edgefirst_decoder/

byte.rs

// SPDX-FileCopyrightText: Copyright 2025 Au-Zone Technologies
// SPDX-License-Identifier: Apache-2.0

//! Quantized-domain NMS, the int8 / uint8 mirror of [`crate::float`].
//!
//! These kernels suppress and post-process candidates while the scores are
//! still in their raw integer units, which is the point: dequantizing every
//! anchor costs more than dequantizing the handful that survive. The threshold
//! crosses the other way instead — [`quantize_score_threshold`] maps a float
//! score threshold into the tensor's quantized domain so the comparison stays
//! an integer one.
//!
//! [`postprocess_boxes_quant`] and [`postprocess_boxes_index_quant`] handle
//! candidate selection; [`nms_int`], [`nms_class_aware_int`], and their
//! `_extra` payload-carrying variants handle suppression. Boxes are
//! dequantized only after suppression.

#[cfg(target_arch = "aarch64")]
use crate::arg_max_i8;
use crate::{
    arg_max, float::jaccard, BBoxTypeTrait, BoundingBox, DecoderError, DecoderResult,
    DetectBoxQuantized, Quantization,
};
use ndarray::{
    parallel::prelude::{IntoParallelIterator, ParallelIterator as _},
    Array1, ArrayView1, ArrayView2, Zip,
};
use num_traits::{AsPrimitive, PrimInt};
use rayon::slice::ParallelSliceMut;

/// NEON-accelerated column max update for the column-major argmax path.
///
/// Processes 16 elements per iteration using SIMD max + bitwise select.
/// Handles both unsigned (u8) and signed (i8) comparison semantics.
///
/// # Safety
///
/// - `col_ptr` must point to at least `n` valid bytes.
/// - `max_ptr` must point to at least `n` valid mutable bytes.
/// - `class_ptr` must point to at least `n` valid mutable bytes.
#[cfg(target_arch = "aarch64")]
unsafe fn column_max_update_neon(
    col_ptr: *const u8,
    max_ptr: *mut u8,
    class_ptr: *mut u8,
    n: usize,
    class_idx: u8,
    signed: bool,
) {
    unsafe {
        use std::arch::aarch64::*; // NOSONAR

        let class_vec = vdupq_n_u8(class_idx);
        let chunks = n / 16;
        let remainder = n % 16;

        if signed {
            // Signed i8 comparison: interpret bytes as i8.
            for chunk in 0..chunks {
                let offset = chunk * 16;
                let col = vld1q_s8(col_ptr.add(offset) as *const i8);
                let cur_max = vld1q_s8(max_ptr.add(offset) as *const i8);
                // mask[i] = 0xFF where col[i] >= cur_max[i], else 0x00
                let mask = vcgeq_s8(col, cur_max);
                // new_max = max(col, cur_max)
                let new_max = vmaxq_s8(col, cur_max);
                vst1q_s8(max_ptr.add(offset) as *mut i8, new_max);
                // Select class_idx where mask is set, keep old class otherwise.
                let cur_class = vld1q_u8(class_ptr.add(offset));
                let new_class = vbslq_u8(mask, class_vec, cur_class);
                vst1q_u8(class_ptr.add(offset), new_class);
            }
            // Scalar tail.
            for i in (chunks * 16)..n {
                let val = *(col_ptr.add(i) as *const i8);
                let cur = *(max_ptr.add(i) as *const i8);
                if val >= cur {
                    *(max_ptr.add(i) as *mut i8) = val;
                    *class_ptr.add(i) = class_idx;
                }
            }
        } else {
            // Unsigned u8 comparison.
            for chunk in 0..chunks {
                let offset = chunk * 16;
                let col = vld1q_u8(col_ptr.add(offset));
                let cur_max = vld1q_u8(max_ptr.add(offset));
                let mask = vcgeq_u8(col, cur_max);
                let new_max = vmaxq_u8(col, cur_max);
                vst1q_u8(max_ptr.add(offset), new_max);
                let cur_class = vld1q_u8(class_ptr.add(offset));
                let new_class = vbslq_u8(mask, class_vec, cur_class);
                vst1q_u8(class_ptr.add(offset), new_class);
            }
            // Scalar tail.
            for i in (chunks * 16)..n {
                let val = *col_ptr.add(i);
                let cur = *max_ptr.add(i);
                if val >= cur {
                    *max_ptr.add(i) = val;
                    *class_ptr.add(i) = class_idx;
                }
            }
        }
        let _ = remainder; // suppress unused warning
    }
}

/// NEON-accelerated column max update with software prefetch for DMA-BUF.
///
/// Adds PRFM (prefetch for load, L1 data cache, streaming) hints 2 cache
/// lines ahead of the current read position to hide CMA memory latency.
/// On Cortex-A55 (64B cache lines) this prefetches 128 bytes ahead; on
/// Cortex-A76 (64B L1, 128B L2 lines) the same offset covers one L2 line.
///
/// # Safety
///
/// - `col_ptr` must point to at least `n` valid bytes.
/// - `max_ptr` must point to at least `n` valid mutable bytes.
/// - `class_ptr` must point to at least `n` valid mutable bytes.
#[cfg(target_arch = "aarch64")]
#[allow(dead_code)] // Reserved for DMA-BUF CMA paths where latency hiding helps.
unsafe fn column_max_update_neon_prefetch(
    col_ptr: *const u8,
    max_ptr: *mut u8,
    class_ptr: *mut u8,
    n: usize,
    class_idx: u8,
    signed: bool,
) {
    unsafe {
        use std::arch::aarch64::*; // NOSONAR

        const PREFETCH_AHEAD: usize = 128; // 2 cache lines on A55 (64B each)

        let class_vec = vdupq_n_u8(class_idx);
        let chunks = n / 16;

        if signed {
            for chunk in 0..chunks {
                let offset = chunk * 16;
                // Software prefetch: hint the next read 2 cache lines ahead.
                if offset + PREFETCH_AHEAD < n {
                    core::arch::asm!(
                        "prfm pldl1strm, [{ptr}]",
                        ptr = in(reg) col_ptr.add(offset + PREFETCH_AHEAD),
                        options(nostack, preserves_flags),
                    );
                }
                let col = vld1q_s8(col_ptr.add(offset) as *const i8);
                let cur_max = vld1q_s8(max_ptr.add(offset) as *const i8);
                let mask = vcgeq_s8(col, cur_max);
                let new_max = vmaxq_s8(col, cur_max);
                vst1q_s8(max_ptr.add(offset) as *mut i8, new_max);
                let cur_class = vld1q_u8(class_ptr.add(offset));
                let new_class = vbslq_u8(mask, class_vec, cur_class);
                vst1q_u8(class_ptr.add(offset), new_class);
            }
            for i in (chunks * 16)..n {
                let val = *(col_ptr.add(i) as *const i8);
                let cur = *(max_ptr.add(i) as *const i8);
                if val >= cur {
                    *(max_ptr.add(i) as *mut i8) = val;
                    *class_ptr.add(i) = class_idx;
                }
            }
        } else {
            for chunk in 0..chunks {
                let offset = chunk * 16;
                if offset + PREFETCH_AHEAD < n {
                    core::arch::asm!(
                        "prfm pldl1strm, [{ptr}]",
                        ptr = in(reg) col_ptr.add(offset + PREFETCH_AHEAD),
                        options(nostack, preserves_flags),
                    );
                }
                let col = vld1q_u8(col_ptr.add(offset));
                let cur_max = vld1q_u8(max_ptr.add(offset));
                let mask = vcgeq_u8(col, cur_max);
                let new_max = vmaxq_u8(col, cur_max);
                vst1q_u8(max_ptr.add(offset), new_max);
                let cur_class = vld1q_u8(class_ptr.add(offset));
                let new_class = vbslq_u8(mask, class_vec, cur_class);
                vst1q_u8(class_ptr.add(offset), new_class);
            }
            for i in (chunks * 16)..n {
                let val = *col_ptr.add(i);
                let cur = *max_ptr.add(i);
                if val >= cur {
                    *max_ptr.add(i) = val;
                    *class_ptr.add(i) = class_idx;
                }
            }
        }
    }
}

