flashraw-core 0.1.1

//! Parallel execution helpers that degrade to sequential execution when the
//! `rayon` feature is disabled (e.g. single-threaded WASM) or parallelism is off.

/// Minimum number of samples before splitting work across threads pays off.
pub const PAR_MIN_SAMPLES: usize = 1 << 18;

/// Run `f(first_row, rows_chunk)` over `data` split into chunks of `rows_per_chunk`
/// rows of `row_len` elements each.
pub fn for_each_row_chunk<T, F>(data: &mut [T], row_len: usize, rows_per_chunk: usize, parallel: bool, f: F)
where
    T: Send,
    F: Fn(usize, &mut [T]) + Sync + Send,
{
    if row_len == 0 || data.is_empty() {
        return;
    }
    let chunk = row_len * rows_per_chunk.max(1);
    #[cfg(feature = "rayon")]
    if parallel && data.len() >= PAR_MIN_SAMPLES && threads() > 1 {
        use rayon::prelude::*;
        data.par_chunks_mut(chunk).enumerate().for_each(|(i, c)| f(i * rows_per_chunk.max(1), c));
        return;
    }
    let _ = parallel;
    for (i, c) in data.chunks_mut(chunk).enumerate() {
        f(i * rows_per_chunk.max(1), c);
    }
}

/// Workers worth using for memory-bound kernels (copies, byte swaps, packed unpack,
/// lookup curves). Measured on a 16-thread x86 server: such loops peak at ~4 threads and
/// get 1.7-1.9x slower at 16 from memory-controller and page-fault contention. Machines
/// with more bandwidth per core scale further, so the cap grows with the thread count.
pub fn memory_bound_workers() -> usize {
    (threads() / 4).clamp(2, 8)
}

/// Like [`for_each_row_chunk`] but split into at most `max_workers` tasks (used for
/// memory-bound kernels where more threads only add contention).
pub fn for_each_row_chunk_bounded<T, F>(data: &mut [T], row_len: usize, max_workers: usize, parallel: bool, f: F)
where
    T: Send,
    F: Fn(usize, &mut [T]) + Sync + Send,
{
    if row_len == 0 || data.is_empty() {
        return;
    }
    let rows = data.len().div_ceil(row_len);
    let workers = max_workers.clamp(1, threads().max(1));
    for_each_row_chunk(data, row_len, rows.div_ceil(workers), parallel && workers > 1, f)
}

/// Map `f` over `items` in parallel (when enabled), preserving order.
pub fn map<I, O, F>(items: Vec<I>, parallel: bool, f: F) -> Vec<O>
where
    I: Send,
    O: Send,
    F: Fn(I) -> O + Sync + Send,
{
    #[cfg(feature = "rayon")]
    if parallel && items.len() > 1 && threads() > 1 {
        use rayon::prelude::*;
        return items.into_par_iter().map(f).collect();
    }
    let _ = parallel;
    items.into_iter().map(f).collect()
}

/// Run `f(i)` for `i in 0..n` in parallel when enabled.
pub fn for_each_index<F>(n: usize, parallel: bool, f: F)
where
    F: Fn(usize) + Sync + Send,
{
    #[cfg(feature = "rayon")]
    if parallel && n > 1 && threads() > 1 {
        use rayon::prelude::*;
        (0..n).into_par_iter().for_each(f);
        return;
    }
    let _ = parallel;
    (0..n).for_each(f);
}

#[cfg(feature = "rayon")]
std::thread_local! {
    /// Set while [`sequential`] runs on this thread.
    static SEQUENTIAL: std::cell::Cell<bool> = const { std::cell::Cell::new(false) };
}

