strided-basic 0.4.6

//! Kernel iteration engine ported from Julia's Strided.jl/src/mapreduce.jl
//!
//! This module implements the core iteration engine that follows Julia's
//! `_mapreduce_kernel!` pattern for cache-optimized strided array operations.

use crate::fuse::{compress_dims, fuse_dims};
use crate::{block, order, Result, StridedError};

/// Maximum total elements for the small tensor fast path.
/// Tensors at or below this size skip `compute_order` and `compute_block_sizes`
/// since they fit in L1 cache and blocking provides no benefit.
pub const SMALL_TENSOR_THRESHOLD: usize = 1024;

/// Blocking metadata produced by the execution planners; not a layout-validity proof.
#[derive(Debug)]
pub struct KernelPlan {
    #[cfg_attr(not(test), allow(dead_code))]
    pub(crate) order: Vec<usize>, // outer -> inner
    pub block: Vec<usize>,
}

/// Build an execution plan for strided iteration (used only in tests).
///
/// This follows Julia's `_mapreduce_fuse!` -> `_mapreduce_order!` -> `_mapreduce_block!` pipeline:
/// 1. Fuse contiguous dimensions
/// 2. Compute optimal iteration order
/// 3. Compute block sizes for cache efficiency
#[cfg(test)]
pub(crate) fn build_plan(
    dims: &[usize],
    strides_list: &[&[isize]],
    dest_index: Option<usize>,
    elem_size: usize,
) -> KernelPlan {
    let order = order::compute_order(dims, strides_list, dest_index);
    let block = block::compute_block_sizes(dims, &order, strides_list, elem_size);
    KernelPlan { order, block }
}

/// Build an execution plan with dimension fusion.
///
/// Pipeline: order -> reorder -> fuse -> block.
///
/// Ordering first ensures that dimensions are sorted by stride importance
/// (smallest stride innermost). Fusing *after* ordering catches contiguous
/// dimensions regardless of the original memory layout (column-major,
/// row-major, or any permutation).
///
/// Returns `(fused_dims, ordered_strides, KernelPlan)` where dimensions
/// and strides are already in iteration order (plan.order is identity).
pub(crate) fn build_plan_fused(
    dims: &[usize],
    strides_list: &[&[isize]],
    dest_index: Option<usize>,
    elem_size: usize,
) -> (Vec<usize>, Vec<Vec<isize>>, KernelPlan) {
    // 1. Compute optimal iteration order on original dims
    let order = order::compute_order(dims, strides_list, dest_index);

    // 2. Reorder dims and strides
    let ordered_dims: Vec<usize> = order.iter().map(|&d| dims[d]).collect();
    let ordered_strides: Vec<Vec<isize>> = strides_list
        .iter()
        .map(|strides| order.iter().map(|&d| strides[d]).collect())
        .collect();
    let ordered_strides_refs: Vec<&[isize]> =
        ordered_strides.iter().map(|s| s.as_slice()).collect();

    // 3. Fuse contiguous dimensions in ordered space
    let fused_dims = fuse_dims(&ordered_dims, &ordered_strides_refs);

    // 4. Compress: remove size-1 dimensions to reduce loop depth
    let (compressed_dims, compressed_strides) = compress_dims(&fused_dims, &ordered_strides);
    let compressed_strides_refs: Vec<&[isize]> =
        compressed_strides.iter().map(|s| s.as_slice()).collect();

    // 5. Compute blocks with identity ordering (already ordered)
    let identity: Vec<usize> = (0..compressed_dims.len()).collect();
    let block = block::compute_block_sizes(
        &compressed_dims,
        &identity,
        &compressed_strides_refs,
        elem_size,
    );

    (
        compressed_dims,
        compressed_strides,
        KernelPlan {
            order: identity,
            block,
        },
    )
}

/// Simplified plan for small tensors that fit in L1 cache.
///
/// Skips `compute_order` and `compute_block_sizes` since blocking is
/// unnecessary for small data. Only fuses and compresses dimensions
/// to reduce loop depth.
///
/// Returns `(fused_dims, fused_strides, KernelPlan)` where the plan's
/// order is identity and block sizes equal the full dimension sizes.
pub(crate) fn build_plan_fused_small(
    dims: &[usize],
    strides_list: &[&[isize]],
) -> (Vec<usize>, Vec<Vec<isize>>, KernelPlan) {
    let strides_owned: Vec<Vec<isize>> = strides_list.iter().map(|s| s.to_vec()).collect();