/// Fast argmax dispatching to NEON-optimized path for i8 on aarch64.
#[inline(always)]
fn fast_arg_max<T: PrimInt + Copy>(score: ArrayView1<T>) -> (T, usize) {
    #[cfg(target_arch = "aarch64")]
    {
        // Check if this is an i8 slice and contiguous.
        if std::mem::size_of::<T>() == 1 && score.as_slice().is_some() {
            let slice = score.as_slice().unwrap();
            // Safety: T is i8 when size_of::<T>() == 1 and PrimInt.
            // PrimInt covers i8, u8, i16, etc. We only want to use the
            // i8 NEON path for signed i8.
            let ptr = slice.as_ptr() as *const i8;
            let i8_slice = unsafe { std::slice::from_raw_parts(ptr, slice.len()) };
            // Only valid for signed i8 (not u8). Check sign bit behavior:
            // PrimInt for i8 means min_value() is negative.
            if T::min_value() < T::zero() {
                let (max_val, idx) = arg_max_i8(i8_slice);
                // Safety: transmute i8 back to T (they have the same size and
                // representation for i8).
                let result: T = unsafe { std::mem::transmute_copy(&max_val) };
                return (result, idx);
            }
        }
    }
    arg_max(score)
}

/// Post processes boxes and scores tensors into quantized detection boxes,
/// filtering out any boxes below the score threshold. The boxes tensor
/// is converted to XYXY using the given BBoxTypeTrait. The order of the boxes
/// is preserved.
#[doc(hidden)]
pub fn postprocess_boxes_quant<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> Vec<DetectBoxQuantized<Scores>> {
    assert_eq!(scores.dim().0, boxes.dim().0);
    assert_eq!(boxes.dim().1, 4);

    // Use column-major path for transposed DMA-BUF views (see postprocess_boxes_index_quant).
    if scores.strides()[0] == 1 && scores.as_slice().is_none() {
        return postprocess_boxes_quant_column_major::<B, _, _>(
            threshold,
            boxes,
            scores,
            quant_boxes,
        );
    }

    Zip::from(scores.rows())
        .and(boxes.rows())
        .into_par_iter()
        .filter_map(|(score, bbox)| {
            let (score_, label) = fast_arg_max(score);
            if score_ < threshold {
                return None;
            }

            let bbox_quant = B::ndarray_to_xyxy_dequant(bbox.view(), quant_boxes);
            Some(DetectBoxQuantized {
                label,
                score: score_,
                bbox: BoundingBox::from(bbox_quant),
            })
        })
        .collect()
}

/// Column-major optimized path for `postprocess_boxes_quant`.
fn postprocess_boxes_quant_column_major<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> Vec<DetectBoxQuantized<Scores>> {
    let (n_candidates, n_classes) = scores.dim();

    // Column-major NEON path uses u8 class indices; fall back for >255 classes.
    if n_classes > 255 {
        return Zip::from(scores.rows())
            .and(boxes.rows())
            .into_par_iter()
            .filter_map(|(score, bbox)| {
                let (score_, label) = fast_arg_max(score);
                if score_ < threshold {
                    return None;
                }
                let bbox_quant = B::ndarray_to_xyxy_dequant(bbox.view(), quant_boxes);
                Some(DetectBoxQuantized {
                    label,
                    score: score_,
                    bbox: BoundingBox::from(bbox_quant),
                })
            })
            .collect();
    }
    let mut max_scores = vec![Scores::min_value(); n_candidates];
    let mut max_classes = vec![0u8; n_candidates];

    for class_idx in 0..n_classes {
        let col = scores.column(class_idx);
        if let Some(slice) = col.as_slice() {
            #[cfg(target_arch = "aarch64")]
            {
                if std::mem::size_of::<Scores>() == 1 {
                    unsafe {
                        // Non-prefetch variant for heap-backed tensors;
                        // prefetch variant reserved for DMA-BUF CMA paths.
                        column_max_update_neon(
                            slice.as_ptr() as *const u8,
                            max_scores.as_mut_ptr() as *mut u8,
                            max_classes.as_mut_ptr(),
                            n_candidates,
                            class_idx as u8,
                            Scores::min_value() < Scores::zero(),
                        );
                    }
                    continue;
                }
            }
            for (i, &val) in slice.iter().enumerate() {
                if val >= max_scores[i] {
                    max_scores[i] = val;
                    max_classes[i] = class_idx as u8;
                }
            }
        } else {
            for (i, &val) in col.iter().enumerate() {
                if val >= max_scores[i] {
                    max_scores[i] = val;
                    max_classes[i] = class_idx as u8;
                }
            }
        }
    }

    // Copy boxes column-by-column if also transposed.
    let boxes_buf: [Vec<Boxes>; 4] = if boxes.strides()[0] == 1 && boxes.as_slice().is_none() {
        let mut cols: [Vec<Boxes>; 4] = [
            vec![Boxes::zero(); n_candidates],
            vec![Boxes::zero(); n_candidates],
            vec![Boxes::zero(); n_candidates],
            vec![Boxes::zero(); n_candidates],
        ];
        for (dim, col_buf) in cols.iter_mut().enumerate() {
            let col = boxes.column(dim);
            if let Some(slice) = col.as_slice() {
                col_buf.copy_from_slice(slice);
            } else {
                for (i, &val) in col.iter().enumerate() {
                    col_buf[i] = val;
                }
            }
        }
        cols
    } else {
        [vec![], vec![], vec![], vec![]]
    };
    let boxes_copied = !boxes_buf[0].is_empty();

    let mut result = Vec::new();
    for i in 0..n_candidates {
        if max_scores[i] >= threshold {
            let bbox_quant = if boxes_copied {
                let raw = [
                    boxes_buf[0][i],
                    boxes_buf[1][i],
                    boxes_buf[2][i],
                    boxes_buf[3][i],
                ];
                B::to_xyxy_dequant(&raw, quant_boxes)
            } else {
                B::ndarray_to_xyxy_dequant(boxes.row(i), quant_boxes)
            };
            result.push(DetectBoxQuantized {
                label: max_classes[i] as usize,
                score: max_scores[i],
                bbox: BoundingBox::from(bbox_quant),
            });
        }
    }

    result
}

/// Post processes boxes and scores tensors into quantized detection boxes,
/// filtering out any boxes below the score threshold. The boxes tensor
/// is converted to XYXY using the given BBoxTypeTrait. The order of the boxes
/// is preserved.
///
/// This function is very similar to `postprocess_boxes_quant` but will also
/// return the index of the box. The boxes will be in ascending index order.
///
/// When scores originate from a transposed DMA-BUF view (stride-1 along axis 0),
/// an optimized column-major scan is used to avoid catastrophic strided reads on
/// uncacheable memory.
#[doc(hidden)]
pub fn postprocess_boxes_index_quant<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> Vec<(DetectBoxQuantized<Scores>, usize)> {
    assert_eq!(scores.dim().0, boxes.dim().0);
    assert_eq!(boxes.dim().1, 4);