/// Run `f` on the calling thread with every parallel helper sequential: inside `f`,
/// [`threads`] returns 1 and no work is handed to a thread pool. This is how a
/// single-threaded decode runs, so it stays on the caller's thread (and on the CPU core
/// class the operating system picked for it) instead of moving to a pool worker.
pub fn sequential<R>(f: impl FnOnce() -> R) -> R {
    #[cfg(feature = "rayon")]
    {
        struct Restore(bool);
        impl Drop for Restore {
            fn drop(&mut self) {
                SEQUENTIAL.with(|s| s.set(self.0));
            }
        }
        let _restore = Restore(SEQUENTIAL.with(|s| s.replace(true)));
        f()
    }
    #[cfg(not(feature = "rayon"))]
    {
        f()
    }
}

/// Number of worker threads available to parallel helpers (1 inside [`sequential`]).
pub fn threads() -> usize {
    #[cfg(feature = "rayon")]
    {
        if SEQUENTIAL.with(|s| s.get()) { 1 } else { rayon::current_num_threads() }
    }
    #[cfg(not(feature = "rayon"))]
    {
        1
    }
}

/// Geometry of a [`wavefront`] pass over a row-major buffer.
#[derive(Clone, Copy, Debug)]
pub struct Wavefront {
    /// Samples per row.
    pub row_len: usize,
    /// Rows above a row that it reads (the band above must be at least this tall).
    pub halo: usize,
    /// Columns left of a block that its first samples read from rows above.
    pub margin: usize,
    /// Block width in samples (block boundaries are multiples of it).
    pub block: usize,
}

/// Process a buffer whose sample `(r, x)` depends only on samples left of it in row `r`
/// and on columns `<= x` of the `halo` rows above it. Rows are cut into bands (one per
/// thread) and columns into blocks; block `j` of a band runs once block `j - 1` of the
/// same band and block `j` of the band above are done, so the bands advance as a
/// wavefront without waiting for each other row by row.
///
/// `f(first_row, band, above, lo, a, b)` must process columns `a..b` of every row of
/// `band` (whole rows, `first_row` being the index of its first row in `data`) in row
/// order. `above` holds the rows directly above the band (up to `halo`, oldest first),
/// restricted to columns `lo..b` (`lo = a - margin`, clamped at 0); for the first band
/// they come from `top` (which may hold fewer than `halo` rows, or none).
#[allow(clippy::type_complexity)]
pub fn wavefront<T, F>(data: &mut [T], top: &[T], geo: Wavefront, parallel: bool, f: F)
where
    T: Copy + Default + Send + Sync,
    F: Fn(usize, &mut [T], &[T], usize, usize, usize) + Sync,
{
    use std::sync::Mutex;
    use std::sync::atomic::{AtomicU8, Ordering};

    let Wavefront { row_len: rl, halo, block, .. } = geo;
    if rl == 0 || block == 0 || data.len() < rl {
        return;
    }
    let rows = data.len() / rl;
    let halo = halo.max(1);
    let workers = if parallel { threads() } else { 1 };
    let nbands = workers.min(rows / halo).max(1);
    let nblocks = rl.div_ceil(block);
    let bounds: Vec<(usize, usize)> =
        (0..nbands).map(|k| (rows * k / nbands * rl, rows * (k + 1) / nbands * rl)).collect();
    if nbands == 1 {
        // One band: whole rows, no blocks.
        let top_rows = (top.len() / rl).min(halo);
        f(0, &mut data[..rows * rl], &top[top.len() - top_rows * rl..], 0, 0, rl);
        return;
    }
    let Some(parts) = split_disjoint(&mut data[..rows * rl], &bounds) else { return };

    struct Ctx<'a, T, F> {
        /// A band is used by one task at a time (its blocks run in order); the locks only
        /// make that visible to the compiler and are never contended.
        bands: Vec<Mutex<&'a mut [T]>>,
        first_rows: Vec<usize>,
        /// The rows above each band, filled block by block as the band above finishes.
        edges: Vec<Mutex<Vec<T>>>,
        /// Unfinished dependencies of each block (`band * nblocks + block`).
        pending: Vec<AtomicU8>,
        nblocks: usize,
        geo: Wavefront,
        halo: usize,
        f: F,
    }

