PHORONIX: OpenCL, GLSL Back-End For LLVM May Soon Open Up
ohartman at zedat.fu-berlin.de
Mon Aug 29 13:31:58 UTC 2011
On 08/29/11 14:31, Yamagi Burmeister wrote:
> On Mon, 29 Aug 2011 20:12:54 +0800
> Adrian Chadd<adrian at freebsd.org> wrote:
>> http://forums.nvidia.com/index.php?showtopic=38242 Post 18
>> This indicates the driver supports CUDA somehow. What's missing is a
>> FreeBSD runtime.
>> Can someone please do some legwork with this and see if it's possible
>> to bring the Linux CUDA SDK up in the linuxulator?
> Yes that's possible. jhb@ wrote a little tutorial some time ago:
With beginnig of 2011 I tried everything possible to make CUDA SDK
running on FreeBSD 9-CUR/amd64.
For those interested in this subject, the mentioned forae entries and
that mentioned tutorial are
a bit outdated.
I tried the tutorial mentioned by Yamagi, but I had never real success
installing SDK 3.2/32bit. For serious
scientific usage 32bit isn't a great deal, my model code highly depends
on 64 bit in the portions were the
model is calculated on the cpu, not the GPU. Anyway, having CUDA working
in the Linuxulator would be
But having something OpenCL-ish natively would be a real quantum leap
... And 64bit.
My lectures in "building compilers" I liestend to is quite a time ago,
so as far as I know, nVidias
driver BLOB is executing some binary code that is only applicable for
the GPU. Further, the OpenCL
code or even CUDA code of a C/C++ software is runtime compiled when
accessed and transfered to the
GPU/driver, so the nvcc compiler is a "sine conditio qua non" in this
game, so far.
My thinking: eliminating the need of the nvcc producing PTX code/binary
code via LLVM could be a solution. But at this point
I do not know whether PTX has to translated again in binary code via
nvcc to make the GPU understand
what I want.
By the beginning of this year Pathscale introduced a kind of compiler
capable of HMPP, a very smart model
like OpenMP. This compiler seems not to be ready by now and I never got
access to a beta version to test
whether it was capable of compiling CUDA ready code. As far as I know,
HMPP is also an "open" standard.
Well, conclusively, there seems no real solution to be present for
FreeBSD by now (as I mentioned, the
Linuxulator is no real alternative due to its 32bit limitations).
More information about the freebsd-performance