Prof. Bora Nikolic
TAs: Daniel Grubb, Nayiri Krzysztofowicz, Zhaokai Liu
Department of Electrical Engineering and Computer Science
College of Engineering, University of California, Berkeley
Like last week, this lab has two parts. For the first part, we will continue to develop our GCD coprocessor by improving its performance. After that, we will continue the physical design flow by performing routing.
To begin this lab, get the project files and set up your environment by typing the following commands:
git clone /home/ff/eecs151/fa21/sky130/lab5-sky130.git
cd lab5One way we can improve the performance of our GCD coprocessor is by parallelizing the compute. We can do this by including multiple GCD units in our design, and routing traffic to them as they become available.
You will find that the solution to last week’s lab (fifo.v and gcd_coprocessor.v) is included. The
test has been modified to check the total number of cycles taken by the coprocessor to complete the
tests. Run make sim-rtl to run the new testbench on the solution code. Take note of the number
of cycles that the tests take without modification, as you will need it to calculate your speedup.
Your task is to edit gcd_coprocessor.v to improve the performance below 225 cycles. We will do
this by using two instances of GCD.
You will find RTL that connects the datapath and controller into one module in gcd_unit.v. You
may find this useful when refactoring the gcd_coprocessor, since you will need fewer wires to place
both GCD instances.
You will also find stub code for an arbiter, which you should complete. We will use the arbiter to route traffic to GCD units and preserve the response ordering. Most of your design can be implemented with combinational logic, but you will need some state to remember which GCD block contains the earliest data to preserve ordering.
a.) Submit your code (gcd_coprocessor.v and gcd_arbiter.v) with your lab assignment.
b.) How many cycles did your simulation take? What was the % speedup?
In the last lab, we only focused on the PAR flow through CTS. In this lab, we will go through the full flow. Routing is the next major flow step. Prior to the actual routing step, Innovus uses a basic routing engine with errors and shorts, but ignores these errors and simply tries to get an estimate of delays and parasitics. Once post-CTS optimization is done, it switches to a different tool that actually legalizes routing and tries to eliminate shorts while meeting timing. Routing is one of the most computationally heavy tasks of digital IC design and can take days to complete for complicated designs. This will be reflected in the runtime in this lab.
After routing is complete, a post-Route optimization is run to ensure no timing violations remain. Post-Route optimization typically has little freedom to move cells around, and it tries to meet the timing constraints mostly by tweaking the length of the routings. You may see some DRC (Design Rule Check) errors caused by the 7nm technology library, after routing.
First, synthesize the design:
make synThen, simulate the synthesized design to make sure it still works:
make sim-gl-synOnce your synthesized design passes the test, you can start the PAR flow:
make parThe PAR command will take a long time to complete, as it runs through all stages of PAR.
Check out the iterations that Innovus runs through during optimization. You can see some of the metrics that Innovus is using.
Once it completes, take a look at the build directory as in the previous labs. You might see additional files
compare to the syn-rundir, and that’s because the PAR flow incorporates the RC and parasitic delays, in addition to the cell delays. Open build/par-rundir/gcd_coprocessor.setup.par.spef
and search for the first occurrence of D_NET. What does it say about the first net? You may find
this wiki page helpful. (thought experiment #1 : get a sense of the units at the top and orders of magnitude of the RC parasitics in the SPEF file. If we used a 5nm technology library, do you expect the resistance to generally increase or decrease? How about the capacitance?)
a.) Check the post-Synthesis timing report
(syn-rundir/reports/final_time_ss_100C_1v60.setup_view.rpt) and post-PAR timing report (par-rundir/timingReports/gcd_coprocessor_postRoute_all.tarpt).
What are the critical paths of your post-PAR and post-Synthesis designs?
Are they the same path?
How does this critical path compare to your single-unit critical path?
b.) Iterate on your design by modifying design.yml to find a rough estimate (no need to be too
precise) for the clock period until you start running into setup errors.
Given the number of cycles it takes to complete the testbench, what is the shortest time your design can finish the computation?
c.) Open the post-CTS timing report(par-rundir/hammer_cts_debug/hammer_cts_all.tarpt) and the post-PAR
timing report(par-rundir/timingReports/gcd_coprocessor_postRoute_all.tarpt).
Find a common path (same start and end sequential elements). What differences do you notice within the paths?
As in the previous lab, we will look at the contents of par.tcl that Hammer generates and follow
along using Innovus.
(thought experiment #2 : open the par.tcl and search for the command set_db add_fillers_cells. Based on the names of the cells specified by this command, what do you think is the function of the filler cells?)
Navigate to the directory build/par-rundir and type:
innovus -common_uiThis will open the Innovus shell. Next, type read_db gcd_coprocessor_FINAL to load the current design
database from the latest PAR flow. This will help us to avoid re-running the entire flow. To see
all the reporting commands, type help report* in the Innovus shell and read through the options
available to you.
a.) What is the area consumed by your design? What percentage of the total area does the arbiter occupy?
b.) Submit a screenshot of your setup slack histogram. Compared with the histogram you obtained in Lab 4, does your new slack distribution support the observed performance improvements you obtained in your coprocessor?
After you are done with the flow, it is time to simulate our newly printed post-PAR netlist. Type the following command:
make sim-gl-parThis will use the same testbench, but will now use the post-PAR netlist of your design, backannotated with delays and parasitics from PAR. Make sure to adjust the CLOCK_PERIOD variable in sim-gl-par.yml to match the clock period you obtained from PAR. Note, however, that the exact
clock period may not work and you may need to relax it slightly.
After running make sim-gl-par you can run power analysis using:
make power-parNavigate to power-rundir/activePowerReports and open ss_100C_1v60.setup_view.rpt. Do
the power estimation numbers match your expectation?
a.) Re-run the flow using your old design.
To prevent your build directory from being overwritten, set the OBJ_DIR Make variable to a different name (i.e. make par OBJ_DIR=build2).
Using the area and power values from Innovus,
how does the performance improvement from the dual-unit design compare to area occupation and power consumption increase compared to your old design?
b.) Modify your gcd_coprocessor.v to take an input parameter in terms of number of clock cycles we
want our design to meet (parameter TARGET_NUMBER_OF_CYCLES) for this given testbench. Your
code should generate a low area, low power design if the number is greater than that your simple
gcd coprocessor can achieve, and it should generate the dual-unit design if it is lower.
Submit your code.
c.) (Optional) Using a rough estimate of target number of cycles versus number of units in the design, write a code that will generate 1-8 cores depending on the performance demand. Do NOT do this by writing out every possible case explicitly. You can limit the number of units to powers of two (1,2,4,8) if it makes your life easier.
- Submit a written report with all 4 questions answered to Gradescope
- Checkoff with an ASIC lab TA
This lab is the result of the work of many EECS151/251 GSIs over the years including:
- Nathan Narevsky (2014, 2017)
- Brian Zimmer (2014)
- Cem Yalcin (2019)
Modified By:
- John Wright (2015,2016)
- Ali Moin (2018)
- Arya Reais-Parsi (2019)
- Cem Yalcin (2019)
- Tan Nguyen (2020)
- Harrison Liew (2020)
- Sean Huang (2021)
- Daniel Grubb, Nayiri Krzysztofowicz, Zhaokai Liu (2021)