Typical timing closure strategies involve running many implementation
strategies and picking the best one to take to the lab. ML strategies are an alternative
to this, requiring you to run only three strategies to achieve a similar QoR benefit.
They use machine learning to examine features from a post-route design to predict the
performance of different strategies on the same design. The best three strategies are
captured in RQS files generated by report_qor_suggestions
(and write_qor_suggestions
) and can be applied moving forward. Considerably
less server power is required as a result.
When the directive is set to RQS on implementation commands, the command references the RQS file for the directive and other tool command options. The flow is shown in the following figure:
There are four key points to this flow:
- The
report_qor_suggestions
command must be run on a fully routed design that is generated using eitherDefault
orExplore
directives. For complete details about the requirements, see ML Strategy Availability. - The
write_qor_suggestions -strategy_dir <dir>
command generates the required RQS files in the directory specified. By default, three strategies are generated. For each strategy generated, a single RQS file contains all the suggestion objects as well as the strategy suggestions object. The RQS file specified usingwrite_qor_suggestions -file <fn>.rqs
can be discarded as the information is replicated in each strategy RQS file.Note: To generate more strategies, increase the number using the following command:report_qor_suggestions -max_strategies <n>
- The generated RQS file must be read in to the new implementation run.
- The directive RQS must be set and the script must contain a call to
opt_design
,place_design
,phys_opt_design
androute_design
.