SH4ZAM - Accelerated Vector Math on the Sega Dreamcast

The talk starts with the motivation for SH4ZAM: a Dreamcast port of Grand Theft Auto 3 that ran poorly because of slow floating-point math. The speaker explains how the library grew from hand-written SH4 assembly into a standalone project with scalar, trigonometry, vector and matrix APIs, each offered in C and C++ headers. Examples include a hardware reciprocal square root used for fast division and square roots, and a single instruction that computes sine and cosine together, which replaced lookup tables and freed memory in a game that used them.

The middle section covers the vector and matrix paths. A four-component dot product instruction and a transform instruction that multiplies a vector by a 4x4 matrix held in a register bank give large speedups over a standard C math library. Demos from Star Fox 64, Mario Kart 64 and Diddy Kong Racing show frame-rate and audio gains, including vectorized audio synthesis. A software back end lets the same math code build for non-Dreamcast targets.

The closing takeaway is that careful use of a console's vector hardware, inline assembly and modern compilers can pay off on old machines. The speaker argues the same lessons apply to other game consoles and to performance-focused math libraries generally. The library is bundled with a community Dreamcast SDK, and the speaker invited questions afterward.