Unspecified Launch Failure in cudaMemcpyToSymbol

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m facing a weird problem concerning cudaMemcpyToSymbol() on my 64bit Ubuntu 9.10 (Kernel 2.6.31-22) with CUDA 3.2 installed. When calling init() in the following code cudaMemcpyToSymbol() brings up an “unspecified launch failure”. I’ve no clue why since the exactly same code runs perfectly fine on my other system which is an Ubuntu 10.10 (Kernel 2.6.35-25, also CUDA 3.2). Also, both system use the 260.19.26 NVidia driver.

__device__ __constant__ int KRNL_RADIUS;

__host__ void init() {

	int radius = 10;

	safecall(cudaMemcpyToSymbol( KRNL_RADIUS, &radius, sizeof(int), 0, cudaMemcpyHostToDevice));

}

Does anybody know if this could be related with the older kernel? Of course I’m also happy about other hints. Thanks!

2

There must be a kernel execution that is failing before your are calling that function. Are the GPUs in both machines the same?

3

You are right. I’m quite new to CUDA and missed it to read section “3.2.8 Error Handling” in the CUDA Programming Guide in advance ;-) Thanks for the support!