Feature: Heuristics for choosing a SIMD variant over the corresponding scalar non-SIMD version

[nordlow]

Describe what you are looking for

For a given function f, are there any (typically architecture-specific) heuristics for when a SIMD version opposite to a serial version of an algorithm should be chosen?

For instance, for

sz_find(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length) {

one such heuristics could be

bool useSIMD(size_t h_length, size_t n_length) {
    if (h_length > 1024) {
        return true;
    } else if (h_length / n_length >= 4) {
        return true;
    }
}

.

Can you contribute to the implementation?

• I can contribute

Is your feature request specific to a certain interface?

It applies to everything

Contact Details

No response

Is there an existing issue for this?

• I have searched the existing issues

Code of Conduct

• I agree to follow this project's Code of Conduct

[ashvardanian]

There are definitely spots where platform-specific performance tuning can be a great idea. This can be a topic for a proper research project, potentially exposing threshold variables with extern and tuning our sz_find, sz_copy, and other kernels for the target platform. It get's a bit trickier, if the same kernel depends on multiple variables and we are forced to grid-search...

Are you familiar with similar projects, @nordlow?

[nordlow]

Unfortunately not.

Thanks for the response anyway.