Fast math and approximation

Accept bounded error for speed: polynomial approximations, fast inverse square root, reduced precision and fast-math flags.

511proven techniques

23pitfalls: tried, then slower

238repositories

210from papers

Examples from real commits

xtensor-stack/xsimd 6.8× Add 2^52 to round to integer, subtract 2^52, and adjust for floor using bitwise ops. proof: commit → blender/cycles 2× Stochastically sample a single trilinear tap to approximate tricubic interpolation, reducing taps from 64 to 8. proof: commit →

Languages

C++180
C39
CUDA34
Python20
Rust17

Hardware

CPU111
x86111
NVIDIA GPU74
GPU63
ARM36

Fields

graphics143
math128
GPU71
compilers45
networking39

Where these come from

Repositories with the most techniques on this topic in the database:

genn-team/genn · 4google/highway · 4kornia/kornia-rs · 4google/swiftshader · 4Redot-Engine/redot-engine · 4mrdoob/three.js · 3google/brotli · 3Dao-AILab/flash-attention · 3ARM-software/optimized-routines · 3tenstorrent/tt-metal · 3

Ask Claude for the full answer

Each technique comes with its bottleneck, measured speedup, trade-offs, conditions and the commit or paper behind it, served to Claude through MCP.

claude mcp add --transport http adtomic https://www.adtomic.co/mcp

Connect Claude Pricing