3D NoC

[soheilshahrouz]

Description

Generalizing the NoC data structures, architecture description and routing algorithms to 3D (multi-layer) FPGAs.

Motivation and Context

VTR is mostly 3D now, but the NoCs are still only 2D.

How Has This Been Tested?

Created a new arch file and ran synthetic benchmarks on this new 3D NoC arch within the 3D Stratix IV arch. The 10x10 2D mesh is replaced by two 7x7 meshes stacked on top of each other. All the benchmarks passed, and the QoR metrics are slightly better as the 3D NoC topology shortens some traffic flows.

CI has been updated to run some tests on this 3D architecture.

Types of changes

• New feature (change which adds functionality)

Checklist:

• My change requires a change to the documentation
• I have updated the documentation accordingly
• I have added tests to cover my changes
• All new and existing tests passed

[vaughnbetz]

Some CI test failures due to vtr::isclose getting one float and one double; you'll have to explicitly cast or change a constant or some such (@soheilshahrouz tells me he has float constants and double variables so that is probably it).

[vaughnbetz]

Until CI is sorted please run the major nightly tests on wintermute and list the results here.

[vaughnbetz]

Looks good. A few changes. Also:

• Check UP/DOWN are consistently changed to NORTH/SOUTH
• We should have at least one test for each routing algorithm (can be separate PR).

[vaughnbetz]

re-launched CI. I also suggest one QoR test (Titan or VTR, could be just a circuit or two if you wish) to confirm no unexpected slowdown, and a test on the NoC designs to confirm the same.
Other than that I think this is ready to merge -- is that right?

[vaughnbetz]

@soheilshahrouz tells me there is a strange behaviour: the NoC centroid move is working, but when enabled the initial temperature is very low (3 orders of magnitude lower). This did not occur in the earlier, FPL 2024 code. Only affects NoC designs. The QoR is very bad due to this lower temperature, so we aren't just finding a very good solution quickly.
It's not clear when the regression occurred; could have been when the 3D NoC support was added or could have been earlier.

Action after this is figured out: turn this move type on by default, or have a CI test for it.

[vaughnbetz]

Almost ready for merge ... NoC centroid move problem is resolved (was a memory bug).
Please get QoR results (2D and 3D) and we can merge this.

[soheilshahrouz]

QoR results on a subset of synthetic benchmarks on a 2D noc to measure if the introduction of 3D NoC support has slowed down NoC optimization in 2D architectures

Metric	Master	3D	Ratio
Pack Time	333.1751111	339.1172979	1.018
Place WL	550050.0964	543046.6124	0.987
Place Time	372.4883381	365.6559439	0.981
Total Swap	4183327.301	4189977.075	1.002
Place CPD	7.164124389	7.170530009	1.001
Agg BW	1.22E+07	1.22E+07	0.997
Agg Lat	2.94E-07	2.93E-07	0.999
Latency Overrun	0.00000000024	0.00000000024	1
Congestion	0.016	0.016	1
Congested Links	0.16	0.16	1

[vaughnbetz]

3D seems to have slightly improved things :). Merging!