edgefirst_image/cpu/

mod.rs

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

use crate::{
    Crop, Error, Flip, FunctionTimer, ImageProcessorTrait, Rect, ResolvedCrop, Result, Rotation,
};
use edgefirst_decoder::{DetectBox, ProtoData, Segmentation};
use edgefirst_tensor::{
    DType, PixelFormat, Tensor, TensorDyn, TensorMapTrait, TensorMemory, TensorTrait,
};

mod convert;
mod masks;
mod resize;
mod simd;
mod tests;

// bilinear_dot removed — masks.rs now uses slice-native bilinear_dot_slice
// closure-based kernel, invoked through the local dtype dispatch below.

/// Resolved colorimetry parameters for a format conversion, computed once at
/// the dispatch site (`convert_impl`) from the appropriate tensor:
/// - YUV→RGB conversions use the **source** tensor's colorimetry.
/// - RGB→YUV conversions use the **destination** tensor's colorimetry.
///
/// `matrix`/`range` feed the `yuv` crate kernels; `src_full_range`/
/// `dst_full_range` gate the hand-rolled luma (limited↔full) expansion in the
/// grey/luma copy helpers.
#[derive(Debug, Clone, Copy)]
pub(crate) struct ColorParams {
    /// Matrix resolved from the YUV side (src for decode, dst for encode).
    pub matrix: yuv::YuvStandardMatrix,
    /// Range resolved from the YUV side (src for decode, dst for encode).
    pub range: yuv::YuvRange,
    /// The same matrix/range in HAL's own terms, so the hand-rolled fixed-point
    /// encoders draw their `(kr, kb)` weights and luma/chroma swings from the
    /// canonical [`edgefirst_tensor::colorimetry`] source rather than a private
    /// duplicate table.
    pub encoding: edgefirst_tensor::ColorEncoding,
    pub range_kind: edgefirst_tensor::ColorRange,
    /// True when the **source** tensor's resolved range is full (decode-side
    /// luma extraction: copy luma directly instead of limited→full expansion).
    pub src_full_range: bool,
    /// True when the **destination** tensor's resolved range is full
    /// (encode-side luma copies).
    pub dst_full_range: bool,
}

/// CPUConverter implements the ImageProcessor trait using the fallback CPU
/// implementation for image processing.
#[derive(Debug)]
pub struct CPUProcessor {
    resizer: fast_image_resize::Resizer,
    options: fast_image_resize::ResizeOptions,
    colors: [[u8; 4]; 20],
    /// Reusable scratch tensor for the U8→float widen path.
    ///
    /// Holds a U8 image in the destination's pixel format and dimensions.
    /// Reallocated only when the format or dimensions change, amortising
    /// the heap allocation across repeated same-size conversions.
    widen_scratch: Option<TensorDyn>,
    /// Reusable scratch for de-striding a padded source before resize.
    ///
    /// `fast_image_resize` needs a tightly-packed input; a padded source
    /// (64-aligned stride from codec decode / `create_image`) is copied here
    /// row-by-row first. Kept on the processor so the steady-state resize loop
    /// reuses the allocation instead of a fresh `Vec` per call (the stride
    /// alignment in this release makes that de-stride copy fire far more often).
    resize_destride_scratch: Vec<u8>,
    /// Reusable scratch for de-striding a padded resize *destination* before
    /// resize.
    ///
    /// `fast_image_resize` needs a tightly-packed output buffer; a padded
    /// destination (a DMA pitch-aligned tensor, or a `view()` narrower than
    /// its parent's row stride) is resized into this tight scratch first,
    /// then copied back into the real destination row-by-row at its true
    /// stride. Kept on the processor so the steady-state resize loop reuses
    /// the allocation instead of a fresh `Vec` per call — mirrors
    /// `resize_destride_scratch` on the source side.
    resize_dst_destride_scratch: Vec<u8>,
    /// Reusable cache-resident scratch for the strip-fused NV→planar path
    /// (`convert_nv_to_planar_fused`): holds a single row strip of packed RGB
    /// (`STRIP_ROWS * width * 3` bytes). Keeping it on the processor avoids a
    /// per-frame allocation and lets each strip stay hot in L2 between the YUV
    /// decode and the deinterleave. Grown on demand; never shrunk.
    nv_strip_scratch: Vec<u8>,
    /// Reusable cache-resident scratch for the fused NV→planar path's
    /// **region** case: a tight (stride == width) copy of the current
    /// strip's luma rows, starting at column 0. The `yuv` crate's row
    /// iteration processes fixed-`stride`-sized chunks internally, using
    /// only the first `width` bytes of each — feeding it a slice that starts
    /// at a mid-row column offset (a source-crop `region` with a nonzero
    /// `left`) reads past the *source* buffer's true end for the region's
    /// last row when that row is also the source's last row (the unused
    /// chunk tail has no next row to alias into there). Packing each row
    /// into a `stride == width` buffer first sidesteps that read. Unused
    /// (and zero-cost) for any row-aligned read (`region.left == 0`,
    /// including the whole-frame `region: None` case): with no column
    /// shift, every row chunk the crate sees is a real, fully-owned source
    /// row, so those reads stay fully zero-copy. Grown on demand; never
    /// shrunk.
    nv_strip_y_pack: Vec<u8>,
    /// Same as [`Self::nv_strip_y_pack`], for the chroma (UV) rows.
    nv_strip_uv_pack: Vec<u8>,
    /// Test-only counter of `convert` calls that took the fused NV→planar
    /// strip path (`convert_nv_to_planar_fused`), incremented at the gate in
    /// `convert_u8`. Exists only under `#[cfg(test)]` — a field that's
    /// written but (outside tests) never read trips clippy's dead-code lint
    /// on both OS lanes. Read via [`Self::fused_hits`] so path-selection
    /// tests assert on the actual gate decision rather than on timing.
    #[cfg(test)]
    fused_hits: u64,
    /// Test-only record of the dimensions of the pre-resize intermediate the
    /// most recent `convert_u8` call allocated, or `None` when that call did
    /// not need one. Lets the allocation-proportionality tests assert that a
    /// cropped convert sizes its intermediate to the crop rather than to the
    /// whole frame, without reaching into the allocator. Reset at the top of
    /// every `convert_u8`.
    #[cfg(test)]
    last_tmp_dims: Option<(usize, usize)>,
    /// Reusable intermediate buffers for the multi-step convert pipeline
    /// (pre-resize format-convert and the resized-RGB scratch). Reused across
    /// frames when the dimensions/format match so the steady-state
    /// letterbox/resize loop does not reallocate — and `alloc_zeroed`-clear — a
    /// full-frame buffer per call. The region each consumer reads is always
    /// fully overwritten first, so reused (non-zeroed) contents are never read.
    convert_tmp: Option<Tensor<u8>>,
    convert_tmp2: Option<Tensor<u8>>,
    /// Reusable crop-sized copy of the source's own pixel format, holding the
    /// halo-grown crop rectangle extracted by [`Self::extract_nv_region`] for
    /// the crop-sized pre-resize intermediate. Only allocated for a cropped
    /// convert that takes that path; the whole-frame path never touches it.
    convert_src_sub: Option<Tensor<u8>>,
}

// `CPUProcessor` was `#[derive(Clone)]` before the `widen_scratch` field was
// added; `TensorDyn` is not `Clone`, so the derive no longer applies. Restore
// the public `Clone` impl by hand to avoid a breaking API change. The scratch
// cache is a private allocation amortiser, not part of the logical value, so a
// clone starts empty rather than sharing or duplicating it.
impl Clone for CPUProcessor {
    fn clone(&self) -> Self {
        Self {
            resizer: self.resizer.clone(),
            options: self.options,
            colors: self.colors,
            widen_scratch: None,
            resize_destride_scratch: Vec::new(),
            resize_dst_destride_scratch: Vec::new(),
            nv_strip_scratch: Vec::new(),
            nv_strip_y_pack: Vec::new(),
            nv_strip_uv_pack: Vec::new(),
            #[cfg(test)]
            fused_hits: 0,
            #[cfg(test)]
            last_tmp_dims: None,
            convert_tmp: None,
            convert_tmp2: None,
            convert_src_sub: None,
        }
    }
}

unsafe impl Send for CPUProcessor {}
unsafe impl Sync for CPUProcessor {}

impl Default for CPUProcessor {
    fn default() -> Self {
        Self::new_bilinear()
    }
}

/// Write the base layer of `dst` before mask rendering.
///
/// This is the terminal fallback: on CPU we have no 2D hardware, so a
/// direct buffer write is the appropriate primitive. The invariant is that
/// every call to the CPU draw_* entry points fully initialises dst — we
/// never rely on "whatever was in the buffer" from the caller.
///
/// - `background == Some(bg)` → byte-for-byte copy bg → dst (after shape /
///   format validation).
/// - `background == None` → fill dst with 0x00 (transparent black).
fn prepare_dst_base_cpu(dst: &mut TensorDyn, background: Option<&TensorDyn>) -> Result<()> {
    match background {
        Some(bg) => {
            if bg.shape() != dst.shape() {
                return Err(Error::InvalidShape(
                    "background shape does not match dst".into(),
                ));
            }
            if bg.format() != dst.format() {
                return Err(Error::InvalidShape(
                    "background pixel format does not match dst".into(),
                ));
            }
            let bg_u8 = bg.as_u8().ok_or(Error::NotAnImage)?;
            let dst_u8 = dst.as_u8_mut().ok_or(Error::NotAnImage)?;
            let bg_map = bg_u8.map_read()?;
            let mut dst_map = dst_u8.map_mut()?;
            let bg_slice = bg_map.as_slice();
            let dst_slice = dst_map.as_mut_slice();
            if bg_slice.len() != dst_slice.len() {
                return Err(Error::InvalidShape(
                    "background buffer size does not match dst".into(),
                ));
            }
            dst_slice.copy_from_slice(bg_slice);
        }
        None => {
            let dst_u8 = dst.as_u8_mut().ok_or(Error::NotAnImage)?;
            let mut dst_map = dst_u8.map_mut()?;
            dst_map.as_mut_slice().fill(0);
        }
    }
    Ok(())
}

