-
Notifications
You must be signed in to change notification settings - Fork 10
Get Started!
We tested this workflow on Ubuntu 20.10, and it should basically work on other platforms. You need these packages to compile the whole framework. You can also use the docker image we provided below.
$ sudo apt install cmake build-essential autoconf libtool zlib1g zlib1g-dev python3 python-is-python3 python3-pip
$ sudo pip install six
These are the commands used to set up GemForge.
$ git clone --recurse-submodules [this repo]
$ cd [this repo]
# Set up some environment variables.
$ source envs.sh
# Build llvm.
$ cd llvm && bash setup.sh
# Build Gem Forge transformation and gem5, etc.
bash setup.sh
There are some microbenchmarks in transform/benchmark
. Here we use a matrix-vector multiply example to demonstrate the whole workflow. It is vectorized by AVX-512 and parallelized with OpenMP.
It is located at tranform/benchmark/GemForgeMicroSuite/omp_dense_mv_blk
. We also provide a make file in example
that you can easily follow.
Gem Forge takes a single LLVM bitcode as the input. Use this command to compile the program. Since Gem Forge was first designed to process LLVM IR trace, we call the bitcode trace.bc
. But don't be confused by this name, this example does not require any trace information and is purely based on static inforation.
$ cd example && make traced.bc
Next we build two binaries: one for vanilla x86 and one with Stream ISA enabled:
$ make valid.exe
$ make stream.exe
Then we simulate 4 configurations (64 cores with 8x8 mesh):
-
valid.o8.sim
: vanilla OOO8 core without any prefetching support. -
valid.o8-pf.sim
: OOO8 core with a Bingo prefetcher at L1$ and a stride prefetcher at L2$. -
stream.o8.sim
: OOO8 core with Stream Engine at the core. -
stream.o8-float.sim
: OOO8 with Stream Engine both at the core and in the cache (aka. Stream Floating).
You can simulate them all by:
$ make sim-all
That's it! You can verify the results in valid/o8
, etc., and check that Stream Floating indeed improves the performance and reduces NoC traffic.
You can also try this docker file:
FROM ubuntu:20.10
ARG USER_ID
ARG GROUP_ID
RUN groupadd -g ${GROUP_ID} gf &&\
useradd -l -u ${USER_ID} -g gf gf &&\
usermod -aG sudo gf &&\
install -d -m 0755 -o gf -g gf /home/gf &&\
apt update &&\
apt install -y \
cmake \
build-essential \
autoconf \
libtool \
zlib1g \
zlib1g-dev \
python3 \
python3-pip \
python-is-python3 \
scons \
git \
curl \
wget \
sudo \
vim \
zsh &&\
pip3 install six
RUN echo 'gf:gf' | chpasswd
USER gf
Save this into Dockerfile
. Then build the image and start the container:
$ docker image build --build-arg USER_ID=$(id -u $USER) --build-arg GROUP_ID=$(id -g $USER) -t gemforge:0.1 .
$ docker run --network host --detach -i --mount type=bind,src=[this repo],dst=[mounted path] --name gemforge gemforge:0.1
$ docker exec -it gemforge zsh
Then you can follow the commands above to build and run!