RadeonSI May Eventually Switch To NIR Completely

[phoronix]

Phoronix: RadeonSI May Eventually Switch To NIR Completely

RadeonSI developers have been working on supporting the NIR intermediate representation within their Gallium3D driver as a means to support ARB_gl_spirv for being able to load SPIR-V shaders in OpenGL and interact with the RADV Vulkan driver code paths, which is making use of NIR. It's looking like in the future the RadeonSI driver could end up using NIR completely by default...

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

[davidbepo]

good, especially if it can lead to performance improvements

[tildearrow]

Typo:

still need to be done for RadeoNSI NIR

[Nille_kungen]

If it makes things easier then why not.

[Mani]

I wonder if this would affect the gallium9 driver. Wouldn't this change the driver model from being a gallium driver to being a “classical“ driver?

[curfew]

You can't say "New Intermediate Representation intermediate representation". It is a primitive linguistical error. Instead you have to write it out: "New Intermediate Representation (NIR)". This way also your readers can understand what the fuck are you talking about.

[Luke_Wolf]

You can't say "New Intermediate Representation intermediate representation". It is a primitive linguistical error. Instead you have to write it out: "New Intermediate Representation (NIR)". This way also your readers can understand what the fuck are you talking about.

Sure you can, New Intermediate Representation is a proper noun, intermediate representation is a common noun, and proper noun common noun is a common linguistic pattern in English. For example: Kleenex tissues, however a proper noun followed by a common noun, where the proper noun includes the common noun certainly does sound silly and is discouraged but it's not wrong.

Now that said what you're proposing is better writing but what Michael did isn't really wrong, just it sounds silly.

[kiffmet]

can someone explain to me why they do GLSL -> NIR -> Backend instead of GLSL -> Backend? Doesn't this cause a lot of overhead on the CPU side?

[dcbaker]

can someone explain to me why they do GLSL -> NIR -> Backend instead of GLSL -> Backend? Doesn't this cause a lot of overhead on the CPU side?

No. The point is that each IR has specific properties that make certain kinds of optimizations easier, and doing the transformation from one IR to another is a very cheap operation, while the optimizations that each IR makes possible can make the shader much more efficient on the GPU. Since the shader is only compiled once with a shader cache, the compilation cost (while not bad anyway) is well worth paying to get better runtime on the GPU.

[marek]

"Completely" meant that we would stop using TGSI in Mesa "completely". It didn't mean that we would stop using LLVM. We will always use LLVM. I admit it might have not been very clear on the mailing list.