sve and avx512 masked intrinsics and sve semi-scalar intrinsics

[jtlap]

sve and avx512 provides many masked intrinsics with 2 or 3 flavours up to now we only use the first one

• merge with first input: m
• merge with 0: z
• merge with unknown: u (sve)

perhaps we have to also support some way the last 2 items with a syntax that could be
pipo(mask) or pipo[mode] mask) where mode could be m z or u with pipomask meaning m.
Perhaps m z and u could be upper_case (M,Z,U) to avoid most ambiguities

Also sve provides many mixed vector scalar intrinsics. The elementwise_callable way we define many functors does not allow a call to these intrinsics as pipo(vec,scalar) is converted in pipo(vec,brodcast(scalar) before the call the intrinsic.
Is there something to do ans is the gain sufficient to need an effort ?

[jfalcou]

* merge with 0: z

THis probabyl just need us to special case func[if_(mask).else_(eve::zero)] to be treated specially.
No this for additional markers.

* merge with unknown: u (sve)

Having unspecified values is probably not what we want. Testing will be hell.

Also sve provides many mixed vector scalar intrinsics. The elementwise_callable way we define many functors does not allow a call to these intrinsics as pipo(vec,scalar) is converted in pipo(vec,brodcast(scalar) before the call the intrinsic. Is there something to do ans is the gain sufficient to need an effort ?

THat is what strict_elementwise_callableis for : let you define fine grained overload. No need for more.

[jtlap]

ok just that func[if_(mask).else_(eve::zero)] is a little heavy

[DenisYaroshevskiy]

I also think that we need to make the compiler optimize those a bit