    impl<T: Copy + Default + Send + Sync, F: Fn(usize, &mut [T], &[T], usize, usize, usize) + Sync> Ctx<'_, T, F> {
        /// Run block `j` of band `k`; returns the blocks it made ready.
        fn block(&self, k: usize, j: usize) -> impl Iterator<Item = (usize, usize)> + '_ {
            let (rl, block, halo) = (self.geo.row_len, self.geo.block, self.halo);
            let (a, b) = (j * block, ((j + 1) * block).min(rl));
            let lo = a.saturating_sub(self.geo.margin);
            let above: Vec<T> = {
                let e = self.edges[k].lock().unwrap_or_else(|e| e.into_inner());
                e.chunks_exact(rl).flat_map(|row| row[lo..b].iter().copied()).collect()
            };
            {
                let mut band = self.bands[k].lock().unwrap_or_else(|e| e.into_inner());
                (self.f)(self.first_rows[k], &mut band, &above, lo, a, b);
                if let Some(next) = self.edges.get(k + 1) {
                    let mut e = next.lock().unwrap_or_else(|e| e.into_inner());
                    let n = band.len() / rl;
                    for (h, dst) in e.chunks_exact_mut(rl).enumerate() {
                        let src = &band[(n - halo + h) * rl..][..rl];
                        dst[lo..b].copy_from_slice(&src[lo..b]);
                    }
                }
            }
            // The last dependency to finish starts a block; AcqRel orders this block's
            // writes before it (the locks order the data as well).
            [(k + 1, j), (k, j + 1)].into_iter().filter(move |&(k2, j2)| {
                k2 < self.bands.len()
                    && j2 < self.nblocks
                    && self.pending[k2 * self.nblocks + j2].fetch_sub(1, Ordering::AcqRel) == 1
            })
        }
    }

    let first_rows: Vec<usize> = bounds.iter().map(|&(s, _)| s / rl).collect();
    let top_rows = (top.len() / rl).min(halo);
    let edges = (0..nbands)
        .map(|k| {
            Mutex::new(if k == 0 { top[top.len() - top_rows * rl..].to_vec() } else { vec![T::default(); halo * rl] })
        })
        .collect();
    let pending =
        (0..nbands * nblocks).map(|i| AtomicU8::new((i / nblocks > 0) as u8 + (i % nblocks > 0) as u8)).collect();
    let ctx =
        Ctx { bands: parts.into_iter().map(Mutex::new).collect(), first_rows, edges, pending, nblocks, geo, halo, f };

    #[cfg(feature = "rayon")]
    if nbands > 1 && threads() > 1 {
        fn run<'s, T, F>(s: &rayon::Scope<'s>, ctx: &'s Ctx<'s, T, F>, k: usize, j: usize)
        where
            T: Copy + Default + Send + Sync,
            F: Fn(usize, &mut [T], &[T], usize, usize, usize) + Sync,
        {
            for (k2, j2) in ctx.block(k, j) {
                s.spawn(move |s| run(s, ctx, k2, j2));
            }
        }
        rayon::scope(|s| run(s, &ctx, 0, 0));
        return;
    }
    // In order, band by band: every block's dependencies are done before it runs.
    for k in 0..nbands {
        for j in 0..nblocks {
            ctx.block(k, j).for_each(drop);
        }
    }
}

