Unspecified Launch failure with For loops?

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m a little confused on why I’m getting this error. The following code (its a global function) produces NO errors:

for (int i = 0; i < 4; i++)

{

	float amount = round(error/float(i+1))*increment;

	//*neighbors[i] = amount;

}

float amount;

amount = round(error/float(1))* increment;;

*neighbors[0] = amount;

amount = round(error/float(2))* increment;;

*neighbors[1] = amount;

amount = round(error/float(3))* increment;;

*neighbors[2] = amount;

amount = round(error/float(4))* increment;;

*neighbors[3] = amount;

Now if I uncomment the commented line, then this code FAILS with unspecified launch failure.

for (int i = 0; i < 4; i++)

{

	float amount = round(error/float(i+1))*increment;

	*neighbors[i] = amount;

}

float amount;

amount = round(error/float(1))* increment;;

*neighbors[0] = amount;

amount = round(error/float(2))* increment;;

*neighbors[1] = amount;

amount = round(error/float(3))* increment;;

*neighbors[2] = amount;

amount = round(error/float(4))* increment;;

*neighbors[3] = amount;

I’m not sure whats going on. Any ideas? Could it be something to do with how I’m compiling this? The flags I’m using are -use_fast_math -O3

Any help would be greatly appreciated, thanks.

2

out of bounds memory access?

3

tmurray: While you’ve answered similar questions hundreds of times and seems to know your stuff, I’ve got say I don’t get it either. It looks to me like the functionality is identical. One loop, and the same loop unrolled.

Skyd: Run under the debugger and report back :)

4

This is a little curious. What is the declaration of neighbors? How is it initialized? Double pointers in device code are usually a red flag for possible problems.

5

That FOR loop would be unrolled anyway, so both versions would be the same. Could you check that the numbers of registers are different in these two cases?