AMD's Plans For ARB_gl_spirv Support In RadeonSI

[phoronix]

Phoronix: AMD's Plans For ARB_gl_spirv Support In RadeonSI

Back in May we talked about ARB_gl_spirv / NIR Support Being Worked On For RadeonSI while now we have more details from AMD's Nicolai Hähnle regarding these plans...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-ARB_gl_spirv-Steps

[M@GOid]

I was looking at Mesamatrix and man, there is a lot of extensions post OpenGL 4.5. Khronos should launch 4.6 to put things strait.

[schmidtbag]

I was looking at Mesamatrix and man, there is a lot of extensions post OpenGL 4.5. Khronos should launch 4.6 to put things strait.

Yeah no kidding, especially considering how many of the "Other Extensions" are necessary for certain games. I understand they're trying to push attention toward Vulkan but the reality is, adoption for that will be relatively slow (though, I am a bit surprised how quickly it has been adopted considering how slow things have been in the past).

[Veerappan]

Just read the blog post, pretty straight-forward and I like that nhaenle's re-using a bunch of the radv SPIR-V -> NIR -> LLVM code that Dave/Bas have been maintaining. That may have benefits for the radv driver at some point as well.

I also see some interesting opportunities for comparison/optimization in the testing path that he's highlighted.

If all goes to plan, you could theoretically get radeonsi to compile its glsl shaders using either TGSI or NIR as an IR, which means that you can do side-by-side testing of TGSI/NIR and figure out which one has benefits/optimizations that the other is lacking. It'll also give us a bit of insight as to which has different CPU overhead characteristics.

That being said, TGSI/Gallium has a pile of state trackers available for it besides just GLSL/SPIR-V, which means that it'll be sticking around for a long time to come in some form or another.