Performance on short strings about simdutf8 HOT 11 CLOSED

Indeed:

#[inline]. This suggests that the function should be inlined, including across crate boundaries.

from simdutf8.

@lemire That is what I thought but benchmarks proved me otherwise. I will dig up the branch, run the benchmark again and post the comparison here.

from simdutf8.

Note also that you could just fallback on a scalar implementation when the platform is not supported. :-)

from simdutf8.

Note also that you could just fallback on a scalar implementation when the platform is not supported. :-)

That is what it does already. It works similar to simdjson as far as I understand it. The function pointer is initialized to the get_fastest() function in validate_utf8_basic(). On the first invocation it checks what CPU features are supported and then replaces the function pointer to itself with the fastest available implementation for further invocations.

On architectures for which there currently is no SIMD implementation (e.g. ARM right now), the fallback method is compiled in.

from simdutf8.

So the following advice might not be needed...

If there is no native implementation for your platform (yet), use the standard library instead.

from simdutf8.

So the following advice might not be needed...

If there is no native implementation for your platform (yet), use the standard library instead.

Correct, though there is a slight performance penalty with the default LTO setting as Rust does not inline function calls across crate boundaries while it does inline functions calls to the std library.

from simdutf8.

Ah. So that explains why you would see weak performance on short strings.

Any way to lift this 'no inlining' limitation? It seems quite substantial.

from simdutf8.

Compiling with lto=full or lto=thin enables cross-crate inlining. It is just not the default.

I will benchmark delegating the validation of byte sequences shorter than 64 bytes to the std library again with LTO enabled.

from simdutf8.

You might be able to get away with just slapping a #[inline] on the pub fns to hint on inlining across crate boundaries.

from simdutf8.

You might be able to get away with just slapping a #[inline] on the pub fns to hint on inlining across crate boundaries.

You are both right of course, not sure how I missed that. I have added the #[inline] attribute and will benchmark that and calling std::str::from_utf8() for small strings next.

from simdutf8.

Fixed in v0.1.1.

from simdutf8.