    // Detect transposed C-contiguous layout (e.g., from reversed_axes() on DMA-BUF).
    // In this layout columns are contiguous (stride-1 along axis 0) but rows are not.
    // The column-major path reads memory sequentially, avoiding cache-hostile strides.
    if scores.strides()[0] == 1 && scores.as_slice().is_none() {
        return postprocess_boxes_index_quant_column_major::<B, _, _>(
            threshold,
            boxes,
            scores,
            quant_boxes,
        );
    }

    let indices: Array1<usize> = (0..boxes.dim().0).collect();
    Zip::from(scores.rows())
        .and(boxes.rows())
        .and(&indices)
        .into_par_iter()
        .filter_map(|(score, bbox, index)| {
            let (score_, label) = fast_arg_max(score);
            if score_ < threshold {
                return None;
            }

            let bbox_quant = B::ndarray_to_xyxy_dequant(bbox.view(), quant_boxes);

            Some((
                DetectBoxQuantized {
                    label,
                    score: score_,
                    bbox: BoundingBox::from(bbox_quant),
                },
                *index,
            ))
        })
        .collect()
}

/// Multi-label variant of [`postprocess_boxes_index_quant`]: one entry per
/// class at or above `threshold`, sharing the anchor's bbox and index.
///
/// Intended for validation decode; deployment uses the argmax variant.
///
/// # Errors
/// [`DecoderError::InvalidShape`] unless boxes are `[N, 4]` and scores `[N, C]`.
pub fn postprocess_boxes_multilabel_index_quant<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> DecoderResult<Vec<(DetectBoxQuantized<Scores>, usize)>> {
    check_candidate_shapes(boxes.dim(), scores.dim())?;
    if scores.strides()[0] == 1 && scores.as_slice().is_none() {
        return Ok(postprocess_boxes_multilabel_index_quant_column_major::<
            B,
            _,
            _,
        >(threshold, boxes, scores, quant_boxes));
    }
    let indices: Array1<usize> = (0..boxes.dim().0).collect();
    Ok(Zip::from(scores.rows())
        .and(boxes.rows())
        .and(&indices)
        .into_par_iter()
        .flat_map(|(score_row, bbox_row, &anchor_idx)| {
            let bbox = BoundingBox::from(B::ndarray_to_xyxy_dequant(bbox_row.view(), quant_boxes));
            score_row
                .iter()
                .enumerate()
                .filter(|(_, &s)| s >= threshold)
                .map(move |(label, &score)| (DetectBoxQuantized { label, score, bbox }, anchor_idx))
                .collect::<Vec<_>>()
        })
        .collect())
}

/// [`postprocess_boxes_multilabel_index_quant`] without anchor indices.
///
/// # Errors
/// [`DecoderError::InvalidShape`] unless boxes are `[N, 4]` and scores `[N, C]`.
pub fn postprocess_boxes_multilabel_quant<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> DecoderResult<Vec<DetectBoxQuantized<Scores>>> {
    Ok(
        postprocess_boxes_multilabel_index_quant::<B, _, _>(threshold, boxes, scores, quant_boxes)?
            .into_iter()
            .map(|(b, _)| b)
            .collect(),
    )
}

/// Column-major path for [`postprocess_boxes_multilabel_index_quant`].
///
/// Reads each class column contiguously (the transposed DMA-BUF layout),
/// then orders hits by `(anchor, class)` to match the row-major output and
/// dequantizes each anchor's bbox once.
fn postprocess_boxes_multilabel_index_quant_column_major<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> Vec<(DetectBoxQuantized<Scores>, usize)> {
    let mut hits: Vec<(usize, usize, Scores)> = (0..scores.dim().1)
        .into_par_iter()
        .flat_map_iter(|class_idx| {
            let col = scores.column(class_idx);
            let above = |(i, &s): (usize, &Scores)| (s >= threshold).then_some((i, class_idx, s));
            match col.as_slice() {
                Some(slice) => slice
                    .iter()
                    .enumerate()
                    .filter_map(above)
                    .collect::<Vec<_>>(),
                None => col.iter().enumerate().filter_map(above).collect(),
            }
        })
        .collect();
    hits.par_sort_unstable_by_key(|&(anchor, class, _)| (anchor, class));

    let mut out = Vec::with_capacity(hits.len());
    let mut current: Option<(usize, BoundingBox)> = None;
    for (anchor, label, score) in hits {
        let bbox = match current {
            Some((a, bbox)) if a == anchor => bbox,
            _ => {
                let bbox =
                    BoundingBox::from(B::ndarray_to_xyxy_dequant(boxes.row(anchor), quant_boxes));
                current = Some((anchor, bbox));
                bbox
            }
        };
        out.push((DetectBoxQuantized { label, score, bbox }, anchor));
    }
    out
}

/// Candidate generators need `[N, 4]` boxes and `[N, C]` scores.
pub(crate) fn check_candidate_shapes(
    boxes: (usize, usize),
    scores: (usize, usize),
) -> DecoderResult<()> {
    if boxes.1 != 4 || boxes.0 != scores.0 {
        return Err(DecoderError::InvalidShape(format!(
            "candidate selection needs boxes [N, 4] and scores [N, C]; got boxes {boxes:?}, scores {scores:?}"
        )));
    }
    Ok(())
}

/// Column-major optimized path for `postprocess_boxes_index_quant`.
///
/// When scores come from a transposed DMA-BUF view ([N_candidates, N_classes]
/// with strides [1, N_candidates]), row iteration causes N_classes reads each
/// N_candidates bytes apart — catastrophic on uncacheable memory. Instead, this
/// iterates over classes (columns, which are contiguous), maintaining a running
/// argmax per candidate in cacheable heap buffers.
fn postprocess_boxes_index_quant_column_major<
    B: BBoxTypeTrait,
    Boxes: PrimInt + AsPrimitive<f32> + Send + Sync,
    Scores: PrimInt + AsPrimitive<f32> + Send + Sync,
>(
    threshold: Scores,
    boxes: ArrayView2<Boxes>,
    scores: ArrayView2<Scores>,
    quant_boxes: Quantization,
) -> Vec<(DetectBoxQuantized<Scores>, usize)> {
    let (n_candidates, n_classes) = scores.dim();

