cubecl-tile 0.12.0-pre.1

//! The order a plane level spanning the launch hands its tiles to the launch's planes
//! ([`PlaneOrder`]), and the cubes a unit level spanning the launch groups its units in
//! ([`UnitCube`]).

use cubecl::prelude::*;

use super::{ComputeScope, CubeOrder, Level, cube_positions};
use crate::{Axis, Coords, CubeAxis, Integer, IntegerExpand, Space};

/// The most cubes one grid dimension takes on every device a launch may run on.
const CUBES_PER_GRID_DIMENSION: usize = 65_535;

/// The order the planes of a launch take a level's tiles in, where a plane level spans the launch
/// rather than one cube ([`Levels::launch_planes`](crate::Levels::launch_planes)): in blocks, each
/// a box of tiles held by consecutive planes, the blocks taken in a [`CubeOrder`] over their grid.
///
/// A plane of the launch is numbered by its cube and its place in it, `cube · planes_per_cube +
/// UNIT_POS_Y`, the cube read off the grid's `x` and `y`. Where a block holds a cube's planes,
/// as it does unless [`cubes_of`](PlaneOrder::cubes_of) says otherwise, each cube takes one block,
/// and the launch deals its tiles as a cube level with a plane level inside it would. Any order
/// is right: the planes share nothing, so the order only says which tiles the planes of one cube
/// take together.
///
/// Every plane the launch spawns walks one tile, as every plane of a cube level's cube does: a
/// plane past the last tile, of a block the edge cuts short or of a cube past the last block,
/// walks one past the grid, which the binding masks (`PlaneOrder::overruns`). So the
/// planes of a cube all reach any barrier one of them meets.
#[derive(Clone, PartialEq, Eq, Hash, Debug)]
pub struct PlaneOrder {
    /// Tiles one block holds along each axis it names; along an axis it does not name, one.
    block: Vec<(Axis, usize)>,
    block_order: CubeOrder,
    /// The launch's `y`.
    planes_per_cube: usize,
}

impl PlaneOrder {
    /// Blocks of `block` tiles, one tile a plane, taken in `block_order`, one block a cube: what a
    /// cube level holding a plane level of `block` hands out.
    pub fn new(block: &[(Axis, usize)], block_order: CubeOrder) -> Self {
        for &(axis, tiles) in block {
            assert!(
                tiles > 0,
                "PlaneOrder: a block holds no tile along {axis:?}"
            );
        }
        let block = block.to_vec();
        let planes_per_cube = block.iter().map(|&(_, tiles)| tiles).product();
        Self {
            block,
            block_order: block_order.canonicalize(),
            planes_per_cube,
        }
    }

    /// This order with the launch's cubes holding `planes` planes each, which need not be a
    /// block's: a block then spans several cubes, or a cube several blocks.
    pub fn cubes_of(self, planes: usize) -> Self {
        assert!(
            planes > 0,
            "PlaneOrder::cubes_of: a cube holds at least one plane"
        );
        Self {
            planes_per_cube: planes,
            ..self
        }
    }

    pub fn planes_per_cube(&self) -> usize {
        self.planes_per_cube
    }

    /// Tiles one block holds along `axis`.
    pub fn block_along(&self, axis: Axis) -> usize {
        self.block
            .iter()
            .find(|&&(a, _)| a == axis)
            .map_or(1, |&(_, tiles)| tiles)
    }

    pub fn block_order(&self) -> CubeOrder {
        self.block_order
    }

    pub(crate) fn block_axes(&self) -> impl Iterator<Item = Axis> + '_ {
        self.block.iter().map(|&(axis, _)| axis)
    }

    /// Planes one block holds, a tile each.
    pub(crate) fn block_planes(&self) -> usize {
        self.block.iter().map(|&(_, tiles)| tiles).product()
    }

    /// Whether each cube takes one block.
    pub(crate) fn one_block_a_cube(&self) -> bool {
        self.block_planes() == self.planes_per_cube
    }

    /// The grid's `x` and `y` for `level`'s blocks over `space`: every block's planes in cubes of
    /// [`planes_per_cube`](Self::planes_per_cube), laid along `x` and spilling into `y` past what
    /// one dimension takes. At least one cube, so an empty problem divides by no empty dimension.
    /// Static extents only.
    ///
    /// A spill lays the cubes out exactly where it can: `x` is the widest count one dimension
    /// takes that divides the cubes, so no cube runs past the last block. Only a count with no
    /// such divisor rounds `y` up, and leaves the spare cubes [`overruns`](Self::overruns) masks.
    pub(crate) fn grid(&self, level: &Level, space: &Space) -> (usize, usize) {
        let planes = self.blocks(level, space) * self.block_planes();
        let cubes = planes.div_ceil(self.planes_per_cube).max(1);
        if cubes <= CUBES_PER_GRID_DIMENSION {
            return (cubes, 1);
        }
        let rows = cubes.div_ceil(CUBES_PER_GRID_DIMENSION);
        let x = (rows..=CUBES_PER_GRID_DIMENSION)
            .find(|&rows| cubes.is_multiple_of(rows) && cubes / rows <= CUBES_PER_GRID_DIMENSION)
            .map_or(CUBES_PER_GRID_DIMENSION, |rows| cubes / rows);
        (x, cubes.div_ceil(x))
    }