/// Whether a source-crop origin is safe for the fused NV→planar strip path
/// (`convert_nv_to_planar_fused`'s `region` parameter). NV12 (4:2:0)
/// subsamples chroma both directions, so an odd `top` would desync the
/// luma/chroma row pairing; NV16 (4:2:2) subsamples only horizontally, so
/// only `left` must be even; NV24 (4:4:4) carries full-resolution chroma and
/// has no alignment constraint. Misaligned origins fall through to the
/// general path rather than snapping or shifting.
fn chroma_alignment_ok(fmt: PixelFormat, r: Rect) -> bool {
    match fmt {
        PixelFormat::Nv12 => r.left.is_multiple_of(2) && r.top.is_multiple_of(2),
        PixelFormat::Nv16 => r.left.is_multiple_of(2),
        PixelFormat::Nv24 => true,
        _ => false,
    }
}

/// Compute row stride for a packed-format Tensor<u8> image given its format.
fn row_stride_for(width: usize, fmt: PixelFormat) -> usize {
    use edgefirst_tensor::PixelLayout;
    match fmt.layout() {
        PixelLayout::Packed => width * fmt.channels(),
        PixelLayout::Planar | PixelLayout::SemiPlanar => width,
        _ => width, // fallback for non-exhaustive
    }
}

/// Read the effective row stride from a tensor, falling back to the computed
/// minimum stride if the tensor has no explicit stride set. This correctly
/// handles tensors with GPU pitch-alignment padding (e.g., from
/// `ImageProcessor::create_image()` or codec strided decode).
fn tensor_row_stride(tensor: &Tensor<u8>) -> usize {
    tensor.effective_row_stride().unwrap_or_else(|| {
        let w = tensor.width().unwrap_or(0);
        let fmt = tensor.format().unwrap_or(PixelFormat::Rgb);
        row_stride_for(w, fmt)
    })
}

/// The `(rows, row_bytes)` of an image tensor's **logical** pixel surface: how
/// many rows a converter may write, and how many bytes of each row are pixels.
///
/// Everything between `row_bytes` and [`tensor_row_stride`] is off-limits. For
/// an allocation-padded tensor those bytes are dead padding; for a
/// [`Tensor::view`](edgefirst_tensor::Tensor::view) destination they are the
/// **parent image's neighbouring columns**, so a writer that treats the mapped
/// slice as one flat buffer both mis-places its own rows and corrupts pixels
/// that are not its to write.
fn logical_surface(tensor: &Tensor<u8>) -> Result<(usize, usize)> {
    use edgefirst_tensor::PixelLayout;
    let fmt = tensor.format().ok_or(Error::NotAnImage)?;
    let w = tensor.width().ok_or(Error::NotAnImage)?;
    let h = tensor.height().ok_or(Error::NotAnImage)?;
    Ok(match fmt.layout() {
        PixelLayout::Packed => (h, w * fmt.channels()),
        PixelLayout::Planar => (fmt.channels() * h, w),
        PixelLayout::SemiPlanar => (fmt.combined_plane_height(h).unwrap_or(h), w),
        // `PixelLayout` is non-exhaustive; treat any future layout as one row
        // per shape row at the tensor's own pitch (see `row_stride_for`).
        _ => (h, row_stride_for(w, fmt)),
    })
}

/// Iterate `(src_row, dst_row)` pairs over a pair of image surfaces, each row
/// clipped to its own logical width so neither side's stride padding — nor a
/// view destination's neighbouring parent pixels — is read or written.
///
/// Callers must have validated both surfaces with [`guard_plane`] first.
fn packed_row_pairs<'s, 'd>(
    src: &'s [u8],
    src_stride: usize,
    src_row_bytes: usize,
    dst: &'d mut [u8],
    dst_stride: usize,
    dst_row_bytes: usize,
    rows: usize,
) -> impl Iterator<Item = (&'s [u8], &'d mut [u8])> {
    src.chunks(src_stride)
        .zip(dst.chunks_mut(dst_stride))
        .take(rows)
        .map(move |(s, d)| (&s[..src_row_bytes], &mut d[..dst_row_bytes]))
}

/// Split a contiguous semi-planar (NV12/NV16/NV24) tensor's mapped bytes into
/// `(luma, chroma)` planes at the `stride * src_h` boundary, validating the map
/// holds the full combined plane first. A bare `split_at` would panic if an
/// imported tensor's caller-supplied dimensions/stride exceed its actual buffer
/// (untrusted input), so this returns `Error::InvalidShape` instead.
fn split_semi_planar(
    bytes: &[u8],
    stride: usize,
    src_h: usize,
    fmt: PixelFormat,
) -> Result<(&[u8], &[u8])> {
    let total_h = fmt.combined_plane_height(src_h).unwrap_or(src_h);
    let need = stride.checked_mul(total_h).ok_or_else(|| {
        Error::InvalidShape(format!(
            "{fmt:?} plane size overflow (stride={stride}, h={src_h})"
        ))
    })?;
    if bytes.len() < need {
        return Err(Error::InvalidShape(format!(
            "{fmt:?} source has {} bytes but needs {need} (stride={stride}, h={src_h})",
            bytes.len()
        )));
    }
    Ok(bytes.split_at(stride * src_h))
}

/// Mutable mirror of [`split_semi_planar`]: split a contiguous semi-planar
/// destination's mapped bytes into `(luma, chroma)` planes at the
/// `stride * dst_h` boundary, validating the map holds the full combined plane
/// first. A bare `split_at_mut` would panic if a caller-supplied (untrusted)
/// destination's declared dimensions/stride exceed its actual buffer, so this
/// returns `Error::InvalidShape` instead.
fn split_semi_planar_mut(
    bytes: &mut [u8],
    stride: usize,
    dst_h: usize,
    fmt: PixelFormat,
) -> Result<(&mut [u8], &mut [u8])> {
    let total_h = fmt.combined_plane_height(dst_h).unwrap_or(dst_h);
    let need = stride.checked_mul(total_h).ok_or_else(|| {
        Error::InvalidShape(format!(
            "{fmt:?} plane size overflow (stride={stride}, combined_h={total_h})"
        ))
    })?;
    if bytes.len() < need {
        return Err(Error::InvalidShape(format!(
            "{fmt:?} destination has {} bytes but needs {need} (stride={stride}, combined_h={total_h})",
            bytes.len()
        )));
    }
    Ok(bytes.split_at_mut(stride * dst_h))
}

/// Validate that a mapped plane buffer of `buf_len` bytes can hold `rows` rows
/// of `stride` bytes each, with `row_bytes` valid bytes per row. Returns
/// `Error::InvalidShape` (instead of letting a later row slice panic) when a
/// caller-supplied stride/shape exceeds the actual allocation. `what` labels the
/// buffer in the error message.
fn guard_plane(
    buf_len: usize,
    stride: usize,
    rows: usize,
    row_bytes: usize,
    what: &str,
) -> Result<()> {
    let need = stride.checked_mul(rows).ok_or_else(|| {
        Error::InvalidShape(format!(
            "{what} plane size overflow (stride={stride}, rows={rows})"
        ))
    })?;
    if row_bytes > stride || buf_len < need {
        return Err(Error::InvalidShape(format!(
            "{what} buffer too small: {buf_len} bytes, need {need} (stride={stride}, rows={rows}, row_bytes={row_bytes})"
        )));
    }
    Ok(())
}

/// Apply XOR 0x80 bias to color channels only, preserving alpha.
///
/// Matches GL int8 shader behavior: `vec4(int8_bias(c.rgb), c.a)`.
/// For formats without alpha, XORs every byte (fast path).
pub(crate) fn apply_int8_xor_bias(data: &mut [u8], fmt: PixelFormat) {
    use edgefirst_tensor::PixelLayout;
    if !fmt.has_alpha() {
        for b in data.iter_mut() {
            *b ^= 0x80;
        }
    } else if fmt.layout() == PixelLayout::Planar {
        // Planar with alpha (e.g. PlanarRgba): XOR color planes, skip alpha plane.
        let channels = fmt.channels();
        let plane_size = data.len() / channels;
        for b in data[..plane_size * (channels - 1)].iter_mut() {
            *b ^= 0x80;
        }
    } else {
        // Packed with alpha (Rgba, Bgra): XOR color bytes, skip alpha byte.
        let channels = fmt.channels();
        for pixel in data.chunks_exact_mut(channels) {
            for b in &mut pixel[..channels - 1] {
                *b ^= 0x80;
            }
        }
    }
}

