Skip to content

Latest commit

 

History

History
184 lines (154 loc) · 13 KB

0066-2023-04-26.md

File metadata and controls

184 lines (154 loc) · 13 KB
Raw

26 Apr 2023

Previous journal: Next journal:
0065-2023-04-23.md 0067-2023-04-27.md

Anton's Tiny Tapeout 3 (TT03) submission

  • I successfully submitted a very simple design to Tiny Tapeout 3: anton1-tt03 (or see it on GitHub)
  • GDS can be viewed in 3D here: https://algofoogle.github.io/anton1-tt03/
  • I managed to scrape in at the last minute; 3:30am my time, right on the deadline! Waiting for GitHub Action processing was nerve-wracking.
  • I paid to receive the chip on a board, and hopefully that will be somewhere in October 2023 or maybe November... assuming Efabless gets the chip completed by the estimated mid-August date (for the 2304C shuttle).
  • The design itself is not very interesting; I had very little time.
    • I first tried an RLE encoder, but realised I needed to come up with a good spec for an RLE protocol that would not lead to a buffer overrun.
    • I then tried a fixed-point reciprocal approximator, but after multiple attempts to cut it down I found that it was generating just too many cells to fit. Maybe it would work OK if in future I dropped it from Q6.10 to Q6.6 or something.
    • Then I had a multiplier that takes two 16-bit inputs and gives a 32-bit product, but that was a bit too big.
    • With very little time left I landed on a multiplier that takes two 8-bit inputs and gives a 16-bit product.
  • Since then, I've revised the design in another branch (config-width) to support a configurable number of operand nibbles, and it fits given the range of 1..3, but not 4+.

Raybox

  • NOTE: On the matter of reciprocals: for raydir reciprocals, we'll typically only be dealing with an input range of something like 0..1.414 (for 90deg FOV?). Hence we'd only need Q1.X on input, but maybe full precision on output.
  • Think more about Rational maths.

Notes

Trying to run TT GDS builder for myself

  1. Go into MPW8 VM, then cd ~/anton
  2. Clone my repo: git clone [email protected]:algofoogle/anton1-tt03 then cd anton1-tt03
  3. In it, then clone tt-support-tools: git clone [email protected]:tinytapeout/tt-support-tools tt
  4. We already have $OPENLANE_ROOT, $PDK_ROOT and $PDK defined.
  5. Make sure we also define these others (which don't seem to otherwise be defined by the MPW8 VM) -- Note that maybe we could use whatever our default is, instead? 
    # This one might not be needed, but I'm setting it anyway:
export OPENLANE_TAG=2023.02.14

# Instead of this one, we could use the default...
export OPENLANE_IMAGE_NAME=efabless/openlane:4cd0986b3ae550cdf7a6d0fba4e0657012f635d8-amd64
# ...default is: efabless/openlane:cb59d1f84deb5cedbb5b0a3e3f3b4129a967c988-amd64
  6. python3 --version shows 3.8.10; not the same as the 3.10 required by the GHA workflow, but I'll live with what we've got for now.
  7. Install Python stuff needed by TT: pip install -r tt/requirements.txt
  8. ./tt/tt_tool.py --create-user-config
    • Outputs: 2023-04-26 06:08:07,817 - project - INFO - creating include file
    • Contents of src/user_config.tcl: 
      set ::env(DESIGN_NAME) algofoogle_product
set ::env(VERILOG_FILES) "\
    $::env(DESIGN_DIR)/product.v"
  9. ./tt/tt_tool.py --harden:
    1. Pulls the specified container image.
    2. It fails to run:
    OpenLane 4cd0986b3ae550cdf7a6d0fba4e0657012f635d8
(with mounted scripts from cb59d1f84deb5cedbb5b0a3e3f3b4129a967c988)
All rights reserved. (c) 2020-2022 Efabless Corporation and contributors.
Available under the Apache License, version 2.0. See the LICENSE file for more details.

[ERROR]: The version of netgen installed in the environment does not match the one required by the OpenLane flow scripts (installed: 28a29504390d53cd3748ff2636be112ef299da0b, expected: 2292ab813b54cc6d6fa41368c4e0e7492fb627c9)
The version of klayout installed in the environment does not match the one required by the OpenLane flow scripts (installed: 44a2aa9ca17c2b1c154f9c410ded063de9ed3e12, expected: 428d0fe8c941faece4eceebc54170cc04d916c03)
The version of openroad_app installed in the environment does not match the one required by the OpenLane flow scripts (installed: 1f720d3b442e2cd8dc6c5372535320b18a105e8d, expected: 7f00621cb612fd94e15b35790afe744c89d433a7)
[ERROR]: Please update your environment. OpenLane will now quit.
2023-04-26 06:10:28,954 - project    - ERROR    - harden failed
  10. Let's try again, but this time using our existing (known good) OpenLane image: 
    unset OPENLANE_TAG
export OPENLANE_IMAGE_NAME=efabless/openlane:cb59d1f84deb5cedbb5b0a3e3f3b4129a967c988-amd64
./tt/tt_tool.py --harden
    ...which seems to run OK.
  11. Results are in runs/wokwi/, with these warnings: 
    [WARNING]: Current core area is too small for the power grid settings chosen. The power grid will be scaled down.
[WARNING]: There are max fanout violations in the design at the typical corner. Please refer to '../work/runs/wokwi/reports/signoff/25-rcx_sta.slew.rpt'.
    ...the fanout is just an 11/10 on a clock buffer. Seems fine.
  12. Summaries:
    • ./tt/tt_tool.py --print-warnings: No output
    • ./tt/tt_tool.py --print-stats: 18.35% utilisation (just for routing?) with wire length of 13563um.
    • ./tt/tt_tool.py --print-cell-category: Hard to read because it's a Markdown table with URLs, but it appears to be the same numbers as what the official build did in GHA.
    • ./tt/tt_tool.py --create-png

