Created attachment 175576 [details]
svn diff for math/clfft
Please find attached the trivial update to update math/clfft to 2.12.2. Release highlights:
* now works with pocl OpenCL runtime
* validation of sizes (powers of 2,3,5 and combinations thereof) to work up to huge data sets (up to 25GB)
Passes portlint. Tested on 10.1, 10.2, 10.3, and 12-CURRENT amd64 w/ poudriere.
@Dieterich, please set maintainer-approval to + on attachments for ports you are maintainer of. Attachment -> Details -> maintainer-approval [+]
A commit references this bug:
Date: Mon Oct 10 13:03:24 UTC 2016
New revision: 423661
- Update to 2.12.2
Submitted by: firstname.lastname@example.org(maintainer)
I found clfft build well on FreeBSD-9.3-amd64, but I am not sure it run well or not.
Dear Wen, thanks for committing and checking on 9.3. Indeed, it should compile there, however, with OpenCL it is a bit more complex than that. For initial compilation, one only uses libOpenCL and we have the stub library in ports for this. For actual at-runtime compilation, one needs an OpenCL runtime loaded by the stub.
Currently, we have
* freeocl (broken, doesn't work with libOpenCL loading mechanism, doesn't compile w/ GCC 4.9)
* clover (for AMD GPUs, requiring a late KMS)
* beignet (for Intel GPUs, also requiring a late KMS)
So on 9.x, one cannot execute anything. Now, there is a patch here for pocl, a CPU runtime that actually works. The submitter has tried to get it into the tree for a long time, it'd be a nice addition as it would actually allow us to run OpenCL code on all releases but, alas, no commit.