/// Row-confined [`apply_int8_xor_bias`] for a destination whose row pitch may
/// exceed its pixel bytes — an allocation-padded tensor, or a `Tensor::view()`
/// whose pitch is the parent's and whose trailing bytes are the parent's
/// neighbouring pixels.
fn apply_int8_xor_bias_rows(tensor: &mut Tensor<u8>, fmt: PixelFormat) -> Result<()> {
    use edgefirst_tensor::PixelLayout;
    let (rows, row_bytes) = logical_surface(tensor)?;
    let stride = tensor_row_stride(tensor);
    // A planar alpha channel is a whole trailing plane, so the rows to bias are
    // the leading `channels - 1` planes rather than a prefix of each row.
    let color_rows = if fmt.has_alpha() && fmt.layout() == PixelLayout::Planar {
        rows / fmt.channels() * (fmt.channels() - 1)
    } else {
        rows
    };
    // Within a colour row the packed-alpha skip still applies; a planar row is
    // all colour, so bias it as a format without alpha.
    let row_fmt = if fmt.layout() == PixelLayout::Planar {
        PixelFormat::Grey
    } else {
        fmt
    };
    let mut map = tensor.map_mut()?;
    let buf = map.as_mut_slice();
    guard_plane(buf.len(), stride, rows, row_bytes, "int8 bias dst")?;
    for row in buf.chunks_mut(stride).take(color_rows) {
        apply_int8_xor_bias(&mut row[..row_bytes], row_fmt);
    }
    Ok(())
}

impl CPUProcessor {
    /// Creates a new CPUConverter with bilinear resizing.
    pub fn new() -> Self {
        Self::new_bilinear()
    }

    /// Creates a new CPUConverter with bilinear resizing.
    fn new_bilinear() -> Self {
        let resizer = fast_image_resize::Resizer::new();
        let options = fast_image_resize::ResizeOptions::new()
            .resize_alg(fast_image_resize::ResizeAlg::Convolution(
                fast_image_resize::FilterType::Bilinear,
            ))
            .use_alpha(false);

        log::debug!("CPUConverter created");
        Self {
            resizer,
            options,
            colors: crate::DEFAULT_COLORS_U8,
            widen_scratch: None,
            resize_destride_scratch: Vec::new(),
            resize_dst_destride_scratch: Vec::new(),
            nv_strip_scratch: Vec::new(),
            nv_strip_y_pack: Vec::new(),
            nv_strip_uv_pack: Vec::new(),
            #[cfg(test)]
            fused_hits: 0,
            #[cfg(test)]
            last_tmp_dims: None,
            convert_tmp: None,
            convert_tmp2: None,
            convert_src_sub: None,
        }
    }

    /// Creates a new CPUConverter with nearest neighbor resizing.
    pub fn new_nearest() -> Self {
        let resizer = fast_image_resize::Resizer::new();
        let options = fast_image_resize::ResizeOptions::new()
            .resize_alg(fast_image_resize::ResizeAlg::Nearest)
            .use_alpha(false);
        log::debug!("CPUConverter created");
        Self {
            resizer,
            options,
            colors: crate::DEFAULT_COLORS_U8,
            widen_scratch: None,
            resize_destride_scratch: Vec::new(),
            resize_dst_destride_scratch: Vec::new(),
            nv_strip_scratch: Vec::new(),
            nv_strip_y_pack: Vec::new(),
            nv_strip_uv_pack: Vec::new(),
            #[cfg(test)]
            fused_hits: 0,
            #[cfg(test)]
            last_tmp_dims: None,
            convert_tmp: None,
            convert_tmp2: None,
            convert_src_sub: None,
        }
    }

    /// Test-only accessor for the `fused_hits` counter (see the field doc).
    #[cfg(test)]
    pub(super) fn fused_hits(&self) -> u64 {
        self.fused_hits
    }

    /// Test-only accessor for the dimensions of the pre-resize intermediate the
    /// last `convert_u8` allocated (see the `last_tmp_dims` field doc).
    #[cfg(test)]
    pub(super) fn last_tmp_dims(&self) -> Option<(usize, usize)> {
        self.last_tmp_dims
    }

    pub(crate) fn support_conversion_pf(src: PixelFormat, dst: PixelFormat) -> bool {
        use PixelFormat::*;
        matches!(
            (src, dst),
            (Nv12, Rgb)
                | (Nv12, Rgba)
                | (Nv12, Grey)
                | (Nv16, Rgb)
                | (Nv16, Rgba)
                | (Nv16, Bgra)
                | (Nv24, Rgb)
                | (Nv24, Rgba)
                | (Nv24, Grey)
                | (Nv24, Bgra)
                | (Yuyv, Rgb)
                | (Yuyv, Rgba)
                | (Yuyv, Grey)
                | (Yuyv, Yuyv)
                | (Yuyv, PlanarRgb)
                | (Yuyv, PlanarRgba)
                | (Yuyv, Nv16)
                | (Vyuy, Rgb)
                | (Vyuy, Rgba)
                | (Vyuy, Grey)
                | (Vyuy, Vyuy)
                | (Vyuy, PlanarRgb)
                | (Vyuy, PlanarRgba)
                | (Vyuy, Nv16)
                | (Rgba, Rgb)
                | (Rgba, Rgba)
                | (Rgba, Grey)
                | (Rgba, Yuyv)
                | (Rgba, PlanarRgb)
                | (Rgba, PlanarRgba)
                | (Rgba, Nv16)
                | (Rgb, Rgb)
                | (Rgb, Rgba)
                | (Rgb, Grey)
                | (Rgb, Yuyv)
                | (Rgb, PlanarRgb)
                | (Rgb, PlanarRgba)
                | (Rgb, Nv16)
                | (Grey, Rgb)
                | (Grey, Rgba)
                | (Grey, Grey)
                | (Grey, Yuyv)
                | (Grey, PlanarRgb)
                | (Grey, PlanarRgba)
                | (Grey, Nv16)
                | (Nv12, Bgra)
                | (Yuyv, Bgra)
                | (Vyuy, Bgra)
                | (Rgba, Bgra)
                | (Rgb, Bgra)
                | (Grey, Bgra)
                | (Bgra, Bgra)
                | (PlanarRgb, Rgb)
                | (PlanarRgb, Rgba)
                | (PlanarRgba, Rgb)
                | (PlanarRgba, Rgba)
                | (PlanarRgb, Bgra)
                | (PlanarRgba, Bgra)
        )
    }

    /// Format conversion dispatch for Tensor<u8> with PixelFormat metadata.
    pub(crate) fn convert_format_pf(
        src: &Tensor<u8>,
        dst: &mut Tensor<u8>,
        src_fmt: PixelFormat,
        dst_fmt: PixelFormat,
        cp: ColorParams,
    ) -> Result<()> {
        let _timer = FunctionTimer::new(format!(
            "ImageProcessor::convert_format {} to {}",
            src_fmt, dst_fmt,
        ));

        use PixelFormat::*;
        match (src_fmt, dst_fmt) {
            (Nv12, Rgb) => Self::convert_nv12_to_rgb(src, dst, cp),
            (Nv12, Rgba) => Self::convert_nv12_to_rgba(src, dst, cp),
            (Nv12, Grey) => Self::convert_nv12_to_grey(src, dst, cp),
            (Yuyv, Rgb) => Self::convert_yuyv_to_rgb(src, dst, cp),
            (Yuyv, Rgba) => Self::convert_yuyv_to_rgba(src, dst, cp),
            (Yuyv, Grey) => Self::convert_yuyv_to_grey(src, dst, cp),
            (Yuyv, Yuyv) => Self::copy_image(src, dst),
            (Yuyv, PlanarRgb) => Self::convert_yuyv_to_8bps(src, dst, cp),
            (Yuyv, PlanarRgba) => Self::convert_yuyv_to_prgba(src, dst, cp),
            (Yuyv, Nv16) => Self::convert_yuyv_to_nv16(src, dst),
            (Vyuy, Rgb) => Self::convert_vyuy_to_rgb(src, dst, cp),
            (Vyuy, Rgba) => Self::convert_vyuy_to_rgba(src, dst, cp),
            (Vyuy, Grey) => Self::convert_vyuy_to_grey(src, dst, cp),
            (Vyuy, Vyuy) => Self::copy_image(src, dst),
            (Vyuy, PlanarRgb) => Self::convert_vyuy_to_8bps(src, dst, cp),
            (Vyuy, PlanarRgba) => Self::convert_vyuy_to_prgba(src, dst, cp),
            (Vyuy, Nv16) => Self::convert_vyuy_to_nv16(src, dst),
            (Rgba, Rgb) => Self::convert_rgba_to_rgb(src, dst),
            (Rgba, Rgba) => Self::copy_image(src, dst),
            (Rgba, Grey) => Self::convert_rgba_to_grey(src, dst),
            (Rgba, Yuyv) => Self::convert_rgba_to_yuyv(src, dst, cp),
            (Rgba, PlanarRgb) => Self::convert_rgba_to_8bps(src, dst),
            (Rgba, PlanarRgba) => Self::convert_rgba_to_prgba(src, dst),
            (Rgba, Nv16) => Self::convert_rgba_to_nv16(src, dst, cp),
            (Rgb, Rgb) => Self::copy_image(src, dst),
            (Rgb, Rgba) => Self::convert_rgb_to_rgba(src, dst),
            (Rgb, Grey) => Self::convert_rgb_to_grey(src, dst),
            (Rgb, Yuyv) => Self::convert_rgb_to_yuyv(src, dst, cp),
            (Rgb, PlanarRgb) => Self::convert_rgb_to_8bps(src, dst),
            (Rgb, PlanarRgba) => Self::convert_rgb_to_prgba(src, dst),
            (Rgb, Nv16) => Self::convert_rgb_to_nv16(src, dst, cp),
            (Grey, Rgb) => Self::convert_grey_to_rgb(src, dst),
            (Grey, Rgba) => Self::convert_grey_to_rgba(src, dst),
            (Grey, Grey) => Self::copy_image(src, dst),
            (Grey, Yuyv) => Self::convert_grey_to_yuyv(src, dst, cp),
            (Grey, PlanarRgb) => Self::convert_grey_to_8bps(src, dst),
            (Grey, PlanarRgba) => Self::convert_grey_to_prgba(src, dst),
            (Grey, Nv16) => Self::convert_grey_to_nv16(src, dst, cp),

