Skip to content

Latest commit

 

History

History
560 lines (436 loc) · 19.8 KB

File metadata and controls

560 lines (436 loc) · 19.8 KB

Overview

In this demo we will look at:

  • Building LLVM with the correct settings to allow for model training
  • Collecting a training corpus for the regalloc model based on Chromium
  • Training the regalloc model on the collected corpus
  • Compiling the trained model into LLVM

Preliminaries

Set up some environment variables according to where you want to clone/build all of the code:

export WORKING_DIR=~

Change the directory to wherever you'd like to put everything.

Get repositories

ml-compiler-opt (this repository)

Clone the github repo:

cd $WORKING_DIR
git clone https://github.com/google/ml-compiler-opt

LLVM

Grabbing LLVM should be as simple as running the below command, but if something goes awry, make sure to check the official documentation.

git clone https://github.com/llvm/llvm-project.git

Chromium

Grabbing Chromium is a bit more involved. The official documentation for Linux based systems (the only platform currently supported by MLGO) is available at that link. However, cloning the code should be as simple as downloading depot_tools and adding them to your $PATH:

git clone https://chromium.googlesource.com/chromium/tools/depot_tools.git
export PATH="$PATH:$WORKING_DIR/depot_tools"

Creating a folder for Chromium and cloning it using the fetch utility:

mkdir chromium
cd chromium
fetch --nohooks --no-history chromium

Note: Running fetch with the --no-history flag makes your local checkout significantly smaller and speeds up the underlying git clone by a significant amount. However, if you actually need the history for any reason (eg to revert to a previous commit or to work on the Chromium side with MLGO stuff), make sure to omit this flag to do a full checkout.

The fetch command will take at least a couple minutes on a fast internet connection and much longer on slower ones.

Next, we need to modify the .gclient file that fetch creates in the directory that you run it in to make sure that the Chromium PGO profiles get checked out:

sed -i 's/"custom_vars": {},/"custom_vars": { "checkout_pgo_profiles" : True },/' .gclient

This sed command will set the necessary variable correctly. After this, you can move into the src directory that fetch created that contains the actual Chromium codebase.

Now that this is all in place, you need to run the Chromium hooks in order to get the development environment ready for a full compilation:

gclient runhooks

Install Dependencies

If you're working in a Debian based docker container, it will most likely not come by default with sudo. It isn't stricly necessary to install, but it makes it easier to copypaste the installation commands below and it also enables the use of the Chromium dependency auto-installation script:

apt-get install sudo

First, install some base dependencies that will be needed when building LLVM:

sudo apt-get install cmake ninja-build lld

Now, install the Chromium dependencies using the auto-installation script:

$WORKING_DIR/chromium/src/build/install-build-deps.sh

Note: These installation commands are all designed to be run on Debian based distros. However, adapating to other distros with alternative package management systems should not be too difficult. The packages for the first command should be very similarly named and the official Chromium documentation has info on dependency installation for their build process on other common distros.

Also make sure that you install the Python dependencies for the ml-compiler-opt repository:

cd $WORKING_DIR/ml-compiler-opt
pip3 install pipenv
pipenv sync --system

If you plan on doing development work on this checkout of ml-compiler-opt, use the /ml-compiler-opt/requirements-ci.txt requirements file to install some additional development dependencies.

Building Chromium

WARNING: Currently, Chromium only builds with protobuf 4.x.x while Tensorflow requires protobuf 3.x.x. In order to make sure that Chromium compiles correctly you can either use a virtual environment and install protobuf 4.x.x there (currently just the latest version), or you can install protobuf 4.x.x over the currently installed version and then undo it later after the compile is complete. This tutorial assumes no usage of virtual environments. Install a compatible version of protobuf:

pip3 install protobuf==4.21.7

To build Chromium, make sure you are in the /chromium/src directory and then run the following command to open a CLI text editor that will allow you to configure build settings:

cd $WORKING_DIR/chromium/src
gn args ./out/Release

Then, you need to specify the configuration options to use when compiling Chromium. This will depend upon the type of training corpus that you want to extract. If you want to extract a non-thinLTO corpus, you can use the configuration listed below:

is_official_build=true
use_thin_lto=false
is_cfi=false
clang_embed_bitcode=true
is_debug=false
symbol_level=0
enable_nacl=false

But if you want to extract a thinLTO corpus, you need to use the following config:

is_official_build=true
lld_emit_indexes_and_imports=true
is_debug=false
symbol_level=0
enable_nacl=false

Immedaitely after closing the editor, gn will generate all of the files necessary so that ninja can execute all the necessary compilation steps. However, to extract a corpus for ML training, we also need a database of compilation commands. This can be obtained by running the following command:

gn gen ./out/Release --export-compile-commands

Then you can build Chromium with the autoninja utility:

autoninja -C ./out/Release

A full Chromium compile will take at least an hour on pretty well specced hardware (ie 96 thread work station) and much longer on lower specced hardware.

