Announcement

Collapse
No announcement yet.

Linux Gaming Performance With Radeon Vulkan NGG Culling

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

  • clouddrop
    replied
    Shadow of the Tomb Raider on my 6900 XTXH at 3440x1440 resolution, highest settings:

    default (2614 sclk) GPU score:
    no nggc:
    Min - 111
    Max - 261
    Avg - 150
    95% - 123

    with nggc:
    Min - 113
    Max - 262
    Avg - 153
    95% - 125

    and overclocked to 2860 sclk:
    no nggc:
    Min - 118
    Max - 280
    Avg - 161
    95% - 131

    with nggc:
    Min - 122
    Max - 276
    Avg - 163
    95% - 134

    Leave a comment:


  • Venemo
    replied
    Originally posted by ernstp View Post
    Ideally 3D engines should do perfect culling by themselves, it's a very important optimization. The better job the engine does, the less benefit this will have.
    Yes, a very well written compute based culling solution can implement some stuff (eg. depth based culling) that would be extremely difficult to implement in the driver. However, not every developer is willing to go to such lengths, and for them, a driver based solution does well.

    However the problem with engine-side solutions is that they have to do this work either on the CPU or in compute shaders. On the other hand a driver side solution can compile this directly into vertex (or tess eval) shaders. Ultimately, no one knows which approach scales and performs better.

    Mesh shaders will add some more possibilities for engine developers to implement culling more efficiently.

    Leave a comment:


  • ernstp
    replied
    Ideally 3D engines should do perfect culling by themselves, it's a very important optimization. The better job the engine does, the less benefit this will have.
    Last edited by ernstp; 17 July 2021, 03:59 PM.

    Leave a comment:


  • Venemo
    replied
    Originally posted by CochainComplex View Post
    got tricked by the placebo effect. unfortunately nggc is slightly worsening the performance on RDNA1 NAVI10 5700XT in CP2077.
    Its such a pitty that CP2077 does not have a proper ingame Benchmark. A scene loads allways slightly different but the fps was consistent anyway.
    However Settings set to Ultra - res 1920x1200 (proton-ge-6.12-GE1)

    FPS in this scene Cherry Blossom Market*

    Nggc: 52-55
    no Nggc: 54-56
    Yeah, this is similar to what I saw in Cyberpunk as well. The game has a different bottleneck and doesn't benefit from the increased primitive throughput, OTOH the added latency regresses it a little.

    Originally posted by CochainComplex View Post
    Keep in mind mesa is still compiled with the mr
    radv: Use 128-sized vertex grouping for NGG shaders.
    https://gitlab.freedesktop.org/mesa/...requests/11810
    That should be merged into main by now.


    Leave a comment:


  • CochainComplex
    replied
    Originally posted by shmerl View Post
    thx

    VK_INSTANCE_LAYERS=VK_LAYER_MESA_overlay VK_LAYER_MESA_OVERLAY_CONFIG=fps=1

    Leave a comment:


  • tildearrow
    replied
    Originally posted by CochainComplex View Post
    *for some reason I'm not allowed to upload attachments - so no screenshots? tildearrow
    Yep. Perhaps to not overload Phoronix's servers with images.

    I suggest you to use a different service for uploading screenshots, like your own website or Postimage (beware the latter might insert advertisement links on its BBCode option).

    Leave a comment:


  • CochainComplex
    replied
    got tricked by the placebo effect. unfortunately nggc is slightly worsening the performance on RDNA1 NAVI10 5700XT in CP2077.
    Its such a pitty that CP2077 does not have a proper ingame Benchmark. A scene loads allways slightly different but the fps was consistent anyway.
    However Settings set to Ultra - res 1920x1200 (proton-ge-6.12-GE1)

    FPS in this scene Cherry Blossom Market*

    Nggc: 52-55

    no Nggc: 54-56


    Keep in mind mesa is still compiled with the mr
    radv: Use 128-sized vertex grouping for NGG shaders.
    https://gitlab.freedesktop.org/mesa/...requests/11810

    *for some reason I'm not allowed to upload attachments - so no screenshots? tildearrow


    edit: tildearrow thx for clearifing the upload situation. makes sense.
    Last edited by CochainComplex; 17 July 2021, 06:21 AM.

    Leave a comment:


  • theriddick
    replied
    Originally posted by Venemo View Post

    It is beneficial in games that render a lot of primitives, where the fixed-function culling becomes a bottleneck. What this does is remove triangles in the shader so the fixed-function hardware has less work to do.
    Wonder what it does for star citizen, it renders allot of things... had major bottleneck under Linux, models, objects, and textures take extremely long time to load in.

    Leave a comment:


  • Venemo
    replied
    Originally posted by middy View Post
    any reason why its not going to be enabled by default once released? isn't this enabled by default in the windows drivers? so why wouldn't it be on linux? from what i can tell there doesn't really appear to be a universal downside to it.
    I'd like to eventually enable it by default on RDNA 2, but we're not comfortable enabling it until we got some solid performance data that proves it's worth it.

    Leave a comment:


  • middy
    replied
    any reason why its not going to be enabled by default once released? isn't this enabled by default in the windows drivers? so why wouldn't it be on linux? from what i can tell there doesn't really appear to be a universal downside to it.

    Leave a comment:

Working...
X