Since Plasticine currently does not have a runtime and area model, the only way to perform DSE of a
application is run various configuration of the appilcation through the compiler and simulation.
spatial/pir/bin/dse_util.py
provides some helper function for easy generation of these parameters.
Example usage in here.
from dse_util import *
def gen_apps():
with dseopen("src/DSE.scala") as f:
dseApp("BlackScholes",
ip=16,
op=[1,2],
N=1024*100,
pirArgs=["--split-algo=dfs --retime-glob=true " + \
f"--pcu-stage={pcuStage}" for pcuStage in [5,6]],
)
dse = DSE(logdir="logs", project="spatialApps")
dse.setup_dse(gen_apps)
Run bin/dse -g will generates src/DSE.scala
import spatial.dsl._
import forge.tags._
@virtualize
class Dummy2
class BlackScholes_D0 extends BlackScholes(ip=16,op=1,N=102400){ override def pirArgs = super.pirArgs + " --split-algo=dfs --retime-glob=true --pcu-stage=5"; }
class BlackScholes_D1 extends BlackScholes(ip=16,op=1,N=102400){ override def pirArgs = super.pirArgs + " --split-algo=dfs --retime-glob=true --pcu-stage=6"; }
class BlackScholes_D2 extends BlackScholes(ip=16,op=2,N=102400){ override def pirArgs = super.pirArgs + " --split-algo=dfs --retime-glob=true --pcu-stage=5"; }
class BlackScholes_D3 extends BlackScholes(ip=16,op=2,N=102400){ override def pirArgs = super.pirArgs + " --split-algo=dfs --retime-glob=true --pcu-stage=6"; }
To run apps, run bin/dse -a BlackScholes_D* [options]
Options used by bin/test can be used in bin/dse options. To override default pir options,
set dse.pirArgs = 'new options' before calling dse.setup_dse
dseApp will create a cross product space between key value pairs of its named arguments. In this
app, they are ip, op, and N. These arguments can be a single element or a list.
If the corss product space is too large,
dseApp can also take another parameter---filterNot that can take in a lambda or a function to
prune the space.
The input of the filterNot is a dictionary of parameter names to parameter values, and the lambda returns
a boolean for whether this design point should be dropped from DSE.
For example filterNot=lambda param: param['ip'] != param['op'] will drop design points whose ip ==
op.