Analyze FPGA tool performance (MHz, resources, runtime, etc)
fpga-tool-perf uses the Miniconda (conda) package manager to install and get all the required tools. Currently, the following tools that are available in conda are:
- vtr
- nextpnr-xilinx
- yosys (+ yosys-plugins)
- prjxray
Prior to setting up the conda environment, the symbiflow and quicklogic data files need to be installed through the following commands:
make install_symbiflow
make install_quicklogic
The install packages are extracted in the env/<toolchain>/
location.
To setup the conda environment, run the following commands:
TOOLCHAIN=symbiflow make env
TOOLCHAIN=quicklogic make env
TOOLCHAIN=nextpnr make env
fpga-tool-perf can also run the Vivado EDA tool. The tool is not available in the conda environment and it needs to be installed by the user.
The user needs also to set the VIVADO_SETTINGS
environmental variable, which points to the settings64.sh
file to enable Vivado.
With the conda environment correctly installed, run the following to activate the environment:
source env.sh
Once the environment settings has been sourced, you are ready to proceed with the tests
python3 fpgaperf.py --toolchain vivado --project oneblink --board arty-a35t
or
python3 fpgaperf.py --toolchain vpr --project oneblink --board basys3
For example to compare pure Vivado flow and Yosys -> Vivado flow for an xc7z device the following commands can be run:
# Yosys -> Vivado
python3 fpgaperf.py --toolchain yosys-vivado --project oneblink --board basys3
# Pure Vivado
python3 fpgaperf.py --toolchain vivado --project oneblink --board basys3
Use --help
to see additional parameters for the fpgaperf.py
script.
Supported toolchains can be queried as follows:
$ python3 fpgaperf.py --list-toolchains
nextpnr-ice40
nextpnr-xilinx
nextpnr-xilinx-fasm2bels
vivado
vpr
vpr-fasm2bels
yosys-vivado
You can check if you have the toolchain environments correctly installed as follows:
$ python3 fpgaperf.py --check-env --toolchain vpr
vpr
yosys: True
vpr: True
prjxray-config: True
Supported projects can be queried as follows:
$ python3 fpgaperf.py --list-projects
baselitex
blinky
bram-n1
bram-n2
bram-n3
dram-test-64x1d
hamsternz-hdmi
ibex
murax
oneblink
picorv32
picosoc
picosoc-simpleuart
picosoc-spimemio
vexriscv
vexriscv-smp
Use exhaust.py
to automatically test all projects, toolchain and boards supported
python3 exhaust.py
Its also possible to run a test against specific project(s), toolchain(s), and/or board(s):
python3 exhaust.py --project blinky oneblink --toolchain vpr
See build
directory for output. Note in particular all.json
.
This section describes the structure of this project to better understand its mechanisms.
-
the
project
directory contains all the information relative to a specific test. This includes:- srcs: all the source files needed to run the test
- top: top level module name of the design
- name: project name
- data: all of the data/memory files needed to run the test
- clocks: all the input clocks of the design
- required_toolchains: all the toolchains that are required to correctly run to completion.
- vendors: all the vendors that are enabled for this project (e.g. xilinx, lattice). Each vendor requires a list of boards enabled for the test project.
-
the
src
directory contains all the source files needed to build the test project. It also contains the constraints files relative to the various boards supported. -
the
other
directory contains two json files, describing all the supported boards and vendors in this test suite. -
the
toolchains
directory contains the python scripts that enable a toolchain to be run within fpga-tool-perf. -
the
infrastructure
directory contains python scripts to control the fpga-tool-perf framework to run the tests
wrapper.py
creates a simple verilog interface against an arbitrary verilog module.
This allows testing arbitrary verilog modules against a standard pin configuration. The rough idea is taken from project x-ray.
Run wrappers.sh
to regenerate all wrappers. Requires pyverilog
wrapper.py
(iverilog based) has the following known limitations:
- Bidrectional ports are not supported
- Spaces inside numbers are not supported (ex: 8' h00 vs 8'h00)
- Attributes (sometimes?) are not supported (ex: (* LOC="HERE" *) )
As a result, sometimes the module definition is cropped out to make running the tool easier (ex: src/picorv32/picosoc/spimemio.v was cropped to src/picosoc_spimemio_def.v).
Projects are .json files in the project directory. Project names shouldn't contain underscores such that they are clearly separated from other fields when combined into folder names.
These are the basic steps to inserting an existing project into fpga-tool-perf:
Add a folder within fpga-tool-perf/src
under the name of the project (make sure there are no underscores '_' in the name). For example, for the project named counter:
cd ~/fpga-tool-perf/src
mkdir counter
cd counter
Add the source (verilog) and data/memory files to this directory.
Create a constr
subdirectory, and within it, add the project's .pcf
(for symbiflow) and .xdc
(for vivado) files under the name of the board it uses.
mkdir constr
touch constr/basys3.pcf
touch constr/basys3.xdc
If you don't have both the .pcf
and .xdc
files, You can look at the other projects for examples of how the .xdc
and .pcf
code correspond.
Within the project
directory, create a .json
file under the name of the project. Copy the contents of another project's .json
file and modify it to match your project's specs. It will look like somthing like this:
{
"srcs": [
"src/counter/counter.v"
],
"top": "top",
"name": "counter",
"clocks": {
"clk": 10.0
},
"required_toolchains": [
"vpr", "vpr-fasm2bels", "vivado", "yosys-vivado"
],
"vendors": {
"xilinx": ["arty-a35t", "basys3"]
}
}
Test the newly added project with vpr and vivado. For example:
python3 fpgaperf.py --project counter --toolchain vpr --board basys3
python3 fpgaperf.py --project counter --toolchain vivado --board basys3
There may be errors if your .xdc
or .pcf
files have the wrong syntax. Debug, modify, and run until it works, and you have successfully added a new project to fpga-tool-perf.