search

AdamNiederer / faster

SIMD for humans
Mozilla Public License 2.0
1.56k stars 51 forks source link

PMOVMSKB operation equivalent #26

Open deepinthebuild opened 6 years ago

deepinthebuild commented 6 years ago

It would be useful to me if there were an operation like fn bytemask<T: Packed>(v: T) -> usize that set the bits of the usize to 1 where the packed vector's bytes are all 1, corresponding to the asm PMOVMSKB or VPMOVMSKB. I don't know how possible this would be for faster, but it would be nice.

AdamNiederer commented 6 years ago

That would definitely be feasible, but I'm wondering if there's a way to make this portable for all vector lengths, though. Perhaps a bit_vec::BitVec would be a more suitable return type?

deepinthebuild commented 6 years ago

a u64 is large enough for all existing vector types, including AVX-512. 512 bits is only u8x64, after all.

To be clear about what I'm after, I want an operation that takes a vector like: u8x8(255, 255, 0, 255, 255, 0, 0, 0) and returns a u64: 0b00011011 where the nth least significant bit of the output is set to 1 if the nth byte of the input is 255.