On Friday 27 February 2004 15:14, Grendel wrote: > On Fri, 27 Feb 2004, Sami Näätänen uttered the following immortal > > On Friday 27 February 2004 08:21, Grendel wrote: > > > For example under mandrake mencoder takes the same time to rip a > > > dvd as under gentoo, there is no improvement obtained by > > > compiling everything from source. > > > > Mencoder is not a good example, because it doesn't use the set > > cflags at all, but simply uses the hand written assembly routines > > and compiles C-sources with pre set cflags. > > Lame when used to encode a large wav file doesnt offer any advantage > either :( > > So what are the examples you suggest?
Heavily math oriented software without the hand optimized assembler. The obvious example is raytracers. Those have sometimes code that is wrote to the vector extensions of CPU's in mind. These pieces of code are faster by magnitude not just few percents, when compared to normal instruction sets. > What most of these people who say that "compiling from source with > architecture specific flags is faster than generic binaries" and > hence gentoo is faster is not correct for most (>90%) applications > the bottleneck isnt the cpu, ie most applications are event driven, > they spend most of there time waiting for something to happen. like > waiting for a button press, waiting for data to reach a socket etc. > So in those cases no amount or source compiling will afford any > benefit as the bottleneck is not the cpu, its the wait loop. In fact most people over optimize and in fact get slower systems than the generic binaries used by binary distributions. After all CPU's have very limited level 1 cache and expecially Intel CPU's are very low on it. By poluting that cache when using -O3 can result in fraction of the speed of the same program compiled in -O2, which doesn't unroll all loops. -- [EMAIL PROTECTED] mailing list
