RadeonSI Picks Up Primitive Culling With Async Compute For Performance Wins

[marek]

Not seeing much of a difference in gaming tests: https://twitter.com/phoronix/status/1095825661556654080

Did you test with the kernel from amd-staging-drm-next and LLVM 8.0 from the svn release branch or newer? You need both, otherwise the optimization will be silently disabled.

Run: AMD_DEBUG=info glxgears|grep -e ordered -e scheduled

You should see:
has_gds_ordered_append = 1
has_scheduled_fence_dependency = 1

The original Mesa branch broke mplayer and mpv. The latest version is fixed.

[Michael]

Did you test with the kernel from amd-staging-drm-next and LLVM 8.0 from the svn release branch or newer? You need both, otherwise the optimization will be silently disabled.

Run: AMD_DEBUG=info glxgears|grep -e ordered -e scheduled

You should see:
has_gds_ordered_append = 1
has_scheduled_fence_dependency = 1

The original Mesa branch broke mplayer and mpv. The latest version is fixed.

I did with the very latest LLVM 8.0 bits but not with amd-staging-drm-next, wasn't aware newer kernel bits were needed so was using Linux 5.0. Looks like I'll re-test in the morning.

[marek]

Could be worse. I read that Pixar's RenderMan would tessellate well below the size of a pixel - until the triangle was too small to influence the color of the pixel.

A triangle has to intersect at least 1 sample point to be visible. Even with 8x MSAA, if a triangle falls exactly in the empty space between sample points, it's invisible (it doesn't contribute to the color).

[coder]

I did with the very latest LLVM 8.0 bits but not with amd-staging-drm-next, wasn't aware newer kernel bits were needed so was using Linux 5.0. Looks like I'll re-test in the morning.

Probably best to start by reproducing the Paraview result. That way, you know all the correct bits are in place.

[clapbr]

Did you test with the kernel from amd-staging-drm-next and LLVM 8.0 from the svn release branch or newer? You need both, otherwise the optimization will be silently disabled.

Run: AMD_DEBUG=info glxgears|grep -e ordered -e scheduled

You should see:
has_gds_ordered_append = 1
has_scheduled_fence_dependency = 1

The original Mesa branch broke mplayer and mpv. The latest version is fixed.

I confirm the mpv bug indeed, and also had has_gds_ordered_append = 1/has_scheduled_fence_dependency = 1 on terminal, but at least in unigine benchmarks I had no difference, gonna test more tomorrow with the newer code.

[spstarr]

Looking at the code, it just looks for "PRO" anywhere in the cards marketing name.

yeah, typical laptops like my Dell:

01:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Saturn XT [FirePro M6100]

So that code path would trigger for this CIK chip,

[dungeon]

Wondering what couple of Radeon Instinct MI60 could do there

Hm, are APUs supported by these patches?

[Peter Fodrek]

Environment variable.

Thank you very much

[tuxd3v]

Wondering what couple of Radeon Instinct MI60 could do there

Hm, are APUs supported by these patches?

Yeah that would be awesome,
I think that Marek talked about GCN 1.1 Sea Islands or above..

Kavery APUs have 8 ACEs units I think..?
I don't know if they have 2 dma units( for bidireccional usage), but I believe so, has they use System Ram memory, as main memory

Can Someone more Educated, on the matter than me, help on this questions?

[nuetzel]

I confirm the mpv bug indeed, and also had has_gds_ordered_append = 1/has_scheduled_fence_dependency = 1 on terminal, but at least in unigine benchmarks I had no difference, gonna test more tomorrow with the newer code.

Hello clapbr,

have a look, here:

https://lists.freedesktop.org/archives/mesa-dev/2019-February/215240.html

Greetings,
Dieter