OpenCL backend for LLVM
ohartman at mail.zedat.fu-berlin.de
Fri Mar 9 17:31:24 UTC 2012
On 03/09/12 14:35, Eduardo Morras wrote:
> At 12:36 09/03/2012, you wrote:
>> Sorry if you feel boring by those messages, but soem of us still get wet
>> eyes when it comes to OpenCL and LLVm (LLVM is supposed to become soon
>> the backend compiler in FreeBSD, as I understand). On PHORONIX I read
>> this message days ago:
>> I think this is promising - in a way. Maybe some compiler specialists
>> will have merci (as well as nVidia and AMD) and present soemday in the
>> future OpenCL-based GPGPU capabilities for HPC on FreeBSD.
> I agree with you, it will allow to run some parts of llvm compiled apps
> to run in gpu, perhaps even parts of FreeBSD. A fullpower driver is the
> last piece of the puzzle.
> Don't know who has to do the first step, FreeBSD who changes the kernel
> to allow easier port of the linux drivers or nVidia/AMD begin the port
> and tell us what must be changed to make it.
Well, having to pick up existing ideas and incarntions of those for
Linux is always a pain in the ass, but necessary at the moment. The
"experts" neglected long time the need for keeping FBSD on par with KMS
stuff or all the other development done in that area. Hope "we"
(FreeBSD) will get such a thing on a high performance base soon in
FreeBSD. Or things change again and there is a real platform-independent
standard, not necessaryly bound to Linux (which is not very realistic).
>> I still have hope that the important part HPC will become again an
>> option using FreeBSD as it was 10 years ago.
> The part of my hpc app in FreeBSD is the scheduler and data collector.
> The "real" hpc number crunching part runs on Linux with cuda code.
The part of our numerical stuff is now completely running on Linux.
FreeBSD is just my "hobby" in the lab and the backbone for my purposes
since Linux outperforms FreeBSD in nearly every aspect. And OpenCL (not
CUDA) is the backend of our numerical modelling. It is amazing watching
the Linux boxes calculation solutions on GPUs or start doing rendering
DTMs in blender with the help of a present graphics card via CUDA or
OpenCL (try blender 2.62, you will be amazed how fast a nVidia GTX570 or
even a smaller GTX560Ti is rendering a scene of a planetary surface
compared to the raw processing performance of even a modern 6-core Core
i7 Sandy Bridge-E).
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 488 bytes
Desc: OpenPGP digital signature
Url : http://lists.freebsd.org/pipermail/freebsd-performance/attachments/20120309/43c44cc9/signature.pgp
More information about the freebsd-performance