Host Machine has CUDA 6.5. Is this a problem for cross compiling for Jetson TK1?
I haven’t bother to try as of yet.
Most likely.
You might have issue trying to link their libs.
Either issue with alignment or you used 6.5 method calls that doesn’t exist in 6.0
FYI, I suspect there is a problem for “cross display” (versus cross compile). So I’m curious as to what prompts the question, as this may be related.
A while back I tried executing a CUDA 6.0 program I had built on Jetson (natively compiled on Jetson), but could not execute because of missing CUDA error of some sort. The error was tied to me logging in to Jetson with ssh -Y (X11 forwarding) and running the CUDA program on Jetson but displaying on my linux workstation. The error goes away if displaying and running locally on Jetson.
Here’s what happens…normal X11 from Jetson could run glxinfo or xdpyinfo and obtain information about Jetson. Direct hardware rendering is possible here. Remote login via ssh -Y is intended to display on the workstation instead, but execute on Jetson…thus xdpyinfo and especially direct rendering listings of glxinfo will change. Direct rendering is no longer possible, Jetson does not have workstation video directly wired. So here is where the CUDA software has become confused…apparently Jetson believes CUDA is remote video, and wants the workstation to use its CUDA abilities during remote display.
What makes this most interesting is that if both workstation and Jetson have compatible CUDA, then use of X11 forwarding (ssh -Y) will silently do all CUDA work on the workstation and not on Jetson. You would probably never know it is doing this…could make for some interesting benchmarks!
So far as I know this bug where X11 display interferes with where CUDA runs has never been addressed.
We’re all waiting for CUDA 6.5 support on the Jetson TK1.
Well, for the cross display, my workaround for x11 so far is just define my application to always display on the Jetson.
And just remote desktop to the Jetson to see the results.
Kind of defeats the purpose though…some people will not have monitors on compute nodes. Mine is a smaller monitor with hard-to-reach keyboard/mouse, etc. Display on my workstation should be separate from compute on my Jetson. I wish nVidia would address this bug, but I don’t think they even know it exists. But think about people who are doing performance tests of one kind or another…wouldn’t that be inconvenient if it were really running on their workstation? Also, what about cases where software detects if CUDA is available, uses it if it is, but goes pure software if CUDA is not detected?