# SPDX-License-Identifier: Apache-2.0
"""C14 image pre-processing: bit-exact host emulations of the Autoware camera nodes' resize kernels, driven by
per-source-size lookup tables (LUTs) that are computed once and cached.

Each preset reproduces one deployed CUDA / OpenCV routine exactly (integer outputs equal, not "close"), because a
standard resize in its place changes the network input enough to fail the PCC gates (YOLOX: ``cv2.resize`` drops the
detection-output PCC to 0.94, research/yolox/SPEC.md section 3). The tables hold everything that depends only on the
source and destination sizes (source indices, interpolation weights, letterbox geometry); applying them is a few
vectorised numpy passes over the image. numpy only; no ttnn, no torch, no OpenCV.

Presets (PLAN.md C14; the other camera ports add theirs here):

==========================  =========================================================================================
``yolox_letterbox``         autoware_tensorrt_yolox ``resize_bilinear_letterbox_nhwc_to_nchw32_batch_kernel``
                            (autoware_universe @ 9ceaccf, ``perception/autoware_tensorrt_yolox/src/preprocess.cu:43-129``,
                            launched by ``tensorrt_yolox.cpp:243-332``): a "bilinear" resize with inverted,
                            half-magnitude weights, float -> int truncation between the horizontal and vertical passes,
                            ``lroundf``, top-left letterbox padded with 114, BGR kept, no normalisation.
``bevdet_nearest_crop``     the ``bevdet::Preprocess`` TensorRT plugin of autoware_tensorrt_bevdet (``bevdet_vendor``
                            0.2.1; research/bevdet/SPEC.md section 3.2): nearest resize by r = 0.44f + crop (140, 0)
                            of the 1600x900 image, ``roundf(i / r + crop_h / r)`` rows / columns in float32, planar
                            B, G, R uint8 (the normalisation ``bevdet_normalize`` is separate: device-side in the TT
                            port). The node's ``cv::resize`` of non-1600x900 images to 1600x900 stays with the caller.
``bevformer_preprocess``    autoware_tensorrt_bevformer's ``preprocessing_pipeline`` (autoware_universe @ 9ceaccf,
                            ``src/preprocessing/normalize_multiview_image.cpp:70-95`` +
                            ``augmentation_transforms.cpp:90-222``; research/bevformer/SPEC.md section 3.1):
                            normalise FIRST (BGR uint8 -> float, ``(x - mean_c) / std_c``, BGR mean, std 1), then
                            ``cv::resize`` x0.8 with INTER_LINEAR **on float data** (1600x900 -> 1280x720), then pad
                            bottom / right with 0 (normalised space) to a multiple of 32 (1280x736), CHW. The node's
                            ``cv::resize`` of non-1600x900 uint8 images to 1600x900 stays with the caller.
==========================  =========================================================================================

Exactness notes for ``yolox_letterbox`` (all arithmetic is float32 in the kernel):

- ``scale = min(dst_w / (float)src_w, dst_h / (float)src_h)``; ``r_h = (int)(scale * src_h)``,
  ``r_w = (int)(scale * src_w)``; the kernel receives ``k = (float)(1.0 / (double)scale)`` (``preprocess.cu:121-128``).
- per output row ``h``: ``c = k * (float)(h + 0.5)``, ``src_row = clamp(lroundf(c) - 1, 0, src_h - 2)``,
  weight ``t = (1 + lroundf(c) - c) / 2`` (float32), and the same per column (``preprocess.cu:48-59, 77-88``).
- ``lerp1d(a, b, t) = fma(t, b, fma(-t, a, a))`` with float operands (CUDA's float overload, i.e. ``fmaf``:
  one rounding per call); the horizontal results are truncated to int before the vertical ``lerp1d``
  (``lerp1d`` takes ``int a, int b``), and the result is ``lroundf``-ed (``preprocess.cu:43-46, 57, 101-103``).
- ``t`` lies in [0.25, 0.75], so ``t`` is a multiple of 2**-25 and ``a - t*a`` / ``t*b + inner`` with integer
  ``a, b`` in [0, 255] need at most 34 significant bits: float64 holds them exactly, and one ``astype(float32)`` is
  exactly the single rounding of ``fmaf``.
- Rows ``h >= r_h`` and columns ``w >= r_w`` are 114 (``preprocess.cu:109-110``); the output values are integers in
  [0, 255], so this module returns uint8 (the network input is ``float(value)``, ``norm_factor`` 1.0).

Verified against two independent scalar transliterations of the CUDA source (``common/tests/host/test_image_host.py``
and research/yolox/scripts/{check_preprocess,verify_preproc_scalar}.py: 0 mismatches). Residual uncertainty, from the
SPEC: if nvcc picked the double ``fma`` overload, 0-2 of 3000 samples would differ by 1 (not testable without CUDA).

