videoprocessing

[Guest]

Can I imagine myself ( as user, not ASM coder but Z80 )
typical numbers (range ) of what is speed gain using
one of these CPU optimizations ?
I know, it has to depend on code, but just as rough imagine ?

[Clouded]

kassandro is able to quote scarily precise speed up factors, so I will leave the main answer to him... but for what it's worth, in my last assembly project, x5 speed up. Range varies heavily on either side of that...
_________________
a.k.a. mg262

[kassandro]

The speed gain of using SSE/SSE2/SSE3 depends very much on the subject and probably to a lesser extent also on the CPU.
With SSE you can process 8 8 bit quantities or 4 16 bit quantities simultaneously. With SSE2/SSE3 you can process 16 8 bit quantities or 8 16 bit quantities simultaneously, but there are some alignment problems, which can be very costly especially if you have only SSE2 and not SSE3.
There is one further advantage of SSE/SSE2/SSE3. Namely that programming is essentially jump free. This is also a disadvantage, if the code branches a lot. In C or normal assembler code, only one alternative of a branch is executed. In SSE/SSE2/SSE3 programming the CPU has always to execute both, because for some bytes one alternative has to be chosen, while for the other bytes the other alternative is necessary. In RemoveGrain I would say that the average gain is about 8 times for SSE, 9 times for SSE2 and 13 times for SSE3 on my cpu (Prescott P4 design).

[Libor aka Poutnik]

Thank you for the answer. I have already thought
it could be very fine work in SSE optimization tuning
in "SSE hostile code", where speed gain is not worthy enough for such coder effort.

[fish14765]

[kassandro]

[sade]