Guide to adding a new device to an existing family¶
This documents how to add support for a new device. The running example is the addition of the xc7a100t device to the Artix-7 family.
Adding a new device to an existing family is much simpler than adding a new family, since the building blocks (tiles) are already known. There are just more or fewer of them, arranged differently. No new fuzzers are needed. You just need to rerun some fuzzers for the new device to understand how the tiles are connected to each other and to IOs.
If you are just adding a new package for a device that is already supported, you can skip Steps 2 through 5.
Note: Since this guide was written, the xc7a100t has become the primary device in the database, not a secondary device as it was when it was originally added. Therefore the files currently in the repo don’t match what is described here. But if you look at the original PRs, they match what is described in the examples here.
Fork a copy of https://github.com/SymbiFlow/prjxray on GitHub (go to the page, click “Fork” button, select your own workspace).
Clone your fork, and make a new branch, with a name related to the new device/package:
git clone firstname.lastname@example.org:<yourUserID>/prjxray.git cd prjxray git checkout -b <new_branch_name>
Follow the Project X-Ray developer setup instructions in the
up through Step 7 and choose Option 1 (invoke the
./download-latest-db.sh script). This script will clone the official
prjxray-db database under
database/. The following steps will make
changes under this directory. You may want to put these changes on your
own fork of
prjxray-db for testing. This is explained at the end,
under “Database Updates”.
Add a new settings file. Usually you will start with an existing settings file and modify it. Assuming you’re in prjxray/,
cp settings/<baseline_device>.sh settings/<new_device>.sh git add settings/<new_device>.sh
cp settings/artix7_200t.sh settings/artix7_100t.sh git add settings/artix7_100t.sh
Update the following values in the new settings file:
XRAY_PART– Important: choose a package that is fully bonded (typically the one with the largest number of pins). If the part that you’re actually interested in is different (with fewer bonded pins), it will be handled later. In the running example, the actual part of interest was the xc7a100tcsg324, since that is on the Arty A7-100T board. But here, the xc7a100tfgg676 part is used; the xc7a100tcsg324 is handled later.
XRAY_ROI_TILEGRID– modify the bounding boxes to be a tight fit on your new part.
XRAY_IOI3_TILES– These tiles need special handling for an irregularity in Xilinx 7-series FPGAs. See the comments in the 005 fuzzer for more information.
This is what the new settings file looked like in the example.
Source this new settings file:
Add all informations about the part. YAML files for each family are located at database/<family>/mapping/, which contain the part information (parts.yaml), device to fabric mapping (devices.yaml) or hints about resources (resources.yaml).
The first file contains a mapping between a part number and informations about the device, package and speed grade used by the fuzzers. The complete part number is used as key. Device, package and speedgrade are parts of the part numbers.
device: “xc7a100t” package: “csg324” speedgrade: “1” pins:
0: “N15” 1: “U17” 2: “V17” 3: “V16” 4: “V14” 5: “U14” 6: “U16”
The second file maps devices to fabrics. Because some fabrics are added to multiple devices, they are only generated for one and parts with the same link to the result.
The last file contains information about the information about a resource to support the fuzzers generating the informations. The dictionary pins defines package pins with the following purpose:
00– this must be a clock pin. You can look at the device in the Vivado GUI interactively (click on IOs and check their properties until you find one with IS_CLOCK=true), or run a small clocked design in Vivado and see which pin is assigned to ‘clk’.
01and on – these should be normal data pins on the device.
Edit the top Makefile
Update the Makefile by adding the new device to the correct list, so that the Makefile generates targets for the new device (used in Step 4).
<new_device>is the basename of the new settings file that you just created.
In our running example, we add
Make sure you’ve sourced your new device settings file (see the end of step 2). Now it is time to run some fuzzers to figure out how the tiles on your new device are connected.
Make the following target, with
<new_device> as above, and setting
the parallelism factor
-j<n> appropriate for the number of cores
your host has. The make job can benefit from large numbers of cores.
make -j<n> MAX_VIVADO_PROCESS=<n> db-part-only-<new_device>
<new_device> must match the base name of the new settings
file that was added. For example,
make -j32 MAX_VIVADO_PROCESS=32 db-part-only-artix7_100t
It should run fuzzers 000, 001, 005, 072, 073, 074, and 075.
005 will take a long time. Using multiple cores will help.
074 will fail the first time, since it hasn’t been told to ignore certain wires.