    // Step 1: Column-based argmax — sequential reads over contiguous columns.
    // Use u8 for class indices (max 255 classes) for optimal NEON vectorization.
    // Fall back to row-major path for models with >255 classes.
    if n_classes > 255 {
        let indices: Array1<usize> = (0..n_candidates).collect();
        return Zip::from(scores.rows())
            .and(boxes.rows())
            .and(&indices)
            .into_par_iter()
            .filter_map(|(score, bbox, index)| {
                let (score_, label) = fast_arg_max(score);
                if score_ < threshold {
                    return None;
                }
                let bbox_quant = B::ndarray_to_xyxy_dequant(bbox.view(), quant_boxes);
                Some((
                    DetectBoxQuantized {
                        label,
                        score: score_,
                        bbox: BoundingBox::from(bbox_quant),
                    },
                    *index,
                ))
            })
            .collect();
    }
    let mut max_scores = vec![Scores::min_value(); n_candidates];
    let mut max_classes = vec![0u8; n_candidates];
    for class_idx in 0..n_classes {
        let col = scores.column(class_idx);
        if let Some(slice) = col.as_slice() {
            // Use NEON-accelerated column max update on aarch64 for u8 scores.
            #[cfg(target_arch = "aarch64")]
            {
                if std::mem::size_of::<Scores>() == 1 {
                    // SAFETY: Scores is u8 or i8 (size == 1). We transmute the
                    // slice pointers to the concrete byte type for NEON processing.
                    unsafe {
                        column_max_update_neon(
                            slice.as_ptr() as *const u8,
                            max_scores.as_mut_ptr() as *mut u8,
                            max_classes.as_mut_ptr(),
                            n_candidates,
                            class_idx as u8,
                            Scores::min_value() < Scores::zero(), // signed flag
                        );
                    }
                    continue;
                }
            }
            for (i, &val) in slice.iter().enumerate() {
                if val >= max_scores[i] {
                    max_scores[i] = val;
                    max_classes[i] = class_idx as u8;
                }
            }
        } else {
            for (i, &val) in col.iter().enumerate() {
                if val >= max_scores[i] {
                    max_scores[i] = val;
                    max_classes[i] = class_idx as u8;
                }
            }
        }
    }

    // Step 2: Copy boxes column-by-column into contiguous heap buffer.
    // Boxes view is also transposed [N_candidates, 4] with strides [1, N_candidates],
    // so column reads are sequential while row reads are strided.
    let boxes_buf: [Vec<Boxes>; 4] = if boxes.strides()[0] == 1 && boxes.as_slice().is_none() {
        let mut cols: [Vec<Boxes>; 4] = [
            vec![Boxes::zero(); n_candidates],
            vec![Boxes::zero(); n_candidates],
            vec![Boxes::zero(); n_candidates],
            vec![Boxes::zero(); n_candidates],
        ];
        for (dim, col_buf) in cols.iter_mut().enumerate() {
            let col = boxes.column(dim);
            if let Some(slice) = col.as_slice() {
                col_buf.copy_from_slice(slice);
            } else {
                for (i, &val) in col.iter().enumerate() {
                    col_buf[i] = val;
                }
            }
        }
        cols
    } else {
        // Boxes are contiguous or differently strided — read per-candidate below.
        [vec![], vec![], vec![], vec![]]
    };
    let boxes_copied = !boxes_buf[0].is_empty();

    // Step 3: Threshold filter — collect candidates that pass.
    let mut result = Vec::new();
    for i in 0..n_candidates {
        if max_scores[i] >= threshold {
            let bbox_quant = if boxes_copied {
                let raw = [
                    boxes_buf[0][i],
                    boxes_buf[1][i],
                    boxes_buf[2][i],
                    boxes_buf[3][i],
                ];
                B::to_xyxy_dequant(&raw, quant_boxes)
            } else {
                B::ndarray_to_xyxy_dequant(boxes.row(i), quant_boxes)
            };
            result.push((
                DetectBoxQuantized {
                    label: max_classes[i] as usize,
                    score: max_scores[i],
                    bbox: BoundingBox::from(bbox_quant),
                },
                i,
            ));
        }
    }

    result
}

/// Uses NMS to filter boxes based on the score and iou. Sorts boxes by score,
/// then greedily selects a subset of boxes in descending order of score.
#[doc(hidden)]
#[must_use]
pub fn nms_int<SCORE: PrimInt + AsPrimitive<f32> + Send + Sync>(
    iou: f32,
    max_det: Option<usize>,
    mut boxes: Vec<DetectBoxQuantized<SCORE>>,
) -> Vec<DetectBoxQuantized<SCORE>> {
    // Boxes get sorted by score in descending order so we know based on the
    // index the scoring of the boxes and can skip parts of the loop.

    boxes.par_sort_by(|a, b| b.score.cmp(&a.score));

    // When the iou is 1.0 or larger, no boxes will be filtered so we just return
    // immediately
    if iou >= 1.0 {
        return match max_det {
            Some(n) => {
                boxes.truncate(n);
                boxes
            }
            None => boxes,
        };
    }

    let min_val = SCORE::min_value();
    let cap = max_det.unwrap_or(usize::MAX);
    let mut survivors: usize = 0;
    // Outer loop over all boxes.
    for i in 0..boxes.len() {
        if boxes[i].score <= min_val {
            // this box was merged with a different box earlier
            continue;
        }
        for j in (i + 1)..boxes.len() {
            // Inner loop over boxes with lower score (later in the list).

            if boxes[j].score <= min_val {
                // this box was suppressed by different box earlier
                continue;
            }

            if jaccard(&boxes[j].bbox, &boxes[i].bbox, iou) {
                // suppress this box
                boxes[j].score = min_val;
            }
        }
        survivors += 1;
        if survivors >= cap {
            break;
        }
    }
    // Filter out boxes that were suppressed; cap because boxes after the
    // break may still hold positive scores but are all lower than survivors.
    boxes
        .into_iter()
        .filter(|b| b.score > min_val)
        .take(cap)
        .collect()
}

/// Uses NMS to filter boxes based on the score and iou. Sorts boxes by score,
/// then greedily selects a subset of boxes in descending order of score.
///
/// This is same as `nms_int` but will also include extra information along
/// with each box, such as the index
#[doc(hidden)]
#[must_use]
pub fn nms_extra_int<SCORE: PrimInt + AsPrimitive<f32> + Send + Sync, E: Send + Sync>(
    iou: f32,
    max_det: Option<usize>,
    mut boxes: Vec<(DetectBoxQuantized<SCORE>, E)>,
) -> Vec<(DetectBoxQuantized<SCORE>, E)> {
    // Boxes get sorted by score in descending order so we know based on the
    // index the scoring of the boxes and can skip parts of the loop.
    boxes.par_sort_by(|a, b| b.0.score.cmp(&a.0.score));

    // When the iou is 1.0 or larger, no boxes will be filtered so we just return
    // immediately
    if iou >= 1.0 {
        return match max_det {
            Some(n) => {
                boxes.truncate(n);
                boxes
            }
            None => boxes,
        };
    }

    let min_val = SCORE::min_value();
    let cap = max_det.unwrap_or(usize::MAX);
    let mut survivors: usize = 0;
    // Outer loop over all boxes.
    for i in 0..boxes.len() {
        if boxes[i].0.score <= min_val {
            // this box was merged with a different box earlier
            continue;
        }
        for j in (i + 1)..boxes.len() {
            // Inner loop over boxes with lower score (later in the list).

            if boxes[j].0.score <= min_val {
                // this box was suppressed by different box earlier
                continue;
            }
            if jaccard(&boxes[j].0.bbox, &boxes[i].0.bbox, iou) {
                // suppress this box
                boxes[j].0.score = min_val;
            }
        }
        survivors += 1;
        if survivors >= cap {
            break;
        }
    }

