[XviD-devel] Floating Point Operations
Christoph Lampert
xvid-devel@xvid.org
Tue, 31 Dec 2002 16:32:23 +0100 (CET)
Hi,
I'll do "real" GMC next, which involves some matrix operations using
float(!) values (at least at the current stage).
I hope this won't be eating much CPU, but it involves:
* several scalar products of vectors which are of the length 2*N
where N is the total number of macroblocks (=1620 for full PAL,
=320 for CIF).
* calculating the inverse of a symmetric 4x4 matrix
(or rather solving a system of 4 linear equations for a 4-vector)
* entries between the order of 10^7 and 10^{-7}
Before I create data structures for this, does anyone have suggestion
how to implement this in the fastest possible way? How to make this
compatible for future 3dnow, SSE, SSE2 versions?
gruel