Optimal string copy performance
|
From: Jonathan de Boyne Pollard on 18 Mar 2010 02:01 > > > I need to copy some RAM memory area to the screen buffer with as good > a performance as possible. > > The prerequisites/restrictions are: > > 1.) it has to work under OS/2 - that is with 32-bit legacy - no 64-bit > instructions > > 2.) if possible it should support INTEL and AMD CPUs but I might think > about processor specific implementations > > 3.) it should give preference to low irq latency - that is, if a lot > of IRQs need to be serviced, copy performance (and therefore video > refresh performance) can be sacrified against low irq latency/quick > irq response > > 4.) the solution should outperform the "rep movsd" instruction > Is your VRAM write-back or write-through? Is it write-combining? All of these things will affect whether your choice of data manipulation instructions is even significant. (-:
From: Terje Mathisen "terje.mathisen at on 18 Mar 2010 03:52 Jonathan de Boyne Pollard wrote: >> >> >> I need to copy some RAM memory area to the screen buffer with as good >> a performance as possible. >> >> The prerequisites/restrictions are: [snip] >> 4.) the solution should outperform the "rep movsd" instruction >> > Is your VRAM write-back or write-through? Is it write-combining? All of > these things will affect whether your choice of data manipulation > instructions is even significant. (-: Even more interesting is the issue of why he thinks the code could/should affect hw interrupt response times! This pretty much means that he's thinking about running this copy code in kernel mode, with interrupts disabled, something that stopped making sense around the time of the original EGA/VGA adapters! Terje PS. Why the original cross-poster thinks this is even marginally relevant to comp.arch beats me.:-( -- - <Terje.Mathisen at tmsw.no> "almost all programming can be viewed as an exercise in caching"
|
Pages: 1 Prev: Infinite geometry 3D technology (no more polygons?) Next: ATI Catalyst 10.3 Preview Edition |