CUDA SDK

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi there I am having the following problems with the current version of the CUDA SDK

First of all is the warning stating that Double is not supported, this is worrying because our kernel does not contain any doubles.

Second is ptxas that crashes, with no detail is there any way I can work around this, is there anybody in nVidia or people who have
discovered this crash?

ptxas /tmp/tmpxft_000026bb_00000000-2_DeviceUtility.ptx, line 100; warning : Double is not supported. Demoting to float
ptxas info : Compiling entry function ‘Z18GPU_AdvancingFrontiP13__IndexStrideP10__particleP14
_threadOutputP14__ThreadMemory’
Segmentation fault
make: *** [TetOffensive] Error 255

2

Are there any options that can at least help me hunt down what is causing the crash?

3

Looks like a toolkit bug…

WHich version of CUDA are you using? CUDA 3.0?
You could try 2.3 just for testing purposes…

And, I suggest you to post a minimal repro case so that NV guys can reproduce it in their labs…

4

We are using 2.3 we were thinking of downgrading to 2.2 to see if it happens again

Again I don’t know what part of the code is producing this error, I did get more verbose messages from the compiler

(I cannot share the kernel as its proprietary, not even mine :( )

I have a core dump if that helps nVidia

SNIP

I worked out the issue manually, turns out a college had just a “return” on a piece of code that was meant to return an int type

the compiler flagged this as a warning and the Assembler would crash implementing it.