Exactness notes for ``bevformer_preprocess`` (float32 data; OpenCV's float INTER_LINEAR path):

- normalisation: ``Mat - mean`` / ``std`` evaluates as ``convertTo(alpha = 1/std, beta = -mean/std)`` with float
  ``alpha`` / ``beta`` (one fused multiply-add per value); with the deployed std 1 it is ``x - float(mean_c)``.
- resize coefficients (``resize.cpp``): ``scale = 1.0 / ((double)dst / src)``, ``f = (float)((d + 0.5) * scale -
  0.5)``, ``s = floor(f)``, ``w = f - s`` (float), clamped to the border (``s < 0`` -> ``s = 0, w = 0``;
  ``s >= src - 1`` -> ``s = src - 1, w = 0``); the destination size is ``(int)(src * 0.8f)`` in float.
- each pass is ``fma(x1 - x0, w, x0)`` in float32 (the difference rounded once, then one rounding of the fused
  product-sum), horizontal pass first, then vertical. This is what OpenCV 4.8.1 and 4.11.0 compute on this x86-64
  host (AVX2 + FMA) bit for bit for every x0.8 down-scale, on arbitrary float images (not just uint8 - mean): the
  plain two-product form ``x0 * (1 - w) + x1 * w`` differs from it in about 10 % of the samples by 1 ulp
  (``tests/host/test_image_bevformer_host.py`` holds both checks). The weights of a x0.8 resize are multiples of
  1/8, so float64 holds every fused product-sum exactly and one ``astype(float32)`` is the single rounding. OpenCV
  computes the coefficients of other scales differently, so :func:`opencv_linear_lut` refuses them by default.
"""
from __future__ import annotations

import threading
from collections import OrderedDict
from dataclasses import dataclass
from typing import Any, Callable, Dict, Hashable, Optional, Tuple

import numpy as np

__all__ = [
    "LetterboxGeometry",
    "YoloxLetterboxLUT",
    "yolox_letterbox",
    "yolox_letterbox_geometry",
    "yolox_letterbox_lut",
    "lroundf",
    "f32",
    "LUTCache",
    "PRESETS",
    "YOLOX_PAD_VALUE",
    "NearestCropLUT",
    "bevdet_nearest_lut",
    "bevdet_nearest_crop",
    "bevdet_normalize",
    "BEVDET_MEAN",
    "BEVDET_STD",
    "LinearResizeLUT",
    "opencv_linear_lut",
    "opencv_resize_linear_f32",
    "bevformer_input_geometry",
    "bevformer_normalize",
    "bevformer_preprocess",
    "BEVFORMER_MEAN_BGR",
    "BEVFORMER_STD",
]

F32 = np.float32
YOLOX_PAD_VALUE = 114  # preprocess.cu:109-110 (letterbox fill), the YOLOX training pad value


def f32(x: Any) -> np.ndarray:
    """``x`` rounded once to float32 (round-to-nearest-even), as a numpy float32 scalar or array."""
    return np.asarray(x).astype(F32)


def lroundf(x: Any) -> np.ndarray:
    """C ``lroundf``: round half away from zero, for float32 inputs (exact in float64). Returns int64."""
    x = np.asarray(x, dtype=np.float64)
    return (np.sign(x) * np.floor(np.abs(x) + 0.5)).astype(np.int64)


class LUTCache:
    """A small thread-safe LRU of tables keyed by the source / destination sizes (a camera stream keeps its size,
    so the table is built once per stream; Autoware re-creates its buffers only when the source size changes,
    ``tensorrt_yolox.cpp:253-291``)."""

    def __init__(self, maxsize: int = 8):
        self.maxsize = int(maxsize)
        self._items: "OrderedDict[Hashable, Any]" = OrderedDict()
        self._lock = threading.Lock()
        self.hits = 0
        self.misses = 0

    def get(self, key: Hashable, make: Callable[[], Any]) -> Any:
        with self._lock:
            if key in self._items:
                self._items.move_to_end(key)
                self.hits += 1
                return self._items[key]
        value = make()  # outside the lock: building a table takes a few ms
        with self._lock:
            self.misses += 1
            self._items[key] = value
            self._items.move_to_end(key)
            while len(self._items) > self.maxsize:
                self._items.popitem(last=False)
        return value

    def clear(self) -> None:
        with self._lock:
            self._items.clear()
            self.hits = self.misses = 0

    def __len__(self) -> int:
        return len(self._items)


# ------------------------------------------------------------------------------------------- YOLOX letterbox


@dataclass(frozen=True)
class LetterboxGeometry:
    """Where the source image lands in the network input, and what the decoder and the mask crop use.

    ``scale`` is the float32 scale of ``preprocess.cu:121`` (also ``scales_`` of ``tensorrt_yolox.cpp:285``, used to
    map boxes back, and the mask scale of ``tensorrt_yolox.cpp:528-536``); ``resized_hw`` = ``(r_h, r_w)``, the
    un-letterboxed region at the top-left of the ``dst_hw`` canvas. Autoware's mask crop ``(out_h, out_w)`` =
    ``((int)(src_h * scale), (int)(src_w * scale))`` is the same pair (float multiplication commutes)."""

    src_hw: Tuple[int, int]
    dst_hw: Tuple[int, int]
    scale: float           # exact float32 value, stored as a Python float
    resized_hw: Tuple[int, int]
    pad_value: int = YOLOX_PAD_VALUE

    @property
    def scale_f32(self) -> np.float32:
        return F32(self.scale)