    // Filter out boxes that were suppressed; cap at `max_det`.
    boxes
        .into_iter()
        .filter(|b| b.0.score > min_val)
        .take(cap)
        .collect()
}

/// Class-aware NMS for quantized boxes: only suppress boxes with the same
/// label.
///
/// Sorts boxes by score, then greedily selects a subset of boxes in descending
/// order of score. Unlike class-agnostic NMS, boxes are only suppressed if they
/// have the same class label AND overlap above the IoU threshold.
#[doc(hidden)]
#[must_use]
pub fn nms_class_aware_int<SCORE: PrimInt + AsPrimitive<f32> + Send + Sync>(
    iou: f32,
    max_det: Option<usize>,
    mut boxes: Vec<DetectBoxQuantized<SCORE>>,
) -> Vec<DetectBoxQuantized<SCORE>> {
    boxes.par_sort_by(|a, b| b.score.cmp(&a.score));

    // When the iou is 1.0 or larger, no boxes will be filtered so we just return
    // immediately
    if iou >= 1.0 {
        return match max_det {
            Some(n) => {
                boxes.truncate(n);
                boxes
            }
            None => boxes,
        };
    }

    let min_val = SCORE::min_value();
    let cap = max_det.unwrap_or(usize::MAX);
    let mut survivors: usize = 0;
    for i in 0..boxes.len() {
        if boxes[i].score <= min_val {
            continue;
        }
        for j in (i + 1)..boxes.len() {
            if boxes[j].score <= min_val {
                continue;
            }
            // Only suppress if same class AND overlapping
            if boxes[j].label == boxes[i].label && jaccard(&boxes[j].bbox, &boxes[i].bbox, iou) {
                boxes[j].score = min_val;
            }
        }
        survivors += 1;
        if survivors >= cap {
            break;
        }
    }
    boxes
        .into_iter()
        .filter(|b| b.score > min_val)
        .take(cap)
        .collect()
}

/// Class-aware NMS for quantized boxes with extra data: only suppress boxes
/// with the same label.
///
/// This is same as `nms_class_aware_int` but will also include extra
/// information along with each box, such as the index.
#[doc(hidden)]
#[must_use]
pub fn nms_extra_class_aware_int<
    SCORE: PrimInt + AsPrimitive<f32> + Send + Sync,
    E: Send + Sync,
>(
    iou: f32,
    max_det: Option<usize>,
    mut boxes: Vec<(DetectBoxQuantized<SCORE>, E)>,
) -> Vec<(DetectBoxQuantized<SCORE>, E)> {
    boxes.par_sort_by(|a, b| b.0.score.cmp(&a.0.score));

    // When the iou is 1.0 or larger, no boxes will be filtered so we just return
    // immediately
    if iou >= 1.0 {
        return match max_det {
            Some(n) => {
                boxes.truncate(n);
                boxes
            }
            None => boxes,
        };
    }

    let min_val = SCORE::min_value();
    let cap = max_det.unwrap_or(usize::MAX);
    let mut survivors: usize = 0;
    for i in 0..boxes.len() {
        if boxes[i].0.score <= min_val {
            continue;
        }
        for j in (i + 1)..boxes.len() {
            if boxes[j].0.score <= min_val {
                continue;
            }
            // Only suppress if same class AND overlapping
            if boxes[j].0.label == boxes[i].0.label
                && jaccard(&boxes[j].0.bbox, &boxes[i].0.bbox, iou)
            {
                boxes[j].0.score = min_val;
            }
        }
        survivors += 1;
        if survivors >= cap {
            break;
        }
    }
    boxes
        .into_iter()
        .filter(|b| b.0.score > min_val)
        .take(cap)
        .collect()
}

/// Quantizes a score from f32 to the given integer type, using the following
/// formula `(score/quant.scale + quant.zero_point).ceil()`, then clamping to
/// the min and max value of the given integer type
///
/// # Examples
/// ```rust
/// use edgefirst_decoder::{Quantization, byte::quantize_score_threshold};
/// let quant = Quantization {
///     scale: 0.1,
///     zero_point: 128,
/// };
/// let q: u8 = quantize_score_threshold::<u8>(0.5, quant);
/// assert_eq!(q, 128 + 5);
/// ```
#[doc(hidden)]
pub fn quantize_score_threshold<T: PrimInt + AsPrimitive<f32>>(score: f32, quant: Quantization) -> T
where
    f32: AsPrimitive<T>,
{
    if quant.scale == 0.0 {
        return T::max_value();
    }
    let v = (score / quant.scale + quant.zero_point as f32).ceil();
    let v = v.clamp(T::min_value().as_(), T::max_value().as_());
    v.as_()
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{XYWH, XYXY};
    use ndarray::Array2;

    fn unit_quant() -> Quantization {
        Quantization {
            scale: 0.1,
            zero_point: 0,
        }
    }

    fn sorted_by_anchor_label(
        mut v: Vec<(DetectBoxQuantized<i8>, usize)>,
    ) -> Vec<(DetectBoxQuantized<i8>, usize)> {
        v.sort_by_key(|(b, i)| (*i, b.label));
        v
    }

    #[test]
    fn multilabel_index_quant_emits_every_class_above_threshold() {
        let boxes = ndarray::arr2(&[[10i8, 10, 20, 20], [30, 30, 40, 40]]);
        let scores = ndarray::arr2(&[[60i8, 10, 70], [5, 90, 5]]);
        let got = sorted_by_anchor_label(
            postprocess_boxes_multilabel_index_quant::<XYXY, _, _>(
                50,
                boxes.view(),
                scores.view(),
                unit_quant(),
            )
            .unwrap(),
        );
        let labels: Vec<(usize, usize)> = got.iter().map(|(b, i)| (*i, b.label)).collect();
        assert_eq!(labels, vec![(0, 0), (0, 2), (1, 1)]);
        assert_eq!(got[0].0.bbox, got[1].0.bbox);
        assert_eq!(got[1].0.score, 70);
    }

    #[test]
    fn multilabel_index_quant_column_major_matches_row_major() {
        let boxes = ndarray::arr2(&[[10i8, 10, 20, 20], [30, 30, 40, 40]]);
        let scores_t = ndarray::arr2(&[[60i8, 5], [10, 90], [70, 5]]);
        let scores_cm = scores_t.view().reversed_axes();
        let scores_rm = scores_cm.to_owned();
        let a = postprocess_boxes_multilabel_index_quant::<XYXY, _, _>(
            50,
            boxes.view(),
            scores_cm,
            unit_quant(),
        )
        .unwrap();
        let b = postprocess_boxes_multilabel_index_quant::<XYXY, _, _>(
            50,
            boxes.view(),
            scores_rm.view(),
            unit_quant(),
        )
        .unwrap();
        assert_eq!(sorted_by_anchor_label(a), sorted_by_anchor_label(b));
    }

    #[test]
    fn multilabel_index_quant_rejects_mismatched_shapes() {
        let boxes = ndarray::arr2(&[[10i8, 10, 20, 20]]);
        let scores = ndarray::arr2(&[[60i8, 10], [5, 90]]);
        let r = postprocess_boxes_multilabel_index_quant::<XYXY, _, _>(
            50,
            boxes.view(),
            scores.view(),
            unit_quant(),
        );
        assert!(matches!(r, Err(crate::DecoderError::InvalidShape(_))));
    }

