Open GoogleCodeExporter opened 8 years ago
sorry its not a type defect
Original comment by wawatorr...@gmail.com
on 24 Jul 2011 at 11:35
I know there is ATI. The reason for missing ATI support is not because of the
NV fellowship, which is later than our prototype. It is because:
- I am still not familiar with OpenCL. Specifically, I need equivalences of
CUDA streams(concurrent multiple kernel execution, stream synchronization and
stream state query) and page-locked memory.
- I don't have an ATI card.
In fact, I highly encourage anyone who's interested in providing ATI, or more
generally, OpenCL support of KGPU to fork the github repository and do it! The
one doesn't need to be a kernel hacker. The code involved are all in user
space. They are:
- under kgpu/, gputils.h and gpuops.cu, just replace the CUDA terms by their
OpenCL equivalences -- if they exist.
- under services/*/libsrv*/, any *.cu files, simply change CUDA keywords to
OpenCL ones.
- one more thing is the Makefile to compile cu files: need OpenCL cc to replace
nvcc.
And finally, the one who will make this doesn't need to understand KGPU at
all...
Original comment by EndySun@gmail.com
on 26 Jul 2011 at 5:26
a great project !! - i'd luv to see it maintained in the kernel :-)
Original comment by christop...@gmail.com
on 16 Nov 2013 at 11:59
+1
KGPU is such a great idea, but it being specific to NVIDIA and CUDA 2.0 is
obviously a fairly significant restriction.
As was previously said in comment #2, it would absolutely make sense to
generalize, and extend on, the whole idea of kernel-enhancement via GPGPU and
base everything on OpenCL so that this becomes platform/vendor-agnostic.
If this is something that cannot be done by the developers currently involved
in the project, please just start a kickstarter/crowdfunding project and just
post a follow-up here, I definitely would love to donate to such an effort, and
I am sure that others would help, too. It would be mainly a matter of spreading
the world via various blogs, ycombinator and digg/slashdot to get this started.
There are literally hundreds of open source enthusiasts out there familiar with
OpenCL (especially in the bitcoin/litecoin communitities!) and there's a ton of
CL code that could be borrowed/learnt from for this effort.
Particularly, in the context of disk encryption, GPGPU could be a huge asset -
and AES/blowfish implementations in OpenCL are readily available for these
purposes, eCryptFS with OpenCL support via KGPU would be a major improvement -
likewise, projects like OpenSSL could greatly benefit from OpenCL support.
So there are very real opportunities for KGPU-based approaches, but obviously
OpenCL needs to be used instead of CUDA, so that more hardware/vendors can get
involved.
Original comment by wonco...@googlemail.com
on 15 Aug 2014 at 9:04
Original issue reported on code.google.com by
wawatorr...@gmail.com
on 24 Jul 2011 at 11:34