TODO(boomanaiden154): Investigate the source of this assertion error.

Note: If the build fails in the last couple steps tripping an assertion in the linker when compiling a non-thinLTO corpus, you can safely ignore this. Preparing a corpus for ML training in the non-thinLTO case only requires the object files that get fed to the linker

WARNING: make sure to reinstall a version of protobuf compatible with the current tf-nightly release used by ml-compiler-opt if you changed versions earlier to get the Chromium compile working. Reinstalling using the ml-compiler-opt lockfile should work:

pip3 uninstall protobuf
pipenv sync --system

Building LLVM

To build LLVM to train ML models, we first need to build TFLite and some dependencies so that we can embed it within LLVM to load and execute models on the fly during reinforcement learning. There is a script within this repository that clones and builds everything automatically and prepares a CMake cache file that can be passed to CMake during the LLVM build configuration. Running the script looks like this:

cd $WORKING_DIR
mkdir tflite
cd tflite
$WORKING_DIR/ml-compiler-opt/buildbot/build_tflite.sh

This script should only take a couple minutes to execute as all the libraries that it pulls and builds are relatively small.

Now, create a new folder to do an LLVM build and configure it using CMake:

mkdir $WORKING_DIR/llvm-build
cd $WORKING_DIR/llvm-build
cmake -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLVM_ENABLE_PROJECTS="clang" \
  -C $WORKING_DIR/tflite/tflite.cmake \
  $WORKING_DIR/llvm-project/llvm

Now you can run the actual build with the following command:

cmake --build .

ML training

All of the following example commands assume you are working from within your checkout of the ml-compiler-opt repository:

cd $WORKING_DIR/ml-compiler-opt

To start off training, we need to extract a corpus from the Chromium compile. The procedure for this will depend upon how you built your corpus, particularly whether or not you used thinLTO.

Installing the corpus extraction tooling

Install the corpus extraction tooling:

pip3 install mlgo-utils

Make sure that the local binary directory that Python installs executables into (typically ~/.local/bin) is on your $PATH so that the invocations below will work as expected.

Corpus extraction (non-thinLTO case)

For corpus extraction in the non-thinLTO case, you can simply run the following command:

extract_ir \
  --cmd_filter="^-O2|-O3" \
  --input=$WORKING_DIR/chromium/src/out/Release/compile_commands.json \
  --input_type=json \
  --llvm_objcopy_path=$WORKING_DIR/llvm-build/bin/llvm-objcopy \
  --output_dir=$WORKING_DIR/corpus

This command will extract all the relevant bitcode and compilation flags from the Chromium compile and put them in the $WORKING_DIR/corpus directory. No further processing on the corpus should be needed.

Corpus extraction (thinLTO case)

Corpus extraction for the thinLTO case is slightly more involved. Start off by running the following command to do the initial step in the corpus extraction process:

extract_ir \
  --cmd_filter="^-O2|-O3" \
  --llvm_objcopy_path=$WORKING_DIR/llvm-build/bin/llvm-objcopy \
  --output_dir=$WORKING_DIR/corpus \
  --thinlto_build=local \
  --obj_base_dir=$WORKING_DIR/chromium/src/out/Release/obj

After this, it is necessary to grab the flags passed to the linker and add them to the corpus_description.json file in the $WORKING_DIR/corpus folder. To find this, it is helpful to look at the actual invocation of the linker. To see the linker invocation for a target like chrome, go to the Chromium build directory and run the following command:

cd $WORKING_DIR/chromium/src/out/Release
ninja -t commands chrome

(Note - libchrome instead of chrome if targeting Android)

The last command should look something like this:

