Bottom-up implementation? (divide and conquer)

[alexisfrjp]

Hello!

I have an OOC module which is hard to meet timing. I assigned it to a pblock IS_SOFT=false pblock and I finally met timing.
Now, I want to keep its placement forever to avoid having to deal with it ever again.

1. I want to keep its placement.
2. I want to replicate it vertically (pblock is a clock region)

The DFX is for Top-down implementation and I'm stuck, I can't find how to keep an OOC module. I've tried lock_design and exporting the xdc but it doesn't lock anything (bug?).

Can RapidWright help me for the bottom-up/divide and conquer flow?
If yes, how? any tutorial or blog post or anything showing it?

https://www.rapidwright.io/docs/Merge_Designs.html
The first paragraph seems interesting but doesn't show how to do it.

Thanks!

[eddieh-xlnx]

Why, hello there @alexisfrjp. What you're looking for is what we tend to call "pre-implemented modules". Take a look here:

• https://www.rapidwright.io/docs/PreImplemented_Modules_Part_I.html (which deals with the creation of such a module, which it looks like you've done already)
• https://www.rapidwright.io/docs/PreImplemented_Modules_Part_II.html (this talks about how to relocate/replicate)

Let us know if that hits the spot or not!

[alexisfrjp]

Implementation finished, no error, critical warnings might just have been spurious warnings.
I still have a lot of failing paths in the critical module because the routing wasn't preserved.

Edit:
OOC seems to be no-go for what I want. And even though HD is deprecated for Ultrascale+, it seems mandatory.
#369 (comment)

I'll give another try with it.

[eddieh-xlnx]

Thanks for trying this out! I think I'm able to reproduce what you're seeing. Leave this with me to see if I can get some other way of achieving this.

In the meantime, can you explain why:

OOC seems to be no-go for what I want

please?

[alexisfrjp]

Ok it is more complicated than it seemed.

To keep the routing, HD.RECONFIGURABLE=1 has to be set, but then it complains for a pblock (loose, used as a hint only) that contains partition and non-partition cells...

[DRC HDPR-11] A Partition Pblock cannot have non-partition instances: The Pblock 'pCRITHINT' has a mixture of HD.RECONFIGURABLE partition instances and non-HD.RECONFIGURABLE instances. The Pblock cannot have both the HD.RECONFIGURABLE partition 'critical', and non-HD.RECONFIGURABLE instance 'noncritical'.

HD.RECONFIGURABLE seems to automatically initiate the DFX feature (https://docs.amd.com/r/2021.1-English/ug909-vivado-partial-reconfiguration/Define-a-Module-as-Reconfigurable), which forces me to exclude "static"/other logic from the pblock, which I don't want to. My goal isn't dynamic reconfiguration.

If we solve that issue, it'd open a new faster and reliable way to get larger designs closing timing.

Regarding OOC, it was just a quick comment based on clavin-xlnx (#369 (comment)):

If this succeeds, then you should be able to write out the cell with write_checkpoint -cell. If it fails, you'll want to change the DCP from out-of-context to in-context.

Thanks again!

[clavin-xlnx]

@alexisfrjp thanks for working through the various issues here. We think we might have a solution for you if you don't mind trying #1215. Give it a shot and let us know how it goes.

[alexisfrjp]

Thanks so much @eddieh-xlnx and @clavin-xlnx !

I'll try as soon as possible and I'll keep you updated!