    @property
    def mask_hw(self) -> Tuple[int, int]:
        """``(out_h, out_w)`` of Autoware's mask (``getMaskImageGpu``): the un-letterboxed region."""
        return self.resized_hw

    def to_dict(self) -> Dict[str, Any]:
        return {"src_hw": list(self.src_hw), "dst_hw": list(self.dst_hw), "scale": self.scale,
                "scale_f32_hex": F32(self.scale).tobytes()[::-1].hex(), "resized_hw": list(self.resized_hw),
                "mask_hw": list(self.mask_hw), "pad_value": self.pad_value,
                "letterbox": "top-left (padding at the bottom / right)"}


def yolox_letterbox_geometry(src_h: int, src_w: int, dst_h: int = 960, dst_w: int = 960,
                             pad_value: int = YOLOX_PAD_VALUE) -> LetterboxGeometry:
    """Letterbox geometry of ``preprocess.cu:121-123`` (float32 scale, int truncation)."""
    src_h, src_w, dst_h, dst_w = int(src_h), int(src_w), int(dst_h), int(dst_w)
    if src_h < 2 or src_w < 2:
        raise ValueError(f"the Autoware resize needs a source of at least 2x2 pixels, got {src_w}x{src_h}")
    if dst_h < 1 or dst_w < 1:
        raise ValueError(f"bad destination size {dst_w}x{dst_h}")
    scale = min(F32(dst_w) / F32(src_w), F32(dst_h) / F32(src_h))   # float32 divisions
    scale = F32(scale)
    r_h = int(F32(scale * F32(src_h)))
    r_w = int(F32(scale * F32(src_w)))
    return LetterboxGeometry((src_h, src_w), (dst_h, dst_w), float(scale), (min(r_h, dst_h), min(r_w, dst_w)),
                             int(pad_value))


def _axis_table(n_dst: int, n_src: int, k: np.float32) -> Tuple[np.ndarray, np.ndarray]:
    """Source index and weight of every destination row (or column): ``preprocess.cu:48-51, 77-88``."""
    c = (k * (np.arange(n_dst) + 0.5).astype(F32)).astype(F32)          # centroid = scale * (float)(h + 0.5)
    rc = lroundf(c)
    idx = rc - 1
    idx = np.where(idx < 0, 0, idx)
    idx = np.where(idx >= n_src - 1, n_src - 2, idx)
    weight = ((F32(1) + rc.astype(F32)).astype(F32) - c).astype(F32)     # 1 + lroundf(c) - c (float32)
    weight = (weight / F32(2)).astype(F32)
    return idx.astype(np.int64), weight


@dataclass(frozen=True)
class YoloxLetterboxLUT:
    """The per-size tables of the YOLOX letterbox: source row / column of every destination row / column and the
    float32 weights. Build with :func:`yolox_letterbox_lut` (cached); apply with :meth:`apply`."""

    geometry: LetterboxGeometry
    rows: np.ndarray       # (r_h,) int64: source row hi (hi + 1 is the second tap)
    row_w: np.ndarray      # (r_h,) float32: vertical weight
    cols: np.ndarray       # (r_w,) int64: source column wi
    col_w: np.ndarray      # (r_w,) float32: horizontal weight
    k: float               # the kernel's float32 ``scale`` argument (1 / scale)

    def apply(self, image: np.ndarray, *, out: Optional[np.ndarray] = None, chunk_rows: int = 96) -> np.ndarray:
        """Letterbox one BGR uint8 image ``(src_h, src_w, 3)`` (any channel count) -> uint8 ``(dst_h, dst_w, C)``
        HWC, channel order unchanged. ``out`` (uint8, that shape) is written in place when given."""
        g = self.geometry
        img = np.asarray(image)
        if img.ndim != 3 or img.dtype != np.uint8 or tuple(img.shape[:2]) != g.src_hw:
            raise ValueError(f"expected a uint8 ({g.src_hw[0]}, {g.src_hw[1]}, C) image, got {img.dtype} {img.shape}")
        dst_h, dst_w = g.dst_hw
        r_h, r_w = g.resized_hw
        shape = (dst_h, dst_w, img.shape[2])
        if out is None:
            out = np.empty(shape, np.uint8)
        elif out.shape != shape or out.dtype != np.uint8:
            raise ValueError(f"out must be uint8 {shape}, got {out.dtype} {out.shape}")
        out[r_h:] = g.pad_value
        out[:r_h, r_w:] = g.pad_value
        if r_h == 0 or r_w == 0:
            return out
        tw = self.col_w.astype(np.float64)[None, :, None]   # exact float32 values
        cols0, cols1 = self.cols, self.cols + 1
        for start in range(0, r_h, max(1, int(chunk_rows))):
            stop = min(r_h, start + int(chunk_rows))
            hi = self.rows[start:stop]
            # horizontal pass on the source rows this chunk needs (hi and hi + 1), each row once
            need = np.unique(np.concatenate([hi, hi + 1]))
            src = img[need]                                                  # (R, src_w, C) uint8
            a = src[:, cols0].astype(np.float64)                             # (R, r_w, C)
            b = src[:, cols1].astype(np.float64)
            inner = (a - tw * a).astype(F32)                                 # fmaf(-t, a, a)
            horiz = (tw * b + inner).astype(F32)                             # fmaf(t, b, inner)
            horiz = np.trunc(horiz).astype(np.float64)                       # lerp1d(int a, int b, ...): truncation
            pos0 = np.searchsorted(need, hi)
            pos1 = np.searchsorted(need, hi + 1)
            a1, b1 = horiz[pos0], horiz[pos1]                                # (rows, r_w, C)
            th = self.row_w[start:stop].astype(np.float64)[:, None, None]
            inner = (a1 - th * a1).astype(F32)
            r = (th * b1 + inner).astype(F32)                                # second lerp1d, float32
            out[start:stop, :r_w] = np.floor(r.astype(np.float64) + 0.5)     # lroundf (r >= 0)
        return out


