x86-64 Compatability

[Ericson2314]

No description provided.

[iamgreaser]

Very important because 64-bit is awesome and you should all be porting everything to 64-bit.

With that said, porting the SSE stuff to compiler intrinsics would make this part much easier.

[aponigricon]

Approved

On Friday, April 5, 2013, iamgreaser wrote:

Very important because 64-bit is awesome and you should all be porting
everything to 64-bit.

With that said, porting the SSE stuff to compiler intrinsics would make
this part much easier.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/17#issuecomment-15981326
.

[Ericson2314]

Only v5.asm NEEDs that for x86-64, for the rest the C equivalents will do as a stop gap. On the other hand fixing subtle errors in the C types may be harder than the assembly

[Lensman]

Seems worthwhile, especially converting to intrin.

[Ericson2314]

In the end, I certainly want intrinsics and C working on x86-64

[Lensman]

Will have to be careful of __m64 datatypes. Probably have to convert them to __128 equivalents, which may mean wasted ops unless algorithms are also converted to take advantage of extra bandwidth eg;

static inline void mmxcolorsub(long _color){
*color = _m_to_int(_mm_subs_pu8(_m_from_int(_color), _m_from_int(flashbrival)));
}

[Ericson2314]

What exactly is the problem with __m64?

[Lensman]

Nothing is wrong with it, but intrinsics that support that datatype are not supported on x86_64, so it will mean hand coded assembly. I have converted hrendzsse to intrinsics, and surprisingly it runs faster than the original code. Go figure ;)
It's all 128 bit xxm registers now, and on x86_64 there are 8 more to play with too. Thought this might be agood place to tackle in my spare time as it covers 6 asm functions;
hrendzsse/3dn, hrendsse/3dn vrendsse/3dn etc. I haven't finished the fog versions, but it's trivial to do so. This should collapse 8 functions into one more manageble one. I also had a play with mem mapped files, and now have voxlap working with 4096x4096x256 faster than the original. I'd say that is some progress.

[Ericson2314]

Wow, I'll be back working on Voxlap about a week, but it looks like you might have it done before then!

[iamgreaser]

The point you made about __m64 not working on x86-64 is incorrect - it DOES work, at least in GCC.
But yeah, if you're willing to do SSE2 support, then using an __m128 version of the code should be a bit faster.

[aponigricon]

The point I made?

[Ericson2314]

Lensman I think.

[aponigricon]

Oh, well... my computer is broken... and I wont get my fingers on it's keyboard before at least two weeks, sooo...
What have you guys actually been doing?

[Ericson2314]

https://github.com/Ericson2314/henet
https://github.com/Ericson2314/hasnip

[aponigricon]

Haskell? Dude... NO... please... NOOOO...

[Ericson2314]

Yup. I may use Voxlap, but everything else will be Haskell. Seriously C is a shitty language. Rust or ATS is much better for manual memory.