Reactant.jl is an innovative package designed to enhance the performance of Julia functions through the integration of MLIR and XLA, facilitating high-performance execution across various devices, including CPU, GPU, and TPU. This project is currently under active development and may undergo changes, making it essential for users to stay updated with the latest releases.

Key Features

• Automatic Differentiation: Reactant employs EnzymeMLIR for automatic differentiation, optimizing functions with minimal effort.
• Optimized Control Flow: The system compiles Julia functions while preserving the original control flow, enhancing overall performance by eliminating type instabilities.
• New Array Types: At the core of Reactant are two new array types: ConcreteRArray for device data storage and TracedRArray that ensures structured data handling.

Usage Examples

Creating a ConcreteRArray from a Julia array is straightforward:

using Reactant

julia_data = ones(2, 10)
reactant_data = Reactant.ConcreteRArray(julia_data)

For custom data structures, Reactant provides a method to trace structures automatically:

struct Pair{A,B}
   x::A
   y::B
end

pair = Pair(ones(3), ones(10))
reactant_pair = Reactant.to_rarray(pair)

Compilation of functions using ConcreteRArray is seamless:

input1 = Reactant.ConcreteRArray(ones(10))
input2 = Reactant.ConcreteRArray(ones(10))

function sinsum_add(x, y)
   return sum(sin.(x) .+ y)
end

f = @compile sinsum_add(input1, input2)

# Execute the compiled program
f(input1, input2)

Device Compatibility

Reactant simplifies the process of running programs on accelerators. For instance, to execute code on a GPU:

using Reactant
Reactant.set_default_backend("gpu")

With this setup, all operations are executed on the GPU without the need for additional dependencies like CUDA.jl, making Reactant a powerful tool for optimizing Julia functions in high-performance computing environments.