_YOLOX_LUTS = LUTCache(maxsize=8)


def yolox_letterbox_lut(src_h: int, src_w: int, dst_h: int = 960, dst_w: int = 960,
                        pad_value: int = YOLOX_PAD_VALUE) -> YoloxLetterboxLUT:
    """The (cached) tables for one source size."""

    def make() -> YoloxLetterboxLUT:
        g = yolox_letterbox_geometry(src_h, src_w, dst_h, dst_w, pad_value)
        k = F32(1.0 / np.float64(g.scale_f32))           # 1.0 / scale in double, passed as a float argument
        rows, row_w = _axis_table(dst_h, g.src_hw[0], k)
        cols, col_w = _axis_table(dst_w, g.src_hw[1], k)
        r_h, r_w = g.resized_hw
        return YoloxLetterboxLUT(g, rows[:r_h].copy(), row_w[:r_h].copy(), cols[:r_w].copy(), col_w[:r_w].copy(),
                                 float(k))

    return _YOLOX_LUTS.get(("yolox", int(src_h), int(src_w), int(dst_h), int(dst_w), int(pad_value)), make)


def yolox_letterbox(image: np.ndarray, dst_hw: Tuple[int, int] = (960, 960), *, pad_value: int = YOLOX_PAD_VALUE,
                    layout: str = "hwc", out: Optional[np.ndarray] = None) -> Tuple[np.ndarray, LetterboxGeometry]:
    """Autoware YOLOX letterbox of one BGR8 image ``(H, W, 3)`` -> ``(uint8 image, geometry)``.

    ``layout``: ``"hwc"`` -> ``(dst_h, dst_w, 3)``; ``"nchw"`` -> ``(1, 3, dst_h, dst_w)`` (the ONNX ``images``
    layout; ``.astype(np.float32)`` is exactly the network input). The channel order is unchanged (Autoware feeds
    BGR)."""
    img = np.asarray(image)
    lut = yolox_letterbox_lut(img.shape[0], img.shape[1], int(dst_hw[0]), int(dst_hw[1]), pad_value)
    if layout == "hwc":
        return lut.apply(img, out=out), lut.geometry
    if layout == "nchw":
        hwc = lut.apply(img)
        res = np.ascontiguousarray(hwc.transpose(2, 0, 1)[None])
        if out is not None:
            out[...] = res
            return out, lut.geometry
        return res, lut.geometry
    raise ValueError(f"layout must be 'hwc' or 'nchw', got {layout!r}")


# ------------------------------------------------------------------------------------------- BEVDet nearest crop

# Per-plane statistics of the BEVDet Preprocess plugin, applied to the B, G, R planes in this order: training loaded
# images as RGB (PIL) and its mmlabNormalize swapped them to BGR before applying the "RGB" ImageNet statistics, and
# the Autoware node feeds BGR planes with the same numbers (research/bevdet/SPEC.md section 3.1).
BEVDET_MEAN = (123.675, 116.28, 103.53)
BEVDET_STD = (58.395, 57.12, 57.375)


def _roundf(x: Any) -> np.ndarray:
    """C ``roundf`` (half away from zero) of float32 values -> int64."""
    x = np.asarray(x, F32)
    return (np.sign(x) * np.floor(np.abs(x) + F32(0.5))).astype(np.int64)


@dataclass(frozen=True)
class NearestCropLUT:
    """The ``bevdet::Preprocess`` gather: output row ``i`` / column ``j`` reads source row ``rows[i]`` / column
    ``cols[j]`` of the (already 1600x900) image. Build with :func:`bevdet_nearest_lut` (cached); apply with
    :meth:`apply` (one HWC image) or :meth:`apply_planes` (planar ``[..., C, H, W]``)."""

    src_hw: Tuple[int, int]
    dst_hw: Tuple[int, int]
    crop_hw: Tuple[int, int]
    resize: float          # exact float32 ratio dst_w / src_w, stored as a Python float
    rows: np.ndarray       # (dst_h,) int64
    cols: np.ndarray       # (dst_w,) int64

    def apply(self, image: np.ndarray, *, channels: str = "rgb", out: Optional[np.ndarray] = None) -> np.ndarray:
        """One ``(src_h, src_w, 3)`` uint8 image -> ``(3, dst_h, dst_w)`` uint8 B, G, R planes. ``channels`` is the
        input's order: ``"rgb"`` (PIL / ``ttaw.io.load_image``) or ``"bgr"`` (OpenCV, cv_bridge ``bgr8``)."""
        img = np.asarray(image)
        if img.ndim != 3 or img.shape[2] != 3 or img.dtype != np.uint8 or tuple(img.shape[:2]) != self.src_hw:
            raise ValueError(f"expected a uint8 ({self.src_hw[0]}, {self.src_hw[1]}, 3) image, got {img.dtype} "
                             f"{img.shape} (resize it to the source size first, like the node's cv::resize)")
        if channels not in ("rgb", "bgr"):
            raise ValueError(f"channels must be 'rgb' or 'bgr', got {channels!r}")
        sub = img[self.rows][:, self.cols]                                  # (dst_h, dst_w, 3)
        order = [2, 1, 0] if channels == "rgb" else [0, 1, 2]              # -> B, G, R
        res = sub[:, :, order].transpose(2, 0, 1)
        if out is None:
            return np.ascontiguousarray(res)
        shape = (3,) + tuple(self.dst_hw)
        if out.shape != shape or out.dtype != np.uint8:
            raise ValueError(f"out must be uint8 {shape}, got {out.dtype} {out.shape}")
        out[...] = res
        return out

