pub fn split<D, S>(data: D, splitter: S) -> Split<D, S>where
D: Send,
S: Fn(D) -> (D, Option<D>) + Sync,
Expand description
The split
function takes arbitrary data and a closure that knows how to
split it, and turns this into a ParallelIterator
.
Examples
As a simple example, Rayon can recursively split ranges of indices
use rayon::iter;
use rayon::prelude::*;
use std::ops::Range;
// We define a range of indices as follows
type Range1D = Range<usize>;
// Splitting it in two can be done like this
fn split_range1(r: Range1D) -> (Range1D, Option<Range1D>) {
// We are mathematically unable to split the range if there is only
// one point inside of it, but we could stop splitting before that.
if r.end - r.start <= 1 { return (r, None); }
// Here, our range is considered large enough to be splittable
let midpoint = r.start + (r.end - r.start) / 2;
(r.start..midpoint, Some(midpoint..r.end))
}
// By using iter::split, Rayon will split the range until it has enough work
// to feed the CPU cores, then give us the resulting sub-ranges
iter::split(0..4096, split_range1).for_each(|sub_range| {
// As our initial range had a power-of-two size, the final sub-ranges
// should have power-of-two sizes too
assert!((sub_range.end - sub_range.start).is_power_of_two());
});
This recursive splitting can be extended to two or three dimensions, to reproduce a classic “block-wise” parallelization scheme of graphics and numerical simulations:
// A two-dimensional range of indices can be built out of two 1D ones
struct Range2D {
// Range of horizontal indices
pub rx: Range1D,
// Range of vertical indices
pub ry: Range1D,
}
// We want to recursively split them by the largest dimension until we have
// enough sub-ranges to feed our mighty multi-core CPU. This function
// carries out one such split.
fn split_range2(r2: Range2D) -> (Range2D, Option<Range2D>) {
// Decide on which axis (horizontal/vertical) the range should be split
let width = r2.rx.end - r2.rx.start;
let height = r2.ry.end - r2.ry.start;
if width >= height {
// This is a wide range, split it on the horizontal axis
let (split_rx, ry) = (split_range1(r2.rx), r2.ry);
let out1 = Range2D {
rx: split_rx.0,
ry: ry.clone(),
};
let out2 = split_rx.1.map(|rx| Range2D { rx, ry });
(out1, out2)
} else {
// This is a tall range, split it on the vertical axis
let (rx, split_ry) = (r2.rx, split_range1(r2.ry));
let out1 = Range2D {
rx: rx.clone(),
ry: split_ry.0,
};
let out2 = split_ry.1.map(|ry| Range2D { rx, ry, });
(out1, out2)
}
}
// Again, rayon can handle the recursive splitting for us
let range = Range2D { rx: 0..800, ry: 0..600 };
iter::split(range, split_range2).for_each(|sub_range| {
// If the sub-ranges were indeed split by the largest dimension, then
// if no dimension was twice larger than the other initially, this
// property will remain true in the final sub-ranges.
let width = sub_range.rx.end - sub_range.rx.start;
let height = sub_range.ry.end - sub_range.ry.start;
assert!((width / 2 <= height) && (height / 2 <= width));
});