-
Notifications
You must be signed in to change notification settings - Fork 26
The Spatial Use Case
This page explains how to run HyperMapper with the Spatial language (https://spatial-lang.org). Here Spatial is the program that calls HyperMapper. HyperMapper replies with the points that have to be run by Spatial. Both software are writing on stdio and stdout and following a precise protocol of communication, like explained here. This is what in HyperMapper terminology is called "client-server mode". The developer responsible for Spatial is Matt Feldman.
For a general introduction on how to use Spatial and HyperMapper see the Design Space Exploration tutorial at PLDI 2020.
We provide a quick installation setup for HyperMapper here. We recommend you use spatial-quickstart, but full spatial works too.
First, install the following Spatial dependencies:
-
Scala SBT -
sudo apt-get install sbt # see www.scala-sbt.org/release/docs/Setup.html if sbt isn't found -
Java JDK 8 -
sudo apt-get install openjdk-8-jdk -
Integer Set Library -
sudo apt-get install libisl-dev - Python 3 - Make sure
pythonpoints to python3 and not python2
Then, install Spatial:
git clone https://github.com/stanford-ppl/spatial-quickstart
cd spatial-quickstart/
export SPATIAL_HOME=`pwd`
mkdir dse_hm
git submodule update --initInstall HyperMapper following our installation guide. Be sure to install from the source repo, rather than Conda, as Spatial currently relies on this old method of installation.
Set the HYPERMAPPER_HOME environment variable:
export HYPERMAPPER_HOME=/path/to/hypermapperThis variable is mandatory for this use case, Spatial will use it to call HyperMapper.
This branch conveniently contains a GEMM example, annotated with 13 DSE parameters.
bin/spatial GEMM --tune --hypermapper --synth --fpga=ZCU --hypermapper_doeSamples=14 --hypermapper_iters=20In general, Spatial apps should live in spatial-quickstart/src/main/scala
After the script finishes, the result of the optimization will be saved in spatial/gen/<benchmark>/<benchmark>_output_dse_samples.csv. E.g., for GEMM, the results will be saved in spatial/gen/GEMM/GEMM_output_dse_samples.csv.
hm-compute-pareto $SPATIAL_HOME/dse_hm/GEMM.json
hm-plot-pareto $SPATIAL_HOME/dse_hm/GEMM.json
cd $SPATIAL_HOME/gen/GEMM/
# hm-compute-pareto and hm-plot-pareto are located in $HYPERMAPPER_HOME/scripts/compute_pareto.py
# and $HYPERMAPPER_HOME/scripts/plot_pareto.py, if you are working without a pip installation of HypermapperA pareto front plot will be stored as all_GEMM_output_pareto.pdf in the app's generated directory.