import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Voxel-grid pooling for 3-D molecular point clouds.  Discretizes
    continuous 3-D coordinates into a regular grid and pools per-atom
    features into each voxel.  Used in 3-D-CNN approaches for molecular
    property prediction and ligand docking (e.g., PointVoxel-CNN,
    3D-Infomax).
    """

    def __init__(self, feat_dim, grid_size=32, voxel_size=1.0):
        super().__init__()
        self.feat_dim = feat_dim
        self.grid_size = grid_size
        self.voxel_size = voxel_size
        self.linear = nn.Linear(feat_dim, feat_dim)

    def forward(
        self, coords: torch.Tensor, features: torch.Tensor
    ) -> torch.Tensor:
        """
        Args:
            coords:   (B, N, 3) – atom coordinates
            features: (B, N, feat_dim) – per-atom features
        Returns:
            voxel_grid: (B, feat_dim, G, G, G) – pooled voxel features
        """
        B, N, _ = coords.shape
        G = self.grid_size

        features = self.linear(features)

        # Center coordinates and discretize
        center = coords.mean(dim=1, keepdim=True)
        shifted = (coords - center) / self.voxel_size + G / 2.0
        indices = shifted.long().clamp(0, G - 1)  # (B, N, 3)

        voxel_grid = coords.new_zeros(B, self.feat_dim, G, G, G)
        ix = indices[:, :, 0]
        iy = indices[:, :, 1]
        iz = indices[:, :, 2]

        batch_idx = torch.arange(B, device=coords.device).unsqueeze(1).expand(-1, N)
        for d in range(self.feat_dim):
            feat_d = features[:, :, d]
            voxel_grid[:, d].index_put_(
                (batch_idx.reshape(-1), ix.reshape(-1), iy.reshape(-1), iz.reshape(-1)),
                feat_d.reshape(-1),
                accumulate=True,
            )

        return voxel_grid


feat_dim = 32
grid_size = 32
voxel_size = 1.0
num_atoms = 256
batch_size = 4


def get_inputs():
    p = popcorn_pri
    bs = p.jitter_int(batch_size)
    na = p.jitter_int(num_atoms)
    coords = torch.randn(bs, na, 3) * 5.0
    features = torch.randn(bs, na, feat_dim)
    return [coords, features]


def get_init_inputs():
    return [feat_dim, grid_size, voxel_size]

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

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

__global__ void voxel_pool_kernel(
    const float* __restrict__ features,   // (BN, D)
    const long* __restrict__ indices,     // (BN, 3)
    const long* __restrict__ batch_idx,   // (BN)
    float* __restrict__ voxel_grid,       // (B, D, G, G, G)
    int BN, int D, int G
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= BN) return;

    long b = batch_idx[idx];
    long ix = indices[idx * 3];
    long iy = indices[idx * 3 + 1];
    long iz = indices[idx * 3 + 2];

    if (ix < 0 || ix >= G || iy < 0 || iy >= G || iz < 0 || iz >= G) return;

    for (int d = 0; d < D; ++d) {
        float val = features[idx * D + d];
        int offset = ((b * D + d) * G + ix) * G * G + iy * G + iz;
        atomicAdd(&voxel_grid[offset], val);
    }
}

torch::Tensor voxel_pool_cuda(torch::Tensor features, torch::Tensor indices, torch::Tensor batch_idx, int G, int B) {
    int BN = features.size(0);
    int D = features.size(1);
    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(features.device());
    auto voxel_grid = torch::zeros({B, D, G, G, G}, options);

    int threads = 256;
    int blocks = (BN + threads - 1) / threads;

    voxel_pool_kernel<<<blocks, threads>>>(
        features.data_ptr<float>(),
        indices.data_ptr<long>(),
        batch_idx.data_ptr<long>(),
        voxel_grid.data_ptr<float>(),
        BN, D, G
    );

    return voxel_grid;
}
'''

cpp_source = r'''
torch::Tensor voxel_pool_cuda(torch::Tensor features, torch::Tensor indices, torch::Tensor batch_idx, int G, int B);
'''

voxel_pool = load_inline(
    name='voxel_pool',
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=['voxel_pool_cuda'],
    verbose=False,
    extra_cflags=['-O3']
)

class ModelNew(nn.Module):
    def __init__(self, feat_dim, grid_size=32, voxel_size=1.0):
        super().__init__()
        self.feat_dim = feat_dim
        self.grid_size = grid_size
        self.voxel_size = voxel_size
        self.linear = nn.Linear(feat_dim, feat_dim)

    def forward(self, coords: torch.Tensor, features: torch.Tensor) -> torch.Tensor:
        B, N, _ = coords.shape
        G = self.grid_size

        features = self.linear(features)  # (B, N, D)

        # Center coordinates and discretize
        center = coords.mean(dim=1, keepdim=True)
        shifted = (coords - center) / self.voxel_size + G / 2.0
        indices = shifted.long().clamp(0, G - 1)  # (B, N, 3)

        # Prepare for kernel: flatten to (B*N, ...)
        BN = B * N
        features_flat = features.reshape(BN, -1).contiguous()
        indices_flat = indices.reshape(BN, 3).contiguous()
        batch_idx = torch.arange(B, device=coords.device).unsqueeze(1).expand(B, N).reshape(BN).contiguous()

        return voxel_pool.voxel_pool_cuda(features_flat, indices_flat, batch_idx, G, B)