pouët.net

That was interesting. It might be fun to look at the assembly dump to see how much gets optimized out, though, before dismissing the branchless clamping completely =)

use unsigned comparison and CMOVcc

mmx indeed

Using no conditionals?

// if (number < 0) number = 0;
number &= ~number >>31;
// if (number > 0xFF) number = 0xFF;
number = number & 0xFF | ((-(((unsigned)number) >> 8)) >> 24);

Alternatively:

// if (number < 0) number = 0;
number &= ~number >>31;
// if (number > 0xFF) number = 0xFF;
number = 0xFF-number;
number &= ~number >>31;
number = 0xFF-number;

Whether this is faster than conditionals should depend on your platform. Platforms with built-in clamping, min/max or "set register if condition is true"-like instructions are unlikely to be faster with this code.

las: Now try benchmarking again with something that isn't trivially branch-predictable by the CPU. :-)

Sesse: Does it really matter? The compiler probably turned those branches into CMOVs anyway...

yeh.... does it really matter? :) seems a theme on pouet these days.. in this day and age of optimised compilers why are people still discussing this? I'm sure there are many other flaws in your code that impact perf more :)

Sesse:


Benchmark it yourself ;)

Of course the real power of nonconditional version is that you can do simd without simd..

sum0 = a0+b0 | (a0+b0>>8&0x801008)*255;
sum1 = a1+b1 | (a1+b1>>8&0x801008)*255;
*dest = sum0<<1&0xf81f0000 | sum0<<16&0x7e00000 | sum1>>15&0xf81f | sum1&0x3e0;