    #[test]
    fn multilabel_quant_rejects_non_four_column_boxes() {
        let boxes = ndarray::arr2(&[[10i8, 10, 20, 20, 0]]);
        let scores = ndarray::arr2(&[[60i8]]);
        let r = postprocess_boxes_multilabel_quant::<XYXY, _, _>(
            50,
            boxes.view(),
            scores.view(),
            unit_quant(),
        );
        assert!(matches!(r, Err(crate::DecoderError::InvalidShape(_))));
    }

    #[test]
    fn multilabel_quant_drops_indices_only() {
        let boxes = ndarray::arr2(&[[10i8, 10, 20, 20], [30, 30, 40, 40]]);
        let scores = ndarray::arr2(&[[60i8, 10, 70], [5, 90, 5]]);
        let with_idx = postprocess_boxes_multilabel_index_quant::<XYXY, _, _>(
            50,
            boxes.view(),
            scores.view(),
            unit_quant(),
        )
        .unwrap();
        let without = postprocess_boxes_multilabel_quant::<XYXY, _, _>(
            50,
            boxes.view(),
            scores.view(),
            unit_quant(),
        )
        .unwrap();
        let stripped: Vec<_> = with_idx.into_iter().map(|(b, _)| b).collect();
        assert_eq!(stripped, without);
    }

    /// Verify that the column-major path produces identical results to the
    /// row-major path for a transposed (non-contiguous) score array.
    /// Multi-label on a transposed (DMA-BUF style) score view must match the
    /// contiguous row-major result exactly, including order.
    fn multilabel_layouts_agree<S>(n_classes: usize, n_candidates: usize, threshold: S)
    where
        S: PrimInt + AsPrimitive<f32> + Send + Sync + std::fmt::Debug,
        usize: AsPrimitive<S>,
    {
        let mut scores_physical = Array2::<S>::zeros((n_classes, n_candidates));
        for c in 0..n_classes {
            for i in 0..n_candidates {
                scores_physical[[c, i]] = ((c * 3 + i * 7) % 100).as_();
            }
        }
        let mut boxes_physical = Array2::<i16>::zeros((4, n_candidates));
        for i in 0..n_candidates {
            boxes_physical[[0, i]] = (i * 10) as i16;
            boxes_physical[[1, i]] = (i * 20) as i16;
            boxes_physical[[2, i]] = (i * 10 + 50) as i16;
            boxes_physical[[3, i]] = (i * 20 + 100) as i16;
        }
        let quant = Quantization {
            scale: 0.00390625,
            zero_point: 0,
        };
        let scores_rm = scores_physical.clone().reversed_axes().to_owned();
        let boxes_rm = boxes_physical.clone().reversed_axes().to_owned();
        let row = postprocess_boxes_multilabel_index_quant::<XYWH, _, _>(
            threshold,
            boxes_rm.view(),
            scores_rm.view(),
            quant,
        )
        .unwrap();
        let scores_cm = scores_physical.view().reversed_axes();
        assert!(scores_cm.as_slice().is_none() && scores_cm.strides()[0] == 1);
        let col = postprocess_boxes_multilabel_index_quant::<XYWH, _, _>(
            threshold,
            boxes_physical.view().reversed_axes(),
            scores_cm,
            quant,
        )
        .unwrap();
        assert!(!row.is_empty());
        assert_eq!(row, col);
    }

    #[test]
    fn multilabel_column_major_matches_row_major_u8() {
        multilabel_layouts_agree::<u8>(80, 100, 60);
    }

    #[test]
    fn multilabel_column_major_matches_row_major_i8_many_classes() {
        multilabel_layouts_agree::<i8>(300, 64, 90);
    }

    #[test]
    fn candidate_shape_error_names_candidate_selection() {
        let err = check_candidate_shapes((3, 5), (3, 2)).unwrap_err();
        let msg = format!("{err}");
        assert!(msg.contains("candidate selection"), "{msg}");
        assert!(!msg.contains("multi-label"), "{msg}");
    }

    #[test]
    fn column_major_matches_row_major() {
        // Create scores in "model output" layout: [num_classes, num_candidates]
        let n_classes = 80usize;
        let n_candidates = 100usize;
        let mut scores_physical = Array2::<u8>::zeros((n_classes, n_candidates));
        // Fill with known pattern: class c, candidate i → (c * 3 + i * 7) % 256
        for c in 0..n_classes {
            for i in 0..n_candidates {
                scores_physical[[c, i]] = ((c * 3 + i * 7) % 256) as u8;
            }
        }

        // Create boxes: [4, num_candidates] i16
        let mut boxes_physical = Array2::<i16>::zeros((4, n_candidates));
        for i in 0..n_candidates {
            boxes_physical[[0, i]] = (i * 10) as i16; // x
            boxes_physical[[1, i]] = (i * 20) as i16; // y
            boxes_physical[[2, i]] = (i * 10 + 50) as i16; // w
            boxes_physical[[3, i]] = (i * 20 + 100) as i16; // h
        }

        let quant = Quantization {
            scale: 0.00390625,
            zero_point: 0,
        };

        let threshold: u8 = 10;

        // Row-major path: contiguous [n_candidates, n_classes] array
        let scores_contiguous = scores_physical.clone().reversed_axes().to_owned();
        let boxes_contiguous = boxes_physical.clone().reversed_axes().to_owned();
        let row_result = postprocess_boxes_index_quant::<XYWH, _, _>(
            threshold,
            boxes_contiguous.view(),
            scores_contiguous.view(),
            quant,
        );

        // Column-major path: non-contiguous reversed view
        let scores_view = scores_physical.view().reversed_axes();
        let boxes_view = boxes_physical.view().reversed_axes();
        assert!(scores_view.as_slice().is_none(), "should be non-contiguous");
        assert_eq!(scores_view.strides()[0], 1);
        let col_result =
            postprocess_boxes_index_quant::<XYWH, _, _>(threshold, boxes_view, scores_view, quant);

        // Both paths should produce the same results
        assert_eq!(
            row_result.len(),
            col_result.len(),
            "different number of results: row={}, col={}",
            row_result.len(),
            col_result.len()
        );
        for (i, (row, col)) in row_result.iter().zip(col_result.iter()).enumerate() {
            assert_eq!(
                row.0.label, col.0.label,
                "candidate {i}: label mismatch row={} col={}",
                row.0.label, col.0.label
            );
            assert_eq!(row.0.score, col.0.score, "candidate {i}: score mismatch");
            assert_eq!(row.1, col.1, "candidate {i}: index mismatch");
            assert_eq!(row.0.bbox, col.0.bbox, "candidate {i}: bbox mismatch");
        }
    }