/// Split a mutable slice into disjoint sub-slices at the given ascending offsets.
/// `bounds` are `(start, end)` pairs that must not overlap and be sorted. Returns
/// `None` if they are not valid for `data`.
pub fn split_disjoint<'a, T>(mut data: &'a mut [T], bounds: &[(usize, usize)]) -> Option<Vec<&'a mut [T]>> {
    let mut out = Vec::with_capacity(bounds.len());
    let mut consumed = 0usize;
    for &(s, e) in bounds {
        if s < consumed || e < s {
            return None;
        }
        let rest = std::mem::take(&mut data);
        if e - consumed > rest.len() {
            return None;
        }
        let (_, tail) = rest.split_at_mut(s - consumed);
        let (mid, tail) = tail.split_at_mut(e - s);
        out.push(mid);
        data = tail;
        consumed = e;
    }
    Some(out)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn sequential_runs_everything_on_the_calling_thread() {
        let me = std::thread::current().id();
        let outer = threads();
        let r = sequential(|| {
            assert_eq!(threads(), 1);
            let ids = map((0..64).collect::<Vec<_>>(), true, |_| std::thread::current().id());
            assert!(ids.iter().all(|&i| i == me));
            for_each_index(64, true, |_| assert_eq!(std::thread::current().id(), me));
            let mut v = vec![0u8; PAR_MIN_SAMPLES * 2];
            for_each_row_chunk(&mut v, 1024, 1, true, |_, _| assert_eq!(std::thread::current().id(), me));
            // Nested calls keep the setting and restore it on the way out.
            sequential(|| assert_eq!(threads(), 1));
            assert_eq!(threads(), 1);
            7
        });
        assert_eq!(r, 7);
        assert_eq!(threads(), outer);
    }

    #[test]
    fn row_chunks_cover_everything() {
        let mut v = vec![0u32; 1 << 19];
        for_each_row_chunk(&mut v, 1000, 7, true, |first, c| {
            for (i, x) in c.iter_mut().enumerate() {
                *x = (first * 1000 + i) as u32;
            }
        });
        assert!(v.iter().enumerate().all(|(i, &x)| x == i as u32));
    }

    /// `v[r][x] += v[r][x-1] + 3 v[r-1][x] + 5 v[r-1][x-1] + 7 v[r-2][x]` (missing terms 0).
    fn recurrence(get: impl Fn(isize, isize) -> u32, own: u32) -> u32 {
        own.wrapping_add(get(0, -1))
            .wrapping_add(get(-1, 0).wrapping_mul(3))
            .wrapping_add(get(-1, -1).wrapping_mul(5))
            .wrapping_add(get(-2, 0).wrapping_mul(7))
    }

    #[test]
    fn wavefront_matches_row_order() {
        for (rl, rows, block) in [(1000usize, 97usize, 64usize), (130, 9, 16), (64, 3, 64), (50, 40, 7)] {
            let seed: Vec<u32> = (0..rl * rows).map(|i| (i as u32).wrapping_mul(2654435761) >> 7).collect();
            let mut want = seed.clone();
            for r in 0..rows {
                for x in 0..rl {
                    let v = recurrence(
                        |dr, dx| {
                            let (rr, xx) = (r as isize + dr, x as isize + dx);
                            if rr < 0 || xx < 0 { 0 } else { want[rr as usize * rl + xx as usize] }
                        },
                        want[r * rl + x],
                    );
                    want[r * rl + x] = v;
                }
            }
            for parallel in [false, true] {
                let mut got = seed.clone();
                let geo = Wavefront { row_len: rl, halo: 2, margin: 1, block };
                wavefront(&mut got, &[], geo, parallel, |first, band, above, lo, a, b| {
                    let w = b - lo;
                    let na = above.len() / w;
                    for i in 0..band.len() / rl {
                        for x in a..b {
                            let get = |dr: isize, dx: isize| {
                                let xx = x as isize + dx;
                                if xx < 0 || (first + i) as isize + dr < 0 {
                                    return 0;
                                }
                                let xx = xx as usize;
                                let ri = i as isize + dr;
                                if ri >= 0 {
                                    band[ri as usize * rl + xx]
                                } else {
                                    above[(na as isize + ri) as usize * w + xx - lo]
                                }
                            };
                            let v = recurrence(get, band[i * rl + x]);
                            band[i * rl + x] = v;
                        }
                    }
                });
                assert_eq!(got, want, "{rl}x{rows} block {block} parallel {parallel}");
            }
        }
    }