    /// Whether some plane the launch spawns for `level` over `space` walks a tile past the last
    /// along `axis`, so the binding masks it there: where the edge cuts a block short along
    /// `axis`, or where the grid holds more planes than the blocks do and the blocks past the last
    /// land past `axis` ([`block_coords`]). `axis` is static; where an extent the grid counts is
    /// only the launch's, the planes may run past it, and it is masked.
    pub(crate) fn overruns(&self, level: &Level, space: &Space, axis: Axis) -> bool {
        let ragged = !level
            .tiles(space, axis)
            .is_multiple_of(self.block_along(axis));
        if ragged {
            return true;
        }
        if Self::spare_blocks_land_past(level, self.block_order) != axis {
            return false;
        }
        if level.axes().into_iter().any(|axis| space.is_dynamic(axis)) {
            return true;
        }
        let (x, y) = self.grid(level, space);
        x * y * self.planes_per_cube > self.blocks(level, space) * self.block_planes()
    }

    /// Blocks `level` holds over `space`, the edge's counted.
    fn blocks(&self, level: &Level, space: &Space) -> usize {
        level
            .axes()
            .into_iter()
            .map(|axis| level.tiles(space, axis).div_ceil(self.block_along(axis)))
            .product()
    }

    /// The axis a block index past the last lands past ([`block_coords`]): the one axis; of two,
    /// the one `order` runs out along; of more, the outermost, which is never wrapped.
    fn spare_blocks_land_past(level: &Level, order: CubeOrder) -> Axis {
        let axes = level.axes();
        match axes.len() {
            1 => axes[0],
            2 => match order.canonicalize() {
                CubeOrder::SwizzleRow(_) => axes[0],
                CubeOrder::RowMajor | CubeOrder::SwizzleCol(_) => axes[1],
            },
            _ => axes[axes.len() - 1],
        }
    }
}

/// The cubes a unit level spanning the launch groups its units in
/// ([`Levels::launch_units`](crate::Levels::launch_units)): `planes` planes of `plane_width` units
/// each. The units share nothing, so the cube is only how the launch groups them.
#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
pub struct UnitCube {
    pub plane_width: usize,
    pub planes: usize,
}

impl UnitCube {
    /// Units one cube holds.
    pub fn units(&self) -> usize {
        self.plane_width * self.planes
    }

    /// The order the launch's units take a level's tiles in: one tile a unit, the tiles in row
    /// order, [`units`](Self::units) a cube.
    pub(crate) fn order(&self) -> PlaneOrder {
        PlaneOrder::new(&[], CubeOrder::RowMajor).cubes_of(self.units())
    }
}

/// This plane's or unit's tile of `level`, a level spanning the launch, along each axis of
/// `space`: `(positions, counts)`: along each of the level's axes the tile's index and `1`, an
/// index past the last for an instance past the last tile; along each other axis `0` and every
/// tile. `grid` is the tiles along each axis.
#[cube]
pub(crate) fn launch_tiles(
    #[comptime] space: Space,
    #[comptime] level: Level,
    grid: &Coords<usize>,
) -> (Coords<usize>, Coords<usize>) {
    let order = comptime!(
        level
            .launch_order()
            .expect("a level spanning the launch states the order its instances take")
    );
    // An instance's place in its cube: a plane's row, or a unit's place among all of its units.
    let in_cube = match comptime!(level.coverage().scope()) {
        Some(ComputeScope::Unit) => UNIT_POS as usize,
        _ => UNIT_POS_Y as usize,
    };
    let rank = comptime!(space.rank());
    // The level's axes in the order it names them: the first rides fastest across the blocks, as
    // a cube level's `X` does.
    let named = comptime!(
        level
            .axes()
            .into_iter()
            .map(|axis| space.position(axis))
            .collect::<Vec<_>>()
    );
    // The grid's `x` and `y`, as `Partitioning::cube_count` lays the cubes out:
    // its `z` is always one.
    let cube = CubeAxis::position(CubeAxis::X)
        .plus(CubeAxis::position(CubeAxis::Y).times(CUBE_COUNT_X as usize));
    let (block, local) = if comptime!(order.one_block_a_cube()) {
        (cube, in_cube)
    } else {
        let flat = cube
            .times(comptime!(order.planes_per_cube()).runtime())
            .plus(in_cube);
        let planes = comptime!(order.block_planes()).runtime();
        (flat.divided_by(planes), flat.remainder(planes))
    };

