Add docker image for the artifact

.dockerignore

@@ -0,0 +1,4 @@
+.hidet_cache
+build
+cmake-build-debug
+**/outs

Dockerfile

@@ -0,0 +1,40 @@
+FROM nvidia/cuda:11.6.0-devel-ubuntu20.04 as base
+
+COPY . /root/hidet
+
+ENV TZ=America/Toronto
+RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone
+
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    wget \
+    git \
+    vim \
+    python3 \
+    llvm-10 \
+    ccache \
+    software-properties-common \
+    python3-pip \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN python3 -m pip install --upgrade pip && \
+    python3 -m pip install --no-cache-dir -r /root/hidet/requirements.txt && \
+    python3 -m pip install --no-cache-dir --upgrade cmake && \
+    python3 -m pip uninstall -y onnxruntime-gpu==1.11.1 && \
+    python3 -m pip install --no-cache-dir onnxruntime-gpu==1.11.1 && \
+    hash -r  # refresh the hash
+
+RUN ln -s /usr/bin/python3 /usr/bin/python
+
+RUN cd /root/hidet && \
+    mkdir build && \
+    cd build && \
+    cmake .. && \
+    make -j 8 && \
+    rm -rf /root/hidet/build/3rdparty
+
+RUN echo export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/root/hidet/build/lib >> ~/.bashrc && \
+    echo export PYTHONPATH=$PYTHONPATH:/root/hidet/python:/root/hidet/3rdparty/tvm/python >> ~/.bashrc && \
+    echo alias python=python3 >> ~/.bashrc
+
+WORKDIR /root/hidet/artifacts

README.md

@@ -4,9 +4,7 @@ This repository contains the artifacts for the paper
 
 "Hidet: Task Mapping Programming Paradigm for Deep Learning Tensor Programs".
 
-## Installation
-
-### Requirements
+## Hardware requirements
 
 We did experiment on the following hardware platform
 
@@ -16,13 +14,76 @@ We did experiment on the following hardware platform
 
 Other workstation equipped with a modern NVIDIA GPU should also be able to run the experiments.
 
-On the software side, we require the following software to be installed
+The experiments require a Linux system with NVIDIA GPU driver installed.
+
+## Run experiments via docker
+
+We provide a docker image to run the experiments, with pre-configured environment.
+
+### Install docker and nvidia-docker
+
+Please follow the instructions on the [official website](https://github.com/NVIDIA/nvidia-docker) to install docker and 
+nvidia-docker.
+
+### Prepare the image
+
+We provide two ways to get the docker image to use. Please choose the one you like.
+
+#### Option 1: use the prebuilt image 
+
+```bash
+docker pull yyding/hidet-artifact:latest 
+```
+
+#### Option 2: build docker image from Dockerfile
+
+```bash
+git clone --recursive git@github.com:yaoyaoding/hidet-artifacts hidet
+cd hidet
+docker build -t yyding/hidet-artifact:latest .
+```
+
+### Run experiments inside docker container
+
+After above step, you can see a docker image named `yyding/hidet-artifact:latest` in your local docker image list via:
+
+```bash
+docker image ls
+```
+
+You should be able to see something like this:
+
+```text
+REPOSITORY              TAG       IMAGE ID       CREATED          SIZE
+yyding/hidet-artifact   latest    1074c09962c0   33 minutes ago   13.7GB
+```
+
+To run experiments, you need to start a docker container from the image:
+
+```bash
+# the following command will start a container based on the image
+# you will enter the container after the command at working directory /root/hidet/artifacts
+docker run -it --gpus all --rm yyding/hidet-artifact:latest
+# when you are in the container, run
+bash run.sh
+# to run all experiments
+```
+
+You can also run experiments in the container one by one. See [here](#run-the-experiments) for details.
+
+## Build and run experiments from source
+
+You can also build and run experiments from source on your host environment.
+
+### Installation
+
+We require the following software to be installed
 
 - cmake 3.19+
 - llvm (required by TVM, we used llvm-10)
 - ccache (used to accelerate duplicated compilation)
 
-### NVIDIA CUDA Toolkit
+#### NVIDIA CUDA Toolkit
 
 Please follow https://developer.nvidia.com/cuda-downloads guide to install the CUDA toolkit. 
 