python3 "../../build/toolchain/gcc_link_wrapper.py" --output="./chrome" -- ../../third_party/llvm-build/Release+Asserts/bin/clang++ -Wl,--version-script=../../build/linux/chrome.map -Werror -fuse-ld=lld -Wl,--fatal-warnings -Wl,--build-id=sha1 -fPIC -Wl,-z,noexecstack -Wl,-z,relro -Wl,-z,now -Wl,--icf=all -Wl,--color-diagnostics -Wl,-mllvm,-instcombine-lower-dbg-declare=0 -Wl,--save-temps=import -Wl,--thinlto-emit-index-files -flto=thin -Wl,--thinlto-jobs=all -Wl,--thinlto-cache-dir=thinlto-cache -Wl,--thinlto-cache-policy=cache_size=10\%:cache_size_bytes=40g:cache_size_files=100000 -Wl,-mllvm,-import-instr-limit=30 -fwhole-program-vtables -Wl,--undefined-version -m64 -no-canonical-prefixes -Wl,-O2 -Wl,--gc-sections -rdynamic -Wl,-z,defs -Wl,--as-needed -nostdlib++ --sysroot=../../build/linux/debian_bullseye_amd64-sysroot -fsanitize=cfi-vcall -fsanitize=cfi-icall -pie -Wl,--disable-new-dtags -Wl,--lto-O2 -o "./chrome" -Wl,--start-group @"./chrome.rsp"  -Wl,--end-group  -ldl -lpthread -lrt -lgmodule-2.0 -lgobject-2.0 -lgthread-2.0 -lglib-2.0 -lnss3 -lnssutil3 -lsmime3 -lplds4 -lplc4 -lnspr4 -latk-1.0 -latk-bridge-2.0 -lcups -lgio-2.0 -ldrm -ldbus-1 -latspi -lresolv -lm -lX11 -lXcomposite -lXdamage -lXext -lXfixes -lXrender -lXrandr -lXtst -lgbm -lEGL -lexpat -luuid -lxcb -lxkbcommon -lXi -lpci -l:libffi_pic.a -lpangocairo-1.0 -lpango-1.0 -lharfbuzz -lcairo -lasound -lz -lstdc++ -lxshmfence

From this, we can take out a couple flags that we need for our corpus_description.json. A more precise description of the flags that are needed is available in the MLGO ThinLTO documentation. Given the linker command above, the flags that we need consist of the following:

-fPIC
-mllvm,-instcombine-lower-dbg-declare=0
-mllvm,-import-instr-limit=30
-no-canonical-prefixes
-O2
-nostdlib++
--sysroot=../../build/linux/debian_bullseye_amd64-sysroot
-c

Make sure to rewrite the --sysroot flag to be an absolute path. Setting it to the output of the following should work:

echo $WORKING_DIR/chromium/src/build/linux/debian_bullseye_amd64-sysroot

Now, add the commands to the global_command_override section in the corpus_description.json file. Afterwards, the global_command_override section in the file should look something like the following:

"global_command_override": [
  "-fPIC",
  "-mllvm",
  "-instcombine-lower-dbg-declare=0",
  "-mllvm",
  "-import-instr-limit=30",
  "-no-canonical-prefixes",
  "-O2",
  "-nostdlib++",
  "--sysroot=/path/to/workdir/build/linux/debian_bullseye_amd64-sysroot",
  "-c"
]

Now you should have a properly prepared Chromium thinLTO corpus.

PGO Path Rewriting

NOTE: This step is only necessary if you are working on a non-thinLTO corpus. If you are working on a thinLTO corpus, making changes outlined in the section below will result in an error when you try and generate a default trace.

It is essential to have PGO data when training the regalloc model. However, the Chromium build process uses relative paths when referencing PGO profiles. This needs to be fixed by replacing the default -fprofile-instrument-use-path flag with one that uses an absolute path. First we need to know the profile being used. The profiles are located in $WOKRING_DIR/chromium/src/chrome/build/pgo_profiles and they will have a *.profdata extension. If you have never resynced your checkout, there should be just one file. Then to find the absolute path of the file, run the following command:

realpath -s $WORKING_DIR/chromium/src/chrome/build/pgo_profiles/*.profdata

This should output an absolute path to a *.profdata file. Then, open the gin config for the regalloc problem which should be located at compiler_opt/rl/regalloc/gin_configs/common.gin within the ml-compiler-opt root. Then, replace the line

problem_config.flags_to_replace.replace_flags={}

With this:

problem_config.flags_to_replace.replace_flags = {
  '-fprofile-instrument-use-path': '<path to profdata from above command>'
}

eg:

problem_config.flags_to_replace.replace_flags = {
  '-fprofile-instrument-use-path': '/home/aiden/chromium/src/chrome/build/pgo_profiles/chrome-linux-main-1665359392-180123bdd1fedc45c1cbb781e2ad04bd98ab1546.profdata'
}

Adding a flag to remove warnings related to PGO profdata hash mismatches can also be helpful to not clutter up your output. These can be added in by adjusting this line:

problem_config.flags_to_add.add_flags=()

to

problem_config.flags_to_add.add_flags=('-Wno-backend-plugin',)

Collect the Default Trace and Generate Vocab

Before we run reinforcement training, it is best to train the model using behavioral cloning on the default heuristic. First off, we need to collect a trace of the decisions the default heuristic is making:

PYTHONPATH=$PYTHONPATH:. python3 compiler_opt/tools/generate_default_trace.py \
  --data_path=$WORKING_DIR/corpus \
  --output_path=$WORKING_DIR/default_trace \
  --gin_files=compiler_opt/rl/regalloc/gin_configs/common.gin \
  --gin_bindings=clang_path="'$WORKING_DIR/llvm-build/bin/clang'" \
  --sampling_rate=0.2

This will compile 20% of the corpus and save all of the regalloc eviction problem instances it encounters into the /default_trace file.

