gridDim and blockDim in CUDA 4

I have always thought that gridDims could only be 2D while the blockDims could be 3D, however,
I recently tried setting gridDims.z to be above 1 and using CUDA4 on a C2050 it seems to work
as I would expect. Has something changed in CUDA 4 in this regard ?

If I run the same code linked against CUDA 3.2 on the c2050 I don’t get a kernel error but I do
get rubbish results.

The same code, using either CUDA 3.2 or 4 on a gtx285 produces a kernel launch error.

Anyone seen this ?


Yes, 3D grids (i.e. gridDim.z > 1) were introduced in the Fermi architecture (e.g. Tesla C2050 and later) and are supported in CUDA 4.0.

Cool, thanks.