Fast reading for primitive types

[KeKsBoTer]

Hello,

I often want to read the whole data from an array (using .into_vec()).
I have noticed that this is considerably slower for large arrays compared to numpy.

Numpy:

This crates takes 6 times as long just for reading:

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let file = File::open("big.npy")?;

    let npy = npyz::NpyFile::new(file)?;
    let mut sum:f32 = 0.;
    let start = std::time::Instant::now();
    let data = npy.into_vec::<f32>()?;

    println!("reading took: {:?}", start.elapsed());
    
    // do something
    println!("length {:?}", data.len()  );
    

    for arr in data {
        sum+=arr;
    }
    println!("{:.4}", sum);
    Ok(())
}

output:

This boils down to the reader reading and parsing every primitive one by one.
In many cases, we can copy the data into memory and reinterpret it (this is also what Numpy does).

I added a fast read functionality for primitive types using the bytemuck crate.
This makes it about 10 times faster:

What are your thoughts about this?
My solution adds minimal code and only speeds up the reads for the primitives where it is safe.

Sorry for the convoluted git history...please squash it on merge in GitHub.

Best
Simon

[ExpHP]

I considered adding block-reading functionality to the crate but concluded I was just reinventing NpyFile<BufReader<File>>. How does this compare to that in benchmarks?