Kernel not processing half of the elements on GPU, works fine on CPU

Hello everyone,

I have a simple JOCL demo [http://goo.gl/aYBMi] adapted from the HelloJOCL sample [http://goo.gl/cVR6m] (I shall mention that the HelloJOCL sample works OK on the same GPU).

I am having a weird problem where half of the elements in my queue are not being processed (or so it seems) when running on GPU, but everything works ok when switching to CPU.

I tried to change the number of elements and no matter how few I use only the first half gets processed correctly, while the rest is apparently picking up default values for the output buffer. Even though the init vectors are correctly initialized when I peek at them, it almost looks like the second half of the vector values are being passed to the kernel as default values (doubles), but I am not really sure that those items are being processed by the kernel at all.

The kernel is quite simple [http://goo.gl/KiyyK], it's just a pass-through for one of the input buffers (I am only testing that I can invoke it OK) and I just assign the value from one of the input buffers to the relevant output buffer. I am inclined to think the kernel is not the problem since it works fine on CPU.

I know it's quite a bit of code to go through, so I am just asking if any of this rings a bell.

Any help appreciated!

Weirdest finding: if I change all the doubles to floats the problem seems to go away.

Does this make any sense at all?

I'll try to take a look at it this evening, at the first glance your code looks good.
Are you running a 64bit JVM?

best regards,
michael

On 05/15/2011 04:37 AM, John_Idol [via jogamp] wrote:

Weirdest finding: if I change all the doubles to floats the problem seems to go away.

Does this make any sense at all?

---

If you reply to this email, your message will be added to the discussion below:

http://forum.jogamp.org/Kernel-not-processing-half-of-the-elements-on-GPU-works-fine-on-CPU-tp2941599p2942405.html

To start a new topic under jogamp, email [hidden email]
To unsubscribe from jogamp, click here.

Thanks for having a look - after posting on the khronos forums about another issue I was having I was forced to switch to floats (apparently my GPU doesn't support doubles). After that it seems to work OK. I am kind of thinking that the issue was caused by lack of double support too, but it's very strange nonetheless, and I don't understand why were half of the stuff being processed correctly. Weird!

Now everything is working OK anyway ... but I already have another question that I will post soon :)

  yep i saw your post on the message board already :)

if you want to switch between double and single precision take a look at
the multi device demo and how the kernel was written.
https://github.com/mbien/jocl-demos/blob/master/src/com/jogamp/opencl/demos/fractal/Mandelbrot.cl

it uses doubles in host code (java app) and sets
kernel.setForce32BitArgs(true) for double->float conversion in the
kernel.putArg() methods. (+ it recompiles the program with different
compiler options)

best regards,
michael

On 05/17/2011 07:42 PM, Giovanni Idili [via jogamp] wrote: