CUDA Shared memory out-of-bound access

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi All,

I did a simple test to see if there is checking on out-of-bound memory access on shared memory. What i did was quite simple that I allocated an integer array on shared memory and I tried to access to an element that already exceeds the size of the array (a lot). On my old Geforce 9400M there was nothing happened. I was able to access that invalid address and print the element out. I was guessing the hardware did something like modulus on the shared memory for out-of-bound access but based on my experiment the valid elements of the array were not affected.

Any idea of that?

Thank you in advance.

2

The Fermi cards have more strict rules and the kernel crashes more often when there is out of bounds access. But in general you need to check with cuda-memcheck or put explicit checking. For my 8800m gtx I had out of bounds accesses without problems for small sizes, but my kernels were crashing for larger sizes.