Numba Architecture

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.

Core Entry Points

Numba has several modes of use:

  1. As a run-time translator of Python functions into low-level functions.
  2. As a call-time specializer of Python functions into low-level functions.
  3. As a run-time builder of extension types.
  4. As a compile-time translator of Python modules into shared object libraries.
  5. Template instantiation.

The following subsections describe the primary entry points for these modes of use. Each usage mode corresponds to a specific set of definitions provided in the top-level numba module.

Run-time Translation

Users denote run-time translation of a function using the numba.jit() decorator.

Call-time Specialization

Users denote call-time specialization of a function using the numba.autojit() decorator.

Extension Types

Numba supports building extension types using the numba.jit() decorator on a class.

Compile-time Translation

Users denote compile-time translation of a function using the numba.export() and numba.exportmany() decorators.

Translation Internals

Towards More Modular Pipelines

The end goal of building a more modular pipeline is to decouple stages of compilation and make a more modular way of composing transformations.

  • State threaded through the pipeline

    1) AST - Abstract syntax tree, possibly mutated as a side-effect of a pass.

    2) Structured Environment - A dict like object which holds the intermediate forms and data produced as a result of data.

  • Composition of Stages
    • Sequencing
    • Composition Operator
    • Error handling and reporting in pass failure.
  • Pre/Post Condition Checking

    • Stages should have attached pre / post conditions to check the success criterion of the pass for the inputted or resulting ast and environment. Failure to meet this conditions should cause the pipeline to halt.

Modularity

Note: recursive definitions

jit     := parse o link o jit
pycc    := parse o emit o link
autojit := cache o autojit
cache   := pipeline o jit

blaze   := mapast o jit

Diagram

Block diagram:
                 Input
                    |
+----------------------+
|          pass 1      |
+--------|----------|--+
       context     ast
         |          |
  postcondition     |
         |          |
  precondition      |
         |          |
+--------|----------|--+
|          pass 2      |
+--------|----------|--+
       context     ast
         |          |
  postcondition     |
         |          |
  precondition      |
         |          |
+--------|----------|--+
|          pass 3      |
+--------|----------|--+
       context     ast
         |          |
  precondition      |
         |          |
         +----------+-----> Output

Discussion: Pipeline Composition

We can do composition in a functional way:

def compose_stages(stage1, stage2):
  def composition(ast, env):
    return stage2(stage1(ast, env), env)
  return composition

pipeline = compose_stages(...compose_stages(parse, ...), ...)

Or, we can do composition using iteration:

for stage in stages:
  ast = stage(ast, env)

Whether the end result is a function or a class is also still up for discussion.

Proposal 1: We replace the Pipeline class to use a list of stages, but these can either be functions or subclasses of the PipelineStage class.

Discussion: Pipeline Environments

Proposal 1: We present an ad hoc environment. This provides the most flexibility for developers to patch the environment as they see fit.

Proposal 2: We present a well defined environment class. The class will have well defined properties that are documented and type-checked when the internal stage checking flag is set.