    // Single fuse + compress pass (no ordering needed)
    let fused = fuse_dims(dims, strides_list);
    let (fused_dims, fused_strides) = compress_dims(&fused, &strides_owned);

    // Block = full dimension (no tiling)
    let block = fused_dims.clone();
    let identity: Vec<usize> = (0..fused_dims.len()).collect();

    (
        fused_dims,
        fused_strides,
        KernelPlan {
            order: identity,
            block,
        },
    )
}

// ============================================================================
// Block-based iteration with inner stride callback
// ============================================================================

/// Iterate over blocks, calling f with (offsets, block_len, inner_strides).
///
/// The callback receives the current byte offsets for each array, the number
/// of elements in the innermost block, and the innermost strides for each array.
/// This allows the caller to implement vectorized inner loops.
#[cfg(test)]
#[inline]
pub(crate) fn for_each_inner_block<F>(
    dims: &[usize],
    plan: &KernelPlan,
    strides_list: &[&[isize]],
    mut f: F,
) -> Result<()>
where
    F: FnMut(&[isize], usize, &[isize]) -> Result<()>,
{
    let rank = dims.len();
    if rank == 0 {
        let offsets = vec![0isize; strides_list.len()];
        return f(&offsets, 1, &[]);
    }

    // Reorder dimensions and strides according to plan
    let ordered_dims: Vec<usize> = plan.order.iter().map(|&d| dims[d]).collect();
    let ordered_blocks: Vec<usize> = plan.block.clone();

    // Build stride vectors for each array, ordered by plan
    let num_arrays = strides_list.len();
    let mut ordered_strides: Vec<Vec<isize>> = Vec::with_capacity(num_arrays);
    for strides in strides_list {
        let s: Vec<isize> = plan.order.iter().map(|&d| strides[d]).collect();
        ordered_strides.push(s);
    }

    // Initial offsets (all zero)
    let mut offsets = vec![0isize; num_arrays];

    // Call the specialized kernel based on rank
    match rank {
        1 => kernel_1d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        2 => kernel_2d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        3 => kernel_3d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        4 => kernel_4d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        5 => kernel_5d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        6 => kernel_6d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        7 => kernel_7d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        8 => kernel_8d_inner(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
        _ => kernel_nd_inner_iterative(
            &ordered_dims,
            &ordered_blocks,
            &ordered_strides,
            &mut offsets,
            &mut f,
        ),
    }
}

// ============================================================================
// Macro-generated rank-specialized kernels (inner-block callback)
// ============================================================================

// Every loop below advances an offset only when another position along its
// level follows, and each loop rewinds its own level on completion. Offsets
// therefore only ever hold reachable positions of the validated layouts (plus
// the caller's initial offsets), so no intermediate value can overflow even
// when a layout ends within one stride of `isize::MAX` (issue #243
// follow-up). The rewind distance is the difference of two reachable offsets,
// so it is representable as well.

/// Element-level nested loops for ranks >= 2.
///
/// Iterates over the element block, calling `f` at each innermost position.
/// Levels are listed from outermost to innermost (excluding level 0 which is
/// the callback level). For a rank-N kernel, element levels are N-1, N-2, ..., 1.
/// Each level leaves the offsets where it found them.
macro_rules! elem_loops {
    // Base case: single level (innermost element level).
    // Iterates blens[$lv] times, calling f and advancing by stride[$lv]
    // between calls.
    ($offsets:ident, $strides:ident, $f:ident, $blens:ident, $is:ident; $lv:literal) => {
        for _i in 0..$blens[$lv] {
            $f($offsets, $blens[0], &$is)?;
            if _i + 1 < $blens[$lv] {
                for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
                    *o += s[$lv];
                }
            }
        }
        let _back = $blens[$lv].saturating_sub(1) as isize;
        for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
            *o -= _back * s[$lv];
        }
    };
    // Recursive case: outermost level wraps inner levels, which rewind
    // themselves; this level advances between inner passes.
    ($offsets:ident, $strides:ident, $f:ident, $blens:ident, $is:ident;
     $lv:literal, $next:literal $(, $rest:literal)*) => {
        for _i in 0..$blens[$lv] {
            elem_loops!($offsets, $strides, $f, $blens, $is; $next $(, $rest)*);
            if _i + 1 < $blens[$lv] {
                for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
                    *o += s[$lv];
                }
            }
        }
        let _back = $blens[$lv].saturating_sub(1) as isize;
        for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
            *o -= _back * s[$lv];
        }
    };
}