    def apply_planes(self, planes: np.ndarray) -> np.ndarray:
        """Planar ``[..., C, src_h, src_w]`` (any dtype) -> ``[..., C, dst_h, dst_w]``."""
        p = np.asarray(planes)
        if tuple(p.shape[-2:]) != self.src_hw:
            raise ValueError(f"expected planes [..., {self.src_hw[0]}, {self.src_hw[1]}], got {p.shape}")
        return np.ascontiguousarray(p[..., self.rows, :][..., self.cols])


_BEVDET_LUTS = LUTCache(maxsize=4)


def bevdet_nearest_lut(src_hw: Tuple[int, int] = (900, 1600), dst_hw: Tuple[int, int] = (256, 704),
                       crop_hw: Tuple[int, int] = (140, 0)) -> NearestCropLUT:
    """The (cached) gather of the BEVDet Preprocess plugin: ``r = (float)dst_w / src_w`` (0.44f for the deployed
    config, also the ONNX attribute ``resize_radio``), ``rows[i] = roundf(i / r + crop_h / r)``,
    ``cols[j] = roundf(j / r + crop_w / r)``, every quantity float32 (deployed: rows 318, 320, 323, ..., 898 and
    columns 0, 2, 5, 7, ..., 1598). ``crop_hw`` is the crop offset in RESIZED pixels (the yaml ``crop``, (h, w))."""
    src_h, src_w = int(src_hw[0]), int(src_hw[1])
    dst_h, dst_w = int(dst_hw[0]), int(dst_hw[1])
    crop_h, crop_w = int(crop_hw[0]), int(crop_hw[1])

    def make() -> NearestCropLUT:
        r = F32(F32(dst_w) / F32(src_w))
        off_h = F32(F32(crop_h) / r)
        off_w = F32(F32(crop_w) / r)
        rows = _roundf((np.arange(dst_h, dtype=F32) / r + off_h).astype(F32))
        cols = _roundf((np.arange(dst_w, dtype=F32) / r + off_w).astype(F32))
        if rows.min() < 0 or rows.max() >= src_h or cols.min() < 0 or cols.max() >= src_w:
            raise ValueError(f"crop {crop_hw} with ratio {float(r)} reads outside the {src_w}x{src_h} source")
        return NearestCropLUT((src_h, src_w), (dst_h, dst_w), (crop_h, crop_w), float(r), rows, cols)

    return _BEVDET_LUTS.get(("bevdet", src_h, src_w, dst_h, dst_w, crop_h, crop_w), make)


def bevdet_nearest_crop(image: np.ndarray, *, channels: str = "rgb", src_hw: Tuple[int, int] = (900, 1600),
                        dst_hw: Tuple[int, int] = (256, 704), crop_hw: Tuple[int, int] = (140, 0),
                        out: Optional[np.ndarray] = None) -> np.ndarray:
    """One camera image at the source size (1600x900) -> its ``(3, 256, 704)`` uint8 B, G, R crop: exactly the
    samples the BEVDet network sees before ``bevdet_normalize``."""
    return bevdet_nearest_lut(src_hw, dst_hw, crop_hw).apply(image, channels=channels, out=out)


def bevdet_normalize(crop: np.ndarray, mean: Any = BEVDET_MEAN, std: Any = BEVDET_STD) -> np.ndarray:
    """``(x - mean[c]) / std[c]`` in float32 on B, G, R planes ``[..., 3, H, W]`` (the plugin's fp32 path; its fp16
    mode computes in ``__half``)."""
    m = np.asarray(mean, F32).reshape(3, 1, 1)
    s = np.asarray(std, F32).reshape(3, 1, 1)
    return ((np.asarray(crop).astype(F32) - m) / s).astype(F32)


# ------------------------------------------------------------------------------------------- BEVFormer

# autoware_tensorrt_bevformer config/bevformer.param.yaml data_params (= the BEVFormer training img_norm_cfg,
# DerryHub configs/bevformer/bevformer_small.py:19): caffe-style BGR means, std 1, to_rgb false
BEVFORMER_MEAN_BGR = (103.530, 116.280, 123.675)
BEVFORMER_STD = (1.0, 1.0, 1.0)


@dataclass(frozen=True)
class LinearResizeLUT:
    """OpenCV ``cv::resize(..., INTER_LINEAR)`` of float32 data for one (source, destination) size pair: row / column
    pairs and the float32 weight of the second tap (``resize.cpp`` coefficient loop; module docstring). Build with
    :func:`opencv_linear_lut` (cached); apply with :meth:`apply` (HWC) or :meth:`apply_planes` (``[..., H, W]``)."""