After it fails, go to the build directory
cd fuzzers/074-dump_all/build_<XRAY_PART>(this is the
XRAY_PARTfrom the new settings script; in our example, the build directory is
python3 ../analyze_errors.py --output_ignore_list > new-ignored
Inspect and compare
new-ignoredagainst existing ignored wire files in
If it looks good, copy it to an appropriately-named file:
cp new-ignored ../ignored_wires/artix7/<XRAY_PART>_ignored_wires.txt(in our example, it is
git add ../ignored_wires/artix7/<XRAY_PART>_ignored_wires.txt
Return to prjxray/ directory, and clean up 074 to prepare for the rerun:
make -C fuzzers/074-dump-all clean
Rerun the top make command, e.g.
make -j32 MAX_VIVADO_PROCESS=32 db-part-only-artix7_100t
The next task is handling the extra parts – those not fully bonded out. These are usually the parts you actually have on the boards you buy.
Add a new entry in the appropriate ‘harness’ section for any alternative packages (typically with fewer pins, in this example, 324 versus 676). If any
XRAY_PIN_<XX>values you listed in the settings file are not bonded out on the new part, you must specify alternatives. In this example, we need to specify a new clock pin,
XRAY_PARTis the extra part, and
XRAY_EQUIV_PARTis the original, fully-bonded version:
db-extras-artix7-harness: +source settings/artix7.sh && \ XRAY_PART=xc7a35tftg256-1 $(MAKE) -C fuzzers roi_only + +source settings/artix7_100t.sh && \ + XRAY_PART=xc7a100tcsg324-1 $(MAKE) -C fuzzers roi_only +source settings/artix7_200t.sh && \ XRAY_PIN_00=V10 XRAY_PIN_01=W10 XRAY_PIN_02=Y11 XRAY_PIN_03=Y12 \ XRAY_PART=xc7a200tsbg484-1 XRAY_EQUIV_PART=xc7a200tffg1156-1 \ $(MAKE) -C fuzzers roi_only
Make the appropriate harness target (adjusting for your family):
make -j32 db-extras-artix7-harness
This target will make updates for the extra parts of all of the family devices, not just your new device.
Do a spot check.
Check that there are new part directories in the database under the family subdirectory, for example:
$ ll database/artix7/xc7a* xc7a35tftg256-1: total 48 -rw-rw-r-- 1 daniel daniel 8234 Jan 9 13:01 package_pins.csv -rw-rw-r-- 1 daniel daniel 18816 Jan 9 13:01 part.json -rw-rw-r-- 1 daniel daniel 13099 Jan 9 13:01 part.yaml xc7a50t: total 15480 -rw-rw-r-- 1 daniel daniel 695523 Jan 9 12:53 node_wires.json -rw-rw-r-- 1 daniel daniel 8587682 Jan 9 12:53 tileconn.json -rw-rw-r-- 1 daniel daniel 6562851 Jan 9 10:31 tilegrid.json xc7a50tfgg484-1: total 52 -rw-rw-r-- 1 daniel daniel 13056 Jan 9 09:54 package_pins.csv -rw-rw-r-- 1 daniel daniel 18840 Jan 9 09:58 part.json -rw-rw-r-- 1 daniel daniel 13099 Jan 9 09:58 part.yaml
In this case, the tile grid is the same size since it’s the same chip, but the size of the package pins files differs, since there are different numbers of bonded pins.
Note: These changes/additions under
database/ do not get checked
in. They are in the
prjxray-db repo. This spot check is to make sure
that your changes in
prjxray will do the right thing when the
official database is fully rebuilt. See “Database Updates” below for
Assuming everything looks good, commit to your
You should have a new file under settings/, a new ignored_wires file,
and a modified Makefile (see the initial
of the example for reference).
git add Makefile settings/artix7_100t.sh git status git commit --signoff
Push to GitHub:
git push origin <new_branch_name>
Then make a pull request. Navigate to the GitHub page for your
prjxray fork/branch, and click the “New pull request” button.
Making the pull request will kick off continuous integration tests.
Watch the results and fix any issues.
The process above (steps 4 and 5) will create some new files and modify
some existing files under database/, which is a different repo,
To test these changes before the next official prjxray-db gets built (and even before your PR on prjxray is merged), you can put these changes on your own fork of prjxray-db, and then test them in the context of symbiflow-arch-defs.
To put the db updates on your own fork, create your fork of https://github.com/SymbiFlow/prjxray-db if you haven’t already. Then follow one of the approaches suggested in the checked solution of this StackOverflow post.
You are NEVER going to send a pull request on prjxray-db. The database is always rebuilt from scratch. After your changes on prjxray are merged, they will reflected in the next prjxray-db rebuild. The changes submitted to your prjxray-db fork are only for your own testing.
To use your new repo/branch under
symbiflow-arch-defs/third_party/prjxray-db/, you will need to change the
submodule reference to point to your fork/branch of