GLSL IR To TGSI Translator Ready To Be Merged Into Mesa

[marek]

Isn't the GLSL IR to TGSI traslator one of the things that is blocking Intel from switching to Gallium3d?

I don't think so. They apparently don't want to throw away the codebase of their classic driver.

Also, is it possible to eventually use GLSL directly without translating it?

Maybe. There seems to be an increased interest for Gallium to go in that direction.

[Prescience500]

Isn't the GLSL IR to TGSI traslator one of the things that is blocking Intel from switching to Gallium3d? Also, is it possible to eventually use GLSL directly without translating it?

[Plombo]

It would be good to take a few random shaders those games use and compare the TGSI generated by both. I remember originally you said the new code was not quite as good as the old, but close. Has that been completely fixed, or is it possible some shaders could have regressed?

It is possible that some shaders could have regressed, but I have yet to find one. When I said that the new code was not as good as the old, I was mostly referring to copy propagation from temporary to output registers. But most of the drivers will optimize that out anyway after converting to their own IR. (I know nv50, nvc0, and llvmpipe do, and presumably r300g as well.) For everything I've benchmarked on my NVA5 card using the nv50 driver, there has been no measurable difference in framerate between glsl_to_tgsi and ir_to_mesa+mesa_to_tgsi.

Also, do you know if there has been any improvement in the compile time of shaders? I'm guessing that's all spent inside the actual GLSL compiler and not the translation code.

I haven't measured it, but I don't think there is any difference. I know that there is about the same delay while compiling shaders when I play Portal under Wine.

[oibaf]

It would be good to take a few random shaders those games use and compare the TGSI generated by both. I remember originally you said the new code was not quite as good as the old, but close. Has that been completely fixed, or is it possible some shaders could have regressed?

There is shader-db for this (but I think it currently require an intel card for it): http://lists.freedesktop.org/archive...ay/007694.html

[smitty3268]

Plus, in the 2 months or so I've been developing this, I've been using my custom Mesa build with the GLSL->TGSI translator every time I play an OpenGL game on my computer.

It would be good to take a few random shaders those games use and compare the TGSI generated by both. I remember originally you said the new code was not quite as good as the old, but close. Has that been completely fixed, or is it possible some shaders could have regressed?

Also, do you know if there has been any improvement in the compile time of shaders? I'm guessing that's all spent inside the actual GLSL compiler and not the translation code.

[Plombo]

but are not there any legal blockers? and are all of them a target for someone?

No, there are no legal issues here.

Ohh, small question how did you tested Your code?

I used the Piglit test suite to test for regressions. Plus, in the 2 months or so I've been developing this, I've been using my custom Mesa build with the GLSL->TGSI translator every time I play an OpenGL game on my computer.

[Remco]

I was looking up interesting patents

You're a masochist, you!

[bridgman]

I was looking up interesting patents (unrelated to graphics) a few weeks ago for an internal presentation and happened to find the patent you are talking about :

http://www.faqs.org/patents/app/20090081079

[AnonymousCoward]

Stop whining about patents already! If you are worried about patents you might as well stop coding period.

Hell, you may even want to consider to stop breathing, as there's no doubt someone has already patented a "method and apparatus for extracting oxygen from the atmosphere and transporting it into the bloodstream".

[whizse]

but are not there any legal blockers? and are all of them a target for someone?
?

AIUI the only patented parts of GL3 have been dealt with:
http://cgit.freedesktop.org/mesa/mes...cs/patents.txt

But considering that any non-trivial codebase probably infringes on some or other patent out there other patents might pop up, but I would guess that they would be dealt with in the same way, or as with S3TC stuff by an external library.