From SIMD Wrappers to SIMD Ranges (pt 2).

Joel Falcou, Denis Yaroshevskiy

This talk (pt2)

• TODO: link to pt2 slides
• Some plan thingy here

No Magic Compiler for this talk

• let's not assume that the compilers do much more than today
• inlining/ collapsing instructions with exactaly the same semantics
• std::unseq is not really a thing (yet?)
• there is some code in libc++ - that I coudn't do anything with

Elephant in the ROOM (is this too long? I have no idea)

• _this is many slides_
• SVE/RVV don't know the compile time register size
• Mention streaming extensions
• show how this affects find from part 1 (not much)
• show how this affects inclusive_scan from part 1, it's tricky
• Why std::simd doesn't support this (super new)
• Why eve doesn't support this (the sizeless_struct that is no more)
• How google highway manages to do it by splitting meta information and intrinsics
• If a major compiler gives us a sizeless struct, we promise to experiment
• DLLs approach
• One binary approach that is probably not modules friendly
• Hey standard ppl, can we compile multiple versions of the same function in std c++ please?

Importance of extensibility

• GF2P8AFFINEQB is "Compute an affine transformation in the Galois Field 2^8"
• Harold Aptroot: Histogramming bytes with positional popcount (GF2P8AFFINEQB edition)
• eve is implictly convertible
• std::simd supports std:bit_cast

Importance of tuning algorithms

• People will want to try things
• eve exposes unrolling/aligning/cardinal changes

Importance of knowing parameter limits

• How limiting sin inputs affects generated code
• Probably `__builtin_assume` won't be enough