"""
PRISM Custom DiT (Diffusion Transformer) Architecture Definition.
Implements a modular Diffusion Transformer with AdaLN-Zero conditioning,
2D spatial patchification, and cross-attention text guidance.
"""

import math
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F


def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
    """Modulates input tensor using affine scale and shift parameters."""
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)


class TimestepEmbedder(nn.Module):
    """
    Embeds scalar timesteps into vector representations via sinusoidal embeddings
    followed by a two-layer MLP.
    """

    def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
        self.frequency_embedding_size = frequency_embedding_size

    @staticmethod
    def sinusoidal_embedding(t: torch.Tensor, dim: int, max_period: int = 10000) -> torch.Tensor:
        half = dim // 2
        freqs = torch.exp(
            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half
        )
        args = t[:, None].float() * freqs[None]
        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
        if dim % 2 == 1:
            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
        return embedding

    def forward(self, t: torch.Tensor) -> torch.Tensor:
        t_freq = self.sinusoidal_embedding(t, self.frequency_embedding_size)
        return self.mlp(t_freq)


class PatchEmbed(nn.Module):
    """2D Image/Latent to Patch Embedding."""

    def __init__(self, patch_size: int = 2, in_channels: int = 16, embed_dim: int = 1152):
        super().__init__()
        self.patch_size = patch_size
        self.proj = nn.Conv2d(
            in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # x: (B, C, H, W) -> proj -> (B, D, H/p, W/p) -> flatten -> (B, N, D)
        x = self.proj(x)
        x = x.flatten(2).transpose(1, 2)
        return x


class PrismDiTBlock(nn.Module):
    """
    DiT Block with AdaLN-Zero conditioning and Cross-Attention for text conditioning.
    """

    def __init__(self, hidden_size: int, num_heads: int, context_dim: int = 1024, mlp_ratio: float = 4.0):
        super().__init__()
        self.hidden_size = hidden_size
        self.num_heads = num_heads
        self.head_dim = hidden_size // num_heads

        # Normalization layers
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm_cross = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)

        # Self-Attention
        self.self_attn = nn.MultiheadAttention(hidden_size, num_heads, batch_first=True)

        # Cross-Attention for text embeddings
        self.cross_attn = nn.MultiheadAttention(
            embed_dim=hidden_size,
            kdim=context_dim,
            vdim=context_dim,
            num_heads=num_heads,
            batch_first=True,
        )

        # Feed-Forward Network
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        self.mlp = nn.Sequential(
            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
            nn.GELU(approximate="tanh"),
            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
        )

        # AdaLN modulation table: 6 modulation scales (shift, scale, gate for self-attn and mlp)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True),
        )

    def forward(
        self,
        x: torch.Tensor,
        c: torch.Tensor,
        context: Optional[torch.Tensor] = None,
        context_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        # Modulation vectors from timestep / conditioning vector c
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
            self.adaLN_modulation(c).chunk(6, dim=1)
        )

        # 1. Self Attention with AdaLN modulation
        normed_x = modulate(self.norm1(x), shift_msa, scale_msa)
        attn_out, _ = self.self_attn(normed_x, normed_x, normed_x)
        x = x + gate_msa.unsqueeze(1) * attn_out

        # 2. Cross Attention with text context (if provided)
        if context is not None:
            normed_cross = self.norm_cross(x)
            cross_out, _ = self.cross_attn(
                query=normed_cross,
                key=context,
                value=context,
                key_padding_mask=context_mask,
            )
            x = x + cross_out

        # 3. Feed Forward with AdaLN modulation
        normed_mlp = modulate(self.norm2(x), shift_mlp, scale_mlp)
        mlp_out = self.mlp(normed_mlp)
        x = x + gate_mlp.unsqueeze(1) * mlp_out

        return x


class FinalLayer(nn.Module):
    """Final AdaLN layer and linear unpatchify projection."""

    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, bias=True),
        )

    def forward(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
        x = modulate(self.norm_final(x), shift, scale)
        return self.linear(x)


class PrismDiT(nn.Module):
    """
    PRISM Diffusion Transformer Backbone.
    
    Supports:
    - Flow-matching or v-prediction / epsilon diffusion
    - Flexible latent channels (default: 16 channels for high-fidelity VAE)
    - Full multimodal context cross-attention
    """

    def __init__(
        self,
        input_size: int = 64,
        patch_size: int = 2,
        in_channels: int = 16,
        hidden_size: int = 1152,
        depth: int = 28,
        num_heads: int = 16,
        mlp_ratio: float = 4.0,
        context_dim: int = 1024,
    ):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = in_channels
        self.patch_size = patch_size
        self.num_heads = num_heads
        self.hidden_size = hidden_size

        self.x_embedder = PatchEmbed(patch_size, in_channels, hidden_size)
        self.t_embedder = TimestepEmbedder(hidden_size)

        # 2D Positional embeddings (learnable or fixed sinusoidal)
        num_patches = (input_size // patch_size) ** 2
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size))

        # Transformer blocks
        self.blocks = nn.ModuleList([
            PrismDiTBlock(hidden_size, num_heads, context_dim=context_dim, mlp_ratio=mlp_ratio)
            for _ in range(depth)
        ])

        # Final projection layer
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)
        self.initialize_weights()

    def initialize_weights(self):
        # Initialize patch embedding like linear projection
        w = self.x_embedder.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.x_embedder.proj.bias, 0)

        # Initialize pos_embed
        nn.init.normal_(self.pos_embed, std=0.02)

        # Zero-out AdaLN modulation layers for identity start
        for block in self.blocks:
            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)

        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)

    def unpatchify(self, x: torch.Tensor, h: int, w: int) -> torch.Tensor:
        """
        x: (B, N, patch_size**2 * C)
        Returns: (B, C, H, W)
        """
        c = self.out_channels
        p = self.patch_size
        ph, pw = h // p, w // p
        assert ph * pw == x.shape[1]

        x = x.reshape(shape=(x.shape[0], ph, pw, p, p, c))
        x = torch.einsum("nhwpqc->nchpwq", x)
        imgs = x.reshape(shape=(x.shape[0], c, ph * p, pw * p))
        return imgs

    def forward(
        self,
        x: torch.Tensor,
        t: torch.Tensor,
        context: Optional[torch.Tensor] = None,
        context_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """
        Forward pass for PRISM DiT.
        x: (B, C, H, W) Latent tensor
        t: (B,) Timestep tensor
        context: (B, L, context_dim) Text/multimodal conditioning sequence
        """
        b, c, h, w = x.shape
        x = self.x_embedder(x)

        # Add positional embedding (interpolating if resolution differs)
        if x.shape[1] == self.pos_embed.shape[1]:
            x = x + self.pos_embed
        else:
            # Dynamic positional interpolation for arbitrary aspect ratios
            pos_embed = F.interpolate(
                self.pos_embed.transpose(1, 2).reshape(1, self.hidden_size, int(math.sqrt(self.pos_embed.shape[1])), -1),
                size=(h // self.patch_size, w // self.patch_size),
                mode="bicubic",
                align_corners=False,
            ).flatten(2).transpose(1, 2)
            x = x + pos_embed

        c_emb = self.t_embedder(t)

        for block in self.blocks:
            x = block(x, c_emb, context=context, context_mask=context_mask)

        x = self.final_layer(x, c_emb)
        x = self.unpatchify(x, h, w)
        return x