subgroupAll failing

[fhoenig]

subgroupAll gives me the following error when running on iOS:

[mvk-error] VK_ERROR_INITIALIZATION_FAILED: Compute pipeline compile failed (Error code 3):
Encountered unlowered function call to air.quad_all.

Shouldn't it translate to simd_all ?

[cdavis5e]

It depends on your hardware. Older SoCs only support quad operations. Newer ones support simd_*, and we should be using them on those; if they're not, you should probably convert this to a bug.

[fhoenig]

So you are saying subgroupAll mapping to either simd_ or quad_ depending on hardware is expected behavior of the cross compilation chain?

[cdavis5e]

Yes.

[fhoenig]

That is odd.
Because if simd_ intrinsics can be at variable size from 4 to 32, that includes 4-wide. Switching the type of instrinsic for older hardware then is an unnecessary semantic violation of sorts.

Btw, where does that happen? in spir-v to msl or in the msl compiler itself?

[cdavis5e]

That is odd. Because if simd_ intrinsics can be at variable size from 4 to 32, that includes 4-wide. Switching the type of instrinsic for older hardware then is an unnecessary semantic violation of sorts.

We are forced to do this. The Metal compiler backend for A12 and under refuses to compile any simd_* functions; it fails with an ICE.

Btw, where does that happen? in spir-v to msl or in the msl compiler itself?

In SPIRV-Cross, so in the translation from SPIR-V to MSL.

[fhoenig]

But can the A12 and under run "waves" larger than 4? I think it might be. If this is true, then this "fallback" is just wrong or confusing in my opinion.

Anyways, not meaning to complain here. MoltenVK is awesome work and works like a charm otherwise, even for our "esoteric" tech.