Trying Out RadeonSI's Experimental NIR Support

[phoronix]

Phoronix: Trying Out RadeonSI's Experimental NIR Support

Last week AMD developers landed experimental NIR support inside RadeonSI as ultimately a step towards ARB_gl_spirv support for this OpenGL Gallium3D driver in being able to re-use the NIR/RADV code-paths. I decided to take the experimental NIR support through a spin to see if it impacted various titles at this point...

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

[AJenbo]

portal is up by over 3%, how has it regressed?

[Michael]

portal is up by over 3%, how has it regressed?

On Mesa Git it yields odd colors / rendering artifacts.

[R41N3R]

Looks promising even though TGSI seems to be replaced only with NIR and there are not less steps at the moment.

[Masush5]

Looks promising even though TGSI seems to be replaced only with NIR and there are not less steps at the moment.

Getting rid of intermediate steps is not necessarily a good thing. As i understand it, NIR, for example, is particularly suited to do some optimization passes on and thus going for GLSL IR -> NIR (+optimization passes) -> LLVM -> ISA might result in a more performant shader binary than skipping the NIR step.

[BNieuwenhuizen]

Getting rid of intermediate steps is not necessarily a good thing. As i understand it, NIR, for example, is particularly suited to do some optimization passes on and thus going for GLSL IR -> NIR (+optimization passes) -> LLVM -> ISA might result in a more performant shader binary than skipping the NIR step.

Indeed, for radv, we at the time chose to use SPIR-V -> NIR-> LLVM instead of SPIR-V -> LLVM for exactly this reason. LLVM IR throws a lot of information away and some things would be harder to do.

That said this effort is mostly about having SPIR-V support for radeonsi, and possibly eventually having only a NIR->LLVM backend instead of having to maintain both NIR->LLVM and TGSI->LLVM, and not so much about reducing the number of IRs in a compile.

[robclark]

Looks promising even though TGSI seems to be replaced only with NIR and there are not less steps at the moment.

just some background about compilers.. a compiler does many opt/lowering steps, and generally we run many opt passes in a loop until it converges. Converting between IR's is basically less than nothing, it is an O(1) step, compared to passes which are O(n^2) or O(n^n) or worse. So it only sounds like a thing to care about on the surface. (Plus in many cases we can replace less algorithmically efficient glsl IR passes with more efficient NIR passes, which *more* than makes up for the extra conversion between IRs.. iirc turning off the passes that used to be done in glsl IR but can now be done in NIR was something like a ~30% drop in shader-db runtime for freedreno)

[pal666]

Converting between IR's is basically less than nothing, it is an O(1) step

i'm pretty sure you meant o(n)

[robclark]

i'm pretty sure you meant o(n)

yeah, sorry.. I should proof read posts sometimes