import torch
import torch.nn as nn
import math


class Model(nn.Module):
    """
    Outer-product mean: converts MSA representation to a pair
    representation update.  Used in AlphaFold2 to communicate information
    from the MSA stack to the pair stack.  For each pair (i,j), the
    outer product of projected MSA features at positions i and j is
    averaged over the MSA depth dimension.
    """

    def __init__(self, msa_dim, pair_dim, hidden_dim=32):
        super().__init__()
        self.layer_norm = nn.LayerNorm(msa_dim)
        self.proj_a = nn.Linear(msa_dim, hidden_dim)
        self.proj_b = nn.Linear(msa_dim, hidden_dim)
        self.out_proj = nn.Linear(hidden_dim * hidden_dim, pair_dim)

    def forward(self, msa_repr: torch.Tensor, msa_mask: torch.Tensor) -> torch.Tensor:
        """
        Args:
            msa_repr: (B, S, N, msa_dim) – MSA representation
            msa_mask: (B, S, N)          – 1 for valid rows/positions
        Returns:
            pair update (B, N, N, pair_dim)
        """
        x = self.layer_norm(msa_repr)
        a = self.proj_a(x)  # (B, S, N, H)
        b = self.proj_b(x)

        mask = msa_mask.unsqueeze(-1)  # (B, S, N, 1)
        a = a * mask
        b = b * mask

        outer = torch.einsum("bsih,bsjk->bijhk", a, b)  # (B, N, N, H, H)
        denom = msa_mask.sum(dim=1).unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).clamp(min=1)
        outer = outer / denom  # mean over S

        B, N1, N2, H1, H2 = outer.shape
        outer = outer.reshape(B, N1, N2, H1 * H2)
        return self.out_proj(outer)


msa_dim = 64
pair_dim = 64
hidden_dim = 32
seq_len = 32
msa_depth = 8
batch_size = 2


def get_inputs():
    p = popcorn_pri
    msa_repr = torch.randn(p.jitter_int(batch_size), p.jitter_int(msa_depth), p.jitter_int(seq_len), msa_dim)
    msa_mask = torch.ones(p.jitter_int(batch_size), p.jitter_int(msa_depth), p.jitter_int(seq_len))
    return [msa_repr, msa_mask]


def get_init_inputs():
    return [msa_dim, pair_dim, hidden_dim]

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

cuda_source = '''
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

__global__ void outer_mean_kernel(
    const float* __restrict__ a,
    const float* __restrict__ b,
    const float* __restrict__ denom,
    float* __restrict__ out,
    int B, int S, int N, int H
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int total = B * N * N * H * H;
    if (idx >= total) return;

    int h2 = idx % H;
    int h1 = (idx / H) % H;
    int j = (idx / (H*H)) % N;
    int i = (idx / (H*H*N)) % N;
    int batch = idx / (H*H*N*N);

    float sum = 0.0f;
    float d = denom[batch * N + i];
    if (d < 1.0f) d = 1.0f;

    for (int s = 0; s < S; s++) {
        float aval = a[((batch * S + s) * N + i) * H + h1];
        float bval = b[((batch * S + s) * N + j) * H + h2];
        sum += aval * bval;
    }
    out[idx] = sum / d;
}

torch::Tensor outer_mean_cuda(torch::Tensor a, torch::Tensor b, torch::Tensor msa_mask) {
    auto B = a.size(0);
    auto S = a.size(1);
    auto N = a.size(2);
    auto H = a.size(3);

    auto denom = msa_mask.sum(1); // (B, N)
    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(a.device());
    auto out = torch::zeros({B, N, N, H*H}, options);

    int total = B * N * N * H * H;
    int threads = 256;
    int blocks = (total + threads - 1) / threads;

    outer_mean_kernel<<<blocks, threads>>>(
        a.data_ptr<float>(),
        b.data_ptr<float>(),
        denom.data_ptr<float>(),
        out.data_ptr<float>(),
        B, S, N, H
    );

    return out;
}
'''

cpp_source = '''
torch::Tensor outer_mean_cuda(torch::Tensor a, torch::Tensor b, torch::Tensor msa_mask);
'''

module = load_inline(
    name='outer_mean_ext',
    cpp_sources=[cpp_source],
    cuda_sources=[cuda_source],
    functions=['outer_mean_cuda'],
    verbose=False
)

class ModelNew(nn.Module):
    def __init__(self, msa_dim, pair_dim, hidden_dim=32):
        super().__init__()
        self.layer_norm = nn.LayerNorm(msa_dim)
        self.proj_a = nn.Linear(msa_dim, hidden_dim)
        self.proj_b = nn.Linear(msa_dim, hidden_dim)
        self.out_proj = nn.Linear(hidden_dim * hidden_dim, pair_dim)
        self.ext = module

    def forward(self, msa_repr: torch.Tensor, msa_mask: torch.Tensor) -> torch.Tensor:
        x = self.layer_norm(msa_repr)
        a = self.proj_a(x)
        b = self.proj_b(x)

        mask = msa_mask.unsqueeze(-1)
        a = a * mask
        b = b * mask

        outer = self.ext.outer_mean_cuda(a, b, msa_mask)
        return self.out_proj(outer)