Am Montag, 23. Februar 2004 19:32 schrieb Alan Cox: > On Llu, 2004-02-23 at 16:35, Ian Romanick wrote: > > Alternately, this might be a case where writing a little assembly > > conversion routine would help. If we assume at least a 486 instruction > > set (eh-hem!), the core of the BGR to ARGB conversion should take about > > 6 or 7 instructions, including the pointer increments. It may even be > > worthwhile to do an MMX optimized version. Dunno. > > It should actually be hard to write one that doesnt saturate memory > bandwidth. I guess prefetch instructions might help a bit. For raw > memcpy the kernel prefetches about 320 bytes ahead in order for prefetch > to actually kick in and be useful.
What about this (Not only for "AMD"): AMD recommends to perform memory copies with backward read operations instead of prefetch. http://marc.theaimsgroup.com/?l=linux-kernel&m=103548024914815&w=2 Paper: http://208.15.46.63/events/gdc2002.htm http://cdrom.amd.com/devconn/events/gdc_2002_amd.pdf Start at page 18 and see page 31 for a C++ example. > For texture RGB888 is that 24bit packed or with 8bits of empty ? Greetings, Dieter ------------------------------------------------------- SF.Net is sponsored by: Speed Start Your Linux Apps Now. Build and deploy apps & Web services for Linux with a free DVD software kit from IBM. Click Now! http://ads.osdn.com/?ad_id=1356&alloc_id=3438&op=click -- _______________________________________________ Dri-devel mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/dri-devel
