Login
You're viewing the functional.cafe public feed.
  • Nov 15, 2026, 5:18 AM

    @swick @TTimo I'm not involved in this work but I am in Mesa, and your solution of using virgl is OK for non-demanding apps, but adding another process in the command submission path is bound to lower performance and, worst of all, increase jitter : the very reason why modern CPUs come with ginormous caches.

    On top of this, Virgl is barely maintained and has been a thorn in our side for a while now.

    Making the entire flatpak ecosystem depend on it is bound to piss off a lot of people, basically everyone but flatpak devs... even if I know that this is architecturally the right solution until gpu drivers become more embeddable.

    💬 2🔄 0⭐ 1

Replies

  • Nov 15, 2026, 1:41 PM

    @mupuf @swick @TTimo it's been a while since i last used virglrenderer... does flatpak use virgl or venus? the article mixes both... is only virgl unmaintained, or also venus? regarding the extra latency/jitter inherent to virtualization, I'd hope this isn't too impactful since it's used in production by asahi on games (although I can't remember if they use virglrenderer, gfxstream or something else) (edit: virglrenderer with drm forwarding)

    💬 0🔄 0⭐ 0
  • Nov 15, 2026, 2:40 PM

    @mupuf @TTimo Virgl specifically was an exploration on how to provide a fallback when the drivers from the runtime become too old. We could also add libcapsule support, but like I explained in the blog post (and is very conveniently being ignored by Valve) this creates a dependency on the host, specifically on i386. The only reason why Valve gets away with it is because no one wants to drop support for Steam which means no one drops i386 on the host. Except immutable distros where flatpak then provides i386 but there the drivers do come from the Flatpak runtime.

    💬 0🔄 0⭐ 0
  • Nov 15, 2026, 2:46 PM

    @mupuf @TTimo also, what is the maintained solution then? The native context stuff requires matching code on both sides which completely defeats the purpose here.

    💬 0🔄 0⭐ 0
  • Nov 16, 2026, 4:05 AM

    @swick @TTimo The more modern solutions are gfxstream or SRIOV, as far as I know.

    The former shares the same performance impact as Virgl due to the necessary context switch in the command submission path. The latter requires modern hardware and virtual machines AFAIK.

    The big issue with Virgl is the lack of users and the reliance on TGSI, which we definitely are trying hard to get rid of, as it has been superseded by NIR. Gfxstream helps with that part, but still lack wide adoption...

    💬 0🔄 0⭐ 0