    #[test]
    fn disjoint_split() {
        let mut v: Vec<u8> = (0..10).collect();
        let parts = split_disjoint(&mut v, &[(1, 3), (5, 6), (6, 10)]).unwrap();
        assert_eq!(parts[0], &[1, 2]);
        assert_eq!(parts[1], &[5]);
        assert_eq!(parts[2], &[6, 7, 8, 9]);
        assert!(split_disjoint(&mut v, &[(3, 5), (4, 6)]).is_none());
        assert!(split_disjoint(&mut v, &[(3, 50)]).is_none());
    }
}

/// Concurrent writer for a buffer partitioned into a regular tile grid.
///
/// Tiles of a grid share buffer rows, so `chunks_mut` cannot hand them out to
/// threads. `GridWriter` lets each thread [`claim`](GridWriter::claim) a tile and
/// write its rows directly into the final image. Soundness rests on two checked
/// invariants: (1) every claim is recorded in an atomic flag, so a tile can be
/// claimed at most once; (2) grid tiles cover pairwise-disjoint rectangles, and
/// every row slice handed out is clipped to its tile, so no two claims alias.
///
/// The buffer may cover a sub-rectangle (`origin`, `width x height`) of the grid's
/// coordinate space, which is how region decoding writes only intersecting tiles.
pub struct GridWriter<'a, T> {
    ptr: *mut T,
    len: usize,
    width: u32,
    height: u32,
    cpp: usize,
    origin: (u32, u32),
    tile_w: u32,
    tile_h: u32,
    tiles_x: u32,
    tiles_y: u32,
    claimed: Vec<std::sync::atomic::AtomicBool>,
    _marker: std::marker::PhantomData<&'a mut [T]>,
}

// SAFETY: access to the underlying buffer is partitioned by claims (see type docs);
// `T: Send` values may be written from any thread.
unsafe impl<T: Send> Send for GridWriter<'_, T> {}
// SAFETY: shared access only hands out disjoint, exclusively-claimed tiles.
unsafe impl<T: Send> Sync for GridWriter<'_, T> {}

/// Exclusive access to one claimed tile's area within the buffer.
pub struct TileMut<'w, 'a, T> {
    writer: &'w GridWriter<'a, T>,
    /// Clipped area in buffer coordinates.
    x0: usize,
    y0: usize,
    w: usize,
    h: usize,
    /// Offset of the clipped area relative to the tile's own origin.
    pub skip_x: u32,
    pub skip_y: u32,
}

impl<'a, T: Send> GridWriter<'a, T> {
    /// `buf` holds `width x height` pixels of `cpp` samples, covering grid coordinates
    /// starting at `origin`. The grid has `tile_w x tile_h` tiles anchored at (0, 0) and
    /// spans `grid_w x grid_h` pixels.
    #[allow(clippy::too_many_arguments)]
    pub fn new(
        buf: &'a mut [T],
        width: u32,
        height: u32,
        cpp: u32,
        origin: (u32, u32),
        tile_w: u32,
        tile_h: u32,
        grid_w: u32,
        grid_h: u32,
    ) -> Option<Self> {
        if tile_w == 0 || tile_h == 0 || buf.len() < width as usize * height as usize * cpp as usize {
            return None;
        }
        let tiles_x = grid_w.div_ceil(tile_w);
        let tiles_y = grid_h.div_ceil(tile_h);
        let n = tiles_x as usize * tiles_y as usize;
        Some(GridWriter {
            ptr: buf.as_mut_ptr(),
            len: buf.len(),
            width,
            height,
            cpp: cpp as usize,
            origin,
            tile_w,
            tile_h,
            tiles_x,
            tiles_y,
            claimed: (0..n).map(|_| std::sync::atomic::AtomicBool::new(false)).collect(),
            _marker: std::marker::PhantomData,
        })
    }

    pub fn tiles_x(&self) -> u32 {
        self.tiles_x
    }
    pub fn tiles_y(&self) -> u32 {
        self.tiles_y
    }