    src_hw: Tuple[int, int]
    dst_hw: Tuple[int, int]
    rows0: np.ndarray      # (dst_h,) int64
    rows1: np.ndarray      # (dst_h,) int64 (clamped to the last row)
    wy: np.ndarray         # (dst_h,) float32 weight of rows1
    cols0: np.ndarray      # (dst_w,) int64
    cols1: np.ndarray
    wx: np.ndarray         # (dst_w,) float32 weight of cols1

    @staticmethod
    def _lerp(x0: np.ndarray, x1: np.ndarray, w: Any, out: Optional[np.ndarray] = None) -> np.ndarray:
        """``fma(x1 - x0, w, x0)`` in float32: the difference rounded once, the fused product-sum rounded once
        (float64 holds ``d * w + x0`` exactly for weights that are multiples of 1/8). ``x1`` is overwritten."""
        d = np.subtract(x1, x0, out=x1)                                   # float32: one rounding
        t = d.astype(np.float64)
        t *= w
        t += x0
        if out is None:
            return t.astype(F32)
        out[...] = t                                                      # float64 -> float32: one rounding
        return out

    @staticmethod
    def _period(i0: np.ndarray, i1: np.ndarray, w: np.ndarray, n_src: int) -> Optional[Tuple[int, int]]:
        """``(p_in, p_out)`` when the taps repeat every ``p_out`` outputs shifted by ``p_in`` inputs, both taps of
        an output lie in one input block and no border clamp occurs (x0.8: 5 inputs -> 4 outputs), else None."""
        n_dst = len(i0)
        for p_out in range(1, min(n_dst, 64) + 1):
            if n_dst % p_out or (n_src * p_out) % n_dst:
                continue
            p_in = n_src * p_out // n_dst
            base = (np.arange(n_dst) // p_out) * p_in
            reps = n_dst // p_out
            if (np.array_equal(i0 - base, np.tile(i0[:p_out], reps)) and np.array_equal(i1, i0 + 1)
                    and np.array_equal(w, np.tile(w[:p_out], reps)) and int(i1[:p_out].max()) < p_in):
                return p_in, p_out
        return None

    def _pass(self, p: np.ndarray, axis: int, i0: np.ndarray, i1: np.ndarray, w: np.ndarray) -> np.ndarray:
        """One interpolation pass along ``axis`` (-1: columns, -2: rows) of a contiguous ``p`` [..., H, W]: strided
        block slices when the taps are periodic (no transposes), else gathers."""
        period = self._period(i0, i1, w, p.shape[axis])
        if period is None:
            wb = w.astype(np.float64) if axis == -1 else w.astype(np.float64)[:, None]
            return self._lerp(np.take(p, i0, axis=axis), np.take(p, i1, axis=axis), wb)
        p_in, p_out = period
        nb = p.shape[axis] // p_in
        taps0 = [int(t) for t in i0[:p_out]]
        run = taps0 == list(range(taps0[0], taps0[0] + p_out))          # x0.8: taps (0..3) and (1..4) of 5
        wj = np.asarray(w[:p_out], np.float64)
        if axis == -1:
            blk = p.reshape(p.shape[:-1] + (nb, p_in))                    # [..., H, nb, p_in] (a view)
            if run:
                a = taps0[0]
                res = self._lerp(blk[..., a:a + p_out], np.array(blk[..., a + 1:a + 1 + p_out]), wj)
            else:
                res = np.empty(p.shape[:-1] + (nb, p_out), F32)
                for j in range(p_out):
                    self._lerp(blk[..., taps0[j]], np.array(blk[..., taps0[j] + 1]), wj[j], out=res[..., j])
            return res.reshape(p.shape[:-1] + (nb * p_out,))
        blk = p.reshape(p.shape[:-2] + (nb, p_in, p.shape[-1]))          # [..., nb, p_in, W] (a view)
        if run:
            a = taps0[0]
            res = self._lerp(blk[..., a:a + p_out, :], np.array(blk[..., a + 1:a + 1 + p_out, :]), wj[:, None])
        else:
            res = np.empty(p.shape[:-2] + (nb, p_out, p.shape[-1]), F32)
            for j in range(p_out):
                self._lerp(blk[..., taps0[j], :], np.array(blk[..., taps0[j] + 1, :]), wj[j], out=res[..., j, :])
        return res.reshape(p.shape[:-2] + (nb * p_out, p.shape[-1]))

    def apply_planes(self, planes: np.ndarray, *, out: Optional[np.ndarray] = None) -> np.ndarray:
        """Planar float32 ``[..., src_h, src_w]`` -> ``[..., dst_h, dst_w]`` (horizontal pass first, as OpenCV)."""
        p = np.ascontiguousarray(planes)
        if p.dtype != F32 or tuple(p.shape[-2:]) != self.src_hw:
            raise ValueError(f"expected float32 planes [..., {self.src_hw[0]}, {self.src_hw[1]}], got {p.dtype} "
                             f"{p.shape}")
        shape = p.shape[:-2] + tuple(self.dst_hw)
        if out is not None and (out.shape != shape or out.dtype != F32):
            raise ValueError(f"out must be float32 {shape}, got {out.dtype} {out.shape}")
        h = self._pass(p, -1, self.cols0, self.cols1, self.wx)            # [..., src_h, dst_w]
        v = self._pass(h, -2, self.rows0, self.rows1, self.wy)            # [..., dst_h, dst_w]
        if out is None:
            return v
        out[...] = v
        return out