            // the following converts are added for use in testing
            (Nv16, Rgb) => Self::convert_nv16_to_rgb(src, dst, cp),
            (Nv16, Rgba) => Self::convert_nv16_to_rgba(src, dst, cp),
            (Nv24, Rgb) => Self::convert_nv24_to_rgb(src, dst, cp),
            (Nv24, Rgba) => Self::convert_nv24_to_rgba(src, dst, cp),
            (Nv24, Grey) => Self::convert_nv24_to_grey(src, dst, cp),
            (PlanarRgb, Rgb) => Self::convert_8bps_to_rgb(src, dst),
            (PlanarRgb, Rgba) => Self::convert_8bps_to_rgba(src, dst),
            (PlanarRgba, Rgb) => Self::convert_prgba_to_rgb(src, dst),
            (PlanarRgba, Rgba) => Self::convert_prgba_to_rgba(src, dst),

            // BGRA destination: convert to RGBA layout, then swap R and B
            (Bgra, Bgra) => Self::copy_image(src, dst),
            (Nv12, Bgra) => {
                Self::convert_nv12_to_rgba(src, dst, cp)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Nv16, Bgra) => {
                Self::convert_nv16_to_rgba(src, dst, cp)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Nv24, Bgra) => {
                Self::convert_nv24_to_rgba(src, dst, cp)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Yuyv, Bgra) => {
                Self::convert_yuyv_to_rgba(src, dst, cp)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Vyuy, Bgra) => {
                Self::convert_vyuy_to_rgba(src, dst, cp)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Rgba, Bgra) => {
                Self::copy_image(src, dst)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Rgb, Bgra) => {
                Self::convert_rgb_to_rgba(src, dst)?;
                Self::swizzle_rb_4chan(dst)
            }
            (Grey, Bgra) => {
                Self::convert_grey_to_rgba(src, dst)?;
                Self::swizzle_rb_4chan(dst)
            }
            (PlanarRgb, Bgra) => {
                Self::convert_8bps_to_rgba(src, dst)?;
                Self::swizzle_rb_4chan(dst)
            }
            (PlanarRgba, Bgra) => {
                Self::convert_prgba_to_rgba(src, dst)?;
                Self::swizzle_rb_4chan(dst)
            }

            (s, d) => Err(Error::NotSupported(format!("Conversion from {s} to {d}",))),
        }
    }

    /// Tensor<u8>-based fill_image_outside_crop.
    pub(crate) fn fill_image_outside_crop_u8(
        dst: &mut Tensor<u8>,
        rgba: [u8; 4],
        crop: Rect,
    ) -> Result<()> {
        let dst_fmt = dst.format().unwrap();
        let dst_w = dst.width().unwrap();
        let dst_h = dst.height().unwrap();
        // Resolve the YUV fill encoding from the destination tensor so the
        // border color matches the same matrix/range as a later YUV→RGB
        // decode of this image. RGB/Grey fills ignore these params.
        let cm = crate::colorimetry::resolve_colorimetry(dst.colorimetry(), dst.height());
        let cp = ColorParams {
            matrix: crate::colorimetry::yuv_matrix(cm.encoding.unwrap()),
            range: crate::colorimetry::yuv_range(cm.range.unwrap()),
            encoding: cm.encoding.unwrap(),
            range_kind: cm.range.unwrap(),
            src_full_range: cm.range == Some(edgefirst_tensor::ColorRange::Full),
            dst_full_range: cm.range == Some(edgefirst_tensor::ColorRange::Full),
        };
        let dst_stride = tensor_row_stride(dst);
        let mut dst_map = dst.map_mut()?;
        let dst_tup = (dst_map.as_mut_slice(), dst_w, dst_h, dst_stride);
        Self::fill_outside_crop_dispatch(dst_tup, dst_fmt, rgba, crop, cp)
    }

    /// Common fill dispatch by format. The tuple is
    /// `(bytes, logical_width, logical_height, row_stride)` — the stride is the
    /// destination's real byte pitch, which for a `view()` destination is the
    /// parent image's, not `width × bpp`.
    fn fill_outside_crop_dispatch(
        dst: (&mut [u8], usize, usize, usize),
        fmt: PixelFormat,
        rgba: [u8; 4],
        crop: Rect,
        cp: ColorParams,
    ) -> Result<()> {
        use PixelFormat::*;
        match fmt {
            Rgba | Bgra => Self::fill_image_outside_crop_(dst, rgba, crop),
            Rgb => Self::fill_image_outside_crop_(dst, Self::rgba_to_rgb(rgba), crop),
            Grey => Self::fill_image_outside_crop_(dst, Self::rgba_to_grey(rgba), crop),
            Yuyv => {
                let (bytes, w, h, stride) = dst;
                let yuyv = Self::rgba_to_yuyv(rgba, cp);
                // Bulk fill in 4-byte [Y0,U,Y1,V] macropixels; crop bounds in
                // macropixel units so a shared chroma pair is never split.
                Self::fill_image_outside_crop_(
                    (&mut *bytes, w / 2, h, stride),
                    yuyv,
                    Rect::new(crop.left / 2, crop.top, crop.width.div_ceil(2), crop.height),
                )?;
                // An odd width leaves a trailing unpaired pixel that the
                // macropixel walk above (w / 2 units) structurally never
                // reaches — the same trailing-pixel class the odd-width YUYV
                // encoders handle. Its 2 bytes are [Y, U] (a row is w × 2
                // bytes, so there is no room for its V). Fill it on every row
                // where the last column lies outside the crop, judged in
                // pixel units against the original crop.
                if w % 2 == 1 {
                    let last = w - 1;
                    let col_outside = last < crop.left || last >= crop.left + crop.width;
                    for y in 0..h {
                        let row_outside = y < crop.top || y >= crop.top + crop.height;
                        if row_outside || col_outside {
                            let off = y * stride + last * 2;
                            bytes[off] = yuyv[0];
                            bytes[off + 1] = yuyv[1];
                        }
                    }
                }
                Ok(())
            }
            PlanarRgb => Self::fill_image_outside_crop_planar(dst, Self::rgba_to_rgb(rgba), crop),
            PlanarRgba => Self::fill_image_outside_crop_planar(dst, rgba, crop),
            Nv16 => {
                let yuyv = Self::rgba_to_yuyv(rgba, cp);
                Self::fill_image_outside_crop_yuv_semiplanar(dst, yuyv[0], [yuyv[1], yuyv[3]], crop)
            }
            _ => Err(Error::Internal(format!(
                "Found unexpected destination {fmt}",
            ))),
        }
    }
}

impl ImageProcessorTrait for CPUProcessor {
    fn convert(
        &mut self,
        src: &TensorDyn,
        dst: &mut TensorDyn,
        rotation: Rotation,
        flip: Flip,
        crop: Crop,
    ) -> Result<()> {
        let crop = crop.resolve(
            src.width().unwrap_or(0),
            src.height().unwrap_or(0),
            dst.width().unwrap_or(0),
            dst.height().unwrap_or(0),
        )?;
        self.convert_impl(src, dst, rotation, flip, crop)
    }

    fn draw_decoded_masks(
        &mut self,
        dst: &mut TensorDyn,
        detect: &[DetectBox],
        segmentation: &[Segmentation],
        overlay: crate::MaskOverlay<'_>,
    ) -> Result<()> {
        // CPU is the terminal fallback — it must always produce the full
        // output, never assume the caller cleared dst. Every call writes
        // the base layer first (bg copy or zero fill) and then the masks.
        prepare_dst_base_cpu(dst, overlay.background)?;
        let dst = dst.as_u8_mut().ok_or(Error::NotAnImage)?;
        self.draw_decoded_masks_impl(
            dst,
            detect,
            segmentation,
            overlay.opacity,
            overlay.color_mode,
        )
    }

    fn draw_proto_masks(
        &mut self,
        dst: &mut TensorDyn,
        detect: &[DetectBox],
        proto_data: &ProtoData,
        overlay: crate::MaskOverlay<'_>,
    ) -> Result<()> {
        prepare_dst_base_cpu(dst, overlay.background)?;
        let dst = dst.as_u8_mut().ok_or(Error::NotAnImage)?;
        self.draw_proto_masks_impl(
            dst,
            detect,
            proto_data,
            overlay.opacity,
            overlay.letterbox,
            overlay.color_mode,
        )
    }

    fn set_class_colors(&mut self, colors: &[[u8; 4]]) -> Result<()> {
        for (c, new_c) in self.colors.iter_mut().zip(colors.iter()) {
            *c = *new_c;
        }
        Ok(())
    }
}

// Internal methods — dtype-aware dispatch layer.
impl CPUProcessor {
    /// Top-level conversion dispatcher: handles dtype combinations.
    pub(crate) fn convert_impl(
        &mut self,
        src: &TensorDyn,
        dst: &mut TensorDyn,
        rotation: Rotation,
        flip: Flip,
        crop: ResolvedCrop,
    ) -> Result<()> {
        let src_fmt = src.format().ok_or(Error::NotAnImage)?;
        let dst_fmt = dst.format().ok_or(Error::NotAnImage)?;

