Hacker News Reader

Comment by sorenjan | Hacker News Reader

How do you handle arrays of values where the array lengths are not a multiple of 8 in this kind of vectorized code? Do you zero pad them before handling them to the vectorized function, or do you run a second loop element by element on the remaining elements after the main one? What happens if you try to do `_mm256_load_ps(&ys[i])` with < 8 elements remaining?

View on HN

You could pad the input as you said, which avoids a "tail loop," but otherwise you usually do a partial load (load <8 elements into a vector) and store. Some instruction set extensions provide "masked load/store" instructions for this, but there are ways to do it without those too.

To your last question specifically, if you _mm256_load_ps(&ys[i]) and you're at the edge of a page, you'll get a segfault. Otherwise you'll get undefined values (which can be okay, you could ignore them instead of dealing with a partial load).

RISC-V has a solution to this which seems quite elegant - the vector length is set by the program, up to some maximum, and for the final block, it is simply whatever is left:

https://gms.tf/riscv-vector.html

> Do you zero pad them before handling them to the vectorized function, or do you run a second loop element by element on the remaining elements after the main one?

Typically the latter. That's why I find SVEs so interesting. Should improve code density.

Typically you have have the vectorized loop followed by a non-vectorized loop handling the remaining items, or even just explicit if-then-else statements for each possible number of remaining items.

Reply on news.ycombinator.com