3 D thread blocks in a cuda kernel

Accelerated Computing CUDA CUDA Programming and Performance
1

I want to use a 3 dimensional threads in one of my image processing application.
The output image is 512X512X128.
Here is what I am trying to do.
blocks.x = 8, blocks.y = 8, blocks.z = 8
grids.x = 512/8, grids.y = 512/8, grids.z = 128/8.
The kernel in this case times out as the total time taken by it it more than few seconds.
But when I comment grids.z, it takes the default value (1) and kernel runs fine.
To my surprise, the output is also correct.

Please let me know if my understanding of 3-D grids is fine or I am missing something.
I am using a GTX660 (Kepler 3.0 ) and CUDA 5.0.

2

How does Your kernel code looks like? It seems that although You define 3D execution grid You use only two dimentions.

MK

3

You are probably looking at results that are still around from the timed out kernel launch. CUDA doesn’t clear memory between runs.

4

Also if your kernel is timing out (and takes that long) and you’re in Windows, disable TDR.

5

Thanks for all the reply.
I have converted the code to 2D thread-blocks and launching it for 128 times.
It seems like when I specify grids.z = 1 and blocks.z = 2,
I can reduce the number of calls from 128 to 64.

Thanks :)