Jan Wassenberg
Jan Wassenberg
This sounds like good progress! It's nice to hear that the memory-backed approach is working well so far :) An M1 still only supports Neon but Amazon's c7g instances are...
> I noticed that running my application with a lane count of several times the hardware vector width did perform better. Good to have another data point. I also find...
Sounds good! > maybe I'm just being paranoid and it comes down to the same thing. Yes indeed, it is intended to be the same and nice to minimize `#if`....
There is a MaxLanes which is constexpr, but I'm not sure we need it. You can skip the power of two check because the Simd template already ensures that, and...
Nice :) > IsInf and IsNaN FYI the code is done but still under review. You could already patch in #645 if you like. > can you recommend the flags...
Interesting, let's see if there is still a difference after the IsNaN. Sure, LoadInterleaved and StoreInterleaved2 are on the todo list. Would you be using them soon-ish? If so, I...
Yeah, the operators do help readability and I'd love to enable them, but that requires more work in clang+GCC+MSVC (for SVE). OK, I'll bump up the priority of LoadInterleaved/StoreInterleaved2 :)...
:) Ah indeed it's u8 only. Sure, can expand to float32 for the moment if that's enough? LoadInterleaved is also coming soon.
Thanks, that is indeed nice to hear. I like your idea of benchmarking via lux, I haven't seen many that are able to compare SIMD wrappers directly. FYI the StoreInterleaved*...
StoreInterleaved for all types has now landed :)