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 <cuda.h>
#include <cuda_runtime.h>
#include <torch/extension.h>

template <int K>
__device__ void insert_sorted(float* vals, int64_t* idxs, float new_val, int64_t new_idx) {
    // Insert new_val into sorted array (descending order)
    // vals[0] is largest, vals[K-1] is smallest
    for (int i = 0; i < K; i++) {
        if (new_val > vals[i]) {
            // Shift elements down
            for (int j = K - 1; j > i; j--) {
                vals[j] = vals[j-1];
                idxs[j] = idxs[j-1];
            }
            vals[i] = new_val;
            idxs[i] = new_idx;
            break;
        }
    }
}

template <int K>
__global__ void topk_csr_kernel_float(
    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 seg_len = end - start;
    
    // Initialize top-k with -inf values
    float vals[K];
    int64_t idxs[K];
    
    for (int i = 0; i < K; i++) {
        vals[i] = -INFINITY;
        idxs[i] = -1;
    }
    
    // Process segment
    for (int i = start; i < end; i++) {
        float score = edge_scores[i];
        insert_sorted<K>(vals, idxs, score, i);
    }
    
    // Write results
    int curr_k = min(k, seg_len);
    for (int i = 0; i < k; i++) {
        if (i < curr_k) {
            topk_vals[node * k + i] = vals[i];
            topk_idx[node * k + i] = idxs[i];
        } else {
            topk_vals[node * k + i] = -INFINITY;
            topk_idx[node * k + i] = -1;
        }
    }
}

std::tuple<torch::Tensor, torch::Tensor> topk_csr_cuda(
    torch::Tensor row_ptr,
    torch::Tensor edge_scores,
    int k
) {
    int num_nodes = row_ptr.numel() - 1;
    auto options_val = edge_scores.options();
    auto options_idx = edge_scores.options().dtype(torch::kInt64);
    
    auto topk_vals = torch::full({num_nodes, k}, -INFINITY, options_val);
    auto topk_idx = torch::full({num_nodes, k}, -1, options_idx);
    
    const int block_size = 256;
    const int num_blocks = (num_nodes + block_size - 1) / block_size;
    
    // Ensure row_ptr is contiguous int
    auto row_ptr_contig = row_ptr.contiguous().to(torch::kInt);
    
    // For k=4, use the specialized kernel
    topk_csr_kernel_float<4><<<num_blocks, block_size>>>(
        row_ptr_contig.data_ptr<int>(),
        edge_scores.data_ptr<float>(),
        topk_vals.data_ptr<float>(),
        topk_idx.data_ptr<int64_t>(),
        num_nodes,
        k
    );
    
    return std::make_tuple(topk_vals, topk_idx);
}
"""

cpp_source = """
std::tuple<torch::Tensor, torch::Tensor> topk_csr_cuda(torch::Tensor row_ptr, torch::Tensor edge_scores, int k);
"""

topk_ext = load_inline(
    name='topk_csr_ext_v3',
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=['topk_csr_cuda'],
    extra_cuda_cflags=['-O3'],
    verbose=False
)


class ModelNew(nn.Module):
    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]:
        return topk_ext.topk_csr_cuda(row_ptr, edge_scores, self.k)