mikeash.com pyblog/friday-qa-2010-04-02-opencl-basics.html comments

Filip - 2011-10-24 12:50:03

Mon, 24 Oct 2011 12:50:03 GMT

Lets say you are calculating a huge set of numbers, how do you stop the kernel when the user says "no more, i am bored and want to quit, but not the app, only the calculating".

Is that even possible in openCL ?

pyth - 2010-04-06 08:56:14

Tue, 06 Apr 2010 08:56:14 GMT

For a in depth discussion on work sizes and work items check out the excellent opencl tutorial podcast over on macresearch: http://www.macresearch.org/opencl.
It explains what SMUGOpenCL does under the hood and more (probably more then you want to know...)

Kyle Sluder - 2010-04-05 06:20:55

Mon, 05 Apr 2010 06:20:55 GMT

Wouldn't it be nice if we could hand off some LLVM byte code instead of just raw program text? I'm thinking this would be even more useful on the iPhone for pixel shaders, but even on the desktop for OpenCL it would be nice to have already performed the parsing and some static analysis.

mikeash - 2010-04-03 00:27:35

Sat, 03 Apr 2010 00:27:35 GMT

You would not believe how many times I made that typo. Thanks for pointing out the one I missed.

In answer to your question, that probably is not guaranteed. I didn't even think about that. You'd probably need to make the final results array use a larger data type in order to be safe.

natevw - 2010-04-02 17:56:52

Fri, 02 Apr 2010 17:56:52 GMT

A question, and a tip:

Does OpenCL guarantee that setting a single byte in a shared array like your kernel does will be "thread" safe?

During the SL seeds, Ian Ollman posted to the Developer Forums "Why you should vectorize and some tips to get started". Unfortunately, my ADC membership has since expired and it doesn't seem to have been converted to a tech note, but if you have access it could help you get more performance out of OpenCL.

natevw - 2010-04-02 16:41:06

Fri, 02 Apr 2010 16:41:06 GMT

I think you mean "SMUGOpenCL" in the second paragraph. Other than that, this comment is left intentionally useless because I haven't finished the article.