Wrong gridDim ... causes: invalid configuration argument

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I made a really simple kernel, which causes an “invalid configuration argument” (or rather the invoke of the kernel).

__global__ void SimpleKernel (int iLevel) {

	int id = threadIdx.x;

}

// Host code 

int main (...)

{   

	int iLevel = 0;

	SimpleKernel <<< dim3(4,4,4), dim3(16,1,1) >>> (iLevel);

	printf("%s\n",cudaGetErrorString(cudaGetLastError()));

}

Why does my code causes an error? Using “dim3(4,4,1)” instead of the grid “dim3(4,4,4)” works fine.

However, the CUDA manual says there’s just a block-limit of 64 in z-direction - not of 1.

So, anybody can help me out?

2

Grids are restricted to 2D in CUDA at the moment. You can have 3D block indexing, but only 2D grid indexing. It is explained in Appendix A of the program guide.

3

Thank you very much. I possibly skipped that …

So I’ll “emulate” a 3D-grid with a 2D-grid I guess.