    /// Test column-major path with i8 scores (signed, matches NEON argmax path).
    #[test]
    fn column_major_matches_row_major_i8() {
        let n_classes = 80usize;
        let n_candidates = 50usize;
        let mut scores_physical = Array2::<i8>::zeros((n_classes, n_candidates));
        for c in 0..n_classes {
            for i in 0..n_candidates {
                scores_physical[[c, i]] = ((c as i16 * 3 + i as i16 * 7) % 256 - 128) as i8;
            }
        }

        let mut boxes_physical = Array2::<i16>::zeros((4, n_candidates));
        for i in 0..n_candidates {
            boxes_physical[[0, i]] = (i * 10) as i16;
            boxes_physical[[1, i]] = (i * 20) as i16;
            boxes_physical[[2, i]] = (i * 10 + 50) as i16;
            boxes_physical[[3, i]] = (i * 20 + 100) as i16;
        }

        let quant = Quantization {
            scale: 0.0256,
            zero_point: -116,
        };
        let threshold: i8 = -100;

        let scores_contiguous = scores_physical.clone().reversed_axes().to_owned();
        let boxes_contiguous = boxes_physical.clone().reversed_axes().to_owned();
        let row_result = postprocess_boxes_index_quant::<XYWH, _, _>(
            threshold,
            boxes_contiguous.view(),
            scores_contiguous.view(),
            quant,
        );

        let scores_view = scores_physical.view().reversed_axes();
        let boxes_view = boxes_physical.view().reversed_axes();
        let col_result =
            postprocess_boxes_index_quant::<XYWH, _, _>(threshold, boxes_view, scores_view, quant);

        assert_eq!(row_result.len(), col_result.len());
        for (i, (row, col)) in row_result.iter().zip(col_result.iter()).enumerate() {
            assert_eq!(row.0.label, col.0.label, "i8 candidate {i}: label mismatch");
            assert_eq!(row.0.score, col.0.score, "i8 candidate {i}: score mismatch");
            assert_eq!(row.1, col.1, "i8 candidate {i}: index mismatch");
        }
    }

    /// Helper: create `n` non-overlapping boxes with descending u8 scores.
    fn make_nms_boxes_int(n: usize) -> Vec<DetectBoxQuantized<u8>> {
        (0..n)
            .map(|i| DetectBoxQuantized {
                bbox: BoundingBox {
                    xmin: i as f32 * 100.0,
                    ymin: 0.0,
                    xmax: i as f32 * 100.0 + 10.0,
                    ymax: 10.0,
                },
                label: 0,
                score: (200 - i as u32).min(255) as u8,
            })
            .collect()
    }

    /// Four candidates' worth of boxes, one row each, values distinct enough
    /// that a mis-indexed read is visible.
    fn probe_boxes(n: usize) -> Array2<i16> {
        let mut boxes = Array2::<i16>::zeros((n, 4));
        for i in 0..n {
            boxes[[i, 0]] = i as i16 * 10;
            boxes[[i, 1]] = i as i16 * 10 + 1;
            boxes[[i, 2]] = i as i16 * 10 + 2;
            boxes[[i, 3]] = i as i16 * 10 + 3;
        }
        boxes
    }

    fn probe_quant() -> Quantization {
        Quantization {
            scale: 1.0,
            zero_point: 0,
        }
    }

    /// The threshold is a floor: a score equal to it is a detection. Both
    /// entry points share the rule and neither had a test that sat exactly on
    /// it.
    #[test]
    fn postprocess_boxes_quant_keeps_a_score_equal_to_the_threshold() {
        let boxes = probe_boxes(2);
        let scores = ndarray::array![[200_u8, 0], [199, 0]];

        let kept = postprocess_boxes_quant::<XYXY, _, _>(
            200_u8,
            boxes.view(),
            scores.view(),
            probe_quant(),
        );
        assert_eq!(kept.len(), 1, "{kept:?}");
        assert_eq!(kept[0].score, 200);

        let kept = postprocess_boxes_index_quant::<XYXY, _, _>(
            200_u8,
            boxes.view(),
            scores.view(),
            probe_quant(),
        );
        assert_eq!(kept.len(), 1, "{kept:?}");
        assert_eq!(kept[0].1, 0);
    }

    /// Class indices in the column-major scan are `u8`, so beyond 255 classes
    /// it has to hand over to the row-major scan. Running the column scan
    /// anyway wraps the label: class 260 comes back as 4.
    #[test]
    fn column_major_falls_back_beyond_255_classes() {
        let n_classes = 300usize;
        let boxes = probe_boxes(2);
        let mut scores = Array2::<u8>::zeros((2, n_classes));
        scores[[0, 260]] = 200;
        scores[[1, 3]] = 10;

        let kept = postprocess_boxes_quant_column_major::<XYXY, _, _>(
            200_u8,
            boxes.view(),
            scores.view(),
            probe_quant(),
        );
        assert_eq!(kept.len(), 1, "{kept:?}");
        assert_eq!(kept[0].label, 260);
        assert_eq!(kept[0].score, 200);

        let kept = postprocess_boxes_index_quant_column_major::<XYXY, _, _>(
            200_u8,
            boxes.view(),
            scores.view(),
            probe_quant(),
        );
        assert_eq!(kept.len(), 1, "{kept:?}");
        assert_eq!(kept[0].0.label, 260);
        assert_eq!(kept[0].1, 0);
    }

    /// Driving the column-major scan with an ordinary row-major array: its
    /// columns are not contiguous, so the running argmax takes its strided
    /// branch, and the boxes are read in place rather than from the
    /// column-copy buffers.
    #[test]
    fn column_major_scan_handles_row_major_input() {
        let boxes = probe_boxes(3);
        // Row-major (3 candidates, 4 classes); the winner differs per row and
        // is never class 0, so a comparison that keeps the minimum is visible.
        let scores = ndarray::array![[10_u8, 90, 30, 40], [5, 6, 7, 200], [1, 2, 3, 4]];

        let kept = postprocess_boxes_quant_column_major::<XYXY, _, _>(
            50_u8,
            boxes.view(),
            scores.view(),
            probe_quant(),
        );
        assert_eq!(kept.len(), 2, "{kept:?}");
        assert_eq!((kept[0].label, kept[0].score), (1, 90));
        assert_eq!((kept[1].label, kept[1].score), (3, 200));
        // Boxes were read in place, so row 1's values must line up with it.
        assert_eq!(kept[1].bbox.xmin, 10.0);

        let kept = postprocess_boxes_index_quant_column_major::<XYXY, _, _>(
            50_u8,
            boxes.view(),
            scores.view(),
            probe_quant(),
        );
        assert_eq!(kept.len(), 2, "{kept:?}");
        assert_eq!(kept[0].1, 0);
        assert_eq!(kept[1].1, 1);
        assert_eq!(kept[1].0.label, 3);
    }

    /// Two overlapping same-class boxes. `SCORE::min_value()` is the
    /// suppression marker, so for `u8` a score of 0 already means suppressed
    /// and the lower box here is given a real score.
    fn overlapping_pair_int() -> Vec<DetectBoxQuantized<u8>> {
        vec![
            DetectBoxQuantized {
                bbox: BoundingBox {
                    xmin: 0.0,
                    ymin: 0.0,
                    xmax: 10.0,
                    ymax: 10.0,
                },
                label: 0,
                score: 200,
            },
            DetectBoxQuantized {
                bbox: BoundingBox {
                    xmin: 1.0,
                    ymin: 1.0,
                    xmax: 11.0,
                    ymax: 11.0,
                },
                label: 0,
                score: 180,
            },
        ]
    }

