AMD Drops TGSI Support From Their RadeonSI OpenGL Driver

[phoronix]

Phoronix: AMD Drops TGSI Support From Their RadeonSI OpenGL Driver

It was just shy of a month ago in Mesa 20.0-devel code that RadeonSI enabled NIR by default to enjoy OpenGL 4.6 support and now already ahead of the Mesa 20.0 branching the previous TGSI code-path has been removed. TGSI has been the default intermediate representation of RadeonSI and other Gallium3D drivers but these days NIR has become the go-to IR for Mesa drivers...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-RadeonSI-Drops-TGSI

[boxie]

that is certainly a good development - here is to uncovering all the edge cases in the new 20 release :P

[Guest]

Code removal feels so nice

[geearf]

Is there an impact to certain state trackers, like nine?

[Kayden]

Is there an impact to certain state trackers, like nine?

Nine should continue to work—radeonsi will translate the TGSI supplied by nine to NIR via tgsi_to_nir, like iris does. So everything is still supported, but now everything goes through NIR and there's one fewer path in the driver.

(I wondered this myself, so I looked at the patches earlier.)

[geearf]

Nine should continue to work—radeonsi will translate the TGSI supplied by nine to NIR via tgsi_to_nir, like iris does. So everything is still supported, but now everything goes through NIR and there's one fewer path in the driver.

(I wondered this myself, so I looked at the patches earlier.)

Awesome, thank you for looking it up and answering!

[R41N3R]

I believe this is the right decision, the NIR environment variable caused some confusion in the past, so it's good to drop TGSI and to focus on just one path.

[Veerappan]

Nine should continue to work—radeonsi will translate the TGSI supplied by nine to NIR via tgsi_to_nir, like iris does. So everything is still supported, but now everything goes through NIR and there's one fewer path in the driver.

(I wondered this myself, so I looked at the patches earlier.)

I was going to ask basically the same question, but for the video-decoding state trackers. Hopefully the TGSI->NIR pass doesn't have any oddball corner cases that we haven't stumbled upon yet.

[Kayden]

I was going to ask basically the same question, but for the video-decoding state trackers. Hopefully the TGSI->NIR pass doesn't have any oddball corner cases that we haven't stumbled upon yet.

Yeah, that's another reason for TGSI->NIR. If there are any, we'll have to fix them :)

[puleglot]

That's not very impressive. Still waiting for ACO in radeonsi. =P