import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Per-row top-k selection over CSR edge segments.

    Returns the top-k values and their original edge indices for each node's
    neighborhood. Useful for graph sparsification, hard attention, and
    candidate-pruning workloads.
    """

    def __init__(self, k):
        super().__init__()
        self.k = k

    def forward(
        self,
        row_ptr: torch.Tensor,
        edge_scores: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        num_nodes = row_ptr.numel() - 1
        topk_vals = torch.full((num_nodes, self.k), float("-inf"), dtype=edge_scores.dtype, device=edge_scores.device)
        topk_idx = torch.full((num_nodes, self.k), -1, dtype=torch.int64, device=edge_scores.device)

        for node in range(num_nodes):
            start = int(row_ptr[node].item())
            end = int(row_ptr[node + 1].item())
            if end <= start:
                continue
            segment = edge_scores[start:end]
            curr_k = min(self.k, end - start)
            vals, idx = torch.topk(segment, k=curr_k, dim=0)
            topk_vals[node, :curr_k] = vals
            topk_idx[node, :curr_k] = idx.long() + start

        return topk_vals, topk_idx


num_nodes = 320
avg_degree = 12
k = 4




def get_inputs():
    return list(popcorn_pri.make_csr_scalar_edge_scores(num_nodes, avg_degree))



def get_init_inputs():
    return [k]

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>
#include <vector>

__global__ void topk_kernel(
    const int* __restrict__ row_ptr,
    const float* __restrict__ edge_scores,
    float* __restrict__ topk_vals,
    int64_t* __restrict__ topk_idx,
    int num_nodes,
    int k
) {
    int node = blockIdx.x * blockDim.x + threadIdx.x;
    if (node >= num_nodes) return;

    int start = row_ptr[node];
    int end = row_ptr[node + 1];
    int degree = end - start;

    if (degree <= 0) {
        for (int i = 0; i < k; ++i) {
            topk_vals[node * k + i] = -INFINITY;
            topk_idx[node * k + i] = -1;
        }
        return;
    }

    int curr_k = min(k, degree);

    float top_vals[16];
    int64_t top_indices[16];
    for (int i = 0; i < curr_k; ++i) {
        top_vals[i] = -INFINITY;
        top_indices[i] = -1;
    }

    for (int i = 0; i < degree; ++i) {
        float val = edge_scores[start + i];
        for (int j = 0; j < curr_k; ++j) {
            if (val > top_vals[j]) {
                for (int m = curr_k - 1; m > j; --m) {
                    top_vals[m] = top_vals[m-1];
                    top_indices[m] = top_indices[m-1];
                }
                top_vals[j] = val;
                top_indices[j] = start + i;
                break;
            }
        }
    }

    for (int i = 0; i < k; ++i) {
        if (i < curr_k) {
            topk_vals[node * k + i] = top_vals[i];
            topk_idx[node * k + i] = top_indices[i];
        } else {
            topk_vals[node * k + i] = -INFINITY;
            topk_idx[node * k + i] = -1;
        }
    }
}

std::vector<torch::Tensor> topk_cuda(torch::Tensor row_ptr, torch::Tensor edge_scores, int k) {
    int num_nodes = row_ptr.size(0) - 1;
    auto options_val = torch::TensorOptions().dtype(edge_scores.dtype()).device(edge_scores.device());
    auto options_idx = torch::TensorOptions().dtype(torch::kInt64).device(edge_scores.device());
    auto topk_vals = torch::empty({num_nodes, k}, options_val);
    auto topk_idx = torch::empty({num_nodes, k}, options_idx);

    int threads = 256;
    int blocks = (num_nodes + threads - 1) / threads;
    topk_kernel<<<blocks, threads>>>(
        row_ptr.data_ptr<int>(),
        edge_scores.data_ptr<float>(),
        topk_vals.data_ptr<float>(),
        topk_idx.data_ptr<int64_t>(),
        num_nodes,
        k
    );
    cudaDeviceSynchronize();
    return {topk_vals, topk_idx};
}
'''

cpp_source = '''
std::vector<torch::Tensor> topk_cuda(torch::Tensor row_ptr, torch::Tensor edge_scores, int k);
'''

module = load_inline(
    name='topk_module',
    cpp_sources=[cpp_source],
    cuda_sources=[cuda_source],
    functions=['topk_cuda'],
    verbose=False,
    extra_cflags=['-O3'],
    extra_cuda_cflags=['-O3']
)

class ModelNew(nn.Module):
    def __init__(self, k):
        super().__init__()
        self.k = k
        self.topk_cuda = module.topk_cuda

    def forward(self, row_ptr: torch.Tensor, edge_scores: torch.Tensor):
        row_ptr = row_ptr.to(torch.int32).contiguous()
        return self.topk_cuda(row_ptr, edge_scores, self.k)