        // Resolve per-tensor colorimetry once, at the use site, without
        // mutating either tensor. YUV→RGB conversions take their matrix/range
        // from the source; RGB→YUV conversions from the destination. The grey/
        // luma expansion is gated on the resolved range of the relevant side.
        let src_cm = crate::colorimetry::effective_colorimetry(src);
        let dst_cm = crate::colorimetry::effective_colorimetry(dst);
        let src_full = src_cm.range == Some(edgefirst_tensor::ColorRange::Full);
        let dst_full = dst_cm.range == Some(edgefirst_tensor::ColorRange::Full);
        let src_params = ColorParams {
            matrix: crate::colorimetry::yuv_matrix(src_cm.encoding.unwrap()),
            range: crate::colorimetry::yuv_range(src_cm.range.unwrap()),
            encoding: src_cm.encoding.unwrap(),
            range_kind: src_cm.range.unwrap(),
            src_full_range: src_full,
            dst_full_range: dst_full,
        };
        let dst_params = ColorParams {
            matrix: crate::colorimetry::yuv_matrix(dst_cm.encoding.unwrap()),
            range: crate::colorimetry::yuv_range(dst_cm.range.unwrap()),
            encoding: dst_cm.encoding.unwrap(),
            range_kind: dst_cm.range.unwrap(),
            src_full_range: src_full,
            dst_full_range: dst_full,
        };
        match (src.dtype(), dst.dtype()) {
            (DType::U8, DType::U8) => {
                let src = src.as_u8().unwrap();
                let dst = dst.as_u8_mut().unwrap();
                self.convert_u8(
                    src, dst, src_fmt, dst_fmt, rotation, flip, crop, src_params, dst_params,
                )
            }
            (DType::U8, DType::I8) => {
                // Int8 output: reinterpret the i8 destination as u8 (layout-
                // identical), convert directly into it, then XOR 0x80 in-place.
                let src_u8 = src.as_u8().unwrap();
                let dst_i8 = dst.as_i8_mut().unwrap();
                // SAFETY: Tensor<i8> and Tensor<u8> are layout-identical
                // (same element size, no T-dependent drop glue). Same
                // rationale as gl::processor::tensor_i8_as_u8_mut.
                let dst_u8 = unsafe { &mut *(dst_i8 as *mut Tensor<i8> as *mut Tensor<u8>) };
                self.convert_u8(
                    src_u8, dst_u8, src_fmt, dst_fmt, rotation, flip, crop, src_params, dst_params,
                )?;
                // Apply XOR 0x80 bias in-place (u8 → i8 conversion)
                apply_int8_xor_bias_rows(dst_u8, dst_fmt)
            }
            (DType::U8, d @ (DType::F32 | DType::F16)) => {
                let src_u8 = src.as_u8().unwrap();
                let dw = dst.width().ok_or(Error::NotAnImage)?;
                let dh = dst.height().ok_or(Error::NotAnImage)?;
                // Reuse the scratch tensor when format and dimensions match;
                // otherwise reallocate and cache the new scratch.  Take the
                // scratch out of `self` so that `convert_u8` can borrow `self`
                // exclusively, then restore it afterwards.
                let scratch_matches = self.widen_scratch.as_ref().is_some_and(|t| {
                    t.width() == Some(dw) && t.height() == Some(dh) && t.format() == Some(dst_fmt)
                });
                let mut tmp = if scratch_matches {
                    self.widen_scratch.take().unwrap()
                } else {
                    TensorDyn::image(
                        dw,
                        dh,
                        dst_fmt,
                        DType::U8,
                        Some(TensorMemory::Mem),
                        edgefirst_tensor::CpuAccess::ReadWrite,
                    )?
                };
                {
                    let tmp_u8 = tmp.as_u8_mut().unwrap();
                    self.convert_u8(
                        src_u8, tmp_u8, src_fmt, dst_fmt, rotation, flip, crop, src_params,
                        dst_params,
                    )?;
                }
                // Widen the u8 scratch into the float destination, then restore
                // the scratch for reuse on the next call.
                //
                // The scratch is tight, but `dst` need not be: an allocation-
                // padded float tensor, or a `Tensor::view()` whose pitch is the
                // parent's, has a wider row than its pixels. Widen row by row so
                // each output row lands at the destination's own pitch and the
                // bytes past it stay untouched. On a tight destination the two
                // pitches coincide and this is the previous flat widen.
                {
                    let tmp_u8 = tmp.as_u8().unwrap();
                    let (rows, row_len) = logical_surface(tmp_u8)?;
                    let src_stride = tensor_row_stride(tmp_u8);
                    let dst_stride_bytes = dst.effective_row_stride().ok_or(Error::NotAnImage)?;
                    let src_map = tmp_u8.map_read()?;
                    guard_plane(
                        src_map.as_slice().len(),
                        src_stride,
                        rows,
                        row_len,
                        "widen src",
                    )?;
                    let src_rows = src_map.as_slice().chunks(src_stride).take(rows);
                    let elem = d.size();
                    if !dst_stride_bytes.is_multiple_of(elem) {
                        return Err(Error::InvalidShape(format!(
                            "{d} destination row stride {dst_stride_bytes} is not a multiple of \
                             the element size {elem}"
                        )));
                    }
                    let dst_stride = dst_stride_bytes / elem;
                    match d {
                        DType::F32 => {
                            let dst_t = dst.as_f32_mut().unwrap();
                            let mut dst_map = dst_t.map_mut()?;
                            // Counted in elements, not bytes — same invariant.
                            guard_plane(
                                dst_map.as_slice().len(),
                                dst_stride,
                                rows,
                                row_len,
                                "widen f32 dst",
                            )?;
                            // NEON-accelerated u8→f32 `/255` widen (bit-identical
                            // to the scalar `b as f32 / 255.0`); the scalar
                            // iterator form did not vectorise. See cpu::simd.
                            for (s, dr) in
                                src_rows.zip(dst_map.as_mut_slice().chunks_mut(dst_stride))
                            {
                                simd::widen_u8_to_f32_norm(s, &mut dr[..row_len]);
                            }
                        }
                        DType::F16 => {
                            let dst_t = dst.as_f16_mut().unwrap();
                            let mut dst_map = dst_t.map_mut()?;
                            guard_plane(
                                dst_map.as_slice().len(),
                                dst_stride,
                                rows,
                                row_len,
                                "widen f16 dst",
                            )?;
                            // u8→f16 `/255` widen; uses native FP16
                            // (`ucvtf`+`fdiv`) at runtime on FEAT_FP16 CPUs
                            // (Orin), scalar `half::f16::from_f32` elsewhere.
                            // See cpu::simd.
                            for (s, dr) in
                                src_rows.zip(dst_map.as_mut_slice().chunks_mut(dst_stride))
                            {
                                simd::widen_u8_to_f16_norm(s, &mut dr[..row_len]);
                            }
                        }
                        _ => unreachable!(),
                    }
                }
                self.widen_scratch = Some(tmp);
                Ok(())
            }
            (s, d) => Err(Error::NotSupported(format!("dtype {s} -> {d}",))),
        }
    }

    /// Reuse `cached` if it already has dimensions `(w, h)` and pixel format
    /// `fmt`, otherwise allocate a fresh `Mem` image. The returned buffer's
    /// contents are **not** zeroed on reuse — callers must fully overwrite the
    /// region they later read (see the note at the convert pipeline's tmp/tmp2
    /// site). This amortises the per-frame `Tensor::image` `alloc_zeroed`.
    fn reuse_or_alloc_image(
        cached: Option<Tensor<u8>>,
        w: usize,
        h: usize,
        fmt: PixelFormat,
    ) -> Result<Tensor<u8>> {
        if let Some(t) = cached {
            if t.width() == Some(w) && t.height() == Some(h) && t.format() == Some(fmt) {
                return Ok(t);
            }
        }
        Ok(Tensor::<u8>::image(
            w,
            h,
            fmt,
            Some(TensorMemory::Mem),
            edgefirst_tensor::CpuAccess::ReadWrite,
        )?)
    }

    /// The source sub-rectangle to convert into the pre-resize intermediate for
    /// a cropped convert, or `None` to keep converting the whole frame.
    ///
    /// A cropped convert only ever reads the crop rect plus the resize filter's
    /// halo (see [`Self::filter_halo`]), so converting the whole frame into the
    /// intermediate wastes work proportional to the frame — for a 4K frame
    /// tiled into 640x640 crops, ~22x the pixels actually needed. Returning the
    /// halo-grown crop lets the pre-resize convert produce a crop-sized
    /// intermediate instead.
    ///
    /// The returned rect only ever *grows* the crop, and is clamped to the
    /// frame, so the resize reads exactly the same real source pixels — and
    /// clamps at exactly the same frame edges — as it did with a full-frame
    /// intermediate. Output is byte-identical; only the buffer size changes.
    ///
    /// `None` (unchanged full-frame behaviour) for: an uncropped convert, a
    /// crop covering the whole frame, a source format that cannot be extracted
    /// (only the semi-planar NV family can — `Tensor::view` rejects non-packed
    /// layouts, and NV is what the 4K tiling path feeds in), and a resize
    /// algorithm whose filter reach is not modelled.
    fn pre_resize_region(
        &self,
        src_fmt: PixelFormat,
        (src_w, src_h): (usize, usize),
        (dst_w, dst_h): (usize, usize),
        rotation: Rotation,
        crop: ResolvedCrop,
    ) -> Option<Rect> {
        use PixelFormat::{Nv12, Nv16, Nv24};

        if !matches!(src_fmt, Nv12 | Nv16 | Nv24) {
            return None;
        }
        let r = crop.src_rect?;
        let full_src = Rect {
            left: 0,
            top: 0,
            width: src_w,
            height: src_h,
        };
        if r == full_src {
            return None;
        }

