This small project implements a single function, cbrt_ps.
The function computes cubic root of 4 FP32 values in a vector register.
The implementation requires SSE 4.1 instruction set, and can optionally use AVX 1 if available. So far, no NEON version is there.
The implementation is OS agnostic, should compile for all of them as long as the target CPU supports SSE 4.1.
Tested with Visual Studio 2022 on Windows 10, and GCC 7.4 on Linux.
The implementation can be trivially generalized to 32-bytes AVX vectors.
Usage
Copy-paste CbrtPs.hpp header into your project, include the header into your source or header file[s].
As you see from the comment in that header, it comes with a copy paste friendly terms of MIT license.
References
-
The Halley’s refinement method is from “In Search of a Fast Cube Root” article by metamerist
-
Denormals handling and initial estimate ported from FreeBSD’s version of
cbrtf()library function -
Integer division by 3 back-ported from assembly made by clang 13