perf: Speedup writing of Parquet primitive values

[coastalwhite]

This improves the performance writing of primitive arrays into Parquet.

Running the following program, which is made to really test this function.

use std::io::Seek;

use polars::frame::DataFrame;
use polars::prelude::NamedFrom;
use polars::series::Series;
use polars_io::parquet::write::ParquetWriter;
use rand::Rng;

const NUM_VALUES: usize = 10_000_000;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut rng = rand::thread_rng();
    let mut values: Vec<f32> = Vec::with_capacity(NUM_VALUES);

    unsafe { values.set_len(NUM_VALUES) };
    rng.fill(&mut values[..]);

    let values = &values[..];
    let s = Series::new("a", values);

    let mut f = std::fs::OpenOptions::new()
        .write(true)
        .create(true)
        .open("output.parquet")
        .unwrap();

    for _ in 0..100 {
        f.seek(std::io::SeekFrom::Start(0))?;
        let writer = ParquetWriter::new(&mut f)
            .with_compression(polars::prelude::ParquetCompression::Uncompressed);

        let mut df = DataFrame::new(vec![s.clone()]).unwrap();

        writer.finish(&mut df)?;
    }

    Ok(())
}

We run around ~1.8 times faster then before:

Benchmark 1: ./plparbench-before
  Time (mean ± σ):      5.718 s ±  0.804 s    [User: 4.425 s, System: 1.476 s]
  Range (min … max):    4.537 s …  6.458 s    5 runs

Benchmark 2: ./plparbench-after
  Time (mean ± σ):      3.194 s ±  0.034 s    [User: 1.640 s, System: 1.986 s]
  Range (min … max):    3.142 s …  3.228 s    5 runs

Summary
  ./plparbench-after ran
    1.79 ± 0.25 times faster than ./plparbench-before

This also adds specialized implementations for MinMaxKernel::min_max_{ignore, propagate}_nan for the Primitive Arrays. This halves the amount of work needed to calculate the column statistics.

[codecov]

Codecov Report

Attention: Patch coverage is 69.86755% with 91 lines in your changes missing coverage. Please review.

Project coverage is 80.45%. Comparing base (d5265d3) to head (c370891).
Report is 1 commits behind head on main.

Files	Patch %	Lines
crates/polars-compute/src/min_max/simd.rs	67.48%	53 Missing ⚠️
crates/polars-compute/src/min_max/scalar.rs	46.15%	21 Missing ⚠️
.../polars-parquet/src/arrow/write/primitive/basic.rs	84.61%	10 Missing ⚠️
crates/polars-compute/src/min_max/dyn_array.rs	64.70%	6 Missing ⚠️
crates/polars-parquet/src/arrow/write/utils.rs	94.44%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #18020      +/-   ##
==========================================
- Coverage   80.49%   80.45%   -0.05%     
==========================================
  Files        1496     1496              
  Lines      196786   197027     +241     
  Branches     2817     2817              
==========================================
+ Hits       158407   158522     +115     
- Misses      37858    37984     +126     
  Partials      521      521              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ritchie46]

Nice trick with the take_leading_ones/take_leading_zeros alternation. Didn't know we had that. :)