    def apply(self, image: np.ndarray, *, out: Optional[np.ndarray] = None) -> np.ndarray:
        """``(src_h, src_w[, C])`` float32 -> ``(dst_h, dst_w[, C])`` float32 (horizontal pass first, as OpenCV)."""
        img = np.asarray(image)
        if img.dtype != F32 or tuple(img.shape[:2]) != self.src_hw:
            raise ValueError(f"expected a float32 {self.src_hw} image, got {img.dtype} {img.shape}")
        planes = np.moveaxis(img, (0, 1), (-2, -1))                          # [C..., H, W]
        res = np.moveaxis(self.apply_planes(planes), (-2, -1), (0, 1))
        if out is None:
            return np.ascontiguousarray(res)
        if out.shape != res.shape or out.dtype != F32:
            raise ValueError(f"out must be float32 {res.shape}, got {out.dtype} {out.shape}")
        out[...] = res
        return out


def _linear_axis(n_src: int, n_dst: int) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
    """``resize.cpp`` INTER_LINEAR coefficients of one axis: (first tap, second tap, float32 weight of the second)."""
    scale = 1.0 / (float(n_dst) / float(n_src))                       # inv_scale = (double)dst / src; 1. / inv_scale
    f = ((np.arange(n_dst, dtype=np.float64) + 0.5) * scale - 0.5).astype(F32)
    s = np.floor(f).astype(np.int64)
    w = (f - s.astype(F32)).astype(F32)
    low = s < 0
    s[low], w[low] = 0, F32(0)
    high = s >= n_src - 1
    s[high], w[high] = n_src - 1, F32(0)
    return s, np.minimum(s + 1, n_src - 1), w


def _eighths(w: np.ndarray) -> bool:
    x = np.asarray(w, np.float64) * 8.0
    return bool(np.all(x == np.round(x)))


_LINEAR_LUTS = LUTCache(maxsize=8)


def opencv_linear_lut(src_hw: Tuple[int, int], dst_hw: Tuple[int, int], *, strict: bool = True) -> LinearResizeLUT:
    """The (cached) float INTER_LINEAR tables for one size pair.

    Verified bit-exact against OpenCV 4.8.1 / 4.11.0 for every x0.8 down-scale (1600x900 -> 1280x720, the deployed
    one; 1920x1080, 160x90, 80x45), whose weights are multiples of 1/8. OpenCV computes the coefficients of other
    scales differently (up-scales and non-dyadic down-scales differ by up to ~1e-4 relative), so ``strict=True``
    refuses a geometry whose weights are not multiples of 1/8; ``strict=False`` gives the textbook coefficients for
    those (an approximation, not an emulation). Exact 2x down-scales are always refused: OpenCV switches
    INTER_LINEAR to INTER_AREA there (``resize.cpp``)."""
    src_h, src_w = int(src_hw[0]), int(src_hw[1])
    dst_h, dst_w = int(dst_hw[0]), int(dst_hw[1])
    if min(src_h, src_w, dst_h, dst_w) < 1:
        raise ValueError(f"empty resize {src_hw} -> {dst_hw}")
    if src_h == 2 * dst_h and src_w == 2 * dst_w:
        raise ValueError("an exact 2x down-scale runs OpenCV's INTER_AREA path, which is not emulated")

    def make() -> LinearResizeLUT:
        r0, r1, wy = _linear_axis(src_h, dst_h)
        c0, c1, wx = _linear_axis(src_w, dst_w)
        return LinearResizeLUT((src_h, src_w), (dst_h, dst_w), r0, r1, wy, c0, c1, wx)

    lut = _LINEAR_LUTS.get(("linear_f32", src_h, src_w, dst_h, dst_w), make)
    if strict and not (_eighths(lut.wx) and _eighths(lut.wy)):
        raise ValueError(f"{src_hw} -> {dst_hw} is not a geometry this emulation reproduces bit for bit (weights must "
                         "be multiples of 1/8, e.g. a x0.8 down-scale); pass strict=False for an approximation")
    return lut


def opencv_resize_linear_f32(image: np.ndarray, dst_hw: Tuple[int, int], *, strict: bool = True) -> np.ndarray:
    """``cv2.resize(image, (dst_w, dst_h))`` (INTER_LINEAR) of a float32 ``(H, W[, C])`` image: bit-exact for the
    geometries :func:`opencv_linear_lut` accepts with ``strict=True``."""
    img = np.asarray(image)
    return opencv_linear_lut(img.shape[:2], dst_hw, strict=strict).apply(img)


def bevformer_input_geometry(src_hw: Tuple[int, int] = (900, 1600), scale: float = 0.8,
                             pad_divisor: int = 32) -> Dict[str, Tuple[int, int]]:
    """Sizes of the node's pipeline for one source size: ``resized_hw`` = ``(int)(src * scale)`` in float32
    (``augmentation_transforms.cpp:104-105``) and ``padded_hw`` (``PadMultiViewImages``, size divisor). Deployed:
    (900, 1600) -> (720, 1280) -> (736, 1280); the network normalises its UV by the padded size."""
    s = F32(scale)
    h = int(F32(F32(int(src_hw[0])) * s))
    w = int(F32(F32(int(src_hw[1])) * s))
    d = int(pad_divisor)
    if d < 1:
        raise ValueError("pad_divisor must be >= 1")
    return {"src_hw": (int(src_hw[0]), int(src_hw[1])), "resized_hw": (h, w),
            "padded_hw": ((h + d - 1) // d * d, (w + d - 1) // d * d)}