        // The resize maps the source rect onto the destination rect; a quarter
        // turn swaps which destination extent each source axis is scaled onto
        // (see `adjust_dest_rect_for_rotate_flip_dims`), and the filter halo
        // depends on that scale factor.
        let d = crop.dst_rect.unwrap_or(Rect {
            left: 0,
            top: 0,
            width: dst_w,
            height: dst_h,
        });
        let (dst_x, dst_y) = match rotation {
            Rotation::None | Rotation::Rotate180 => (d.width, d.height),
            Rotation::Clockwise90 | Rotation::CounterClockwise90 => (d.height, d.width),
        };
        let halo_x = self.filter_halo(r.width, dst_x)?;
        let halo_y = self.filter_halo(r.height, dst_y)?;

        let mut left = r.left.saturating_sub(halo_x);
        let mut top = r.top.saturating_sub(halo_y);
        let mut right = (r.left + r.width + halo_x).min(src_w);
        let mut bottom = (r.top + r.height + halo_y).min(src_h);

        // NV12 subsamples chroma on both axes, NV16 on the horizontal one, so
        // the extracted origin must land on a chroma sample boundary for the
        // sub-image's chroma to line up with the frame's. Snap the origin DOWN
        // and the far edge UP — growing only, so every pixel inside `r` keeps
        // exactly the neighbours it had, whatever the crop's own parity.
        let (align_x, align_y) = match src_fmt {
            Nv12 => (2, 2),
            Nv16 => (2, 1),
            _ => (1, 1),
        };
        left -= left % align_x;
        top -= top % align_y;
        right = right.next_multiple_of(align_x).min(src_w);
        bottom = bottom.next_multiple_of(align_y).min(src_h);

        let grown = Rect {
            left,
            top,
            width: right - left,
            height: bottom - top,
        };
        // The halo already reaches the whole frame — extracting would be a pure
        // copy with no saving.
        (grown != full_src).then_some(grown)
    }

    /// U8-to-U8 conversion: the full format conversion + resize pipeline.
    #[allow(clippy::too_many_arguments)]
    fn convert_u8(
        &mut self,
        src: &Tensor<u8>,
        dst: &mut Tensor<u8>,
        src_fmt: PixelFormat,
        dst_fmt: PixelFormat,
        rotation: Rotation,
        flip: Flip,
        crop: ResolvedCrop,
        src_params: ColorParams,
        dst_params: ColorParams,
    ) -> Result<()> {
        use PixelFormat::*;

        #[cfg(test)]
        {
            self.last_tmp_dims = None;
        }

        let src_w = src.width().unwrap();
        let src_h = src.height().unwrap();
        let dst_w = dst.width().unwrap();
        let dst_h = dst.height().unwrap();

        crop.check_crop_dims(src_w, src_h, dst_w, dst_h)?;

        // Determine intermediate format for the resize step
        let intermediate = match (src_fmt, dst_fmt) {
            (Nv12, Rgb) => Rgb,
            (Nv12, Rgba) => Rgba,
            (Nv12, Grey) => Grey,
            (Nv12, Yuyv) => Rgba,
            (Nv12, Nv16) => Rgba,
            (Nv12, PlanarRgb) => Rgb,
            (Nv12, PlanarRgba) => Rgba,
            (Nv16, PlanarRgb) => Rgb,
            (Nv16, PlanarRgba) => Rgba,
            (Nv24, PlanarRgb) => Rgb,
            (Nv24, PlanarRgba) => Rgba,
            (Yuyv, Rgb) => Rgb,
            (Yuyv, Rgba) => Rgba,
            (Yuyv, Grey) => Grey,
            (Yuyv, Yuyv) => Rgba,
            (Yuyv, PlanarRgb) => Rgb,
            (Yuyv, PlanarRgba) => Rgba,
            (Yuyv, Nv16) => Rgba,
            (Vyuy, Rgb) => Rgb,
            (Vyuy, Rgba) => Rgba,
            (Vyuy, Grey) => Grey,
            (Vyuy, Vyuy) => Rgba,
            (Vyuy, PlanarRgb) => Rgb,
            (Vyuy, PlanarRgba) => Rgba,
            (Vyuy, Nv16) => Rgba,
            (Rgba, Rgb) => Rgba,
            (Rgba, Rgba) => Rgba,
            (Rgba, Grey) => Grey,
            (Rgba, Yuyv) => Rgba,
            (Rgba, PlanarRgb) => Rgba,
            (Rgba, PlanarRgba) => Rgba,
            (Rgba, Nv16) => Rgba,
            (Rgb, Rgb) => Rgb,
            (Rgb, Rgba) => Rgb,
            (Rgb, Grey) => Grey,
            (Rgb, Yuyv) => Rgb,
            (Rgb, PlanarRgb) => Rgb,
            (Rgb, PlanarRgba) => Rgb,
            (Rgb, Nv16) => Rgb,
            (Grey, Rgb) => Rgb,
            (Grey, Rgba) => Rgba,
            (Grey, Grey) => Grey,
            (Grey, Yuyv) => Grey,
            (Grey, PlanarRgb) => Grey,
            (Grey, PlanarRgba) => Grey,
            (Grey, Nv16) => Grey,
            (Nv12, Bgra) => Rgba,
            (Yuyv, Bgra) => Rgba,
            (Vyuy, Bgra) => Rgba,
            (Rgba, Bgra) => Rgba,
            (Rgb, Bgra) => Rgb,
            (Grey, Bgra) => Grey,
            (Bgra, Bgra) => Bgra,
            (Nv16, Rgb) => Rgb,
            (Nv16, Rgba) => Rgba,
            (Nv16, Bgra) => Rgba,
            (Nv24, Rgb) => Rgb,
            (Nv24, Rgba) => Rgba,
            (Nv24, Grey) => Grey,
            (Nv24, Bgra) => Rgba,
            (PlanarRgb, Rgb) => Rgb,
            (PlanarRgb, Rgba) => Rgb,
            (PlanarRgb, Bgra) => Rgb,
            (PlanarRgba, Rgb) => Rgba,
            (PlanarRgba, Rgba) => Rgba,
            (PlanarRgba, Bgra) => Rgba,
            (s, d) => {
                return Err(Error::NotSupported(format!("Conversion from {s} to {d}",)));
            }
        };

        let need_resize_flip_rotation = rotation != Rotation::None
            || flip != Flip::None
            || src_w != dst_w
            || src_h != dst_h
            || crop.src_rect.is_some_and(|c| {
                c != Rect {
                    left: 0,
                    top: 0,
                    width: src_w,
                    height: src_h,
                }
            })
            || crop.dst_rect.is_some_and(|c| {
                c != Rect {
                    left: 0,
                    top: 0,
                    width: dst_w,
                    height: dst_h,
                }
            });

        // Pick the resolved colorimetry for a single conversion by its YUV
        // side: YUV→RGB decodes use the source side, RGB→YUV encodes the dest.
        let direct_is_yuv_src = matches!(src_fmt, Nv12 | Nv16 | Nv24 | Yuyv | Vyuy);
        let direct_params = if direct_is_yuv_src {
            src_params
        } else {
            dst_params
        };

        // Fused NV→planar: decode the YUV source into packed RGB one
        // cache-resident row strip at a time and NEON-deinterleave each strip
        // straight into the destination planes, so the full-size packed-RGB
        // intermediate never round-trips through DRAM and is not reallocated
        // per frame. JPEG decodes to the NV family and the model wants planar
        // RGB, so this is the hot Orin CPU-preprocess path.
        //
        // Two shapes take this path: the whole-frame case (no crop, dst ==
        // src size — the original hot path, unchanged) and a **scale-
        // identity** source crop (crop size == destination size, no rotate/
        // flip, full destination placement, chroma-aligned origin) — the
        // primary CPU-fallback path for uniform tiling (e.g. SAHI tiles cut
        // from one frame). Any other shape (an actual resize, a partial
        // destination placement, or a chroma-misaligned origin) falls
        // through to the general pipeline below — never snapped or shifted
        // into alignment.
        let full_dst_rect = Rect {
            left: 0,
            top: 0,
            width: dst_w,
            height: dst_h,
        };
        let fused_region = if rotation != Rotation::None || flip != Flip::None {
            None
        } else {
            match crop.src_rect {
                None if src_w == dst_w && src_h == dst_h => Some(None),
                None => None,
                Some(r)
                    if r.width == dst_w
                        && r.height == dst_h
                        && crop.dst_rect.is_none_or(|d| d == full_dst_rect)
                        && chroma_alignment_ok(src_fmt, r) =>
                {
                    Some(Some(r))
                }
                Some(_) => None,
            }
        };
        if let Some(region) = fused_region {
            if matches!(src_fmt, Nv12 | Nv16 | Nv24) && matches!(dst_fmt, PlanarRgb | PlanarRgba) {
                #[cfg(test)]
                {
                    self.fused_hits += 1;
                }
                return self.convert_nv_to_planar_fused(
                    src,
                    dst,
                    src_fmt,
                    dst_fmt,
                    direct_params,
                    region,
                );
            }
        }

        // check if a direct conversion can be done
        if !need_resize_flip_rotation && Self::support_conversion_pf(src_fmt, dst_fmt) {
            return Self::convert_format_pf(src, dst, src_fmt, dst_fmt, direct_params);
        }

        // any extra checks
        if dst_fmt == Yuyv && !dst_w.is_multiple_of(2) {
            return Err(Error::NotSupported(format!(
                "{} destination must have width divisible by 2",
                dst_fmt,
            )));
        }