    /// Claim tile `(tx, ty)`. Returns `None` if out of range, already claimed, or if
    /// the tile does not intersect the buffer area.
    pub fn claim(&self, tx: u32, ty: u32) -> Option<TileMut<'_, 'a, T>> {
        if tx >= self.tiles_x || ty >= self.tiles_y {
            return None;
        }
        let idx = ty as usize * self.tiles_x as usize + tx as usize;
        if self.claimed[idx].swap(true, std::sync::atomic::Ordering::AcqRel) {
            return None;
        }
        // Tile rect in grid coordinates, intersected with the buffer area.
        let (gx, gy) = (tx as u64 * self.tile_w as u64, ty as u64 * self.tile_h as u64);
        let (ox, oy) = (self.origin.0 as u64, self.origin.1 as u64);
        let l = gx.max(ox);
        let t = gy.max(oy);
        let r = (gx + self.tile_w as u64).min(ox + self.width as u64);
        let b = (gy + self.tile_h as u64).min(oy + self.height as u64);
        if r <= l || b <= t {
            return None;
        }
        Some(TileMut {
            writer: self,
            x0: (l - ox) as usize,
            y0: (t - oy) as usize,
            w: (r - l) as usize,
            h: (b - t) as usize,
            skip_x: (l - gx) as u32,
            skip_y: (t - gy) as u32,
        })
    }
}

impl<T> TileMut<'_, '_, T> {
    /// Width (pixels) and height of the clipped area.
    pub fn size(&self) -> (usize, usize) {
        (self.w, self.h)
    }
    /// Mutable samples of row `y` of the clipped area (`w * cpp` samples).
    pub fn row(&mut self, y: usize) -> &mut [T] {
        assert!(y < self.h, "tile row out of range");
        let wr = self.writer;
        let start = ((self.y0 + y) * wr.width as usize + self.x0) * wr.cpp;
        let n = self.w * wr.cpp;
        assert!(start + n <= wr.len);
        // SAFETY: `start..start+n` lies within the buffer (asserted) and within this
        // tile's clipped rectangle, which is disjoint from every other tile's; the tile
        // was claimed exactly once (atomic flag) and `&mut self` prevents aliasing
        // between rows handed out by this same TileMut.
        unsafe { std::slice::from_raw_parts_mut(wr.ptr.add(start), n) }
    }
}

#[cfg(test)]
mod grid_tests {
    use super::*;

    #[test]
    fn grid_writes_and_single_claim() {
        let mut buf = vec![0u32; 10 * 7];
        {
            let g = GridWriter::new(&mut buf, 10, 7, 1, (0, 0), 4, 3, 10, 7).unwrap();
            assert_eq!((g.tiles_x(), g.tiles_y()), (3, 3));
            for_each_index(9, true, |i| {
                let (tx, ty) = (i as u32 % 3, i as u32 / 3);
                let mut t = g.claim(tx, ty).unwrap();
                let (_, h) = t.size();
                for y in 0..h {
                    for v in t.row(y).iter_mut() {
                        *v = i as u32 + 1;
                    }
                }
            });
            assert!(g.claim(0, 0).is_none());
        }
        for y in 0..7 {
            for x in 0..10 {
                assert_eq!(buf[y * 10 + x], (y / 3 * 3 + x / 4) as u32 + 1);
            }
        }
    }

    #[test]
    fn region_buffer() {
        // Buffer covers grid rect (3,2) 5x4; tile 4x4.
        let mut buf = vec![0u8; 5 * 4];
        let g = GridWriter::new(&mut buf, 5, 4, 1, (3, 2), 4, 4, 16, 16).unwrap();
        let t = g.claim(0, 0).unwrap();
        assert_eq!((t.x0, t.y0, t.size(), t.skip_x, t.skip_y), (0, 0, (1, 2), 3, 2));
        let t = g.claim(1, 1).unwrap();
        assert_eq!((t.x0, t.y0, t.size(), t.skip_x, t.skip_y), (1, 2, (4, 2), 0, 0));
        assert!(g.claim(3, 3).is_none());
    }
}