CODEWITHSUNDEEP
multithreadingrust
musako
musako

Reputation: 1267

I want to make rust's for statement multithreaded

I would like to change the following function to parallel processing.

fn cal_v(c: &Array3<f32>, dt: f32) -> Array3<f32> {
  let mut v: Array3<f32> = Array::zeros(c.raw_dim());
  for time in 1..c.shape()[0] {
    let before = c.slice(s![time-1, .., ..]).to_owned();
    let now = c.slice(s![time, .., ..]).to_owned();
    let now_v = (before - now) / dt;
    v.slice_mut(s![time, .., ..]).assign(&now_v);
  }
  v
}

The function c is a 3-dimensional array where the first dimension is the index for time, the second dimension is the index for the particle number, and the third dimension is x,y,z. Personally, I want to learn how to write multi-threaded code, I want to learn how to make it multithreaded, so I want to do the for time in 1..c.shape()[0] part multithreaded. I believe that this process can be parallelized or made asynchronous because there is no dependency in the process at any index of for. What is the best way to write this?

The version of the compiler I'm using is 1.51.0. The version of the library I'm using is ndarray(0.14.0).

Upvotes: 0

Views: 106

Answers (1)

Masklinn
Masklinn

Reputation: 42197

The simplest way to parallelize an iteration is to just use rayon, that's basically its bread and butter. You will have to convert the loop to a more functional style (using for_each for the final loop body), but with that done par_iter() will "magically" distribute the work over its threadpool.

There is an issue I'm not too clear about though because you don't provide any of the code:

v.slice_mut(s![time, .., ..]).assign(&now_v);

is a hard sequential dependency between all the items as-is, and I've no idea what the s macro does, so it may or may not be workable. If each time yields a single slot from v (or at least non-overlapping slots) you would be able to zip the iterators together such that each iteration has a "source" and a "destination" independent from all the others, otherwise you'd have a sequential choke-point in the output, which isn't something rayon supports well.

Upvotes: 1

Related Questions