/// Block-level nested while loops for ranks >= 2.
///
/// Each block level iterates over tiles of the corresponding dimension.
/// At level 0 (innermost), element loops are executed. Each block level
/// advances to the next tile only when one follows and rewinds itself on
/// completion; inner levels rewind themselves.
macro_rules! block_loop {
    // Base case: single block level (the innermost).
    ($dims:ident, $blocks:ident, $strides:ident, $offsets:ident, $f:ident,
     $blens:ident, $is:ident; elem=[$($el:literal),+]; $lv0:literal; top=$top:literal) => {{
        let mut _j = 0usize;
        let mut _advanced = 0usize;
        while _j < $dims[$lv0] {
            $blens[$lv0] = $blocks[$lv0].max(1).min($dims[$lv0]).min($dims[$lv0] - _j);
            elem_loops!($offsets, $strides, $f, $blens, $is; $($el),+);
            _j += $blens[$lv0];
            if _j < $dims[$lv0] {
                for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
                    *o += ($blens[$lv0] as isize) * s[$lv0];
                }
                _advanced = _j;
            }
        }
        for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
            *o -= (_advanced as isize) * s[$lv0];
        }
    }};
    // Recursive case: outer block level wraps inner block levels.
    ($dims:ident, $blocks:ident, $strides:ident, $offsets:ident, $f:ident,
     $blens:ident, $is:ident; elem=[$($el:literal),+];
     $lv:literal, $next:literal $(, $rest:literal)*; top=$top:literal) => {{
        let mut _j = 0usize;
        let mut _advanced = 0usize;
        while _j < $dims[$lv] {
            $blens[$lv] = $blocks[$lv].max(1).min($dims[$lv]).min($dims[$lv] - _j);
            block_loop!($dims, $blocks, $strides, $offsets, $f, $blens, $is;
                elem=[$($el),+]; $next $(, $rest)*; top=$top);
            _j += $blens[$lv];
            if _j < $dims[$lv] {
                for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
                    *o += ($blens[$lv] as isize) * s[$lv];
                }
                _advanced = _j;
            }
        }
        for (o, s) in $offsets.iter_mut().zip($strides.iter()) {
            *o -= (_advanced as isize) * s[$lv];
        }
    }};
}

/// Generate a rank-specialized kernel function.
///
/// For rank 1 there are no element loops; only a single block loop on dim 0.
/// For rank >= 2, block loops nest from outermost to innermost (dim 0), and
/// element loops nest from dim N-1 down to dim 1 inside the innermost block.
/// The offsets are back at their starting values on return.
macro_rules! make_kernel {
    // Rank 1: single block loop, no element nesting.
    ($name:ident, rank=1) => {
        #[inline]
        fn $name<F>(
            dims: &[usize],
            blocks: &[usize],
            strides: &[Vec<isize>],
            offsets: &mut [isize],
            f: &mut F,
        ) -> Result<()>
        where
            F: FnMut(&[isize], usize, &[isize]) -> Result<()>,
        {
            let d0 = dims[0];
            let b0 = blocks[0].max(1).min(d0);
            let inner_strides: Vec<isize> = strides.iter().map(|s| s[0]).collect();

