Sort.sac

module Sort;
export {Sort};

int[n] Sort(double[n] keys) 
{
  /* TODO inefficient for the GPU backend as iota(n) will be created 
     on the device, send back to the host, and then back to the device 
     again. Still 10x faster on a RTX 1650 than on a Ryzen 4600H. */
  iota = with {
    ([0] <= [i] < [n]): i;
  }: genarray([n], 0);
  return TrueSort(keys, iota, n); 
}

/* indices must be initialised to iota(n) */
external int[n] TrueSort(double[n] keys, int[n] indices, int n);
    #pragma linkname "MySortDouble"
    #pragma linksign [2, 1, 2, 3]
#if defined(SAC_TARGET_cuda)
    #pragma linkobj "src/sort_gpu.o"
    #pragma gpumem [0, 1, 2]
#elif defined(SAC_TARGET_default_sbi)
    #pragma linkobj "src/sort_cpu.o"
#elif defined(SAC_TARGET_mt_pth)
    /* TODO: make a multithreaded version work here. Thrust clashes
       with our private heap manager. */
    #pragma linkobj "src/sort_cpu.o"
#else
    #pragma linkobj "src/sort_cpu.o"
#endif