Addition of profiling function for individual kernels

[mattatmet]

Are there any linked Issues or Pull Requests?

MetOffice/lfric_apps#419

Brief description

An additional function can be included in psyclone_tools.py for the profiling of individual kernels. Separating the current implementation which only profiles kernels that have been coloured and tiled, this feature can be expanded to profile all kernels.

Further details of the issue.

The current implementation is restricted to profiling kernels only when tiling is turned on and only for kernels that are tiled. See MetOffice/lfric_apps#107 for more discussion on tiling. As a stand-alone function this would enable more fine grained profiling, separate from tiling. Additionally, this allows the assessment of the impact of tiling on kernels that are not tiled but that may still benefit from inheriting a more beneficial cache state.

[jennyhickson]

Hi Matthew - the review tracker project is just for pull requests :) Thanks.

[mattatmet]

Hi Matthew - the review tracker project is just for pull requests :) Thanks.

Oops, sorry Jenny, thanks for letting me know! :)

[hiker]

While I am all in favour of adding this kind of functionality to the transformations (since the transformations will know best what exactly should be measured), and I don't have the full background what you want to do ... but I just thought to point out that psyclone has an option --profile kernels (and --profile invokes), which automatically adds profiling around all kernels/invokes (in dsl mode only). There is also --profile routine for dsl and transmutation (which is obviously less granular).

[mattatmet]

psyclone has an option --profile kernels (and --profile invokes), which automatically adds profiling around all kernels/invokes (in dsl mode only).

Thanks for highlighting this Joerg (Joerg Henrichs (@hiker)). The implementation I have in mind is slightly more flexible than the -- profile kernels CMD line option allows. I have used the ProfileTrans function which should provide the user with more control over which kernels are profiled, as profiling all of them in a particular application may lead to unnecessary overhead or simply be too much information. This is achieved in two ways. The first being identifying types of loops around kernels (such as colouring or tiling loops) and even types of kernels (such as matrix-vector kernels, although this may be future work). The second being the ability to add it to specific optimisation/transformation scripts to allow the user to target specific algorithms without having to define another psyclone.sh in lfric_apps. It also allows greater control over calliper names.

I probably should have put more of that information in the issue, so thank you for asking about it. Happy to discuss this further if you have other comments or any suggestions. Cheers, Matt

[hiker]

Matthew Walker (@mattatmet) , thanks for the long explanation. I just wasn't sure if you were aware of the command line options (since I felt they might allow you to do a zero-effort options for a comparison), but your approach totally makes sense. If there is anything PSyclone can do for you let me know - I am interested in tools and profiling (and did all the profiling API design and implementation in PSyclone). I am actually quite happy that this is used ;)

[mattatmet]

Joerg Henrichs (@hiker), thank you, I really appreciate you pointing it out as I'm still learning to use Psyclone and LFRic, so very happy to receive any and all advice/tips. I'll keep you updated on how this goes, and let you know if I need help with any Psyclone stuff. Thanks again!