yes but this is not 1 bit matmul, it's 1.58 bits with expensive unpacking
replyThe title and the repo uses 1-bit when it means 1.58 bits tertiary values, it doesn't change any of my arguments (still xors and pop_counts).
replyHow do you do ternary matmul with popcnt on 1.58 bit packed data?
replyAssuming 2 bit per values (first bit is sign and second bit is value).
replyactv = A[_:1] & B[_:1]
sign = A[_:0] ^ B[_:0]
dot = pop_count(actv & !sign) - pop_count(actv & sign)
It can probably be made more efficient by taking a column-first format.
Since we are in CPU land, we mostly deal with dot products that match the cache size, I don't assume we have a tiled matmul instruction which is unlikely to support this weird 1-bit format.
Haven't looked closely, but on modern x86 CPUs it might be possible to do much better with the gf2affineqb instructions, which let us do 8x8 bit matrix multiplications efficiently. Not sure how you'd handle the 2-bit part, of course.
reply