#!/usr/bin/env python3
"""Append offset-0 RGB frames to an existing strict-causal RGB cache.

This is faster than rebuilding the full cache with [-8,-4,-2,-1,0],
because the existing strict cache already stores the negative-offset history.
The output is intended only for current-observation ablations.
"""

from __future__ import annotations

import argparse
import copy
import json
import os
import sys
from collections import Counter, defaultdict
from typing import Any, Dict, List, Tuple

import cv2
import torch

sys.path.insert(0, os.path.dirname(__file__))
from build_rgb_frame_cache import parse_cell_frames, video_t_to_cell  # noqa: E402


def video_path(processed_root: str, boss: str, fight: int) -> str:
    return os.path.join(processed_root, boss, f"video_fight{int(fight)}.mp4")


def read_decord_group(
    path: str,
    frame_indices: List[List[int]],
    height: int,
    width: int,
    chunk_size: int,
) -> torch.Tensor:
    from decord import VideoReader, cpu

    vr = VideoReader(path, ctx=cpu(0), width=width, height=height, num_threads=2)
    n_frames = len(vr)
    flat = [min(max(0, int(x)), n_frames - 1) for row in frame_indices for x in row]
    out = torch.empty((len(flat), 3, height, width), dtype=torch.uint8)
    for start in range(0, len(flat), chunk_size):
        idx = flat[start:start + chunk_size]
        batch = vr.get_batch(idx).asnumpy()
        out[start:start + len(idx)] = torch.from_numpy(batch).permute(0, 3, 1, 2).contiguous()
    return out.reshape(len(frame_indices), len(frame_indices[0]), 3, height, width)


def read_opencv_group(
    path: str,
    frame_indices: List[List[int]],
    height: int,
    width: int,
) -> torch.Tensor:
    cv2.setNumThreads(1)
    cap = cv2.VideoCapture(path)
    if not cap.isOpened():
        raise RuntimeError(f"cannot open video: {path}")
    n_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    rows = []
    for row in frame_indices:
        frames = []
        for frame_idx in row:
            idx = min(max(0, int(frame_idx)), max(0, n_frames - 1))
            cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
            ok, bgr = cap.read()
            if not ok or bgr is None:
                cap.release()
                raise RuntimeError(f"cannot read frame {idx} from {path}")
            rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
            if rgb.shape[0] != height or rgb.shape[1] != width:
                rgb = cv2.resize(rgb, (width, height), interpolation=cv2.INTER_AREA)
            frames.append(torch.from_numpy(rgb).permute(2, 0, 1).contiguous())
        rows.append(torch.stack(frames, dim=0))
    cap.release()
    return torch.stack(rows, dim=0).to(torch.uint8)


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--base_cache", required=True)
    ap.add_argument("--processed_root", default="data/processed")
    ap.add_argument("--frame_in_cell", default="2")
    ap.add_argument("--backend", choices=["decord", "opencv"], default="decord")
    ap.add_argument("--chunk_size", type=int, default=512)
    ap.add_argument("--out", required=True)
    args = ap.parse_args()

    data = torch.load(args.base_cache, map_location="cpu", weights_only=False)
    base_rgb = data["rgb"].contiguous()
    if base_rgb.dtype != torch.uint8:
        raise TypeError(f"expected uint8 rgb cache, got {base_rgb.dtype}")
    if any(int(x) >= 0 for x in data.get("history_offsets", [])):
        raise ValueError("base_cache must be strict-causal; use build_rgb_frame_cache for custom current caches")
    samples: List[Dict[str, Any]] = data["samples"]
    frame_in_cell = parse_cell_frames(args.frame_in_cell)
    n, t, c, h, w = base_rgb.shape
    if c != 3:
        raise ValueError(f"expected RGB channel dimension 3, got {c}")

    groups: Dict[Tuple[str, int], List[Tuple[int, List[int]]]] = defaultdict(list)
    all_first_indices = []
    for i, sample in enumerate(samples):
        boss = sample["boss"]
        fight = int(sample["fight"])
        target_cell = video_t_to_cell(sample["belief"]["time"])
        frame_indices = [4 * target_cell + int(in_cell) for in_cell in frame_in_cell]
        all_first_indices.append(frame_indices[0])
        groups[(boss, fight)].append((i, frame_indices))

    current = torch.empty((n, len(frame_in_cell), 3, h, w), dtype=torch.uint8)
    missing = 0
    missing_by_reason = Counter()
    for group_id, ((boss, fight), entries) in enumerate(sorted(groups.items()), start=1):
        path = video_path(args.processed_root, boss, fight)
        if not os.path.exists(path):
            missing += len(entries)
            missing_by_reason["missing_video"] += len(entries)
            continue
        order = [i for i, _ in entries]
        frame_indices = [idxs for _, idxs in entries]
        try:
            if args.backend == "decord":
                frames = read_decord_group(path, frame_indices, h, w, args.chunk_size)
            else:
                frames = read_opencv_group(path, frame_indices, h, w)
        except Exception as exc:
            missing += len(entries)
            missing_by_reason[f"read_error:{type(exc).__name__}"] += len(entries)
            print(f"warning: failed {boss}/fight{fight}: {type(exc).__name__}: {exc}", flush=True)
            continue
        current[torch.tensor(order, dtype=torch.long)] = frames
        print(
            f"processed_groups {group_id}/{len(groups)} "
            f"group={boss}/fight{fight} samples={len(entries)} filled={n - missing} missing={missing}",
            flush=True,
        )

    if missing:
        raise RuntimeError(f"missing current frames: {missing} {dict(missing_by_reason)}")

    out_rgb = torch.empty((n, t + len(frame_in_cell), 3, h, w), dtype=torch.uint8)
    out_rgb[:, :t] = base_rgb
    out_rgb[:, t:] = current

    payload = dict(data)
    payload["rgb"] = out_rgb.contiguous()
    payload["history_offsets"] = list(data.get("history_offsets", [])) + [0]
    payload["frame_in_cell"] = list(data.get("frame_in_cell", frame_in_cell))
    payload["allow_current_frame"] = True
    payload["current_frame_source"] = "appended_from_raw_processed_fight_video"
    payload["base_cache"] = args.base_cache
    payload["video_backend"] = args.backend
    payload["audit"] = copy.deepcopy(data.get("audit", {}))
    payload["audit"]["append_current_frame"] = {
        "base_cache": args.base_cache,
        "requested": int(n),
        "kept": int(n),
        "missing": int(missing),
        "missing_by_reason": dict(missing_by_reason),
        "groups": int(len(groups)),
        "frame_in_cell": list(frame_in_cell),
        "current_frame_min": int(min(all_first_indices)) if all_first_indices else None,
        "current_frame_max": int(max(all_first_indices)) if all_first_indices else None,
    }
    os.makedirs(os.path.dirname(args.out), exist_ok=True)
    torch.save(payload, args.out)
    print(json.dumps({
        "out": args.out,
        "base_cache": args.base_cache,
        "shape": list(out_rgb.shape),
        "history_offsets": payload["history_offsets"],
        "allow_current_frame": True,
        "audit": payload["audit"]["append_current_frame"],
    }, ensure_ascii=False, indent=2), flush=True)


if __name__ == "__main__":
    main()