Re: [Starpu-devel] Performance with SOCL on multiple devices

[Malcolm Roberts <malcolm.i.w.roberts@gmail.com>]

Hello,

Thanks for taking a look at this.  The relevant branch is socl2 (or 
devel); the master branch doesn't include these optimisations.  I also 
added the option in socl2 to include all of the devices on a platform, 
which is currently (sorry; it's a hack) activated by passing -DSOCL=ON 
to cmake.

Best,

~Malcolm

On 06/01/2016 14:44, Samuel Thibault wrote:
> Hello,
>
> It seems you do not use the CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE
> property. IIRC, this makes SOCL have to sequentialize all tasks queued
> on the same command queue (required by the OpenCL standard). To get
> parallelism, you would need to set
> CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, and use events to synchronize
> the kernels that need to depend on each other.
>
> Samuel

--
http://malcolmiwroberts.com