            let mut j0 = 0usize;
            let mut advanced = 0usize;
            while j0 < d0 {
                let blen0 = b0.min(d0 - j0);
                f(offsets, blen0, &inner_strides)?;
                j0 += blen0;
                if j0 < d0 {
                    for (o, s) in offsets.iter_mut().zip(strides.iter()) {
                        *o += (blen0 as isize) * s[0];
                    }
                    advanced = j0;
                }
            }
            for (o, s) in offsets.iter_mut().zip(strides.iter()) {
                *o -= (advanced as isize) * s[0];
            }
            Ok(())
        }
    };
    // Rank >= 2: nested block loops + element loops.
    //   block=[outermost, ..., 0]  - block loop levels from outermost to innermost
    //   elem=[N-1, ..., 1]         - element loop levels from outermost to innermost
    //   top=N-1                    - topmost element level (= rank - 1)
    ($name:ident, rank=$rank:literal,
     block=[$($blk:literal),+], elem=[$($el:literal),+], top=$top:literal) => {
        #[inline]
        fn $name<F>(
            dims: &[usize],
            blocks: &[usize],
            strides: &[Vec<isize>],
            offsets: &mut [isize],
            f: &mut F,
        ) -> Result<()>
        where
            F: FnMut(&[isize], usize, &[isize]) -> Result<()>,
        {
            let inner_strides: Vec<isize> = strides.iter().map(|s| s[0]).collect();
            let mut blens = [0usize; $rank];
            block_loop!(dims, blocks, strides, offsets, f, blens, inner_strides;
                elem=[$($el),+]; $($blk),+; top=$top);
            Ok(())
        }
    };
}

make_kernel!(kernel_1d_inner, rank = 1);
make_kernel!(
    kernel_2d_inner,
    rank = 2,
    block = [1, 0],
    elem = [1],
    top = 1
);
make_kernel!(
    kernel_3d_inner,
    rank = 3,
    block = [2, 1, 0],
    elem = [2, 1],
    top = 2
);
make_kernel!(
    kernel_4d_inner,
    rank = 4,
    block = [3, 2, 1, 0],
    elem = [3, 2, 1],
    top = 3
);
make_kernel!(
    kernel_5d_inner,
    rank = 5,
    block = [4, 3, 2, 1, 0],
    elem = [4, 3, 2, 1],
    top = 4
);
make_kernel!(
    kernel_6d_inner,
    rank = 6,
    block = [5, 4, 3, 2, 1, 0],
    elem = [5, 4, 3, 2, 1],
    top = 5
);
make_kernel!(
    kernel_7d_inner,
    rank = 7,
    block = [6, 5, 4, 3, 2, 1, 0],
    elem = [6, 5, 4, 3, 2, 1],
    top = 6
);
make_kernel!(
    kernel_8d_inner,
    rank = 8,
    block = [7, 6, 5, 4, 3, 2, 1, 0],
    elem = [7, 6, 5, 4, 3, 2, 1],
    top = 7
);

/// N-dimensional kernel with inner block callback (iterative form).
///
/// Fallback for rank >= 9. Uses a carry-style increment over outer levels
/// instead of compile-time unrolled nested loops.
#[inline]
fn kernel_nd_inner_iterative<F>(
    dims: &[usize],
    blocks: &[usize],
    strides: &[Vec<isize>],
    offsets: &mut [isize],
    f: &mut F,
) -> Result<()>
where
    F: FnMut(&[isize], usize, &[isize]) -> Result<()>,
{
    let rank = dims.len();
    debug_assert!(rank >= 9);

    let d0 = dims[0];
    let b0 = blocks[0].max(1).min(d0);
    let inner_strides: Vec<isize> = strides.iter().map(|s| s[0]).collect();

    // Current position for each outer level (1..rank-1). Level 0 uses block loop.
    let mut idx = vec![0usize; rank];

    // As in the rank-specialized kernels, offsets advance only when another
    // position follows and each level rewinds from its last position, so they
    // never leave the reachable span.
    if dims.contains(&0) {
        return Ok(());
    }
    loop {
        // Level 0: callback over contiguous block fragments.
        let mut j0 = 0usize;
        let mut advanced = 0usize;
        while j0 < d0 {
            let blen0 = b0.min(d0 - j0);
            f(offsets, blen0, &inner_strides)?;
            j0 += blen0;
            if j0 < d0 {
                for (offset, s) in offsets.iter_mut().zip(strides.iter()) {
                    *offset += (blen0 as isize) * s[0];
                }
                advanced = j0;
            }
        }
        for (offset, s) in offsets.iter_mut().zip(strides.iter()) {
            *offset -= (advanced as isize) * s[0];
        }

