#include <torch/extension.h>

void state_step_cuda(torch::Tensor v, torch::Tensor g, torch::Tensor refractory,
                     torch::Tensor arrival, torch::Tensor spikes,
                     torch::Tensor spike_counts, double g_decay, double alpha,
                     double rest, double reset, double threshold,
                     int64_t refractory_steps);
void force_step_cuda(torch::Tensor v, torch::Tensor refractory,
                     torch::Tensor spikes, torch::Tensor spike_counts,
                     torch::Tensor targets, torch::Tensor forced,
                     double reset, int64_t refractory_steps);
void fill_edges_cuda(torch::Tensor indptr, torch::Tensor sources,
                     torch::Tensor output_offsets, torch::Tensor output);
void segmented_sum_cuda(torch::Tensor sorted_edge_ids,
                        torch::Tensor segment_starts,
                        torch::Tensor segment_targets,
                        torch::Tensor weights,
                        torch::Tensor destination);

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("state_step", &state_step_cuda, "Canonical float32 LIF update (CUDA)");
  m.def("force_step", &force_step_cuda, "Canonical external spike injection (CUDA)");
  m.def("fill_edges", &fill_edges_cuda, "Source/edge ordered CSR expansion (CUDA)");
  m.def("segmented_sum", &segmented_sum_cuda, "Fixed-order sequential segmented float32 sum (CUDA)");
}