def _normalize_planes(image: np.ndarray, mean: Any, std: Any) -> np.ndarray:
    """BGR uint8 ``(H, W, 3)`` -> float32 planes ``(3, H, W)`` of ``(x - mean_c) / std_c`` as OpenCV evaluates it:
    ``convertTo`` with float ``alpha = 1/std``, ``beta = -mean/std`` and one fused multiply-add (exact in float64:
    an 8-bit value times a 24-bit alpha plus beta); for alpha == 1 that is the float addition ``x + beta``."""
    img = np.asarray(image)
    if img.ndim != 3 or img.shape[2] != 3 or img.dtype != np.uint8:
        raise ValueError(f"expected a uint8 (H, W, 3) BGR image, got {img.dtype} {img.shape}")
    m = np.asarray(mean, np.float64).reshape(3)
    sd = np.asarray(std, np.float64).reshape(3)
    alpha = (1.0 / sd).astype(F32)
    beta = (-m / sd).astype(F32)
    planes = img.transpose(2, 0, 1).astype(F32)                          # exact: integers 0..255
    for c in range(3):
        if alpha[c] == 1.0:
            planes[c] += beta[c]                                          # float32, one rounding
        else:
            planes[c] = planes[c].astype(np.float64) * float(alpha[c]) + float(beta[c])
    return planes


def bevformer_normalize(image: np.ndarray, mean: Any = BEVFORMER_MEAN_BGR, std: Any = BEVFORMER_STD) -> np.ndarray:
    """One BGR uint8 ``(H, W, 3)`` image -> ``(x - mean_c) / std_c`` as float32 ``(H, W, 3)``, as OpenCV evaluates it
    (``convertTo`` with float ``alpha = 1/std``, ``beta = -mean/std``, one fused multiply-add; std 1 deployed)."""
    return np.ascontiguousarray(_normalize_planes(image, mean, std).transpose(1, 2, 0))


def bevformer_preprocess(images: Any, *, mean: Any = BEVFORMER_MEAN_BGR, std: Any = BEVFORMER_STD,
                         scale: float = 0.8, pad_divisor: int = 32, out: Optional[np.ndarray] = None,
                         workers: int = 1) -> np.ndarray:
    """N BGR uint8 camera images (a sequence of ``(H, W, 3)`` or an array ``[N, H, W, 3]``, one source size; the node
    feeds 1600x900 after its own ``cv::resize``) -> the network input ``[N, 3, Hp, Wp]`` float32: normalise, resize
    x``scale`` (OpenCV float INTER_LINEAR), zero-pad bottom / right to ``pad_divisor``, CHW (B, G, R planes, as
    ``to_rgb`` is false). Deployed: ``[6, 3, 736, 1280]`` (the ONNX input ``image`` is this with a leading 1).
    ``workers`` > 1 processes the cameras in that many threads (numpy releases the GIL in its loops; about 70 ms per
    1600x900 camera single-threaded on the shared host)."""
    imgs = [np.asarray(im) for im in images]
    if not imgs:
        raise ValueError("no images")
    hw = tuple(imgs[0].shape[:2])
    if any(tuple(im.shape[:2]) != hw for im in imgs):
        raise ValueError("every camera image must have the same size (the node resizes them to 1600x900 first)")
    geo = bevformer_input_geometry(hw, scale, pad_divisor)
    (rh, rw), (ph, pw) = geo["resized_hw"], geo["padded_hw"]
    lut = opencv_linear_lut(hw, (rh, rw))            # strict: the verified x0.8 geometries
    shape = (len(imgs), 3, ph, pw)
    if out is None:
        out = np.zeros(shape, F32)
    else:
        if out.shape != shape or out.dtype != F32:
            raise ValueError(f"out must be float32 {shape}, got {out.dtype} {out.shape}")
        out[:, :, rh:, :] = 0.0
        out[:, :, :, rw:] = 0.0
    def one(i: int) -> None:
        lut.apply_planes(_normalize_planes(imgs[i], mean, std), out=out[i, :, :rh, :rw])

    if int(workers) > 1 and len(imgs) > 1:
        from concurrent.futures import ThreadPoolExecutor

        with ThreadPoolExecutor(max_workers=min(int(workers), len(imgs))) as pool:
            list(pool.map(one, range(len(imgs))))
    else:
        for i in range(len(imgs)):
            one(i)
    return out


# name -> (function, one-line description): the presets this module implements
PRESETS: Dict[str, Tuple[Callable[..., Any], str]] = {
    "yolox_letterbox": (yolox_letterbox, "autoware_tensorrt_yolox preprocess.cu:43-129 letterbox (pad 114, BGR)"),
    "bevdet_nearest_crop": (bevdet_nearest_crop, "bevdet::Preprocess nearest resize 0.44 + crop (140, 0) of the "
                                                 "1600x900 image -> B, G, R uint8 planes"),
    "bevformer_preprocess": (bevformer_preprocess, "autoware_tensorrt_bevformer normalise (BGR mean, std 1) -> "
                                                   "cv::resize x0.8 INTER_LINEAR on float -> zero pad to /32, CHW"),
}