import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Fused sparse attention-value aggregation in CSR format.

    For each destination row, compute a softmax over edge logits and then use
    the normalized weights to aggregate source node values. This is a common
    fused graph-attention pattern.
    """

    def __init__(self):
        super().__init__()

    def forward(
        self,
        row_ptr: torch.Tensor,
        col_idx: torch.Tensor,
        edge_scores: torch.Tensor,
        node_value: torch.Tensor,
    ) -> torch.Tensor:
        num_nodes = row_ptr.numel() - 1
        feat_dim = node_value.shape[1]
        out = torch.zeros(num_nodes, feat_dim, dtype=node_value.dtype, device=node_value.device)

        for dst in range(num_nodes):
            start = int(row_ptr[dst].item())
            end = int(row_ptr[dst + 1].item())
            if end > start:
                weights = torch.softmax(edge_scores[start:end], dim=0)
                src = col_idx[start:end].long()
                out[dst] = (node_value[src] * weights.unsqueeze(-1)).sum(dim=0)

        return out


num_nodes = 384
avg_degree = 18
feat_dim = 64


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



def get_init_inputs():
    return []

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 <cmath>

__device__ __forceinline__ float warpReduceMax(float val) {
    #pragma unroll
    for (int offset = 16; offset > 0; offset /= 2) {
        val = fmaxf(val, __shfl_down_sync(0xffffffff, val, offset));
    }
    return val;
}

__device__ __forceinline__ float warpReduceSum(float val) {
    #pragma unroll
    for (int offset = 16; offset > 0; offset /= 2) {
        val += __shfl_down_sync(0xffffffff, val, offset);
    }
    return val;
}

__global__ void sparse_attention_value_kernel(
    const int* __restrict__ row_ptr,
    const int* __restrict__ col_idx,
    const float* __restrict__ edge_scores,
    const float* __restrict__ node_value,
    float* __restrict__ out,
    int num_nodes,
    int feat_dim
) {
    int dst = blockIdx.x;
    if (dst >= num_nodes) return;

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

    if (degree == 0) {
        for (int f = threadIdx.x; f < feat_dim; f += blockDim.x) {
            out[dst * feat_dim + f] = 0.0f;
        }
        return;
    }

    float max_score = -INFINITY;
    for (int i = threadIdx.x; i < degree; i += blockDim.x) {
        float s = edge_scores[start + i];
        max_score = fmaxf(max_score, s);
    }
    __shared__ float shared_max[32];
    int lane = threadIdx.x & 31;
    int warp = threadIdx.x >> 5;
    max_score = warpReduceMax(max_score);
    if (lane == 0) shared_max[warp] = max_score;
    __syncthreads();
    if (warp == 0) {
        max_score = (threadIdx.x < ((blockDim.x + 31) >> 5)) ? shared_max[threadIdx.x] : -INFINITY;
        max_score = warpReduceMax(max_score);
        if (lane == 0) shared_max[0] = max_score;
    }
    __syncthreads();
    max_score = shared_max[0];

    float sum_exp = 0.0f;
    for (int i = threadIdx.x; i < degree; i += blockDim.x) {
        float s = __expf(edge_scores[start + i] - max_score);
        sum_exp += s;
    }
    __shared__ float shared_sum[32];
    sum_exp = warpReduceSum(sum_exp);
    if (lane == 0) shared_sum[warp] = sum_exp;
    __syncthreads();
    if (warp == 0) {
        sum_exp = (threadIdx.x < ((blockDim.x + 31) >> 5)) ? shared_sum[threadIdx.x] : 0.0f;
        sum_exp = warpReduceSum(sum_exp);
        if (lane == 0) shared_sum[0] = sum_exp;
    }
    __syncthreads();
    sum_exp = shared_sum[0];
    float inv_sum = 1.0f / sum_exp;

    for (int f = threadIdx.x; f < feat_dim; f += blockDim.x) {
        float val = 0.0f;
        for (int i = 0; i < degree; i++) {
            float w = __expf(edge_scores[start + i] - max_score) * inv_sum;
            int src = col_idx[start + i];
            val += node_value[src * feat_dim + f] * w;
        }
        out[dst * feat_dim + f] = val;
    }
}

torch::Tensor sparse_attention_value_cuda(
    torch::Tensor row_ptr,
    torch::Tensor col_idx,
    torch::Tensor edge_scores,
    torch::Tensor node_value
) {
    int num_nodes = row_ptr.size(0) - 1;
    int feat_dim = node_value.size(1);
    auto out = torch::zeros({num_nodes, feat_dim}, node_value.options());

    dim3 blocks(num_nodes);
    dim3 threads(64);

    sparse_attention_value_kernel<<<blocks, threads>>>(
        row_ptr.data_ptr<int>(),
        col_idx.data_ptr<int>(),
        edge_scores.data_ptr<float>(),
        node_value.data_ptr<float>(),
        out.data_ptr<float>(),
        num_nodes,
        feat_dim
    );

    return out;
}
"""

cpp_source = """
torch::Tensor sparse_attention_value_cuda(
    torch::Tensor row_ptr,
    torch::Tensor col_idx,
    torch::Tensor edge_scores,
    torch::Tensor node_value);
"""

module = load_inline(
    name="sparse_attn_value",
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=["sparse_attention_value_cuda"],
    verbose=False
)

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

    def forward(self, row_ptr, col_idx, edge_scores, node_value):
        row_ptr = row_ptr.to(torch.int32).contiguous()
        col_idx = col_idx.to(torch.int32).contiguous()
        return self.module.sparse_attention_value_cuda(row_ptr, col_idx, edge_scores, node_value)