Skip to content

Clock period improvements left on the table due to SYNTH_MINIMUM_KEEP_SIZE counting macros #4714

Open

Description

@oharboe

Description

SYNTH_MINIMUM_KEEP_SIZE decides which modules hierarchical synthesis keeps. The expectation is that it measures the logic yosys works on: since @povik's #2552 the decision is keep_hierarchy -min_cost on yosys's cost estimate of each module, in NAND2 equivalents.

A macro is no work for yosys, but it counts. keep_hierarchy -min_cost needs a cost for every instanced blackbox, and convert_liberty_areas gives each Liberty cell, SRAM macros included, the gate equivalent of its area; since #4650 an AUTO_MEMORIES macro gets its generated view's area the same way. A small wrapper around macros therefore looks like a large module of logic and is kept, and its boundary stays wherever a path runs through it.

Example, asap7/riscv32i (SYNTH_MINIMUM_KEEP_SIZE=10000): the only module kept is dmem, four fakeram7_256x32 instances (1,376 µm² of Liberty area) and 56 µm² of address decode. Its boundary sits on the worst path, between the ALU's address adder and the SRAMs' enable decode, which synthesis then optimises apart. Through global route at 780 ps:

min reg2reg period reg2reg endpoints within 10 % of it: faster / slower
today: dmem kept 835.6 ps –
dmem flattened, which is what the rule decides with macro cost 0 791.8 ps 995 / 3

Details and the measurement in #4713 (closed, FYI only).

Suggested Solution

Make the rule measure what it is for: a blackbox macro, Liberty or AUTO_MEMORIES, costs 0 (or 1) gate equivalents towards its parent's size. keep_hierarchy -min_cost still has a cost for every blackbox, so the error that made a cost necessary does not come back.

The designs this can change are those that synthesise hierarchically with a keep threshold and have macros (make print-* on master; the gf12 designs not checked): asap7/coralnpu, asap7/riscv32i; nangate45/ariane133, ariane136, black_parrot, bp_be_top, bp_fe_top, bp_multi_top, bp_quad, mempool_group, tinyRocket. Their keep decisions change only where a macro sits under a module the rule would otherwise flatten, and their QoR may move either way.

Additional Context

variables.yaml describes the threshold as an area in NAND2 units. Stating there which is intended, the logic yosys works on or area including macros, would make the behaviour less surprising either way.

Activity

  1. changed the title [-]SYNTH_MINIMUM_KEEP_SIZE counts macros, which yosys does no work on[/-] [+]Clock period improvements left on the table due to SYNTH_MINIMUM_KEEP_SIZE counting macros[/+] on Oct 21, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions