forcing loop unrolls

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

Since the cuda compiler often shortens the length of the loop unrollled, is there a way i can force it?
My loop length, n , is really large 25,000,000

I use the following:

int i;
#pragma loop unroll
for(int j=0;j<N;j++)
i=j;

It is shortened to just 17 runs of N. How can i better understand why the compiler shortens it by so much?

thank you

2

The CUDA compiler compiles with full optimizations by default. This includes dead code elimination. If you have a sequence of assignments (as you have in your loop):

i = 0;
i = 1;
...
i = N-1;

this is equivalent to simply

i = N-1;

BTW, the pragma as you wrote it doesn’t look correct to me. It should be

#pragma unroll <unrolling-factor>
3

If I am not mistaken, if the compiler doesn’t know the value of N beforehand, the unrolling is limited?
But then it is not really particular to CUDA, but to C/C++.

4

I am posting this from the NVIDIA programming guide:

“If no number is specified after #pragma unroll, the loop is completely unrolled.
if its trip count is constant, otherwise it is not unrolled at all.”

So if N is set to 25,000,000 , why would it not completely unroll?

5

if you are using

pragma loop unroll

that is incorrect

there is no loop

Furthermore, there is almost certainly some unstated limit to loop unrolling.