No Uninitialized variable access warning?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi Folks,

We 've been using OpenCL for a few month now and we had a sticky problem lately:
A kernel compiled without any complaints, but the ptx code for the kernel only contained a trap; instruction causing the GPU to lock up (and also the rest of the system).
After exhaustive troubleshooting we could point out the following error:

float *fTest;

Test[0] = (some value);

This wasn’t that easy to spot as there are a lot of variables in this kernel and we had no clue what to look for. I’ve been wondering why there ain’t any warnings showing up as an uninitialized local variable was used. Is there a switch (like -Wall) to turn on all Compiler warnings to spot such issues or is it still up to NVIDIA to implement this. The inserted trap indicates that the compiler noticed that somethings wrong, but it could have generated errors or warnings instead of a code that locks up the machine.

Thanks in advance

2

Well, after the forum is finally up again: I have found out that in general it’s quite handy during development to set the compiler to -Werror via the options string in clBuildProgramm (unlike NVIDIA, the ATI implementations will also produce error from “unused variable” warnings) or regularly examine the the build log even if there was no error.

Hope this helps someone else.