CASIO DL Application Suite

Link to Data: Google Drive

Structure of Released Data

The released data includes for the following for each platform and each application ($PLAT/$APP/*):

• Walk clock time (bench-.*)
• TF or Pytorch profile output (prof-.*)
• NSYS output (nsys-.*nsys-rep)
• NCU output (ncu-.*)

The NSYS file can be read by NSIGHT Systems. The NCU file is raw text format (prepended with a header row and then data in CSV form. Data follows after a line that says -- PROF --)

The summaries/ directory includes:

• A file for each platform indicating the smallest and largest batch size we ran ($PLAT-<small/large>-batch-list)
• For each platform / application: summaries extracted from the NSYS output (Can be recreated from nsys-rep files)

The postproc/ directory includes framework operator traces produced from framework profiler data ($PLAT/$APP/op-trace-*.csv)

Finally, the data we recorded on GEMM performance can be found here.

How to run these Apps

Environments

Torch

Setup: Make sure you have a conda installation. Create a new environment for casio by following these instructions:

MAKE SURE TO FOLLOW THESE IN ORDER

$ conda create -n casio-torch python=3.9
# Press enter to accept

$ conda activate casio-torch

# Install requirements
$ pip install -r requirements.txt

NOTE: mmcv-full will take a WHILE to install. This is a one-time thing.

TensorFlow

For tensorflow, use the utils/tf1-docker.sh script to launch a docker container.

NOTE: YOU WILL END UP SOURCE'ING env.sh TWICE!

Swin Transformer (Torch)

$ conda activate casio-torch
$ source env.sh
==============================================
REMEMBER: RUN THIS INSIDE THE DOCKER CONTAINER
FOR TENSORFLOW v1 APPLICATIONS!

MIKE WILL NOT ANSWER THIS QUESTION!
==============================================
What platform is this? (cpu, p100, v100, a100): <type of gpu>
What gpu should we use? (cuda:0, cuda:1, ...): cuda:<N>
Path to CASIO: /nobackup/medavies/casio
Platform: <type of gpu>
Device: cuda:<N>

$ cd Swin-Transformer
$ ./runall.sh

MuZero (Torch)

$ conda activate casio-torch
$ source env.sh
...
$ cd muzero
$ ./runall.sh

QD Track (Torch)

DOWNLOAD DATA AND RUN SETUP FIRST

$ conda activate casio-torch
$ cd qdtrack/
$ wget cs.wisc.edu/~davies/qdtrack-data.tar.xz
$ tar xvf qdtrack-data.tar.xz
$ python setup.py develop
$ cd ..

Running qdtrack:

$ conda activate casio-torch
$ source env.sh
...
$ cd qdtrack
$ ./runall.sh

PINN (Tensorflow)

$ source env.sh
$ ./utils/tf1-docker.sh
$ cd /work
$ source env.sh
...
$ pip install pydoe
$ cd PINNs
$ ./runall.sh

tabnet (Tensorflow)

$ source env.sh
$ ./utils/tf1-docker.sh
$ cd /work
$ source env.sh
...
$ cd tabnet
$ ./runall.sh

meshgraphnets (Tensorflow)

DOWNLOAD DATA AND RUN SETUP FIRST

$ cd meshgraphnets
$ wget cs.wisc.edu/~davies/mgn-datasets.tar.xz
$ tar xvf mgn-datasets.tar.xz
$ cd ..

Running meshgraphnets:

$ source env.sh
$ ./utils/tf1-docker.sh
$ cd /work
$ source env.sh
...
$ cd meshgraphnets
$ pip install -r requirements
$ cd /work
$ ./meshgraphnets/runall.sh