from __future__ import annotations

import argparse
import csv
import os
import random
import sys
from pathlib import Path

import h5py
import numpy as np
import torch
import torch.distributed as dist
from tqdm import tqdm

# This generator is bundled directly under <bundle>/src on the server.
PROJECT_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(PROJECT_ROOT / "src"))

from diffusion import create_diffusion
from models_controlnet_two_condition import TwoConditionControlNetSWaG


def distributed_context() -> tuple[int, int, int]:
    world_size = int(os.environ.get("WORLD_SIZE", "1"))
    rank = int(os.environ.get("RANK", "0"))
    local_rank = int(os.environ.get("LOCAL_RANK", "0"))
    if world_size > 1:
        dist.init_process_group("nccl")
    return rank, local_rank, world_size


def shard_path(output_path: Path, rank: int) -> Path:
    return output_path.with_name(f".{output_path.name}.rank{rank}.partial.h5")


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--checkpoint", type=Path, required=True)
    parser.add_argument("--source-h5", type=Path, required=True)
    parser.add_argument("--output-h5", type=Path, required=True)
    parser.add_argument("--output-meta", type=Path, required=True)
    parser.add_argument("--num-samples", type=int, default=100)
    parser.add_argument("--batch-size", type=int, default=256)
    parser.add_argument("--sampling-steps", type=int, default=25)
    parser.add_argument("--sampler", choices=("ddpm", "ddim"), default="ddpm")
    parser.add_argument("--cfg-scale", type=float, default=2.0)
    parser.add_argument("--seed", type=int, default=2026)
    parser.add_argument(
        "--source-label-format",
        choices=("raw", "two_condition"),
        default="raw",
        help="Use raw SWaG labels or labels already transformed for two-condition training.",
    )
    args = parser.parse_args()
    if not torch.cuda.is_available():
        raise RuntimeError("CUDA is required for sampling")

    rank, local_rank, world_size = distributed_context()
    device = torch.device("cuda", local_rank)
    torch.cuda.set_device(device)
    if args.batch_size % world_size:
        raise ValueError(f"batch-size must be divisible by world size ({world_size})")
    local_batch = args.batch_size // world_size
    random.seed(args.seed + rank)
    np.random.seed(args.seed + rank)
    torch.manual_seed(args.seed + rank)
    torch.cuda.manual_seed_all(args.seed + rank)
    rng = np.random.default_rng(args.seed)

    with h5py.File(args.source_h5, "r") as handle:
        labels_all = np.asarray(handle["labels"], dtype=np.float32)
    if args.num_samples > len(labels_all):
        raise ValueError("num-samples exceeds source dataset size")
    indices = np.sort(rng.choice(len(labels_all), size=args.num_samples, replace=False))
    if args.source_label_format == "raw":
        if labels_all.shape[1] < 3:
            raise ValueError(f"Expected raw source labels with at least 3 columns, got {labels_all.shape}")
        raw = labels_all[indices, :3]
        labels = raw[:, 1:3].astype(np.float32, copy=True)
    else:
        if labels_all.shape[1] != 2:
            raise ValueError(f"Expected two-condition source labels with 2 columns, got {labels_all.shape}")
        labels = labels_all[indices, :2].astype(np.float32, copy=True)

    checkpoint = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
    config = checkpoint["config"]
    model = TwoConditionControlNetSWaG(**config["model"]).to(device).eval()
    model.load_state_dict(checkpoint["model"], strict=True)
    del checkpoint
    respacing = f"ddim{args.sampling_steps}" if args.sampler == "ddim" else str(args.sampling_steps)
    diffusion = create_diffusion(timestep_respacing=respacing, **config["diffusion"])
    generator = torch.Generator(device=device).manual_seed(args.seed + rank)

    if rank == 0:
        args.output_h5.parent.mkdir(parents=True, exist_ok=True)
        args.output_meta.parent.mkdir(parents=True, exist_ok=True)
        for shard_rank in range(world_size):
            shard_path(args.output_h5, shard_rank).unlink(missing_ok=True)
    if world_size > 1:
        dist.barrier()
    shape = (args.num_samples, int(config["model"]["length"]), int(config["model"]["in_channels"]))
    assigned = np.array_split(np.arange(args.num_samples, dtype=np.int64), world_size)[rank]
    local_shape = (len(assigned), shape[1], shape[2])
    with h5py.File(shard_path(args.output_h5, rank), "w") as handle:
        output = handle.create_dataset("data", shape=local_shape, dtype="float32", compression="gzip", shuffle=True)
        handle.create_dataset("indices", data=assigned, dtype="int64")
        with torch.no_grad():
            for start in tqdm(
                range(0, len(assigned), local_batch),
                desc=f"sampling rank {rank}",
                disable=rank != 0,
            ):
                end = min(start + local_batch, len(assigned))
                global_positions = assigned[start:end]
                batch_labels = torch.from_numpy(labels[global_positions]).to(device)
                noise = torch.randn(
                    end - start,
                    shape[2],
                    shape[1],
                    device=device,
                    generator=generator,
                )

                def model_fn(x, t, **kwargs):
                    return model.forward_with_cfg(x, t, batch_labels, args.cfg_scale)

                sample_loop = diffusion.ddim_sample_loop if args.sampler == "ddim" else diffusion.p_sample_loop
                with torch.autocast("cuda", dtype=torch.float16):
                    samples = sample_loop(
                        model_fn,
                        noise.shape,
                        noise=noise,
                        device=device,
                        progress=False,
                        clip_denoised=False,
                    )
                output[start:end] = samples.float().transpose(1, 2).cpu().numpy().astype(np.float32)

    if world_size > 1:
        dist.barrier()
    if rank == 0:
        partial = args.output_h5.with_name(args.output_h5.name + ".partial")
        partial.unlink(missing_ok=True)
        with h5py.File(partial, "w") as handle:
            output = handle.create_dataset("data", shape=shape, dtype="float32", compression="gzip", shuffle=True)
            handle.create_dataset("labels", data=labels, dtype="float32")
            handle.attrs.update(
                checkpoint=str(args.checkpoint.resolve()),
                sampling_steps=args.sampling_steps,
                sampler=args.sampler,
                cfg_scale=args.cfg_scale,
                seed=args.seed,
                world_size=world_size,
            )
            for shard_rank in range(world_size):
                current = shard_path(args.output_h5, shard_rank)
                with h5py.File(current, "r") as shard:
                    positions = np.asarray(shard["indices"], dtype=np.int64)
                    output[positions] = shard["data"][:]
                current.unlink()
        partial.replace(args.output_h5)

        with args.output_meta.open("w", newline="", encoding="utf-8") as handle:
            writer = csv.writer(handle)
            writer.writerow(["generated_index", "source_test_index", "p_index", "s_index"])
            writer.writerows(
                (
                    i,
                    int(indices[i]),
                    float(labels[i, 0]),
                    float(labels[i, 1]),
                )
                for i in range(len(labels))
            )
    if world_size > 1:
        dist.barrier()
        dist.destroy_process_group()


if __name__ == "__main__":
    main()