Intel Compiler is now free

[DaMan]

A free 90 day renewable license. Includes VTune too.
It doesn't like GZ's math library. Complains about division by zero. Can be worked around by undefing INFINITIES, NANS & MINUSZERO but it crashes shortly after starting. Software render is faster by 10% on the opening of Entryway though.

[Rachael]

This is definitely interesting and I am very intrigued.

I know you didn't point this out exclusively for the purpose of GZDoom, however that is a consideration a lot of people will have for it. Given the issues I am not 100% sure it is time to make the switch - and that's something Graf would have to decide on, anyway, depending on how the licensing scheme works and if it would work for an open source project such as this.

From what I am reading, using an Intel compiler produces sub-optimal code for AMD processors - which is no surprise, AMD is a competitor, I don't know if this is on purpose on Intel's part or not, but it legitimately might not be - Intel knows their own processors best and they're not paid to worry about how to make a competitor's CPU run better. So who knows if AMD users would even benefit from the speed boost of switching to the Intel compiler.

To be quite honest, 10% is not as big of a performance gain as I expected it to be.

I think over the long term, we definitely could do some things to make GZDoom compatible with the Intel compiler, though, to add it in as an option for those who wish to use it, but I doubt any official builds will use it anytime soon. If the math is in the state that you say it is in, though, keeping determinism between systems could be problematic.

[Cacodemon345]

Will be waiting for Graf Zahl to say anything about this.

---

Take this with a grain of salt, but would it speed up performance of GZDoom CPU operations?

[Graf Zahl]

I once ran a test with the hardware renderer on Frozen Time where it didn't make any difference, apparently the CPU cache is the most limiting factor here, which is also evidenced by seeing no difference between 32/64 bit and fixed point/float. Normally 64 bit should run faster but if the memory accesses are what defines the speed, better optimized code won't do much of a difference.

[Rachael]

In that case, then, it sounds like the only real thing worth doing over the long term is adding some #ifdefs and #defines to get it to compile properly for those who prefer to use it, but seeing as how there's no significant speed boost it doesn't seem like it's worth switching over. It's like using Clang, really.

[leileilol]

I wouldn't bother with the legal issues over a renewable license part limiting the scope to make reproducible builds in the future as there's always the possibility they'd revoke it.

[Matt]

I wouldn't bother with the legal issues over a renewable license part limiting the scope to make reproducible builds in the future as there's always the possibility they'd revoke it.

Yeah, this seems very much like "free with every purchase which price already includes our cost of making this free thing which is totally free by the way" sort of free.