"""Shared CPC encoder helpers: global palette selection + per-mode bit packing.

The CPC has no per-cell colour limit -- a mode draws a flat palette of pens
(16 in mode 0, 4 in mode 1, 2 in mode 2), each pen any of the 27 colours.  So the
job is: choose the best N-colour sub-palette for the whole image, dither to it,
then pack pens into the CPC's (scrambled) screen bytes.
"""
from __future__ import annotations

import numpy as np

from ... import dither, palette as c64pal
from .. import palette as cpcpal
from ..snapshot import screen_offset


def choose_inks(img_lab, plab, n):
    """Greedy forward selection of the ``n`` palette colours (indices into the
    27-colour PALETTE) that minimise nearest-colour error over the image."""
    flat = img_lab.reshape(-1, 3)
    d = np.sum((flat[:, None, :] - plab[None, :, :]) ** 2, axis=-1)   # (px, 27)
    chosen = []
    best = np.full(flat.shape[0], np.inf)
    for _ in range(n):
        # pick the colour that most reduces the running per-pixel min distance
        cand_err = np.minimum(best[:, None], d).sum(0)                # (27,)
        for c in chosen:
            cand_err[c] = np.inf
        c = int(cand_err.argmin())
        chosen.append(c)
        best = np.minimum(best, d[:, c])
    return sorted(chosen)


def render(img_rgb, n, dither_mode, ramp=None):
    """Pick (or use ``ramp``) an n-colour palette, dither the image to it, and
    return (pen_image (H,W) of 0..n-1, inks (n hardware-ink numbers), idx_image
    (H,W) palette indices, plab, prgb)."""
    plab = cpcpal.palette_lab()
    prgb = cpcpal.get_palette().astype(np.uint8)
    img_lab = c64pal.srgb_to_lab(img_rgb)

    pal_idx = ramp if ramp is not None else choose_inks(img_lab, plab, n)
    pal_idx = list(pal_idx)
    allowed = np.tile(np.array(pal_idx), (*img_lab.shape[:2], 1))
    idx = dither.quantize(img_lab, allowed, plab, dither_mode).astype(np.int64)

    # map palette index -> pen (0..n-1)
    lut = {p: k for k, p in enumerate(pal_idx)}
    pen = np.vectorize(lut.get)(idx).astype(np.uint8)
    inks = [cpcpal.ink_byte(p) for p in pal_idx]
    return pen, inks, idx.astype(np.uint16), img_lab, plab, prgb


def _pack_mode0(pen):
    """160x200 pens (0-15) -> 16K screen.  2 pixels/byte, scrambled bits."""
    scr = bytearray(0x4000)
    for y in range(200):
        row = pen[y]
        for bx in range(80):
            L = int(row[bx * 2]); R = int(row[bx * 2 + 1])
            byte = (((L & 1) << 7) | ((R & 1) << 6) | (((L >> 2) & 1) << 5) |
                    (((R >> 2) & 1) << 4) | (((L >> 1) & 1) << 3) |
                    (((R >> 1) & 1) << 2) | (((L >> 3) & 1) << 1) | ((R >> 3) & 1))
            scr[screen_offset(y, bx)] = byte
    return bytes(scr)


def _pack_mode1(pen):
    """320x200 pens (0-3) -> 16K screen.  4 pixels/byte."""
    scr = bytearray(0x4000)
    for y in range(200):
        row = pen[y]
        for bx in range(80):
            p = [int(row[bx * 4 + k]) for k in range(4)]
            byte = 0
            for k in range(4):
                byte |= (p[k] & 1) << (7 - k)            # bit0 of each pen
                byte |= ((p[k] >> 1) & 1) << (3 - k)     # bit1 of each pen
            scr[screen_offset(y, bx)] = byte
    return bytes(scr)


def _pack_mode2(pen):
    """640x200 pens (0-1) -> 16K screen.  8 pixels/byte, MSB leftmost."""
    scr = bytearray(0x4000)
    for y in range(200):
        row = pen[y]
        for bx in range(80):
            byte = 0
            for k in range(8):
                byte |= (int(row[bx * 8 + k]) & 1) << (7 - k)
            scr[screen_offset(y, bx)] = byte
    return bytes(scr)


PACK = {0: _pack_mode0, 1: _pack_mode1, 2: _pack_mode2}