Benchmarking #1729
Benchmarking #1729
Conversation
|
Results from my machine: |
| .unwrap(); | ||
| for file_entry in read_dir { | ||
| match file_entry { | ||
| Ok(entry) => match entry.path().extension() { |
There was a problem hiding this comment.
Couldn't this just be:
Ok(entry) => if entry.path().extension() == extension {
...
}
?
There was a problem hiding this comment.
That would require some shenanigans on going from OSStr to str to be able to compare them
| #[cfg(not(feature = "wgsl-in"))] | ||
| fn gather_modules() -> Vec<naga::Module> { | ||
| Vec::new() | ||
| } |
There was a problem hiding this comment.
It's a little confusing that when necessary features aren't available, we do still run benchmarks and get meaningless numbers. The loop over the modules is inside the b.iter call, not outside it. It would be clearer to somehow just not run benchmarks that require features we don't have.
There was a problem hiding this comment.
that's a bit of an edge case. We don't really care about the numbers when features aren't enabled. What I did was a minimal thing to make it compile for these cases
| .collect::<Vec<_>>() | ||
| }; | ||
| #[cfg(not(feature = "validate"))] | ||
| let inputs = Vec::<(naga::Module, naga::valid::ModuleInfo)>::new(); |
There was a problem hiding this comment.
Again, we shouldn't run benchmarks on empty input lists because a feature was turned off.
There was a problem hiding this comment.
similar reasoning to the last one - this is the minimal change to compile
|
Thanks for the review! Some of the notes are addressed, other left for follow-ups |
Integrates Criterion and also adds support for bincode-encoded IR form. The purpose here is to see the upper bound for shader loading performance. Today, bincode IR deserialization is roughly 7x faster than WGSL parsing.