[XviD-devel] transfer16x16_copy
suxen_drol
suxen_drol at hotmail.com
Mon Mar 3 09:36:14 CET 2003
hi christoph,
On Sun, 2 Mar 2003 21:14:50 +0100 (CET) Christoph Lampert <chl at math.uni-bonn.de> wrote:
> we once had transfer16x16_copy_mmx() routines by Pete.
> http://list.xvid.org/pipermail/xvid-devel/2002-July/000058.html
> Does anyone remember why and when they left us again? Or we simply forget
that first mail was about extending transfer[8x8,16x16]_copy to support
different src and dst strides. the idea was to speed up alban's direct
rending mode. however he reported memcpy() was faster, so i never
bothered to commit.
> to put them into CVS? Because it seems they were already gone in
note: transfer16x16 is only used for cbp=0 macroblocks, and reduced
resolution. there probably is a minor speed boost available.
i also recall a long time ago playing arround with transfer16x16_sse
("NOT SSE2"), which used the sse floating point movs to access 16-bytes
at a time. it would be nice bench the mmx vs. sse on a athlonxp+ddr.
-- pete; life is like a box of ammo
More information about the XviD-devel
mailing list