Skip to content

Benchmarking Enhancements #546

Open
Open
@rfungquicklogic

Description

@rfungquicklogic

I have a couple of suggestions for VTR benchmarking:
(1) Change vtr_reg_qor_chain_depop to have a 2-sided pin pattern, rather than spread (making it more different than vtr_reg_qor_chain). For example:

clb.I1 clb.I2 clb.I3 clb.I4 clb.O clb.clk clb.cin clb.cout
clb.I1 clb.I2 clb.I3 clb.I4 clb.O clb.clk clb.cin clb.cout

(2) Update the Titan benchmarks to be more realistic by depopulating input crossbars, perhaps using a more restrictive pattern than ‘spread’, and disabling output rotation for carry chains (would require code and parameter changes). Output rotation is also wrong when each BLE has multiple outputs (like in Titan) as the outputs can be routed through different BLEs -- implying breaking internal BLE connections; fixing this might be harder though.

Context

The aim is to increase VTR testing diversity for regular check-ins and to make the Titan benchmarks (which are run less frequently) more realistic.
There might be other upgrades that would help achieve these ends that are beyond the scope of what is suggested here.

Metadata

Metadata

Labels

no-staleExclude issue from being automatically marked as stale

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions