[Ffmpeg-devel] int vs. float: Hard Numbers

[Mike Melanson]

Michael Niedermayer wrote:
> if it where so simple ...
> but it isnt, the code is flawed and wont tell you much usefull
> the first problem is that its all executed just once, so the code wont be in 
> the code cache, practically no speed critical piece of code would be executed 
> just once ...

	Could we mitigate this problem by "warming up" the cache by running the 
function 1000 times and then profiling run #1001?

> the second problem is that every instruction uses the result of the previous, 
> while this gives one interresting value for comparission we should also 
> meassure how fast cpus are if the instructions dont depend upon each other

	I wanted to stay away from memory accesses, naturally. So, how about a 
series of instructions that rotated between using different registers so 
that the CPU could perhaps re-order the instructions during execution?

	Thanks...
-- 
	-Mike Melanson