        // Carry-style increment for outer levels.
        let mut level = 1usize;
        loop {
            if idx[level] + 1 < dims[level] {
                idx[level] += 1;
                for (offset, s) in offsets.iter_mut().zip(strides.iter()) {
                    *offset += s[level];
                }
                break;
            }

            let last = (dims[level] - 1) as isize;
            idx[level] = 0;
            for (offset, s) in offsets.iter_mut().zip(strides.iter()) {
                *offset -= last * s[level];
            }
            level += 1;
            if level == rank {
                return Ok(());
            }
        }
    }
}

// ============================================================================
// Pre-ordered iteration (for threaded leaf functions)
// ============================================================================

/// Iterate over blocks with pre-ordered dimensions and initial offsets.
///
/// Unlike `for_each_inner_block`, this function assumes that `dims`, `blocks`,
/// and `strides` are **already in iteration order** (i.e., identity ordering).
/// It also accepts `initial_offsets` which are added to the starting offsets
/// before iteration begins.
///
/// This avoids the redundant re-ordering and per-callback `Vec` allocation
/// that `for_each_inner_block_with_offsets` previously incurred.
///
/// The block-walking machinery is compiled once (`preordered_dyn`): a generic
/// caller only instantiates its own callback, reached through one indirect call
/// per inner block (a run of elements, never per element). Keeping the walk
/// generic would duplicate it for every operation, dtype and element-op
/// combination of every downstream crate.
#[inline]
pub(crate) fn for_each_inner_block_preordered<F>(
    dims: &[usize],
    blocks: &[usize],
    strides: &[Vec<isize>],
    initial_offsets: &[isize],
    mut f: F,
) -> Result<()>
where
    F: FnMut(&[isize], usize, &[isize]) -> Result<()>,
{
    preordered_dyn(dims, blocks, strides, initial_offsets, &mut f)
}

type BlockFn<'a> = &'a mut dyn FnMut(&[isize], usize, &[isize]) -> Result<()>;

#[inline(never)]
fn preordered_dyn(
    dims: &[usize],
    blocks: &[usize],
    strides: &[Vec<isize>],
    initial_offsets: &[isize],
    mut f: BlockFn<'_>,
) -> Result<()> {
    let rank = dims.len();
    if rank == 0 {
        return f(initial_offsets, 1, &[]);
    }

    // Start from initial_offsets (kernel functions reset to starting values at end)
    let mut offsets = initial_offsets.to_vec();

    match rank {
        1 => kernel_1d_inner(dims, blocks, strides, &mut offsets, &mut f),
        2 => kernel_2d_inner(dims, blocks, strides, &mut offsets, &mut f),
        3 => kernel_3d_inner(dims, blocks, strides, &mut offsets, &mut f),
        4 => kernel_4d_inner(dims, blocks, strides, &mut offsets, &mut f),
        5 => kernel_5d_inner(dims, blocks, strides, &mut offsets, &mut f),
        6 => kernel_6d_inner(dims, blocks, strides, &mut offsets, &mut f),
        7 => kernel_7d_inner(dims, blocks, strides, &mut offsets, &mut f),
        8 => kernel_8d_inner(dims, blocks, strides, &mut offsets, &mut f),
        _ => kernel_nd_inner_iterative(dims, blocks, strides, &mut offsets, &mut f),
    }
}

// ============================================================================
// Utility functions
// ============================================================================

/// Check exact shape agreement before prevalidated replay.
///
/// # Examples
///
/// ```
/// use strided_basic::execution::ensure_same_shape;
/// ensure_same_shape(&[2, 3], &[2, 3]).unwrap();
/// assert!(ensure_same_shape(&[2, 3], &[3, 2]).is_err());
/// ```
///
/// # Errors
/// Returns a rank or shape mismatch.
pub fn ensure_same_shape(a: &[usize], b: &[usize]) -> Result<()> {
    if a.len() != b.len() {
        return Err(crate::StridedError::RankMismatch(a.len(), b.len()));
    }
    if a != b {
        return Err(crate::StridedError::ShapeMismatch(a.to_vec(), b.to_vec()));
    }
    Ok(())
}

