Consider using Bencher::bench_refs

[nvzqz]

Thanks for the cool benchmark! Glad to see how folks are using Divan.

I think it might be an improvement to use Bencher::bench_refs instead of Bencher::bench_values because it would:

• Remove the time to deallocate the inputs from benchmarks
• Balance dealloc info with alloc info

[astral4]

Good point! I've updated the benchmark code and results.

Also, thanks for your work on Divan—it's a great library and I look forward to using it in future projects.

[nvzqz]

Thanks for the kind words!

Also, you don't need to black_box inputs passed via bench_refs. 🙂

[smu160]

@nvzqz

Divan is a great library! Thank you for giving me the pre-release demo at RustConf!

Two questions for you:

What is the diffence between bench_local_refs and bench_refs?

Is there a way to get the output in a format that would make it simply to plot results with matplotlib?

[nvzqz]

1. bench_local_refs does not have a Sync requirement. This allows for mutating local variables at the cost of not being able to benchmark across multiple threads. Thanks for asking since I didn't realize I hadn't documented any distinction between them.
2. Not yet. I recommend subscribing to Output format to JSON nvzqz/divan#10 for updates on this.

[smu160]

@nvzqz Thank you for the prompt reply! Does it matter whether I use bench_local_refsor bench_refs if the function being benchmarked is single-threaded?

[nvzqz]

They behave the same in single-threaded benchmarks. I would stick to bench_refs unless you need bench_local_refs.