Skip to content
athralkPublic

About

custom ASIC development tools. Starting off with technology mapper and static timing analysis.

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

7 Commits

Folders and files

Repository files navigation

FastASIC

FastASIC is a Verilog compiler and ASIC back-end toolkit, written from scratch in C++ with no dependencies. It turns Verilog into a netlist of real standard cells for IHP's open 130 nm process (SG13CMOS5L, the one Tiny Tapeout uses) and checks the timing of that netlist. It does the same jobs as Yosys and OpenSTA (the timing engine inside OpenROAD and OpenLane), and it is much faster than both.

Every netlist it produces is mathematically proven to behave like your Verilog before you see it.

What it does

  • Synthesis: Verilog in, a netlist of sg13cmos5l cells out, optimized for area or speed. Large memories become IHP SRAM macros automatically.
  • Timing: lsta reports setup, hold, worst paths and maximum clock speed for any gate-level netlist, including ones from Yosys or OpenLane. Its numbers match OpenSTA's.
  • Verification: a built-in SAT solver proves that two designs are equivalent, or shows an input where they differ.

How each step works is explained in src/.

Setup

You need a C++17 compiler and make. On Ubuntu:

sudo apt install g++ make

Build:

make

This creates two programs: FastASIC (synthesis and verification) and lsta (timing).

The IHP PDK is needed for anything that uses real cells. FastASIC looks for it in ~/ttsetup/pdk, the folder Tiny Tapeout's setup creates. If yours is somewhere else, set PDK_ROOT:

export PDK_ROOT=/path/to/pdk

The PDK must contain ihp-sg13cmos5l. The version used and tested here is IHP-Open-PDK commit 2bbec755dc67ca3db0261c3d6163e15735d66710.

Optional tools. None are needed to use FastASIC, only to run the comparisons and the full test suite:

  • yosys (with yosys-abc) for the synthesis and equivalence benchmarks
  • iverilog for the simulation checks
  • docker with the ghcr.io/efabless/openlane2:2.3.10 image for the OpenSTA comparison

Run all tests:

make test

Using it

Synthesize a design onto sg13cmos5l cells, optimized for area, and write the netlist:

./FastASIC -q --map area -o cells.v design.v

Use --map delay to optimize for speed instead. Add --sta --delay-model nldm --sdc constraints.sdc to check the timing straight away.

Check the timing of any gate-level netlist:

./lsta -sdc constraints.sdc netlist.v

Check whether two designs behave the same:

./FastASIC --equiv first.v second.v

Run ./FastASIC or ./lsta with no arguments to see every option.

Benchmarks

All numbers below were measured on the same machine with the same inputs. Area is in square micrometres, delay in picoseconds.

Synthesis speed against Yosys + ABC

FastASIC and Yosys + ABC each turn the same Verilog into sg13cmos5l cells. Times are the full run, best of three.

Design FastASIC Yosys + ABC FastASIC is
full adder 35 ms 530 ms 15× faster
4-bit ALU 56 ms 986 ms 18× faster
12-bit comparator 45 ms 972 ms 22× faster
64-bit adder 92 ms 1,599 ms 17× faster
32-bit barrel shifter 56 ms 1,171 ms 21× faster
16 × 16 multiplier 934 ms 3,348 ms 3.6× faster
all 11 designs 3.8 s 20.4 s 5.3× faster

Multipliers gain less because FastASIC also proves the result correct, which Yosys does not do.

Run it yourself: tools/bench_synth_speed.sh

Result quality against Yosys + ABC

Smaller area and smaller delay are better.

Design FastASIC Yosys + ABC Result
16 × 16 multiplier, area 12,126 14,570 FastASIC 17% smaller
16 × 16 multiply-add, area 13,709 15,951 FastASIC 14% smaller
32-bit barrel shifter, area 2,578 2,703 FastASIC 5% smaller
32-bit adder, delay 2,450 ps 4,767 ps FastASIC 49% faster
16 × 16 multiplier, delay 4,724 ps 5,942 ps FastASIC 20% faster
small control logic, area 6,777 (total of 9) 6,310 Yosys + ABC 7% smaller

Area numbers come from --map area, delay numbers from --map delay, so they are two different netlists.

Run it yourself: tools/compare_synth.sh and tools/compare_synth.sh bench/arith/*.v

Timing against OpenROAD's OpenSTA

lsta and OpenSTA read the same netlist, cell library and constraints, and report the slack at every endpoint.

Check Result
Accuracy Same answer as OpenSTA at every endpoint, within 0.007 ps, on 6 designs at 3 corners, with ideal and real clock trees and with SRAM macros
52,000-cell design, timing analysis 150 ms for lsta, 1,944 ms for OpenSTA: 13× faster
52,000-cell design, whole run 0.5 s and 92 MB for lsta, 2.6 s for OpenSTA

Run it yourself: tools/compare_opensta.sh bench/sta/netlists/mac32.v bench/sta/default.sdc typ

Equivalence checking against Yosys

Check FastASIC Yosys sat
Two ways of adding three 32-bit numbers 195 ms 14.4 s
8 × 8 multiplier against a shift-and-add multiplier proven in 6.2 s gave up after 120 s

Run it yourself: tools/bench_equiv.sh

What FastASIC does not do yet

It stops at the gate-level netlist. Placement, routing and layout are still done by OpenROAD, which reads FastASIC's netlists directly. Power analysis is not implemented yet.

Licenses

FastASIC is licensed under the Apache License 2.0. You may use, modify and distribute it, including in commercial and closed-source products, as long as you keep the copyright and license notices (see NOTICE). The license also grants a patent license from every contributor.

Third-party files and tools:

  • IHP-Open-PDK (cell libraries and SRAM macros): Apache License 2.0. FastASIC reads the PDK but does not include it.
  • Yosys (ISC license), OpenSTA (GPL-3.0) and iverilog (GPL-2.0) are only run as separate programs for comparisons and tests. None of their code is part of FastASIC.

About

custom ASIC development tools. Starting off with technology mapper and static timing analysis.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages