Skip to content

Latest commit

 

History

History
113 lines (73 loc) · 4.58 KB

README.md

File metadata and controls

113 lines (73 loc) · 4.58 KB

RL4ReAl

RL4ReAl is a retargetable Reinforcement Learning (RL) approach for solving the REgister ALlocation (REAL) problem on diverse architectures.

This repo contains the source code and relevant information described in the paper (arXiv). Please see here for more details.

RL4ReAl: Reinforcement Learning for Register Allocation, S. VenkataKeerthy, Siddharth Jain, Anilava Kundu, Rohit Aggarwal, Albert Cohen and Ramakrishna Upadrasta

Environment Setup

Setup the environment using the model/RL4ReAl/rl4real_env.yml using the following commands

conda env create -f rl4real_env.yml

Inference on Trained Models

Currently three kinds of model runners are supported by us, via the MLCompilerBridge

Note: Run ONNX based Model Runner for inference flow to get light and fast compilation.

Inference flow using gRPC Model Runner

Starting the gRPC Server:

File x86-server-rllib-sm.py under the path ml-llvm-project\model\RL4REAL\src\ contains script to start the servers.

python x86-server-rllib-sm.py --server_port=<server_port>

Make sure the following status is seen in the terminal Server Status is Running

Invoke the RegAlloc pass in gRPC mode:

{BUILD_DIR}/bin/clang -O3 -mllvm -regalloc=greedy -mllvm -mlra-inference -mllvm -ml-config-path=<config path> -mllvm -mlra-server-address=0.0.0.0:<port> <input_data_file>
  • BUILD_DIR: Directory where the project is built
  • config-path: Specify the path to the config directory
  • input_data_file: Specify path to source (.c/.cpp/.ll) files
  • port: specify the port gRPC server in running on

This process starts the communication between the compiler and the RL model.

Inference flow using Pipe Model Runner

RL4ReAl uses the Pipe to communicate with the model and uses json and bytes as data format.

Starting Pipe Server in Json/Bytes mode:

File x86-server-rllib-sm.py under the path ml-llvm-project\model\RL4REAL\src\ contains script to start the servers.

python x86-server-rllib-sm.py --use_pipe=1  --pipe_name=<pipe_name> --data_format=<bytes/json>
  • --use_pipe: Boolean to specify the pipe usage
  • --pipe-name: Specify Name of Pipe
  • --data-format: Specify the data format Bytes/Json

Invoke the RegAlloc pass in Pipe mode:

{BUILD_DIR}/bin/clang++ -O3 -mllvm -regalloc=greedy -mllvm -mlra-inference -mllvm -ml-config-path=<config_path> -mllvm -mlra-use-pipe  -mllvm -mlra-pipe-name=<pipe-name> -mllvm -mlra-data-format=<data-format> <input_data_file>
  • BUILD_DIR: Directory where the project is built
  • config-path: Specify the path to the config directory
  • input_data_file: Specify path to source (.c/.cpp/.ll) files
  • pipe-name: Specify name of the Pipe used in server
  • data-format: Specify the data format used in the server

Inference flow using ONNX Model Runner

ONNX being an in-process model runner is an effective means of compiler deployment. It is light and fast and runs pre-trained model files differing from a communication based model runner.

Invoke the RegAlloc pass in ONNX mode:

{BUILD_DIR}/bin/<clang / clang++> -O3 -mllvm -regalloc=greedy -mllvm -mlra-inference -mllvm -ml-config-path=<config_path>  -mllvm -rl-inference-engine <input_data_file>
  • BUILD_DIR: Directory where the project is built
  • config-path: Specify the path to the config directory
  • input_data_file: Specify path to .c/.cpp/.ll files

Dumping Model files

Start the Pipe or gRPC server along with the flag -dump_onnx_model=1

python x86-server-rllib-sm.py --server_port=50031 --use_pipe=rl4realpipe --data_format=json --pipe_name=default_pipe5 --dump_onnx_model=1

  • The Model files will be dumped in model/rel4real/src in the directories color_node, select_node, select_task and split_node directories respectively.

  • Rename the Model files in each directory as <MODEL_TYPE>_MODEL_PATH.onnx. e.g., COLOR_NODE_MODEL_PATH.onnx

  • Move the Model files from their respective directory to the path config/regalloc/onnx-checkpoint

  • Make sure to regenerate new ONNX models for new checkpoints.

Model Training: Refer to Model Training