After we have collected a default trace, an optional step is to regenerate the vocab that is used to normalize some of the values that get fed to the ML model:

rm -rf ./compiler_opt/rl/regalloc/vocab
PYTHONPATH=$PYTHONPATH:. python3 compiler_opt/tools/generate_vocab.py \
  --input=$WORKING_DIR/default_trace \
  --output_dir=./compiler_opt/rl/regalloc/vocab \
  --gin_files=compiler_opt/rl/regalloc/gin_configs/common.gin

This isn't completely necessary as there are already default values stored within the repository in the ./compiler_opt/rl/regalloc/vocab folder, but it definitely doesn't hurt to regenerate them. If adding/modifying features, it is necessary to regenerate the vocab.

Now that the vocab is present (or has been regenerated) and we have a default trace, we can start to train the model using behavioral cloning to mimic the default heuristic:

PYTHONPATH=$PYTHONPATH:. python3 compiler_opt/rl/train_bc.py \
  --root_dir=$WORKING_DIR/warmstart \
  --data_path=$WORKING_DIR/default_trace \
  --gin_files=compiler_opt/rl/regalloc/gin_configs/behavioral_cloning_nn_agent.gin

This script shouldn't take too long to run on decently powerful hardware. It will output a trained model in the directory specified by the --rootdir flag, in this case $WORKING_DIR/warmstart.

Reinforcement Learning

Now that we have a model that has been warmstarted based on the default heuristic, we can now proceed with RL learning so that the model can improve beyond the performance of the default heuristic:

PYTHONPATH=$PYTHONPATH:. python3 compiler_opt/rl/train_locally.py \
  --root_dir=$WORKING_DIR/output_model \
  --data_path=$WORKING_DIR/corpus \
  --gin_bindings=clang_path="'$WORKING_DIR/llvm-build/bin/clang'" \
  --gin_files=compiler_opt/rl/regalloc/gin_configs/ppo_nn_agent.gin \
  --gin_bindings=train_eval.warmstart_policy_dir=\"$WORKING_DIR/warmstart/saved_policy\"

This script will take quite a while to run. Probably most of the day on pretty powerful hardware (~100+ vCPUs), and potentially many days on less powerful hardware.

Evaluating the Policy

If you interested in seeing how the trained policy performs, you can go through two different avenues. You can run the generate_default_trace.py script to get info on the reward (reduction in number of instructions) over a specific corpus. However, this still doesn't tell the whole story for the regalloc case and actual benchmarking is needed. There is also some tooling available in this repository to run benchmarks in Chromium and the llvm-test-suite using performance counters to track instructions executed, loads, and stores, all of which are metrics that show how the model is performing.

Evaluating the Model With Reward Metrics

To evaluate a trained policy (for example looking at the output from RL training in $WORKING_DIR/output_model/saved_policy), run the generate_default_trace.py script with some flags to tell it to output performance data:

PYTHONPATH=$PYTHONPATH:. python3 compiler_opt/tools/generate_default_trace.py \
  --data_path=$WORKING_DIR/corpus \
  --gin_files=compiler_opt/rl/regalloc/gin_configs/common.gin \
  --gin_bindings=config_registry.get_configuration.implementation=@configs.RegallocEvictionConfig \
  --gin_bindings=clang_path="'$WORKING_DIR/llvm-build/bin/clang'" \
  --output_performance_path=$WORKING_DIR/performance_data.csv \
  --policy_path=$WORKING_DIR/output_model/saved_policy

This will collect reward data over the entire corpus. If you want it to run faster and don't care about collecting data over the whole corpus, you can set the --sample_rate flag to a desired value to only operate over a portion of the corpus.

Evaluating the Model With Benchmarking

See the documentation available here

Deploying the New Policy

To compile the model into LLVM using Tensorflow AOT compilation, create a new folder and run a CMake configuration with the following commands:

mkdir $WORKING_DIR/llvm-release-build
cd $WORKING_DIR/llvm-release-build
cmake -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DTENSORFLOW_AOT_PATH=$(python3 -c "import tensorflow; import os; print(os.path.dirname(tensorflow.__file__))") \
  -DLLVM_ENABLE_PROJECTS="clang" \
  -DLLVM_RAEVICT_MODEL_PATH="$WORKING_DIR/output_model/saved_policy" \
  $WORKING_DIR/llvm-project/llvm

Then run the actual build:

cmake --build .

Now, you should have a build of clang in $WORKING_DIR/llvm-release-build/bin that you can use to compile projects using the ML regalloc eviction heuristic. To compile with the ML regalloc eviction heuristic, all you need to do is make sure to pass the -mllvm -regalloc-enable-advisor=release flag to clang whenever you're compiling something.