Speed comparison of x265 build

[leghorn]

At http://x265.ru/en/builds one can find x265 builds compiled with either GCC, ICC, or VC++.
Does the compiler matter in any way? Or doesn't it have any impact on the conversion speed?

[LigH.de]

Most of the really time critical code for 8 bit internal resolution, and also much of 10 bit and some of 12 bit, is already available as several assembler optimized routines for different CPU instruction set extensions (SSE3/4, AVX, AVX2...).

There is little code left where a C/C++ compiler could still produce optimized code; mostly for cases which are not in the main fields of usage (e.g. 32 bit code for high bit depths, some HBD routines not yet optimized, fallback routines when you disable assembler optimized routines by --asm switch, and not so time critical code outside the encoder core).

[leghorn]

So the compiler actually doesn't matter. Good to know.

Danke!

[LigH.de]

I didn't say "not at all". Just "not much, compared to handcrafted assembly routines", which are already implemented for most common use cases.

[leghorn]

I must have misunderstood things. Nevertheless, I think in the end it's only a matter of some minutes for a 2-hour movie. Or am I going wrong again?

[LigH.de]

That's right. A difference of more than 2% would be a surprise when assembly core routines are available, especially when x265 has a good CPU utilization and can make use of AVX2 and FMA3.