@@ -35,15 +96,12 @@ nvidia-smi
 nvcc --version
 ```
 
-### Install Hidet and baselines
+#### Install Hidet and baselines
 
 ```bash
 # clone hidet repository 
-git clone git@github.com:yaoyaoding/hidet
+git clone --recursive git@github.com:yaoyaoding/hidet-artifacts hidet
 cd hidet
-git checkout artifact
-git submodule init
-git submodule update --recursive --init
 
 # install the dependencies of hidet and the baselines (e.g., TensorRT, PyTorch, Onnx Runtime)
 # the versions of baselines are specified in requirements.txt file.
@@ -70,7 +128,7 @@ python3 -c "import hidet"
 python3 -c "import artifact"
 ```
 
-# Run the experiments
+### Run the experiments
 
 This artifact contains all the experiments in the evaluation section of the paper:
 

artifacts/2_input_sensitivity/main.py

@@ -9,16 +9,16 @@ def main():
     ]:
         for model in [
             '--model op_matmul_nn_4',   # 2048x2048x2048
-            '--model op_matmul_nn_5',   # 2047x2047x2047
-            '--model op_matmul_nn_6',   # 2046x2046x2046
-            '--model op_matmul_nn_7',   # 2045x2045x2045
-            '--model op_matmul_nn_8',   # 2044x2044x2044
-            '--model op_matmul_nn_9',   # 2043x2043x2043
-            '--model op_matmul_nn_10',  # 2042x2042x2042
-            '--model op_matmul_nn_11',  # 2041x2041x2041
+            '--model op_matmul_nn_5',   # 2039x2039x2039
+            '--model op_matmul_nn_6',   # 2047x2047x2047
+            '--model op_matmul_nn_7',   # 2046x2046x2046
+            '--model op_matmul_nn_8',   # 2045x2045x2045
+            '--model op_matmul_nn_9',   # 2044x2044x2044
+            '--model op_matmul_nn_10',  # 2043x2043x2043
+            '--model op_matmul_nn_11',  # 2042x2042x2042
         ]:
-            if '11' in model and ('ansor' in executor or 'autotvm' in executor):
-                # both schedulers failed to find a schedule for this input 2041x2041x2041, skip
+            if '_5' in model and ('ansor' in executor or 'autotvm' in executor):
+                # both schedulers failed to find a schedule for this input 2039x2039x2039, skip
                 # for autotvm, it will fall back to a default schedule.
                 # for ansor, it will fall into a dead loop.
                 continue

artifacts/run.sh

@@ -5,6 +5,16 @@
 export TVM_BIND_THREADS=0
 export TVM_NUM_THREADS=24
 
+# The first run would generate a lot of logs and take a long time
+python ./0_end2end/main.py
+python ./1_latency_distribution/main.py
+python ./2_input_sensitivity/main.py
+python ./3_batch_size/main.py
+python ./4_prologue_epilogue_fusion/main.py
+python ./5_tensorrt/main.py
+
+# The second run would use the cached results and take a short time
+# The output would be clear (not scattered with logs)
 python ./0_end2end/main.py
 python ./1_latency_distribution/main.py
 python ./2_input_sensitivity/main.py

build_docker_image.sh

@@ -0,0 +1,2 @@
+#!/bin/bash
+docker build -t hidet-artifact:latest .

requirements.txt

@@ -23,6 +23,7 @@ nvtx
 
 # for onnx frontend
 onnx==1.10.2
+protobuf==3.20
 
 ################################################################################
 # Optional packages
@@ -31,13 +32,14 @@ onnx==1.10.2
 torch==1.11
 torchvision==0.12
 
-# for language model converting
-transformers==4.19.2
-transformers[onnx]
-
 # for onnx runtime baseline
+onnxruntime==1.11.1
 onnxruntime-gpu==1.11.1
 
+# for language model converting
+transformers==4.19.2
+transformers[onnx]==4.19.2
+
 # for tensor rt baseline
 --extra-index-url https://pypi.ngc.nvidia.com
 nvidia-tensorrt==8.2.5.1

run_docker.sh

@@ -0,0 +1,2 @@
+#!/bin/bash
+docker run --rm --gpus all -it hidet-artifact:latest