Mesa just deleted its anti-lag layer. RADV does the job itself now
Mesa's implicit anti-lag layer has been deleted, and RADV, the open source Radeon Vulkan driver on Linux, now implements VK_AMD_anti_lag itself. Six commits merged to Mesa's main branch on October 8 at 22:21:31 UTC, adding a shared frame pacing framework of 894 lines and removing 1,836 lines of the layer that did the work before.
VK_AMD_anti_lag is a Vulkan extension, and RADV is the open source Radeon Vulkan driver on Linux. Image: AMD
All six are authored by Daniel Schürmann and were pushed by Marge Bot, the merge train Mesa uses, each one carrying the same merge request reference, 42048. They were read at 21:41 UTC on October 10 through a GitHub mirror of the Mesa repository, since gitlab.freedesktop.org answers a bot check from here, and the hashes below are the upstream ones.
894 lines in, 1,836 out
The foundation is 6854f77017cf, "vulkan/runtime: add vk_frame_pacer utility framework", five files and 894 added lines. "This framework provides basic utility for the implementation of low latency frame pacing extensions," it says, measuring frame times, where a frame is "all submissions between two calls to vkQueuePresentKHR", along with queuing delays taken from calibrated timestamps at submission and at the start of execution on the device.
On top of it sits 8b045fd2baad, "radv: implement VK_AMD_anti_lag", 166 lines across four files, which describes itself as "based on the vk_frame_pacer framework and intended as full replacement of the MESA_LAYER_ANTI_LAG". It spells out the arithmetic: "delay = avg_frame_time + queuing_delay - MIN2(deviation, ONE_MS_IN_NS)", with half the queuing delay added because "that's what we try to minimize" and the average frame time deviation subtracted "as slack time, so that we don't drain the GPU from work".
A six line commit advertises the extension, two more move a ring buffer header out of the old layer into Mesa's shared utilities and add a macro to it, and b99595dedb85, "vulkan: delete anti-lag layer", takes out the rest: eight files, one line added and 1,836 removed.
In main, with 26.3.0-rc1 scheduled for October 14
The branch this landed on is Mesa's main, whose VERSION file reads 26.3.0-devel. Mesa's own release calendar puts 26.3.0-rc1 on October 14, rc2 on October 21, rc3 on October 28 and rc4 on November 4, that last row annotated "Or 26.3.0 final.", with Dylan Baker as release manager. None of that is a shipped driver yet, and the calendar's final row is written as a possibility rather than a date.
The timing inside the commits is its own small story. The framework was authored on June 3, the RADV implementation on June 4 and the deletion of the layer on the same day, with the advertisement commit following on June 22. They were merged four months later, in one batch, six days before the release candidate the branch is pointed at.
Nothing here is measured. The commits carry no latency figure, no frame time comparison and no game, so the only claim on the table is the structural one: a Vulkan extension implemented inside the driver instead of a layer loaded around it. VK_AMD_anti_lag is a Vulkan extension and not the Anti-Lag 2 feature AMD ships in its Windows driver, which is a separate implementation on a separate stack.
Vulkan plumbing reaches Radeon users on Linux often enough, usually as a fault, as with a Steam beta overlay that crashed Vulkan games until Valve fixed it. This one is quieter: if it works, the extension is simply there, and the layer nobody wanted to maintain is not.
