CUDA 3d kernels keep crashing
category: general [glöplog]
hello,
i'm going crazy, because my cuda-kernel keeps crashing when using >1 in the z-dimension of the blocks.
i boiled it down to a really simple case:
works, but
crashes. :(
the kernel-func is completely empty, so there can not be any problem with the kernel itself.
ideas, anyone ?
i'm going crazy, because my cuda-kernel keeps crashing when using >1 in the z-dimension of the blocks.
i boiled it down to a really simple case:
Code:
dim3 dim3Blocks(16, 16, 1);
dim3 dim3Threads(1, 1, 1);
myKernelFunc<<<dim3Blocks, dim3Threads>>>();
works, but
Code:
dim3 dim3Blocks(16, 16, 2);
dim3 dim3Threads(1, 1, 1);
myKernelFunc<<<dim3Blocks, dim3Threads>>>();
crashes. :(
the kernel-func is completely empty, so there can not be any problem with the kernel itself.
ideas, anyone ?
driver bug? maybe?
hmm, i use cuda 2.1 and it even crashes with in emulation-mode. but without an understandable reason =/
will try cuda 2.3...
will try cuda 2.3...
that is tricky. do the sample projects work (that's assuming any of them use a blockDim with z > 1)? That'd narrow it down to compile options.
check the other thread once again:
too many inits/registers used !
( what xtr1m said ! )
if ( !not ) ALU-count_exceeded == TRUE;
too many inits/registers used !
( what xtr1m said ! )
if ( !not ) ALU-count_exceeded == TRUE;
that made you puzzle, eh ? ;)
Well, I had that problems too, and I only know that after 1 week of searching (and even exchanging mails with CUDA developers@nVidia) there's no way around it.
- "It's a known bug, it will be fixed in later versions"
That was one year ago... really, if you can avoid CUDA, then do so.
- "It's a known bug, it will be fixed in later versions"
That was one year ago... really, if you can avoid CUDA, then do so.
ehh, yes indeed ;)
guessing what you mean, my answer would be: since absolutely no regs are used in my func (because its empty) that cannot be the case...
;)
guessing what you mean, my answer would be: since absolutely no regs are used in my func (because its empty) that cannot be the case...
;)
i tried the latest drivers & toolkit and still crashing ='(
well, i can use a 2d-kernel and transform to 3d-coords inside it, but these kind of issues are a waste of neurons... ;)
but thanks for the info, xTr1m :)