#[derive(Copy, Clone, Debug, Eq, PartialEq)]
pub(crate) enum ContiguousLayout {
    /// C-like layout: last axis varies fastest.
    RowMajor,
    /// Julia/Fortran-like layout: first axis varies fastest.
    ColMajor,
}

/// Returns the contiguous memory layout kind for the given (dims, strides).
///
/// Notes:
/// - Ignores axes with `dim <= 1` since they do not affect addressability.
/// - Does not treat negative-stride views as contiguous for fast-path purposes.
pub(crate) fn contiguous_layout(dims: &[usize], strides: &[isize]) -> Option<ContiguousLayout> {
    if dims.len() != strides.len() {
        return None;
    }
    if dims.is_empty() {
        return Some(ContiguousLayout::RowMajor);
    }

    // Row-major: check from last to first.
    let mut expected = 1isize;
    let mut row_ok = true;
    for (&dim, &stride) in dims.iter().rev().zip(strides.iter().rev()) {
        if dim <= 1 {
            continue;
        }
        if stride != expected {
            row_ok = false;
            break;
        }
        expected = expected.saturating_mul(dim as isize);
    }
    if row_ok {
        return Some(ContiguousLayout::RowMajor);
    }

    // Col-major: check from first to last.
    let mut expected = 1isize;
    for (&dim, &stride) in dims.iter().zip(strides.iter()) {
        if dim <= 1 {
            continue;
        }
        if stride != expected {
            return None;
        }
        expected = expected.saturating_mul(dim as isize);
    }
    Some(ContiguousLayout::ColMajor)
}

/// Checked element count of `dims`.
///
/// A shape containing a zero extent has zero elements even when the product
/// of its other extents would overflow; such views pass
/// `strided_view::validate_bounds`, so this must not form that product. A
/// nonempty shape whose element count exceeds `usize::MAX` (for example a
/// huge stride-0 broadcast) returns [`StridedError::OffsetOverflow`].
pub(crate) fn total_len(dims: &[usize]) -> Result<usize> {
    if dims.contains(&0) {
        return Ok(0);
    }
    dims.iter()
        .try_fold(1usize, |acc, &dim| acc.checked_mul(dim))
        .ok_or(StridedError::OffsetOverflow)
}

/// Whether the sequential contiguous fast path should be used.
///
/// When the `parallel` feature is enabled and the total element count exceeds
/// the threading threshold, we normally must *not* take the contiguous fast path
/// so that the parallel kernel path can be reached. If the active Rayon pool has
/// only one worker, however, there is no threaded path to reach and the
/// sequential slice loop is the right kernel.
#[inline]
pub(crate) fn use_sequential_fast_path(total: usize) -> bool {
    #[cfg(feature = "parallel")]
    {
        total <= crate::threading::MINTHREADLENGTH || crate::execution_policy::rayon_threads() <= 1
    }
    #[cfg(not(feature = "parallel"))]
    {
        let _ = total;
        true
    }
}

/// Returns the common contiguous layout if **all** provided stride arrays
/// share the same contiguous layout for the given `dims`.
///
/// Returns `None` if `strides_list` is empty, any array is not contiguous,
/// or any two arrays have different contiguous layouts.
#[inline]
pub(crate) fn same_contiguous_layout(
    dims: &[usize],
    strides_list: &[&[isize]],
) -> Option<ContiguousLayout> {
    let first = contiguous_layout(dims, strides_list.first()?)?;
    for strides in &strides_list[1..] {
        if contiguous_layout(dims, strides)? != first {
            return None;
        }
    }
    Some(first)
}

/// Returns the common contiguous layout only when the sequential fast path
/// should be used (total elements <= threading threshold).
///
/// Returns [`StridedError::OffsetOverflow`] when the element count of `dims`
/// does not fit in `usize`.
#[inline]
pub(crate) fn sequential_contiguous_layout(
    dims: &[usize],
    strides_list: &[&[isize]],
) -> Result<Option<ContiguousLayout>> {
    if !use_sequential_fast_path(total_len(dims)?) {
        return Ok(None);
    }
    Ok(same_contiguous_layout(dims, strides_list))
}

#[cfg(test)]
#[path = "kernel/tests/tests.rs"]
mod tests;