import torch
import numpy as np
from tqdm import tqdm
import utils3d
from PIL import Image
from ..renderers import MeshRenderer, VoxelRenderer, PbrMeshRenderer
from ..representations import Mesh, Voxel, MeshWithPbrMaterial, MeshWithVoxel
from .random_utils import sphere_hammersley_sequence


def yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitchs, rs, fovs):
    is_list = isinstance(yaws, list)
    if not is_list:
        yaws = [yaws]
        pitchs = [pitchs]
    if not isinstance(rs, list):
        rs = [rs] * len(yaws)
    if not isinstance(fovs, list):
        fovs = [fovs] * len(yaws)
    extrinsics = []
    intrinsics = []
    for yaw, pitch, r, fov in zip(yaws, pitchs, rs, fovs):
        fov = torch.deg2rad(torch.tensor(float(fov))).cuda()
        yaw = torch.tensor(float(yaw)).cuda()
        pitch = torch.tensor(float(pitch)).cuda()
        orig = torch.tensor([
            torch.sin(yaw) * torch.cos(pitch),
            torch.cos(yaw) * torch.cos(pitch),
            torch.sin(pitch),
        ]).cuda() * r
        extr = utils3d.torch.extrinsics_look_at(orig, torch.tensor([0, 0, 0]).float().cuda(), torch.tensor([0, 0, 1]).float().cuda())
        intr = utils3d.torch.intrinsics_from_fov_xy(fov, fov)
        extrinsics.append(extr)
        intrinsics.append(intr)
    if not is_list:
        extrinsics = extrinsics[0]
        intrinsics = intrinsics[0]
    return extrinsics, intrinsics


def _safe_ssaa(sample, requested_ssaa, resolution, vram_limit_gb=14.0):
    """
    Cap ssaa so the estimated peak VRAM stays under vram_limit_gb.
    Rough model: raster buffers at (resolution*ssaa)^2, 3 envmaps, 8 peel layers.
    Each peel layer: ~160 MB transient (xyz + img + rast).
    Constant mesh overhead: ~400 MB.
    """
    num_faces = 0
    if isinstance(sample, (MeshWithPbrMaterial, MeshWithVoxel)):
        num_faces = sample.faces.shape[0] if hasattr(sample, 'faces') else 0
    for ssaa in [requested_ssaa, requested_ssaa - 1, 1]:
        if ssaa < 1:
            ssaa = 1
        pixels = (resolution * ssaa) ** 2
        # ~160 MB per peel layer (3 envmaps * shaded + rast + xyz + img)
        peel_layers = 8
        est_mb = (pixels * 4 * 4 * (3 + peel_layers) / 1e6) + 400
        if est_mb < vram_limit_gb * 1024:
            return ssaa
    return 1


def get_renderer(sample, **kwargs):
    if isinstance(sample, (MeshWithPbrMaterial, MeshWithVoxel)):
        renderer = PbrMeshRenderer()
        resolution = kwargs.get('resolution', 512)
        requested_ssaa = kwargs.get('ssaa', 1)
        ssaa = _safe_ssaa(sample, requested_ssaa, resolution)
        if ssaa != requested_ssaa:
            import logging
            logging.getLogger(__name__).warning(
                f"[render_utils] ssaa capped {requested_ssaa}→{ssaa} to stay under VRAM limit"
            )
        renderer.rendering_options.resolution = resolution
        renderer.rendering_options.near = kwargs.get('near', 1)
        renderer.rendering_options.far = kwargs.get('far', 100)
        renderer.rendering_options.ssaa = ssaa
        renderer.rendering_options.peel_layers = kwargs.get('peel_layers', 8)
    elif isinstance(sample, Mesh):
        renderer = MeshRenderer()
        renderer.rendering_options.resolution = kwargs.get('resolution', 512)
        renderer.rendering_options.near = kwargs.get('near', 1)
        renderer.rendering_options.far = kwargs.get('far', 100)
        renderer.rendering_options.ssaa = kwargs.get('ssaa', 1)
        renderer.rendering_options.chunk_size = kwargs.get('chunk_size', None)
    elif isinstance(sample, Voxel):
        renderer = VoxelRenderer()
        renderer.rendering_options.resolution = kwargs.get('resolution', 512)
        renderer.rendering_options.near = kwargs.get('near', 0.1)
        renderer.rendering_options.far = kwargs.get('far', 10.0)
        renderer.rendering_options.ssaa = kwargs.get('ssaa', 1)
    else:
        raise ValueError(f'Unsupported sample type: {type(sample)}')
    return renderer


