WORK IN PROGRESS

"Harp" is a transpiler for multidimensional array computation, deep learning, and computer science, inspired by tinygrad and luminal. It automatically generates optimal C kernels for CPUs or GPUs from a computational graph of a multidimensional array.

• Tensor / Autograd Frontend: A frontend that makes it easy to handle tensor calculations, automatic differentiation, etc.
• Lowerer: Responsible for converting the computational graph into a C-like abstract syntax tree (AST).
• Renderer: Renders the AST as source code in the target language.
• Compiler: An interface that abstracts compiler calls.

In Harp, optimization is done in two main stages:

• Computational graph optimization: Tensor-wise graph optimizations are performed (e.g., fusing multiple element-wise operations into a single operation node).
• AST optimization: Removing unnecessary operations found in the abstract syntax tree (such as meaningless operations like a + 0) and unrolling loops.

This repository is dual licensed under Apache 2.0 and MIT, so you can use it either way if you need to.