Bitonic sort for 64 elements

bitonic_sort64.h

template<typename T>
void bitonic_sort64(T vals_[64])
{
  int h=-1;
  do
  {
    for(unsigned i=0; i<32; ++i)
    {
      unsigned idx0=2*i;
      unsigned idx1=4*(i&h)-2*(i+h)-1;
      T v0=vals_[idx0];
      T v1=vals_[idx1];
      if((idx0<idx1)^(v0<v1))
      {
        vals_[idx0]=v1;
        vals_[idx1]=v0;
      }
    }
    for(int hh=h; !(hh&1);)
    {
      hh>>=1;
      for(unsigned i=0; i<32; ++i)
      {
        unsigned idx0=2*(i&hh)+(i&~hh);
        unsigned idx1=idx0+1+~hh;
        T v0=vals_[idx0];
        T v1=vals_[idx1];
        if(v0>v1)
        {
          vals_[idx0]=v1;
          vals_[idx1]=v0;
        }
      }
    }
    h*=2;
  } while(h!=-64);

HLSL implementation: https://godbolt.org/z/xbefr6aMd