Vulkan CTS memory requirements had been steadily increasing over many years until reaching a breaking point when it started creating issues for some Mesa CI runners. We studied the problem, understood the root cause and decreased peak memory needs from 900 MBs to 300 MBs in our test builds. This talk is a quick postmortem that also helps set expectations for CTS users.
Fixing a Vulkan driver bug is important, but making sure it never returns is just as important. This lightning talk will briefly describe the process of turning a real driver bug into a Vulkan CTS test. I will share a few examples of challenges encountered while reproducing bugs in a small, reliable conformance test, and explain why a straightforward test often isn't enough to expose the...
When several new boards could not fit my traditional setup, I had to revert to use of the CI-Tron for automation of those boards. This lightining talk shows the use of CI-Tron for the kernel developent and maintainer duties, emphasising its usage out of the traditional CI area.
It's been well over a decade since NIR landed upstream in Mesa. Since then, a lot of people have worked hard to get rid of TGSI. While there's still a few road blocks left, we're getting close to the home-stretch.
This talk will try to answer the question: How can we finally get rid of TGSI in Mesa?
Short status update about new OpenCL features supported, features work in progress, changes in driver status and what the priorities are going forward.
Some things are worth the wait. After a long, patient climb, the etnaviv project
reached two milestones this year that once felt a long way off.
This lightning talk looks at where the driver finally caught up on conformance, and
at a new capability that opens the door to a whole class of workloads etnaviv could
not run before.
Years in the making, and worth every step
The V3D (OpenGL) and V3DV (Vulkan) drivers power the Broadcom GPU found in the Raspberry Pi 4 and 5.
In this lightning talk we will go through a high-level tour of the latest updates to these drivers that we deliveredwhat things have been done on them: new extensions and features, performance improvements in different paths, refactors, CI, and so on.
After this talk you will have a...
Workgroups are a concept which allow multiple SIMD waves of shaders to run together and share some memory. API features such as compute and mesh shaders expect the driver to support workgroups. Modern hardware typically supports it, but... What if you want to run workgroups on hardware that doesn't really support it, for example you want to run mesh shaders on NVidia? Or maybe you want to...
Modern games and game engines are getting more and more complicated, and show a lot of creativity in inventing new ways to implement memory management issues. Use after free, use before alloc, you name it. You may have seen GPU hangs thanks to page faults during gameplay.
In this lightning talk, I'd like to present a proposal to deal with this type of problems on modern AMD GPUs on Linux,...
On the last XDC the plan for the common DisplayPort framework in DRM was announced. Thihs lightining talk provides a (small) status update over the past year.
Problem Statement
Compositors are unaware of the media content frame rate and therefore present frames using the display's default refresh cadence. This mismatch leads to irregular frame timing, causing judder during low-frame-rate video playback (e.g., 24/30 fps) on both fixed-refresh and VRR displays. A familiar example is 3:2 pulldown, where 24 fps content is displayed on a 60 Hz panel...
A familiar failure motivates this talk: two high-resolution displays on one USB4/Thunderbolt dock, each working alone but not together, because nothing arbitrates the shared link.
A DisplayPort monitor on such a dock looks like an ordinary connector to userspace, but in the kernel the work is split across two subsystems that never call each other. The graphics driver (i915/xe plus the DRM...
Chromium runs thousands of interactive UI tests that need synthesized input. On Wayland
there's no portable way to do that, and the reference compositor (weston) doesn't implement
enough of the real protocol surface to be representative. So we forked mutter, taught it
Chromium's ui-controls protocol, and wired it into our CI to run interactive tests on a
real compositor for both pre- and...
Wayland input latency is a topic discussed in the Linux graphics community and claims that “Wayland adds latency” have been repeated in the past. This talk explores end-to-end latency comparison in an automated A/B testing setup designed to produce more rigorous and reproducible measurements. The session focuses on comparative evaluation across multiple compositors or window managers,...