Resulting PNG, viewed with xdg-open gds_render.png:

Locally-generated render of anton1-tt03 design GDS

Looks very similar to, but not exactly the same as, the version generated by GHA:

GHA-generated render of anton1-tt03 design GDS

What does the gds job do?

This is a description of what happens in the gds job of the gds.yaml GitHub Action workflow.

  1. Sets OpenLane ENV:
    var value
    OPENLANE_TAG 2023.02.14
    OPENLANE_IMAGE_NAME efabless/openlane:4cd0986b3ae550cdf7a6d0fba4e0657012f635d8-amd64
    OPENLANE_ROOT /home/runner/openlane
    PDK_ROOT /home/runner/pdk
    PDK sky130A
  2. Clones the repo (ours, I guess) using actions/checkout@v3 -- I assume this goes into a subdir with the same name as the repo? Or does it go into the current working dir?
  3. Clones the tt-support-tools repo to dir tt
  4. Installs python 3.10 and runs: pip install -r tt/requirements.txt -- See requirements.txt
  5. Runs ./tt/tt_tool.py --create-user-config -- For Wokwi projects, it will grab that data too.
  6. Clones the OpenLane repo at the specified OPENLANE_ROOT path, then builds it with make
  7. The crucial step: Builds the GDS with ./tt/tt_tool.py --harden -- See here for what this does.
    • In particular, it runs the equivalent of: 
      docker run                                  \
    --rm                                    \
    --volume $OPENLANE_ROOT:/openlane       \
    --volume $PDK_ROOT:$PDK_ROOT            \
    --volume $(pwd):/work                   \
    --env    PDK=$PDK                       \
    --env    PDK_ROOT=$PDK_ROOT             \
    --user   $(id -u $USER):$(id -g $USER)  \
    $OPENLANE_IMAGE_NAME                    \
    /bin/bash -c "./flow.tcl -overwrite -design /work/src -run_path /work/runs -tag wokwi"
    • Compare this, below,with What the MPW8 VM does to run the OpenLane docker.
    • It doesn't look like there are any options that tt_tool.py supports to change where it expects the design source to come from, or where the run results will be recorded.
    • So, the key command ./flow.tcl -overwrite -design /work/src -run_path /work/runs -tag wokwi does this:
      • Expects the design (and hence config.tcl or config.json) to be found in absolute path /work/src
        • This /work path in the container is a Docker volume mapping to the current working directory, so if we want to run this directly, we must (I think) ensure the tt-support-tools repo is within our working TT03 submission repo, e.g. in ./tt, and run it from within the base dir of our repo, hence ./tt/tt_tool.py being the convention.
      • Writes the run data into /work/runs which ends up being the ./runs/ subdirectory of the current working directory (i.e. in our repo base).
      • Names the run wokwi instead of timestamp (I think).
      • Overwrites any existing wokwi run.
  8. Shows resulting files: find runs/wokwi/
  9. Runs a bunch of other summaries: 
    ./tt/tt_tool.py --print-warnings      >> $GITHUB_STEP_SUMMARY  # Yosys warnings
./tt/tt_tool.py --print-stats         >> $GITHUB_STEP_SUMMARY  # Routing stats
./tt/tt_tool.py --print-cell-category >> $GITHUB_STEP_SUMMARY  # Cell usage
  10. "Populates caches" -- Not sure what this does. Preserves data between jobs, maybe, or keeps it for final output?
  11. Runs ./tt/tt_tool.py --create-png to create a PNG of the GDS, then caches this too.

What the MPW8 VM does to run the OpenLane docker

If you do cd $OPENLANE_ROOT; make mount, it does the equivalent of:

docker run \
    --rm                                                                       \
    --volume  /home/zerotoasic/asic_tools/openlane:/openlane                   \
    --volume  /home/zerotoasic/asic_tools/openlane/designs:/openlane/install   \
    --volume  /home/zerotoasic/asic_tools/pdk:/home/zerotoasic/asic_tools/pdk  \
    --volume  /tmp/.X11-unix:/tmp/.X11-unix                                    \
    --volume  /home/zerotoasic/.Xauthority:/.Xauthority                        \
    --env     PDK_ROOT=/home/zerotoasic/asic_tools/pdk                         \
    --env     PDK=sky130A                                                      \
    --env     DISPLAY=:0                                                       \
    --user    1000:1000                                                        \
    --network host                                                             \
    --tty                                                                      \
    --interactive                                                              \
    efabless/openlane:cb59d1f84deb5cedbb5b0a3e3f3b4129a967c988-amd64