usage of QUARK in PLASMA

[dobson156]

Hi,

I am using the library version 2.4.5 and after reading the accompanying documentation (ver 2.3) I still have some questions, firstly the documentation mentions (2.1.3) that PLASMA has two scheduling modes; static and dynamic. However further in the documentation (6.1) it says that PLASMA currently only supports static scheduling, is that information out of date?

Am I right in thinking that the dynamic scheduling uses QUARK and the static version does not? Or do they both use QUARK? Reading the source code suggest that static version are not handled with quark as they make `CORE_xyyzz` calls rather than the `QUARK_CORE_xyydd` calls.

After reading the source I am unclear as to you select between static and dynamic scheduling (if dynamic is currently available), I see that there a `PLASMA_Set` switch case for scheduling but I don't know whether that the correct way to set and whether it is supported if used, this documentation makes no mention of it: http://icl.cs.utk.edu/projectsfiles/pla ... plasma_get.

The source I downloaded doesn't seem to come with a reference, is it still maintained for newer versions?
Finally the docs also say:

A pruned search method to obtain "good" parameters is described in [2]. We note that
autotuning is part of the PLASMA's roadmap; unfortunately, as of 2.0.0, the PLASMA
software does not have its autotuning component available for release.

What is that state of autotuning in version 2.4.5?

Sorry to have bombarded you with questions
Thank You

[admin]

I am using the library version 2.4.5 and after reading the accompanying documentation (ver 2.3) I still have some questions, firstly the documentation mentions (2.1.3) that PLASMA has two scheduling modes; static and dynamic. However further in the documentation (6.1) it says that PLASMA currently only supports static scheduling, is that information out of date?

The short answer is "out of date". PLASMA has two scheduling modes. Different functions can support one, or the other, or both. PLASMA switches modes at runtime, depending on what the user requests and what is supported - basically best effort to accommodate the user. Most common functions have both modes. It is more common for a function to have the dynamic mode and not have the static mode than the other way around. It is easier to use dynamic scheduling than to schedule manually. PLASMA_Set is the correct way of declaring the preferred mode.

Am I right in thinking that the dynamic scheduling uses QUARK and the static version does not? Or do they both use QUARK? Reading the source code suggest that static version are not handled with quark as they make `CORE_xyyzz` calls rather than the `QUARK_CORE_xyydd` calls.

Yes, dynamic scheduling uses QUARK and static does not. Static uses progress tables and busy waiting.

The source I downloaded doesn't seem to come with a reference, is it still maintained for newer versions?

Let me look into it. For now, look at the block of Doxygen comments on top of the function body in the sources.

What is that state of autotuning in version 2.4.5?

There is none. This is hard. Research in progress.

Sorry to have bombarded you with questions

Thank you for your questions. If there are no questions, it means that the software is irrelevant. Your questions validate the existence of the software.
Jakub

[dobson156]

Thanks for your reply, you've given me a bit to mull over.

I would be very helpful if the reference could be updated to reflect the current state of things, but I realise that takes effort.

Some thing like the QUARKS doxygen HTML files would be great.
http://icl.cs.utk.edu/quark/docs/group__QUARK.html

Thanks