- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi, MKL team.
We use MKL in multithreaded application. In each our thread we have heavy calculations: otherwise we wouldn't use such a giant like MKL. That is why we need to know where is MKL routine at the moment: is it almost done and we can wait or it is just started and we can't wait util the end so we will shut it down.
From documentation I see that mkl_progress is intended to report progress and finish MKL routine if needed. And I was happily jumping when saw it at the first time. But turned out that it does not provide any information about the context it was invoked from. This is strange (to say the least), because to make result decision on whether we want to finish MKL job or not, usually we need much more information than just thread id (which I can know via the the system/OpenMP call by the way) and stage name (which is not of a big use to).
1. One thing to make user happy with calling callback function (see MPI, c/c++/python libraries and majority of other libraries that provide callbacks) is to let him pass his own information into callback when register.
2. Another thing is to let him register different mkl_progress functions for different MKL routines.
So the question follows.
Are above arguments convincing enough and is it possible that MKL team will upgrade callback workflow with something more actual?
Thank you in advance,
Victor.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Victor,
Thanks you much for the feedback. I will escalate them to MKL developer team.
Thanks
Ying
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page