        // Take the cached intermediates out of `self` so the resize step can
        // borrow `self` exclusively; they are restored before returning on the
        // success path. Reused buffers are not re-zeroed — every consumer below
        // fully overwrites the region it later reads (the pre-resize convert
        // writes all of `tmp`; the resize writes the scaled rect of `tmp2` and
        // its letterbox border is either pre-filled from `dst` or overwritten in
        // `dst` by the final `fill_image_outside_crop_u8`).
        let mut cached_tmp = self.convert_tmp.take();
        let mut cached_tmp2 = self.convert_tmp2.take();

        // For a cropped convert, size the pre-resize intermediate to the crop
        // (grown by the resize filter's halo) instead of the whole frame — see
        // `pre_resize_region`. `None` keeps the previous full-frame behaviour,
        // which is also the uncropped hot path.
        let pre_region = if intermediate != src_fmt {
            self.pre_resize_region(src_fmt, (src_w, src_h), (dst_w, dst_h), rotation, crop)
        } else {
            None
        };

        // create tmp buffer (reusing the cached one when its geometry matches)
        let tmp_holder: Option<Tensor<u8>> = if intermediate != src_fmt {
            let _s = tracing::trace_span!(
                "image.convert.cpu.format_convert",
                from = ?src_fmt,
                to = ?intermediate,
                pass = "pre_resize",
            )
            .entered();
            let (tmp_w, tmp_h) = pre_region.map_or((src_w, src_h), |g| (g.width, g.height));
            let mut t = Self::reuse_or_alloc_image(cached_tmp.take(), tmp_w, tmp_h, intermediate)?;
            #[cfg(test)]
            {
                self.last_tmp_dims = Some((tmp_w, tmp_h));
            }
            match pre_region {
                Some(g) => {
                    let mut sub = Self::reuse_or_alloc_image(
                        self.convert_src_sub.take(),
                        g.width,
                        g.height,
                        src_fmt,
                    )?;
                    {
                        let _s = tracing::trace_span!(
                            "image.convert.cpu.extract_region",
                            region_w = g.width,
                            region_h = g.height,
                        )
                        .entered();
                        Self::extract_nv_region(src, &mut sub, src_fmt, g)?;
                    }
                    Self::convert_format_pf(&sub, &mut t, src_fmt, intermediate, src_params)?;
                    self.convert_src_sub = Some(sub);
                }
                None => Self::convert_format_pf(src, &mut t, src_fmt, intermediate, src_params)?,
            }
            Some(t)
        } else {
            None
        };

        // The intermediate now starts at `grown`'s origin rather than the
        // frame's, so rebase the source crop into its coordinates for the
        // resize step. The crop's size — and therefore the resize scale, the
        // filter coefficients, and every output pixel — is unchanged.
        let crop = match (pre_region, crop.src_rect) {
            (Some(g), Some(r)) => ResolvedCrop {
                src_rect: Some(Rect {
                    left: r.left - g.left,
                    top: r.top - g.top,
                    ..r
                }),
                ..crop
            },
            _ => crop,
        };
        let (tmp, tmp_fmt): (&Tensor<u8>, PixelFormat) = match &tmp_holder {
            Some(t) => (t, intermediate),
            None => (src, src_fmt),
        };

        // format must be RGB/RGBA/GREY
        debug_assert!(matches!(tmp_fmt, Rgb | Rgba | Grey));
        if tmp_fmt == dst_fmt {
            let _s = tracing::trace_span!("image.convert.cpu.resize_flip_rotate").entered();
            self.resize_flip_rotate_pf(tmp, dst, dst_fmt, rotation, flip, crop)?;
        } else if !need_resize_flip_rotation {
            let _s = tracing::trace_span!(
                "image.convert.cpu.format_convert",
                from = ?tmp_fmt,
                to = ?dst_fmt,
                pass = "direct",
            )
            .entered();
            Self::convert_format_pf(tmp, dst, tmp_fmt, dst_fmt, dst_params)?;
        } else {
            let mut tmp2 = Self::reuse_or_alloc_image(cached_tmp2.take(), dst_w, dst_h, tmp_fmt)?;
            if crop.dst_rect.is_some_and(|c| {
                c != Rect {
                    left: 0,
                    top: 0,
                    width: dst_w,
                    height: dst_h,
                }
            }) && crop.dst_color.is_none()
            {
                Self::convert_format_pf(dst, &mut tmp2, dst_fmt, tmp_fmt, dst_params)?;
            }
            {
                let _s = tracing::trace_span!("image.convert.cpu.resize_flip_rotate").entered();
                self.resize_flip_rotate_pf(tmp, &mut tmp2, tmp_fmt, rotation, flip, crop)?;
            }
            {
                let _s = tracing::trace_span!(
                    "image.convert.cpu.format_convert",
                    from = ?tmp_fmt,
                    to = ?dst_fmt,
                    pass = "post_resize",
                )
                .entered();
                Self::convert_format_pf(&tmp2, dst, tmp_fmt, dst_fmt, dst_params)?;
            }
            cached_tmp2 = Some(tmp2);
        }
        // Restore the intermediates to the cache for the next call (`tmp` — a
        // borrow of `tmp_holder` — is no longer used past this point).
        if let Some(t) = tmp_holder {
            cached_tmp = Some(t);
        }
        self.convert_tmp = cached_tmp;
        self.convert_tmp2 = cached_tmp2;

        if let (Some(dst_rect), Some(dst_color)) = (crop.dst_rect, crop.dst_color) {
            let full_rect = Rect {
                left: 0,
                top: 0,
                width: dst_w,
                height: dst_h,
            };
            if dst_rect != full_rect {
                Self::fill_image_outside_crop_u8(dst, dst_color, dst_rect)?;
            }
        }

        Ok(())
    }

    fn draw_decoded_masks_impl(
        &mut self,
        dst: &mut Tensor<u8>,
        detect: &[DetectBox],
        segmentation: &[Segmentation],
        opacity: f32,
        color_mode: crate::ColorMode,
    ) -> Result<()> {
        let dst_fmt = dst.format().ok_or(Error::NotAnImage)?;
        if !matches!(dst_fmt, PixelFormat::Rgba | PixelFormat::Rgb) {
            return Err(crate::Error::NotSupported(
                "CPU image rendering only supports RGBA or RGB images".to_string(),
            ));
        }

        let _timer = FunctionTimer::new("CPUProcessor::draw_decoded_masks");

        let dst_w = dst.width().unwrap();
        let dst_h = dst.height().unwrap();
        let dst_rs = tensor_row_stride(dst);
        let dst_c = dst_fmt.channels();

        let mut map = dst.map_mut()?;
        let dst_slice = map.as_mut_slice();

        self.render_box(dst_w, dst_h, dst_rs, dst_c, dst_slice, detect, color_mode)?;

        if segmentation.is_empty() {
            return Ok(());
        }

        // Semantic segmentation (e.g. ModelPack) has C > 1 (multi-class),
        // instance segmentation (e.g. YOLO) has C = 1 (binary per-instance).
        let is_semantic = segmentation[0].segmentation.shape()[2] > 1;

        if is_semantic {
            self.render_modelpack_segmentation(
                dst_w,
                dst_h,
                dst_rs,
                dst_c,
                dst_slice,
                &segmentation[0],
                opacity,
            )?;
        } else {
            for (idx, (seg, det)) in segmentation.iter().zip(detect).enumerate() {
                let color_index = color_mode.index(idx, det.label);
                self.render_yolo_segmentation(
                    dst_w,
                    dst_h,
                    dst_rs,
                    dst_c,
                    dst_slice,
                    seg,
                    color_index,
                    opacity,
                )?;
            }
        }

        Ok(())
    }

    fn draw_proto_masks_impl(
        &mut self,
        dst: &mut Tensor<u8>,
        detect: &[DetectBox],
        proto_data: &ProtoData,
        opacity: f32,
        letterbox: Option<[f32; 4]>,
        color_mode: crate::ColorMode,
    ) -> Result<()> {
        let dst_fmt = dst.format().ok_or(Error::NotAnImage)?;
        if !matches!(dst_fmt, PixelFormat::Rgba | PixelFormat::Rgb) {
            return Err(crate::Error::NotSupported(
                "CPU image rendering only supports RGBA or RGB images".to_string(),
            ));
        }

        let _timer = FunctionTimer::new("CPUProcessor::draw_proto_masks");

        let dst_w = dst.width().unwrap();
        let dst_h = dst.height().unwrap();
        let dst_rs = tensor_row_stride(dst);
        let channels = dst_fmt.channels();

        let mut map = dst.map_mut()?;
        let dst_slice = map.as_mut_slice();

        self.render_box(
            dst_w, dst_h, dst_rs, channels, dst_slice, detect, color_mode,
        )?;

        if detect.is_empty() {
            return Ok(());
        }
        let proto_shape = proto_data.protos.shape();
        if proto_shape.len() != 3 {
            return Err(Error::InvalidShape(format!(
                "protos tensor must be rank-3, got {proto_shape:?}"
            )));
        }
        let proto_h = proto_shape[0];
        let proto_w = proto_shape[1];
        let num_protos = proto_shape[2];
        let coeff_shape = proto_data.mask_coefficients.shape();
        if coeff_shape.len() != 2 {
            return Err(Error::InvalidShape(format!(
                "mask_coefficients tensor must be rank-2, got {coeff_shape:?}"
            )));
        }
        // Genuine "no detections this frame" → nothing to render.
        if coeff_shape[0] == 0 {
            return Ok(());
        }
        if coeff_shape[1] != num_protos {
            return Err(Error::InvalidShape(format!(
                "mask_coefficients second dimension must match num_protos \
                 ({num_protos}), got {coeff_shape:?}"
            )));
        }

