Rust: allow multiple threads to modify an image (wrapper of a vector)?

Your two questions are actually unrelated.

First the #2 that is easier:

The idea of the Rayon scoped threads is that the threads created inside cannot outlive the scope, so any variable created outside the scope can be safely borrowed and its references sent into the threads. But your variables are created inside the scope, and that buys you nothing.

The solution is easy: move the variables out of the scope:

    let mut v1 = RawImageView::new(&mut img, 0, 0, 10, 10);
    let mut v2 = RawImageView::new(&mut img, 10, 0, 10, 10);
    pool.scope(|s| {
        s.spawn(|_| {
            modify(&mut v1);
        });
        s.spawn(|_| {
            modify(&mut v2);
        });
    });

The #1 is trickier, and you have to go unsafe (or find a crate that does it for you but I found none). My idea is to store a raw pointer instead of a vector and then use std::ptr::write to write the pixels. If you do it carefully and add your own bounds checks it should be perfectly safe.

I'll add an additional level of indirection, probably you could do it with just two but this will keep more of your original code.

The RawImage could be something like:

pub struct RawImage<'a>
{
    _pd: PhantomData<&'a mut Color>,
    data: *mut Color,
    width: u32,
    height: u32,
}
impl<'a> RawImage<'a>
{
    pub fn new(data: &'a mut [Color], width: u32, height: u32) -> Self
    {
        Self {
            _pd: PhantomData,
            data: data.as_mut_ptr(),
            width: width,
            height: height
        }
    }
}

And then build the image keeping the pixels outside:

    let mut pixels = vec![[0.0, 0.0, 0.0]; (20 * 10) as usize];
    let mut img = RawImage::new(&mut pixels, 20, 10);

Now the RawImageView can keep a non-mutable reference to the RawImage:

pub struct RawImageView<'a>
{
    img: &'a RawImage<'a>,
    offset_x: u32,
    offset_y: u32,
    width: u32,
    height: u32,
}

And use ptr::write to write the pixels:

    pub fn set_pixel(&mut self, x: u32, y: u32, color: Color)
    {
        let index = self.img.xy2index(x + self.offset_x, y + self.offset_y);
        //TODO! missing check bounds
        unsafe { self.img.data.add(index).write(color) };
    }

But do not forget to either do check bounds here or mark this function as unsafe, sending the responsibility to the user.

Naturally, since your function keeps a reference to a mutable pointer, it cannot be send between threads. But we know better:

unsafe impl Send for RawImageView<'_> {}

And that's it! Playground. I think this solution is memory-safe, as long as you add code to enforce that your views do not overlap and that you do not go out of bounds of each view.

This does not exactly match your image problem but this might give you some clues.

The idea is that chunks_mut() considers a whole mutable slice as many independent (non-overlapping) mutable sub-slices. Thus, each mutable sub-slice can be used by a thread without considering that the whole slice is mutably borrowed by many threads (it is actually, but in a non-overlapping manner, so it is sound).

Of course this example is trivial, and it should be trickier to divide an image in many arbitrary non-overlapping areas.

fn modify(
    id: usize,
    values: &mut [usize],
) {
    for v in values.iter_mut() {
        *v += 1000 * (id + 1);
    }
}

fn main() {
    let mut values: Vec<_> = (0..8_usize).map(|i| i + 1).collect();

    let pool = rayon::ThreadPoolBuilder::new()
        .num_threads(2)
        .build()
        .unwrap();
    pool.scope(|s| {
        for (id, ch) in values.chunks_mut(4).enumerate() {
            s.spawn(move |_| {
                modify(id, ch);
            });
        }
    });

    println!("{:?}", values);
}

Edit

I don't know the context in which this parallel work on some parts of an image is needed but I can imagine two situations.

If the intent is to work on some arbitrary parts of the image and allow this to take place in several threads that compute many other things, then a simple mutex to regulate the access to the global image is certainly enough. Indeed, if the precise shape of each part of the image is very important, then it is very unlikely that there exist so many of them that parallelisation can be beneficial.

On the other hand, if the intent is to parallelize the image processing in order to achieve high performance, then the specific shape of each part is probably not so relevant, since the only important guaranty to ensure is that the whole image is processed when all the threads are done. In this case a simple 1-D splitting (along y) is enough.

As an example, below is a minimal adaptation of the original code in order to make the image split itself into several mutable parts that can be safely handled by many threads. No unsafe code is needed, no expensive runtime checks nor copies are performed, the parts are contiguous so highly optimisable.

type Color = [f64; 3];

pub struct RawImage {
    data: Vec<Color>,
    width: u32,
    height: u32,
}

impl RawImage {
    pub fn new(
        width: u32,
        height: u32,
    ) -> Self {
        Self {
            data: vec![[0.0, 0.0, 0.0]; (width * height) as usize],
            width: width,
            height: height,
        }
    }

    fn xy2index(
        &self,
        x: u32,
        y: u32,
    ) -> usize {
        (y * self.width + x) as usize
    }

    pub fn mut_parts(
        &mut self,
        count: u32,
    ) -> impl Iterator<Item = RawImagePart> {
        let part_height = (self.height + count - 1) / count;
        let sz = part_height * self.width;
        let width = self.width;
        let mut offset_y = 0;
        self.data.chunks_mut(sz as usize).map(move |part| {
            let height = part.len() as u32 / width;
            let p = RawImagePart {
                part,
                offset_y,
                width,
                height,
            };
            offset_y += height;
            p
        })
    }
}

pub struct RawImagePart<'a> {
    part: &'a mut [Color],
    offset_y: u32,
    width: u32,
    height: u32,
}

impl<'a> RawImagePart<'a> {
    pub fn set_pixel(
        &mut self,
        x: u32,
        y: u32,
        color: Color,
    ) {
        let part_index = x + y * self.width;
        self.part[part_index as usize] = color;
    }
}

fn modify(img: &mut RawImagePart) {
    // Do some heavy calculation and write to the image.
    let dummy = img.offset_y as f64 / 100.0;
    let last = img.height - 1;
    for x in 0..img.width {
        img.set_pixel(x, 0, [dummy + 0.1, dummy + 0.2, dummy + 0.3]);
        img.set_pixel(x, last, [dummy + 0.7, dummy + 0.8, dummy + 0.9]);
    }
}

fn main() {
    let mut img = RawImage::new(20, 10);
    let pool = rayon::ThreadPoolBuilder::new()
        .num_threads(2)
        .build()
        .unwrap();
    pool.scope(|s| {
        for mut p in img.mut_parts(2) {
            s.spawn(move |_| {
                modify(&mut p);
            });
        }
    });
    for y in 0..img.height {
        let offset = (y * img.width) as usize;
        println!("{:.2?}...", &img.data[offset..offset + 3]);
    }
}