Announcement

Collapse
No announcement yet.

RadeonSI NIR Benchmarks Show Great Progress With Mesa 20.0

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

  • RadeonSI NIR Benchmarks Show Great Progress With Mesa 20.0

    Phoronix: RadeonSI NIR Benchmarks Show Great Progress With Mesa 20.0

    With AMD last week having enabled OpenGL 4.6 for their RadeonSI OpenGL Linux driver when enabling the NIR intermediate representation support, you may be wondering how using NIR is stacking up these days compared to the default TGSI route. Here are some benchmarks on Polaris, Vega, and Navi for comparing this driver option that ultimately allows OpenGL 4.6 to be flipped on.

    http://www.phoronix.com/vr.php?view=28594

  • #2
    This is awesome!
    I was sure it'd be still slower but instead it's faster, thanks to everyone involved!

    Comment


    • #3
      Does radeonsi_enable_nir=true require the user to manually clear the shader cache for the NIR IR to take effect?

      Comment


      • #4
        And still over years no in-home streaming encoder support on mesa for amd....

        Comment


        • #5
          The Vega 56 result in Bioshok looks weird, it should beat Polaris by a fair margin. But that has nothing to do with NIR. Also, is there still more potential for better performance with NIR vs. TGSI?

          Comment


          • #6
            Michael typo:
            Radeon RX 580
            But in benchmark results it becomes 590.

            Comment


            • #7
              Originally posted by atomsymbol View Post
              Does radeonsi_enable_nir=true require the user to manually clear the shader cache for the NIR IR to take effect?
              There was at least one fix to address this, so I think you don't have to worry about this anymore.
              I'm btw. completely free of hassle with RadeonSI NIR and recent mesa-git.

              Comment


              • #8
                Originally posted by ms178 View Post
                The Vega 56 result in Bioshok looks weird, it should beat Polaris by a fair margin. But that has nothing to do with NIR. Also, is there still more potential for better performance with NIR vs. TGSI?
                Realistically, no - not unless there's something hopelessly broken in TGSI. "Shaders is shaders", basically: you MIGHT get 1 or 2 %, MAYBE, from different infrastructure, but the IR would have to be utterly broken in a critical way (e.g. doing per-component dots etc rather than vector ops) for there to be any meaningful impact at all, and that means that defect would have had to sit around unnoticed / unfixed in the old code for x years now, which seems unlikely. (Unless it's part of GNOME, of course... :P *ducks*)

                Comment


                • #9
                  Nice results, hope Nir will be enabled soon in radeonsi!

                  Comment


                  • #10
                    Originally posted by atomsymbol View Post
                    Does radeonsi_enable_nir=true require the user to manually clear the shader cache for the NIR IR to take effect?
                    The IR is part of the key in the current master code. So no, flipping between nir and tgsi will create separate cache entries.

                    Comment

                    Working...
                    X