1#include "common.cuh"
 2
 3void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 4
 5#ifdef GGML_CUDA_USE_CUB
 6void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 7                              const float *    x,
 8                              int *            dst,
 9                              const int        ncols,
10                              const int        nrows,
11                              ggml_sort_order  order,
12                              cudaStream_t     stream);
13#endif  // GGML_CUDA_USE_CUB
14void argsort_f32_i32_cuda_bitonic(const float *   x,
15                                  int *           dst,
16                                  const int       ncols,
17                                  const int       nrows,
18                                  ggml_sort_order order,
19                                  cudaStream_t    stream);