        // Widen coefficients to f32 once; shape [N, num_protos].
        let coeff_f32: Vec<f32> = match proto_data.mask_coefficients.dtype() {
            DType::F32 => {
                let t = proto_data.mask_coefficients.as_f32().expect("F32");
                let m = t.map_read()?;
                m.as_slice().to_vec()
            }
            DType::F16 => {
                let t = proto_data.mask_coefficients.as_f16().expect("F16");
                let m = t.map_read()?;
                m.as_slice().iter().map(|v| v.to_f32()).collect()
            }
            DType::I8 => {
                let t = proto_data.mask_coefficients.as_i8().expect("I8");
                let m = t.map_read()?;
                if let Some(q) = t.quantization() {
                    use edgefirst_tensor::QuantMode;
                    let (scale, zp) = match q.mode() {
                        QuantMode::PerTensor { scale, zero_point } => (scale, zero_point as f32),
                        QuantMode::PerTensorSymmetric { scale } => (scale, 0.0),
                        other => {
                            return Err(Error::NotSupported(format!(
                                "I8 mask_coefficients quantization mode {other:?} not supported"
                            )));
                        }
                    };
                    m.as_slice()
                        .iter()
                        .map(|&v| (v as f32 - zp) * scale)
                        .collect()
                } else {
                    m.as_slice().iter().map(|&v| v as f32).collect()
                }
            }
            DType::I16 => {
                let t = proto_data.mask_coefficients.as_i16().expect("I16");
                let m = t.map_read()?;
                if let Some(q) = t.quantization() {
                    use edgefirst_tensor::QuantMode;
                    let (scale, zp) = match q.mode() {
                        QuantMode::PerTensor { scale, zero_point } => (scale, zero_point as f32),
                        QuantMode::PerTensorSymmetric { scale } => (scale, 0.0),
                        other => {
                            return Err(Error::NotSupported(format!(
                                "I16 mask_coefficients quantization mode {other:?} not supported"
                            )));
                        }
                    };
                    m.as_slice()
                        .iter()
                        .map(|&v| (v as f32 - zp) * scale)
                        .collect()
                } else {
                    m.as_slice().iter().map(|&v| v as f32).collect()
                }
            }
            other => {
                return Err(Error::InvalidShape(format!(
                    "mask_coefficients dtype {other:?} not supported"
                )));
            }
        };

        // Precompute letterbox scale/offset for output-pixel → proto-pixel mapping.
        let (lx0, lx_range, ly0, ly_range) = match letterbox {
            Some([lx0, ly0, lx1, ly1]) => (lx0, lx1 - lx0, ly0, ly1 - ly0),
            None => (0.0_f32, 1.0_f32, 0.0_f32, 1.0_f32),
        };

        // Per-dtype dispatch. Map protos once, call the inner draw loop
        // with a dtype-specialized loader closure.
        match proto_data.protos.dtype() {
            DType::F32 => {
                let t = proto_data.protos.as_f32().expect("F32");
                let m = t.map_read()?;
                self.draw_proto_masks_inner(
                    dst_slice,
                    dst_w,
                    dst_h,
                    dst_rs,
                    channels,
                    detect,
                    m.as_slice(),
                    &coeff_f32,
                    proto_h,
                    proto_w,
                    num_protos,
                    opacity,
                    (lx0, lx_range, ly0, ly_range),
                    color_mode,
                    0.0_f32,
                    |p: &f32, _| *p,
                );
            }
            DType::F16 => {
                let t = proto_data.protos.as_f16().expect("F16");
                let m = t.map_read()?;
                self.draw_proto_masks_inner(
                    dst_slice,
                    dst_w,
                    dst_h,
                    dst_rs,
                    channels,
                    detect,
                    m.as_slice(),
                    &coeff_f32,
                    proto_h,
                    proto_w,
                    num_protos,
                    opacity,
                    (lx0, lx_range, ly0, ly_range),
                    color_mode,
                    0.0_f32,
                    |p: &half::f16, _| p.to_f32(),
                );
            }
            DType::I8 => {
                use edgefirst_tensor::QuantMode;
                let t = proto_data.protos.as_i8().expect("I8");
                let m = t.map_read()?;
                let quant = t.quantization().ok_or_else(|| {
                    Error::InvalidShape("I8 protos require quantization metadata".into())
                })?;
                let (scale, zp) = match quant.mode() {
                    QuantMode::PerTensor { scale, zero_point } => (scale, zero_point as f32),
                    QuantMode::PerTensorSymmetric { scale } => (scale, 0.0),
                    QuantMode::PerChannel { axis, .. }
                    | QuantMode::PerChannelSymmetric { axis, .. } => {
                        return Err(Error::NotSupported(format!(
                            "per-channel quantization (axis={axis}) in draw_proto_masks \
                             CPU path not yet supported"
                        )));
                    }
                };
                self.draw_proto_masks_inner(
                    dst_slice,
                    dst_w,
                    dst_h,
                    dst_rs,
                    channels,
                    detect,
                    m.as_slice(),
                    &coeff_f32,
                    proto_h,
                    proto_w,
                    num_protos,
                    opacity,
                    (lx0, lx_range, ly0, ly_range),
                    color_mode,
                    scale,
                    move |p: &i8, _| (*p as f32) - zp,
                );
            }
            other => {
                return Err(Error::InvalidShape(format!(
                    "proto tensor dtype {other:?} not supported"
                )));
            }
        }

        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    fn draw_proto_masks_inner<P: Copy>(
        &self,
        dst_slice: &mut [u8],
        dst_w: usize,
        dst_h: usize,
        dst_rs: usize,
        channels: usize,
        detect: &[DetectBox],
        protos: &[P],
        coeff_all_f32: &[f32],
        proto_h: usize,
        proto_w: usize,
        num_protos: usize,
        opacity: f32,
        letterbox_xy: (f32, f32, f32, f32),
        color_mode: crate::ColorMode,
        acc_scale: f32,
        load_f32: impl Fn(&P, f32) -> f32 + Copy,
    ) {
        let (lx0, lx_range, ly0, ly_range) = letterbox_xy;
        let stride_y = proto_w * num_protos;
        for (idx, det) in detect.iter().enumerate() {
            let coeff = &coeff_all_f32[idx * num_protos..(idx + 1) * num_protos];
            let color_index = color_mode.index(idx, det.label);
            let color = self.colors[color_index % self.colors.len()];
            let alpha = if opacity == 1.0 {
                color[3] as u16
            } else {
                (color[3] as f32 * opacity).round() as u16
            };

            let start_x = (dst_w as f32 * det.bbox.xmin).round() as usize;
            let start_y = (dst_h as f32 * det.bbox.ymin).round() as usize;
            let end_x = ((dst_w as f32 * det.bbox.xmax).round() as usize).min(dst_w);
            let end_y = ((dst_h as f32 * det.bbox.ymax).round() as usize).min(dst_h);

            for y in start_y..end_y {
                for x in start_x..end_x {
                    let px = (lx0 + (x as f32 / dst_w as f32) * lx_range) * proto_w as f32 - 0.5;
                    let py = (ly0 + (y as f32 / dst_h as f32) * ly_range) * proto_h as f32 - 0.5;

                    // Bilinear interpolation with per-load widening. Inline
                    // bilinear-sample since bilinear_dot_slice takes a
                    // different closure shape (no `zp` arg).
                    let x0 = (px.floor() as isize).clamp(0, proto_w as isize - 1) as usize;
                    let y0 = (py.floor() as isize).clamp(0, proto_h as isize - 1) as usize;
                    let x1 = (x0 + 1).min(proto_w - 1);
                    let y1 = (y0 + 1).min(proto_h - 1);
                    let fx = px - px.floor();
                    let fy = py - py.floor();
                    let w00 = (1.0 - fx) * (1.0 - fy);
                    let w10 = fx * (1.0 - fy);
                    let w01 = (1.0 - fx) * fy;
                    let w11 = fx * fy;
                    let b00 = y0 * stride_y + x0 * num_protos;
                    let b10 = y0 * stride_y + x1 * num_protos;
                    let b01 = y1 * stride_y + x0 * num_protos;
                    let b11 = y1 * stride_y + x1 * num_protos;
                    let mut acc = 0.0_f32;
                    for p in 0..num_protos {
                        let v00 = load_f32(&protos[b00 + p], 0.0);
                        let v10 = load_f32(&protos[b10 + p], 0.0);
                        let v01 = load_f32(&protos[b01 + p], 0.0);
                        let v11 = load_f32(&protos[b11 + p], 0.0);
                        let val = w00 * v00 + w10 * v10 + w01 * v01 + w11 * v11;
                        acc += coeff[p] * val;
                    }
                    let final_acc = if acc_scale == 0.0 {
                        acc
                    } else {
                        acc_scale * acc
                    };
                    // Pass-through: acc_scale=0.0 means "no scaling" (f32/f16
                    // native); non-zero means "apply scale once" (i8 with
                    // per-tensor quant).
                    let mask = 1.0 / (1.0 + (-final_acc).exp());
                    if mask < 0.5 {
                        continue;
                    }
                    let dst_index = y * dst_rs + x * channels;
                    for c in 0..3 {
                        dst_slice[dst_index + c] = ((color[c] as u16 * alpha
                            + dst_slice[dst_index + c] as u16 * (255 - alpha))
                            / 255) as u8;
                    }
                }
            }
        }
    }
}