Also, here's a glimpse at my rotoscoping process... #LUNARK #pixelart

The MOVD (MMX register to general purpose register) instruction is really slow in the Cyrix FPU/MMX unit, which is why I think it's much slower than it should be.

More tests! This one hurts: the IBM/Cyrix 6x86MX PR333 (tested at 225 MHz, 3.0×75):
- Cyrix 6x86 code path: 66.147 fps
- MMX code path: 46.160 fps (-30%)
Ouch.

Another CPU that benefits from the MMX code: the Rise mP6. Running the exact same test:
- 486 code path: 37.432 fps
- MMX code path: 49.634 fps (+32.5%)

Optimized it a bit further, and now we're at 57.120 fps (that's a 16% gain!). It's a shame it's not possible to use MMX registers for memory addressing, and also that MOVD can't transfer the high 32 bits of an MMX register to a general purpose register.

I've also tested the IBM 6x86MX at 225 MHz (75×3). It really doesn't like the new MMX column renderer:

Unrolled non-MMX code: 57.460 fps
Rolled 2-pixels-per-loop MMX code: 45.096 fps (-21.5%)

Yikes.

I HAVE DECODED BURGER KING. I HAVE THE RAW HIGH QUALITY MODELS.

AHAHAHAHAHA