit it unless a workload needs it.

When an emulator exits unexpectedly, the loss incident keeps the process exit signal and a bounded, redacted outputTail of emulator output from launch onward. For an emulator AutoMobile did not launch, outputUnavailable: "adopted" records that no output was tracked. Slot recovery (onLoss) stays off by default and never substitutes a different device.

The swangle cell matched the crashing hosts’ flags, emulator build, image and hardware profile. What still differs is the host machine, the experimental broker build (these runs used the published 0.0.85 package), the repeated acquire/release workload, and the cycle count. A longer soak on swangle and swiftshader is the remaining step for #12453; no self-hosted GPU host is available, so the host row stays unexercised.