AMD User Queue Mesa Support Merged For Linux - Submitting Work Directly To The GPU

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • marek
    replied
    Originally posted by Kjell View Post

    Understandable

    In the future, does it make sense to launch the entire Wayland compositor with AMD_USERQ to achieve the best 3D performance/latency with subsequent processes?
    AMD_USERQ is for developers only. It's unlikely to be enabled for current GPU generations.

    Originally posted by Mitch View Post

    Please excuse my stupid question.
    So if you have the Hardware, the Kernel, and the MESA versions, and the switches enabled, what's left? Does software (Desktops, Browsers, Videogames) need to purposefully use this new feature and potentially rewrite some things, or does this instead automatically change how existing things work under the hood?
    Firmware and performance optimizations are missing. There is no impact on apps and APIs. It just replaces the current way Mesa submits work to the GPU. It's a replacement of existing uAPI.

    Originally posted by nuetzel View Post

    Hello mareko ,

    starting with gfx8+ or later?
    gfx13+. gfx11-12 are maybe.

    Leave a comment:


  • coder
    replied
    Originally posted by patrick1946 View Post
    Is that like io_uring minimizing kernel switches or is this actually a driver in the firmware running on the GPU?
    I think dev_null just means the interaction protocol is like that of io_uring. FWIW, io_uring's circular buffers are inspired by a long history of how hardware typically implements FIFOs. So, I wouldn't infer a connection with io_uring based on superficial similarities, here. Maybe someone was indeed inspired by it, or the performance gains it enabled?

    Leave a comment:


  • coder
    replied
    Originally posted by pac2000 View Post
    The applications would need to either be rewritten, or the compiler might be able take advantage of it via libraries, depending on how it is implemented.
    Isn't the whole point of Mesa to abstract the details of how application code interacts with the hardware? If Mesa now enables this new submission mechanism, then I'd bet application code doesn't even need a recompile.

    Leave a comment:


  • patrick1946
    replied
    Is that like io_uring minimizing kernel switches or is this actually a driver in the firmware running on the GPU?

    Leave a comment:


  • dev_null
    replied
    Right move, I believe inspired by io_uring. In the end they will invent shared video memory and msdos b800:0000 address

    Leave a comment:


  • MrCooper
    replied
    Originally posted by Kjell View Post
    In the future, does it make sense to launch the entire Wayland compositor with AMD_USERQ to achieve the best 3D performance/latency with subsequent processes?
    Yes, this will make sense for everything. Presumably it'll become the default at some point once the kinks are ironed out.

    I'm actually pretty psyched about this, it should allow getting the full benefit of my mutter work described in https://blogs.gnome.org/shell-dev/20...nsive-clients/ with AMD GPUs.

    Originally posted by Mitch View Post
    So if you have the Hardware, the Kernel, and the MESA versions, and the switches enabled, what's left?
    Per the post you quoted, getting the needed firmware. Since that's unreleased, nobody outside of AMD can play with this just yet.

    Does software (Desktops, Browsers, Videogames) need to purposefully use this new feature and potentially rewrite some things, or does this instead automatically change how existing things work under the hood?
    The latter.

    Originally posted by nuetzel View Post

    starting with gfx8+ or later?
    AFAIK as things stand now, this will only be supported with RDNA3+. As it happens, I'm in the process of switching to a Zen4 + RDNA3 laptop for my main machine, good timing.

    Leave a comment:


  • pac2000
    replied
    Originally posted by Mitch View Post

    Please excuse my stupid question.
    So if you have the Hardware, the Kernel, and the MESA versions, and the switches enabled, what's left? Does software (Desktops, Browsers, Videogames) need to purposefully use this new feature and potentially rewrite some things, or does this instead automatically change how existing things work under the hood?
    The applications would need to either be rewritten, or the compiler might be able take advantage of it via libraries, depending on how it is implemented. So would need to be recompiled. But even if that was possible, the compiler would have to be rewritten.

    This is mainly for new games or applications to increase their capabilities and speed at which they can interact with the GPU. Imagine a VM manager being able to pass through the graphics capabilities to a virtual machine. Or create a VM driver that translates to the user queue. This would provide greater graphic capabilities within the VM. Which has been lacking, and restricted to basic graphics. So it's very good news when everything catches up, and gives AMD an advantage over Nvidia.

    Leave a comment:


  • nuetzel
    replied
    Originally posted by marek View Post

    It also depends on unreleased firmware. It's still experimental.
    Hello mareko ,

    starting with gfx8+ or later?

    Leave a comment:


  • Mitch
    replied
    Originally posted by marek View Post

    It also depends on unreleased firmware. It's still experimental.
    Please excuse my stupid question.
    So if you have the Hardware, the Kernel, and the MESA versions, and the switches enabled, what's left? Does software (Desktops, Browsers, Videogames) need to purposefully use this new feature and potentially rewrite some things, or does this instead automatically change how existing things work under the hood?

    Leave a comment:


  • Kjell
    replied
    Originally posted by marek View Post

    It also depends on unreleased firmware. It's still experimental.
    Understandable

    In the future, does it make sense to launch the entire Wayland compositor with AMD_USERQ to achieve the best 3D performance/latency with subsequent processes?

    Leave a comment:

Working...
X