Hercules

Hercules is a compiler for productively programming heterogeneous systems. The project is structured as follows:

• hercules_ir: Core Hercules IR definition and analyses.

• hercules_opt: Transformations on Hercules IR.

• hercules_cg: Code generators that compile Hercules IR functions into device-specific code (currently LLVM, CUDA, or Async Rust).

• hercules_rt: Assorted runtime utilities.

• juno_frontend: Frontend for the Juno application language, compiles into Hercules IR.

• juno_scheduler: Frontend for the Juno scheduling language, controls transformations on Hercules IR.

• juno_build: Build utilities for incorporating Hercules code in Rust programs.

• juno_utils: Assorted utilities for the Juno frontends.

• hercules_samples: Samples / tests manually defining Hercules IR textually.

• juno_samples: Samples / tests of Juno programs.

• hercules_test: Assorted tests for the Hercules compiler.

• paper_resources: Assorted figures for publications.

Setup

Hercules is simple to setup. Just clone the repository:

https://gitlab.engr.illinois.edu/llvm/hercules.git

And run all the tests:

cargo test

Or a single sample (see Cargo.toml for a full list):

cargo test -p juno_matmul

If you want to see the output LLVM / CUDA / Rust code from the Hercules compiler, run with the -vv flag.

cargo test -p juno_matmul -vv

The written samples are setup with a cuda feature - if this feature is provided, the Juno program will be targeted onto the GPU, rather than the CPU. This only works if you have the CUDA toolkit installed:

cargo test -p juno_matmul --features=cuda

Some samples also include benchmarks. These use criterion for measurement. For example (edge_detection requires the opencv feature to be enabled to build OpenCV before running):

cargo bench -p juno_edge_detection --features=opencv