FIFO Error on large datasets

Hello All,

I’m currently using a matrix multiply kernel for testing purposes on the TK1. So far I haven’t had any issues running the kernel locally. I also can connect to the device via ssh and run the code no problem. However, if I connect to the TK1 using a serial connection (specifically minicom), I get the following error message intermittently during the execution of the matrix multiply code:

“[ 5732.360040] gk20a gk20a.0: fifo_error_isr: channel reset initated from fifo_error_isr”

Out of 100 iterations, I get the error message maybe 10 times. It only happens during serial execution, and it only happens for input sizes >= 512x512. Note that the first number in brackets changes each time.

Has anyone encountered this error before? Any ideas on what the issue could be, or how to go about solving it?


Hello, mjgough:
Does your test program output a lot of messages in console?
If then, you can try to re-direct the output messages to /dev/null or a file to see whether the error’s gone.
If there’s no too many output messages, is that possible to paste your test program?