In previous (pre-2.3) releases of CUDA, the CUFFT library evidently had some sort of bug that caused it to eventually crash with an ULF if called repeatedly and which affected all the G80/92 hardware we tested on. With 2.3, this problem seemed to have been fixed, however, the attached code continues to fail, but apparently only on GTX260 -216 cards - we’ve not been able to make it crash on anything else.
The code simply runs a batched 1D FFT in a loop, alternating with a call to a trivial kernel. After a few thousand iterations (a few minutes of runtime) it will fail, with the check after the FFT call reporting a ULF. It’s sensitive to the dimensions of the FFT batch and also the use of float2 rather than float1 in the simple kernel.
The crash can be reproduced on a GTX260-216 with CUDA 2.3 /190.18 driver and 64b Fedora OS. With 2.3, we can’t reproduce it on any other hardware.
PS Board admin: Do you not think it would be sensible to allow cuda source file attachments to posts?
silly.tar (836 Bytes)