vectormath

Vectormath for Go

Programming language: Go

License: GNU General Public License v3.0 or later

Tags: Science And Data Analysis    

README

Vectormath for Go

This is an adaptation of the scalar C functions from Sony's Vector Math library, as found in the Bullet-2.79 source code. (Note that the current svn for Bullet retains a subset of this library in a different location.) The full library can be found in the following directory of the Bullet-2.79 source tarball:

Extras/vectormathlibrary/include/vectormath/scalar/c

Special thanks to Sony Computer Entertainment Inc. for allowing the use of their Vector Math library under a BSD-style license, which made this possible.

Approach

Start with a direct conversion of the Sony library to Go. (This was shipping code, so we will assume it's correct.) Then iterate to make it more Go-like, while trying to balance the twin goals of same/better performance and maintaining readability.

The C version of the library, as opposed to the C++ version, was chosen as a starting point mainly to avoid the distraction of function renaming, as the C++ version of Sony's library relies somewhat on operator and function overloads, which have no direct equivalent in Go.

Note that the original C library provides two versions of each file, e.g., vec_aos.h and vec_aos_v.h. I've brought over the former as opposed to the latter, since this simplifies things for beginners. The 'result' argument is always a pointer passed as the first argument to a function. Currently, the only types used in this library that are passed by value are simple types, e.g., float and int. Everything else is a pointer. (This should, in theory, simplify memory thrash considerations as well. In practice.. I have no idea. :)

Future Direction

Further research is required for determining:

• What makes the most sense for optimizing this code for SIMD
• Whether we should be passing vectors, etc., by value instead of reference
• If we should continue to declare vectors as individual values as opposed to an array of values. This would avoid the branching currently required to access members by index, but would disallow accessing members by common name (x, y, z and w) due to Golang's lack of a union equivalent.
• Whether it makes sense to stay with 32-bit, or if we would get the same performance with 64-bit

Feedback on this library is welcome and appreciated, though I make no promises about my ability to deliver on anything beyond what you see here. :)

License

I have licensed my modifications under a license similar to that of the original library on which this is based. See the LICENSE file for more details.