Overview

A development and dispatch library for high-performance ML kernels

Popcorn provides a common API for high-performance machine learning kernels. It includes 96 kernels and more than 100 optimized implementations, selected using hardware-specific benchmark data.

Each kernel has a PyTorch reference. Validation failures and unmeasured cases remain explicit and are never treated as successful results. Popcorn also provides tools to develop, test, benchmark, and deploy first-party kernels.

Terminology

These terms have specific meanings in Popcorn:

TermMeaning
KernelA pure unit of work with a fixed signature and semantics defined by a ground-truth reference.
InputA valid configuration of arguments for a kernel call.
CaseA call configuration defined by tensor metadata and other arguments. Tensor values are excluded.
ImplementationA function that matches the output of a kernel's reference for a subset of valid cases.
BackendA library or collection of kernel implementations.

Guides

See the kernel explorer for kernel definitions, backends, and measured performance.

On this page