Reverse "for loops" count one short with evergreen GPU code.
After hours of debugging I found the reason my sorting algorithm didn't work.
With a given input of A[] = 1, 2, ..., N the following Kernel code gives us for n=16
sum_total1 = 136; // correct sum with faulty loop
sum_total2 = 135; // incorrect sum with correct loop
The error persists for larger N, but not for N=8 for example.
All tested on Juniper.
__private uint sum_total1 = 0ul; __private uint sum_total2 = 0ul; for( int i=N-1; i>=-1; --i ){ sum_total1 += A[ i ]; } for( int i=N-1; i>= 0; --i ){ sum_total2 += A[ i ]; }