def render_frames(sample, extrinsics, intrinsics, options={}, verbose=True, **kwargs):
    renderer = get_renderer(sample, **options)
    # Free stale GPU allocations from the generation phase before rendering starts.
    # On ROCm, driver-level OOM causes a display freeze rather than a Python exception,
    # so we clear proactively rather than waiting for the allocator to evict.
    torch.cuda.empty_cache()
    rets = {}
    for j, (extr, intr) in tqdm(enumerate(zip(extrinsics, intrinsics)), total=len(extrinsics), desc='Rendering', disable=not verbose):
        res = renderer.render(sample, extr, intr, **kwargs)
        for k, v in res.items():
            if k not in rets: rets[k] = []
            if v.dim() == 2: v = v[None].repeat(3, 1, 1)
            rets[k].append(np.clip(v.detach().cpu().numpy().transpose(1, 2, 0) * 255, 0, 255).astype(np.uint8))
    return rets


def render_video(sample, resolution=1024, bg_color=(0, 0, 0), num_frames=120, r=2, fov=40, **kwargs):
    yaws = -torch.linspace(0, 2 * 3.1415, num_frames) + np.pi/2
    pitch = 0.25 + 0.5 * torch.sin(torch.linspace(0, 2 * 3.1415, num_frames))
    yaws = yaws.tolist()
    pitch = pitch.tolist()
    extrinsics, intrinsics = yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitch, r, fov)
    return render_frames(sample, extrinsics, intrinsics, {'resolution': resolution, 'bg_color': bg_color}, **kwargs)


def render_multiview(sample, resolution=512, nviews=30):
    r = 2
    fov = 40
    cams = [sphere_hammersley_sequence(i, nviews) for i in range(nviews)]
    yaws = [cam[0] for cam in cams]
    pitchs = [cam[1] for cam in cams]
    extrinsics, intrinsics = yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitchs, r, fov)
    res = render_frames(sample, extrinsics, intrinsics, {'resolution': resolution, 'bg_color': (0, 0, 0)})
    return res['color'], extrinsics, intrinsics


def render_snapshot(samples, resolution=512, bg_color=(0, 0, 0), offset=(-16 / 180 * np.pi, 20 / 180 * np.pi), r=10, fov=8, nviews=4, **kwargs):
    yaw = np.linspace(0, 2 * np.pi, nviews, endpoint=False)
    yaw_offset = offset[0]
    yaw = [y + yaw_offset for y in yaw]
    pitch = [offset[1] for _ in range(nviews)]
    extrinsics, intrinsics = yaw_pitch_r_fov_to_extrinsics_intrinsics(yaw, pitch, r, fov)
    return render_frames(samples, extrinsics, intrinsics, {'resolution': resolution, 'bg_color': bg_color}, **kwargs)


def make_pbr_vis_frames(result, resolution=1024):
    num_frames = len(result['shaded'])
    frames = []
    for i in range(num_frames):
        shaded = Image.fromarray(result['shaded'][i])
        normal = Image.fromarray(result['normal'][i])
        base_color = Image.fromarray(result['base_color'][i])
        metallic = Image.fromarray(result['metallic'][i])
        roughness = Image.fromarray(result['roughness'][i])
        alpha = Image.fromarray(result['alpha'][i])
        shaded = shaded.resize((resolution, resolution))
        normal = normal.resize((resolution, resolution))
        base_color = base_color.resize((resolution//2, resolution//2))
        metallic = metallic.resize((resolution//2, resolution//2))
        roughness = roughness.resize((resolution//2, resolution//2))
        alpha = alpha.resize((resolution//2, resolution//2))
        row1 = np.concatenate([shaded, normal], axis=1)
        row2 = np.concatenate([base_color, metallic, roughness, alpha], axis=1)
        frame = np.concatenate([row1, row2], axis=0)
        frames.append(frame)
    return frames