Numba Architecture

Contents

• Numba Architecture
  • Introduction
  • Compile entry point
  • Lowering
  • Targets

Introduction

This document serves two purposes: to introduce other developers to the high-level design of Numba’s internals, and as a point for discussion and synchronization for current Numba developers.

Compile entry point

The module numba.compiler contains the compiler entry-point compile_extra. All decorators eventually call this function. It can be broken down into the following phrases:

• 1. analyze bytecode

  • gather control-flow and data-flow information

• 2. translate into an intermediate representation
• 3. infer types using local type inference

• 4a. if (3) succeed, lower to native code (native mode)

• 4b. otherwise, assume object type for all variables in the lowering (object mode)

  • may try to extract loops in the function and recursively compile the extracted loop. (loop-jitting)

Type inference (3) will only succeed if all types have a non-object type.

Lowering

The module numba.lowering defines the lowering of the internal IR into LLVM. The lowering logic is independent of the target. Each target provides a target context that describes the target-specific lowering.

Targets

Each target implementation has two contexts:

• a typing context
• a codegen context

The typing context is equivalent to a C header file. The codegen context is equivalent to a C source file. This allows the same typing rules to be shared across multiple targets, while each of them has a different implementation.

The default type context is defined in the numba.typing subpackage. The default codegen context is defined in the numba.targets subpackage.