    /// A overlaps B, B overlaps C, A does not reach C.
    fn suppression_chain_int() -> Vec<DetectBoxQuantized<u8>> {
        let mk = |xmin: f32, xmax: f32, score: u8| DetectBoxQuantized {
            bbox: BoundingBox {
                xmin,
                ymin: 0.0,
                xmax,
                ymax: 10.0,
            },
            label: 0,
            score,
        };
        vec![mk(0.0, 10.0, 200), mk(2.0, 12.0, 180), mk(4.0, 14.0, 160)]
    }

    #[test]
    fn nms_class_aware_int_suppresses_same_class_without_a_cap() {
        let result = nms_class_aware_int(0.5, None, overlapping_pair_int());
        assert_eq!(result.len(), 1, "{result:?}");
        assert_eq!(result[0].score, 200);
    }

    #[test]
    fn nms_class_aware_int_keeps_overlapping_boxes_of_other_classes() {
        let mut boxes = overlapping_pair_int();
        boxes[1].label = 1;
        let result = nms_class_aware_int(0.5, None, boxes);
        assert_eq!(result.len(), 2, "{result:?}");
    }

    #[test]
    fn nms_class_aware_int_does_not_suppress_on_behalf_of_a_suppressed_box() {
        let result = nms_class_aware_int(0.5, None, suppression_chain_int());
        assert_eq!(result.len(), 2, "{result:?}");
        assert_eq!(result[0].score, 200);
        assert_eq!(result[1].score, 160);
    }

    /// A suppressed box is marked with `min_value()`, and the final filter
    /// drops it. Keeping `>= min_value` instead would hand every suppressed
    /// box back to the caller.
    #[test]
    fn nms_class_aware_int_drops_boxes_already_at_the_minimum_score() {
        let mut boxes = overlapping_pair_int();
        boxes.push(DetectBoxQuantized {
            bbox: BoundingBox {
                xmin: 50.0,
                ymin: 50.0,
                xmax: 60.0,
                ymax: 60.0,
            },
            label: 0,
            score: u8::MIN,
        });
        let result = nms_class_aware_int(0.5, None, boxes);
        assert_eq!(result.len(), 1, "{result:?}");
        assert_eq!(result[0].score, 200);
    }

    /// A box that overlaps nothing, then a pair that overlaps each other. The
    /// first box confirms a survivor without suppressing anything, so a loop
    /// that stops at the first survivor leaves the pair un-deduplicated while
    /// the chain above, where the first box does the suppressing, looks
    /// identical either way.
    fn isolated_then_overlapping_pair_int() -> Vec<DetectBoxQuantized<u8>> {
        let mk = |xmin: f32, xmax: f32, score: u8| DetectBoxQuantized {
            bbox: BoundingBox {
                xmin,
                ymin: 0.0,
                xmax,
                ymax: 10.0,
            },
            label: 0,
            score,
        };
        vec![
            mk(100.0, 110.0, 200),
            mk(0.0, 10.0, 180),
            mk(1.0, 11.0, 160),
        ]
    }

    #[test]
    fn nms_class_aware_int_keeps_suppressing_after_the_first_survivor() {
        let result = nms_class_aware_int(0.5, None, isolated_then_overlapping_pair_int());
        assert_eq!(result.len(), 2, "{result:?}");
        assert_eq!(result[0].score, 200);
        assert_eq!(result[1].score, 180);
    }

    fn with_index_int(boxes: Vec<DetectBoxQuantized<u8>>) -> Vec<(DetectBoxQuantized<u8>, usize)> {
        boxes.into_iter().enumerate().map(|(i, b)| (b, i)).collect()
    }

    #[test]
    fn nms_extra_class_aware_int_keeps_suppressing_after_the_first_survivor() {
        let result = nms_extra_class_aware_int(
            0.5,
            None,
            with_index_int(isolated_then_overlapping_pair_int()),
        );
        assert_eq!(result.len(), 2, "{result:?}");
        assert_eq!(result[1].1, 1);
    }

    #[test]
    fn nms_extra_class_aware_int_suppresses_same_class_without_a_cap() {
        let result = nms_extra_class_aware_int(0.5, None, with_index_int(overlapping_pair_int()));
        assert_eq!(result.len(), 1, "{result:?}");
        assert_eq!(result[0].0.score, 200);
        assert_eq!(result[0].1, 0, "the surviving box keeps its extra data");
    }

    #[test]
    fn nms_extra_class_aware_int_keeps_overlapping_boxes_of_other_classes() {
        let mut boxes = overlapping_pair_int();
        boxes[1].label = 1;
        let result = nms_extra_class_aware_int(0.5, None, with_index_int(boxes));
        assert_eq!(result.len(), 2, "{result:?}");
    }

    #[test]
    fn nms_extra_class_aware_int_does_not_suppress_on_behalf_of_a_suppressed_box() {
        let result = nms_extra_class_aware_int(0.5, None, with_index_int(suppression_chain_int()));
        assert_eq!(result.len(), 2, "{result:?}");
        assert_eq!(result[0].1, 0);
        assert_eq!(result[1].1, 2);
    }

    #[test]
    fn nms_extra_class_aware_int_drops_boxes_already_at_the_minimum_score() {
        let mut boxes = overlapping_pair_int();
        boxes.push(DetectBoxQuantized {
            bbox: BoundingBox {
                xmin: 50.0,
                ymin: 50.0,
                xmax: 60.0,
                ymax: 60.0,
            },
            label: 0,
            score: u8::MIN,
        });
        let result = nms_extra_class_aware_int(0.5, None, with_index_int(boxes));
        assert_eq!(result.len(), 1, "{result:?}");
        assert_eq!(result[0].0.score, 200);
    }

    /// A single box is never compared with itself: the inner loop starts one
    /// past the outer index. Starting at `i` instead makes every box suppress
    /// itself, since a box's IoU with itself is 1.
    #[test]
    fn nms_extra_class_aware_int_keeps_a_lone_box() {
        let boxes = with_index_int(vec![overlapping_pair_int().remove(0)]);
        let result = nms_extra_class_aware_int(0.5, None, boxes);
        assert_eq!(result.len(), 1, "{result:?}");
        assert_eq!(result[0].0.score, 200);
    }

    #[test]
    fn nms_int_max_det_matches_full_truncated() {
        let boxes = make_nms_boxes_int(20);
        let n = 5;
        let full = nms_int(0.5, None, boxes.clone());
        let capped = nms_int(0.5, Some(n), boxes);
        assert_eq!(capped.len(), n);
        assert_eq!(&full[..n], &capped[..]);
    }

    #[test]
    fn nms_int_max_det_zero_returns_empty() {
        let boxes = make_nms_boxes_int(10);
        let result = nms_int(0.5, Some(0), boxes);
        assert!(result.is_empty());
    }

    #[test]
    fn nms_int_max_det_iou_ge_1_returns_sorted_truncated() {
        let boxes = make_nms_boxes_int(10);
        let result = nms_int(1.0, Some(3), boxes);
        assert_eq!(result.len(), 3);
        // Scores should be in descending order (sorted).
        assert!(result[0].score >= result[1].score);
        assert!(result[1].score >= result[2].score);
    }

    #[test]
    fn nms_int_max_det_larger_than_input() {
        let boxes = make_nms_boxes_int(5);
        let full = nms_int(0.5, None, boxes.clone());
        let capped = nms_int(0.5, Some(100), boxes);
        assert_eq!(full.len(), capped.len());
    }
}