RadeonSI NIR Mesa 19.3 Testing Didn't Turn Up Any Bugs, Similar Performance To TGSI

[phoronix]

Phoronix: RadeonSI NIR Mesa 19.3 Testing Didn't Turn Up Any Bugs, Similar Performance To TGSI

Besides RADV ACO compiler testing and AMDGPU bulk moves restored another common test request recently on the Radeon Linux graphics side has been looking at the NIR support...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-NIR-Mesa-19.3-Test

[Venemo]

There is no such thing as AMD_DEBUG=nir in mesa 19.3 so probably both your test cases in fact used TGSI. To see the available AMD_DEBUG options, use AMD_DEBUG=help or look at the start of si_pipe.c where you can find the debug_options ― you will find that nir is not there. You can enable or disable NIR using a drirc config setting instead.

To enable NIR on radeonsi, make a ~/.drirc file like this:

<driconf>  
     <device driver="radeonsi">
        <application name="all">
            <option name="radeonsi_enable_nir" value="true"/>
        </application>
    </device>
</driconf>

[davidbepo]

i dont know about 19.3 but 19.2 sure has bugs with NIR:

https://gitlab.freedesktop.org/mesa/mesa/issues/1911

[skeevy420]

There is no such thing as AMD_DEBUG=nir in mesa 19.3 so probably both your test cases in fact used TGSI. To see the available AMD_DEBUG options, use AMD_DEBUG=help or look at the start of si_pipe.c where you can find the debug_options ― you will find that nir is not there. You can enable or disable NIR using a drirc config setting instead.

Since Mesa 19.0.4 according to the release notes...damn, I'm slippin...

Now I'm wondering if those HM2 crashes I was dealing with was due to AMD_DEBUG=sisched since I was running AMD_DEBUG=nir,sisched at the time and, apparently, nir had no effect...at least my game works now and that's what matters

[clapbr]

NIR crashes mpv+vaapi for me but works for everything else.

edit: actually doesn't crash anymore, just tested

[Azpegath]

I can't help feeling that perhaps it would be wise to just focus more manpower on getting this move completely done, to then be able to focus on other stuff. marek, am I wrong?

[oxwivi]

If a lack of bugs provides equivalent performance to TGSI, what's the point of NIR?

[marek]

Michael please hold off on testing NIR at the moment. Even though AMD_DEBUG=nir doesn't exist and radeonsi_enable_nir=true should be used instead, the shader cache is inefficient with NIR and we need to fix that first.