Fix vectorization with predicates

[kaushikcfd]

TODO:

• The added regression fails with a CL-kernel compilation error. Not (yet) sure how to teach loopy.codegen about such predicates.

[kaushikcfd]

FWIW, the issue here seems to be that the CodeGenerationState.try_vectorized is a broken interface as it isn't in control of the predicates.

[inducer]

try_vectorized is broken

IMO, float4-style vectorization (in CL, but also in gcc vector extensions) does not have a good way to express masking (i.e. lane-varying control flow). I'd suggest we don't try to support it and scalarize instead.

[sv2518]

TJ didn't support vectorisation of predicates either I think.

[kaushikcfd]

TJ didn't support vectorisation of predicates either I think.

Yep!

IMO, float4-style vectorization (in CL, but also in gcc vector extensions) does not have a good way to express masking (i.e. lane-varying control flow). I'd suggest we don't try to support it and scalarize instead.

In a personal meeting, it was discussed that let's xfail this particular regression and hope to land a fix for #615 once #372 lands.

---

Workaround: Currently in the proposed PyOP2 patch (OP2/PyOP2#654) instead of leaving just these instructions unvectorized, we "turn-off" vectorization for all instructions of such kernels.

[inducer]

Thx!