In my cuda program,I have defined some pointers in device functions. when I compile it,there are some warnings:cannot tell what pointer points to ,assuming global memory space.At first,I ingored them. But when I run my program, I cannot get right results.I guess this is because these warnings. How can I do with it?
Thank you for your attentions and very appreciate your help.
The warning means that you are trying to dereference a pointer but the compiler can’t tell whether the pointer is to shared memory or global memory so it assumes the pointer is to global memory. So long as this pointer is always to global memory you can completely ignore the error message. If the pointer is always to shared memory you have a problem but you might be able to trick it by setting up a pointer that the compiler can easily work out is to shared memory and then copying your ambiguous pointer into it before you try to dereference it (I’ve seen this suggested but never had cause to try it). If the pointer can be to shared memory sometimes and global memory other times then you have an even bigger problem. The issue is that there isn’t a unified address space and a regular 32-bit pointer isn’t big enough to handle both address spaces. If you are using the latest version of CUDA and targeting Fermi then it DOES have a unified address space (using 40-bit or 64-bit pointers or something) and I wouldn’t expect you to be getting this warning.