I use this sample code to perform radix sort for unsigned integer array with N size

inline int *indx( int *table){    int *ind=new int[65536];   int TMP=0; ind[0]=0;   for(int i=1;i<65536;i++)ind[i]=TMP+=table[i-1];   return (ind);  }  void  m_radix(unsigned int *array,const int N){    int *count=new int [65536]();   int *lpos=new int[65536]; // hold the last position   int *fpos;                // hold the first position     for(int i=0;i<N;i++)     count[array[i]&0xffff]++;   fpos=indx(count);   for(int i=0;i<65536;i++)     lpos[i]=fpos[i]+count[i];      for(int j=0;j<65536;j++)      for(int i=fpos[j];i<lpos[j];i++)         while((array[i]&0xffff)!=j)           std::swap(array[i],array[fpos[array[i]&0xffff]++]);       count=new int [65536]();   for(int i=0;i<N;i++)     count[array[i]>>16]++;   fpos=indx(count);   for(int i=0;i<65536;i++)     lpos[i]=fpos[i]+count[i];      for(int j=0;j<65536;j++)      for(int i=fpos[j];i<lpos[j];i++)       while((array[i]>>16)!=j)         std::swap(array[i],array[fpos[array[i]>>16]++]);      } 

I think that this algorithm is linear however benchmarking show low performence against algorithm using auxilary array. where is the problem so ?