    // Blocks along each named axis, the last one ragged.
    let mut blocks = Coords::<usize>::new();
    #[unroll]
    for p in comptime!(named.clone()) {
        let along = comptime!(order.block_along(space.axis_at(p)));
        blocks.push(
            grid.at(p)
                .plus(comptime!(along - 1).runtime())
                .divided_by(along.runtime()),
        );
    }
    let at = block_coords(
        block,
        &blocks,
        comptime!(named.len()),
        comptime!(order.block_order()),
    );

    let mut positions = Coords::<usize>::new();
    let mut counts = Coords::<usize>::new();
    #[unroll]
    for p in 0..rank {
        match comptime!(named.iter().position(|&q| q == p)) {
            None => {
                positions.push(0usize);
                counts.push(grid.at(p));
            }
            Some(i) => {
                let along = comptime!(order.block_along(space.axis_at(p)));
                // The last named axis rides fastest within a block, as a plane level's does.
                let inner = comptime!(
                    named[i + 1..]
                        .iter()
                        .map(|&q| order.block_along(space.axis_at(q)))
                        .product::<usize>()
                );
                let within = local.divided_by(inner.runtime()).remainder(along.runtime());
                let tile = at.at(i).times(along.runtime()).plus(within);
                positions.push(tile);
                counts.push(1usize);
            }
        }
    }
    (positions, counts)
}

/// Block `block`'s coordinates over `blocks`, `named` axes of them: the first two in `order`, the
/// others outermost, first fastest, the last never wrapped, so a block past the grid lands off it.
#[cube]
fn block_coords(
    block: usize,
    blocks: &Coords<usize>,
    #[comptime] named: usize,
    #[comptime] order: CubeOrder,
) -> Coords<usize> {
    let mut out = Coords::<usize>::new();
    if comptime!(named == 1) {
        out.push(block);
    } else {
        let plane = blocks.at(0).times(blocks.at(1));
        let first_two = if comptime!(named == 2) {
            block
        } else {
            block.remainder(plane)
        };
        let (x, y) = cube_positions(first_two, (blocks.at(0), blocks.at(1)), order);
        out.push(x);
        out.push(y);
        let mut rest = block.divided_by(plane);
        #[unroll]
        for i in 2..named {
            if comptime!(i + 1 == named) {
                out.push(rest);
            } else {
                out.push(rest.remainder(blocks.at(i)));
                rest = rest.divided_by(blocks.at(i));
            }
        }
    }
    out
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{Levels, Partitioning};

    const N: Axis = Axis(0);

    /// The grid's `x`, `y` and `z`.
    fn counts(partitioning: &Partitioning) -> (u32, u32, u32) {
        let CubeCount::Static(x, y, z) = partitioning.cube_count() else {
            unreachable!("a static space lays a static grid");
        };
        (x, y, z)
    }

    /// One plane a tile of two along `N`, one plane a cube, over `columns`.
    fn one_plane_a_cube(columns: usize) -> Partitioning {
        Partitioning::new(
            Space::new(&[(N, columns)]),
            Levels::leaf(&[(N, 2)])
                .launch_planes(&[N], PlaneOrder::new(&[(N, 1)], CubeOrder::RowMajor))
                .build(),
        )
    }

    /// More cubes than one dimension takes are laid out exactly where a count one dimension takes
    /// divides them: a vocabulary of columns, two a plane, spills into a second row of the same
    /// width, and no plane runs past the last tile. A count with no such divisor, a prime past
    /// one dimension, still spills into a row it does not fill, and overruns it.
    #[test]
    fn a_spill_divides_the_cubes_where_it_can() {
        let vocabulary = one_plane_a_cube(151_936);
        assert_eq!(counts(&vocabulary), (37_984, 2, 1));
        assert!(!vocabulary.overhangs(N));

        let prime = one_plane_a_cube(2 * 65_537);
        assert_eq!(counts(&prime), (CUBES_PER_GRID_DIMENSION as u32, 2, 1));
        assert!(prime.overhangs(N));
    }

    /// A grid one dimension holds stays one row.
    #[test]
    fn a_grid_one_dimension_holds_is_one_row() {
        assert_eq!(counts(&one_plane_a_cube(2 * 1000)), (1000, 1, 1));
    }
}