Tutorial manually specifies number of threads

[ArnoStrouwen]

In the tutorial the number of threads is always manually specified. For some users it might be difficult to find the suitable number for their GPU. Could an example be added to the documentation on how to automate this?

[maleadt]

Thanks!

Could an example be added to the documentation on how to automate this?

Sure, but I think it needs some more explanation about why (amount of threads is not unbounded, also affects performance) and how (pass a closure instead, use the occupancy API). You also can just pass config=configurator, see e.g. uses in CuArrays.

[ArnoStrouwen]

What is the reason you do not directly use the number of blocks suggested by launch_configuration? For the example in the tutorial this gives me blocks = 16 and is about 10% faster for me, than blocks = 1024 from:

    function configurator(kernel)
        config = launch_configuration(kernel.fun)

        threads = min(total_threads, config.threads)
        blocks = cld(total_threads, threads)

        return (threads=threads, blocks=blocks)
    end

[maleadt]

Sorry, forgot about this. I've included a reworked version of your suggestion in 179498a.

And to answer your question: the block count returned by the occupancy API is the required amount of blocks to fully saturate the GPU. But it obviously depends on the input size whether you should launch as many. It can be useful to change the launch configuration (e.g. decide to use a loop inside your kernel or not) based on this value, but it's generally not possible to use it directly as the number of blocks to launch (unless your kernel is very generic).