Calculation stalls when using more than one block

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi! I have a really stubborn problem and I’ve been trying to fix it for weeks. Finally I have managed to narrow it down and think it is related to the number of blocks but the fight continues! External Image

In the host code there is a loop and the kernels are executed inside it.(Maybe a bad idea to begin with??) There are also some cudaMemcpy()'s inside. When NUM_BLOCK == 1 the code is working fine. When I change NUM_BLOCKS to anything higher, say 2, it stalls inside Kernel2 or directly after.

I think it has to with synhronization but I have no idea what else to do except from using cudaThreadSynchronize() after the kernels and __syncthreads() inside when ever possible. Maybe it has to do with something completely different.

Please give me your thoughts about it. Anything is appreciated!

I use a GTX 260 and Visual Studio 2008 Pro with 64-bit Vista.

//Host code
.
.
.
for(int k = 0; k < 10; k++){
cutilSafeCall( cudaMemcpy(…) );
cudaThreadSynchronize();
cutilCheckMsg(“Execution failed\n”);

Kernel1<<<NUM_BLOCKS, 32>>>(…);
cudaThreadSynchronize();
cutilCheckMsg(“Execution failed\n”);

cutilSafeCall( cudaMemcpy(…) );
cudaThreadSynchronize();
cutilCheckMsg(“Execution failed\n”);

Kernel2<<<NUM_BLOCKS, 32>>>(…);
cudaThreadSynchronize();
cutilCheckMsg(“Execution failed\n”);

cutilSafeCall( cudaMemcpy(…) );
cudaThreadSynchronize();
cutilCheckMsg(“Execution failed\n”);
}
.
.
.

2

I don’t think anyone is going to be able to help you without some indicative kernel code. You probably have a race or synchronisation problem, but it is a bit unrealistic to expect a diagnosis without the code in question…

3

I don’t think anyone is going to be able to help you without some indicative kernel code. You probably have a race or synchronisation problem, but it is a bit unrealistic to expect a diagnosis without the code in question…

4

THANK YOU!! You made my week!

Sorry for not posting the kernel code but a thought the problem was in the host code. I went looking in the kernel code for some race condition after reading your post and I have already found the one causing troubles. I have been searching for the error at the wrong places. Should have ask this question long ago!!

Thank you again!

5

THANK YOU!! You made my week!

Sorry for not posting the kernel code but a thought the problem was in the host code. I went looking in the kernel code for some race condition after reading your post and I have already found the one causing troubles. I have been searching for the error at the wrong places. Should have ask this question long ago!!

Thank you again!