Can't write out large rr_graphs in capnp format

[duck2]

direct_rf from titan_quick_qor with default settings creates a rr_graph with ~236M edges. Trying to write it out will result in a crash with the message failed: total size of struct list is larger than max segment size.

Expected Behaviour

If the rr_graph fits in memory, we should be able to write it out to the disk.

Current Behaviour

We can't

Possible Solution

Apparently Cap'n Proto has a hard limit [1] [2] for words in a single segment. A similar limit applies to lists of structs. From [2]:

D (29 bits) = Size of the list:
when C <> 7: Number of elements in the list.
when C = 7: Number of words in the list, not counting the tag word
(see below).
[...]
A struct list must always be written using C = 7.

In my understanding, this means the size of a list of structs is limited to 8 bytes * 2^29 = 2 GiB. When the list of edges is bigger than that capnp just refuses to serialize it.

A serialized rr_graph edge is 3 uint32s plus a metadata list (which I suppose is a "list tag" of 64 bits) so it's 20 bytes. 236M edges will require ~4.7 GB.

Any ideas on how to go about this? I can't think of a way to solve it without manually tweaking the schema, which somewhat ruins the purpose of having a auto-generated one.

(Maybe this can be an excuse to finally drop uxsdcxx and provide a standalone converter for XML rr graph owners, but I need to progress on the parallel router first...)

Steps to Reproduce

1. Run titan_quick_qor
2. Go to run directory for directrf_stratixiv_arch_timing.blif
3. Get run command from vpr.out, append --write_rr_graph rr_graph.bin and execute it

[duck2]

@amin1377 @kimiatkh Any clue about this?

[amin1377]

@amin1377 @kimiatkh Any clue about this?

I have never used capnp format to serialize the RR graph. So far, I have only used XML to serialize it, and it has been faster than building the graph from scratch (at least for my test cases which usually involve big circuits)

[github-actions]

This issue has been inactive for a year and has been marked as stale. It will be closed in 15 days if it continues to be stale. If you believe this is still an issue, please add a comment.

[github-actions]

This issue has been marked stale for 15 days and has been automatically closed.