0 as number of threads per block behaviour

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I’ve been testing some CUDA features/behaviours, and i couldnt find any documentation on this matter. If someone who experimented with this before could tell me his experience would be great.

kernel_name<<< num_blocks, 0 >>> (params);

What happens if i call a kernel with this configuration?? (i have a test program, and it works fine :S, but i dont know what CUDA does for real). Is not that i’m calling kernels with it, but im writting some kind of utility for CUDA and i would like to know if i should give support to 0 as a parameter.

External Image

2

Why would you like to run 0 threads per block? It seems very stupid. You should not support 0 as a parameter.

3

I dont know, maybe 0 has some special meaning like… “let cuda guess what’s best for your architecture”, that’s what i would like to know, if it has some special meaning, or its just some non-supported random behaviour number.

4

if you put o threads per block you get a launch error. The kernel will not be executed.

5

humm okay thanks, i’ll have to check my output again.

Damn true, forget this topic ._., output was exactly the same than input (usually only minor changes).

6

The correct way to detect errors like this is to check the return codes from CUDA functions. Don’t rely on changes in output to discover that a function somewhere has failed.

7

Id you use some error checking this will automatically stop the program with error.