[Libvisual-devel] MMX benchmarks.

[Dennis S. <sy...@yo...>]

Heya everyone!

Alright, with the ongoing work on VisCPU we also started
with writing MMX routines for some heavy duty routines.

Jeko (the goom rockstar) kicked of with a 32 bits bilinear
filterer written in MMX, I partly stole/written alpha
blend code:

Everything is 640x400x4
__
Bilinear scaler (32bits)
C version:	Scale bench overlay 500 times, depth 8, interpol 1
		0m6.256s

MMX version:    Scale bench overlay 500 times, depth 8, interpol 1
	 	0m2.056s
__
Alpha overlay blitter:

C version:	Blit overlay 500 times
		0m2.058s

MMX version:	Blit overlay 500 times
		0m1.352s
__

There are many many many more areas that should be transformed
into vectorized code!, but I think this is a nice start.

Also plugins can be optimized by a great extend using mmx/sse.

But for now, I think it's most important to have mmx/sse versions
of libvisual core routines, like depth transformation, more
scalers.

Of course this is a very time consuming job, but hey ! :)

Cheers,
Dennis