# Branchless Rust: benchmarks

Companion benchmarks for the article [Branchless Rust](https://www.greyblake.com/blog/branchless-rust/).

The task: filter a slice of one million random `f64` values, keeping the elements
greater than a given threshold. There are three implementations in
[`src/lib.rs`](src/lib.rs):

* `filter_iter` - the idiomatic iterator version
* `filter_prealloc` - a loop with a preallocated output vector
* `filter_branchless` - the branchless version (always write, conditionally advance)

The benchmarks in [`benches/filter.rs`](benches/filter.rs) run every implementation
at different selectivities (1%, 25%, 50%, 75%, 99% of elements kept), plus one extra
experiment: the idiomatic version on the same data sorted and shuffled.

## How to run

```
cargo bench
```

The whole run takes a few minutes. Criterion prints the results to the terminal and
generates an HTML report at `target/criterion/report/index.html`.

## Results

Measured on Intel Core i7-10875H, Rust 1.96.0:

| kept | `filter_iter` | `filter_prealloc` | `filter_branchless` |
|-----:|--------------:|------------------:|--------------------:|
|   1% |       0.59 ms |           0.68 ms |             1.09 ms |
|  25% |       2.69 ms |           2.39 ms |             1.05 ms |
|  50% |       3.94 ms |           3.87 ms |             1.03 ms |
|  75% |       2.75 ms |           2.72 ms |             1.02 ms |
|  99% |       1.49 ms |           1.23 ms |             1.11 ms |

Sorted vs shuffled input (`filter_iter`, 50% kept):

| input    |    time |
|----------|--------:|
| shuffled | 4.15 ms |
| sorted   | 0.93 ms |

## License

MIT