Q3Map3 - Discussion and Development

[Eraser]

AFAIK not yet

[VolumetricSteve]

How are you handling the math for this? Is it going to be mostly single precision or double?

[Silicone_Milk]

@DaEngineer - It'll be done... when it's done. This isn't something I'm actively working on full-time. It's just a hobby project.

@Yatta_Yatta-O_o - Glad to see more interest in this. There isn't a working build that makes immediate use of OpenCL available right now.

@VolumetricSteve - Single precision. And, yes, I was seeing an approximately 7x increase in speed when handling float multiplication (dealing with a large number of floats). The more data there is for the gpu to throw itself at, the greater the speed increase we'll see.

[VolumetricSteve]

http://www.amd.com/us/products/desktop/ ... iew.aspx#2

"Processing power (single precision): 4.64 TeraFLOPS "

My body is ready.

[Silicone_Milk]

Working on some vector math functions and scratching my head over differences in pointer address space.

Nothing huge to report.

[^misantropia^]

SM, you should put it on GitHub so that people who want to can follow it.

[Silicone_Milk]

I'm working on it. Have an account set up on GitHub with a q3map3 repo just no files.

I'm writing some rules for git to ignore certain file types.

[VolumetricSteve]

I'm curious about precision management on the GPU, since you're using single precision, does that mean, (in the case of a GTX 580) you'd have 512 cores all doing 32-bit calculations, but...if they're built to handle double precision, does that mean the a single precision calulcation takes up a core that could be handling twice as much....or is it the other way around? Do the GPUs soft-handle double precision, by using two single precision cores in conjunction? I'm just curious because, depending on how it's all managed, maybe there's a way to force more single precision calculations down a double precision pipe?

[^misantropia^]

if they're built to handle double precision, does that mean the a single precision calulcation takes up a core that could be handling twice as much....

Yes. That 'other way around' thing makes my head hurt just thinking about the logic gates you'd need to farm out a unit of work over two cores.

[Silicone_Milk]

Haven't had time to work on this. Joined a startup and have been working full time on an XBLA project.

Will still be working on this though.

[VolumetricSteve]

I guess this has kinda fallen to the bottom of the "to do" list, but I was wondering if you'd posted the code you had up anywhere or if you had plans to?
I'd be really curious to take a look and see what I could get working.

[o40]

Is this project dead?

[^misantropia^]

No but it's pining for the fjords.

[fKd]

release the hounds!

[Silicone_Milk]

That made me chuckle misantropia

I really didn't want to post in this thread until I finished something that I could show visually. Otherwise it's just stringing you guys along on empty "it's coming" statements.

I had written a small opengl scene that is basically a box with smaller boxes in it with the intention of re-implementing some basic software lightmap generation which I could use as a testing point for breaking it down on the GPU before finally moving to continuing the work on q3map3.

But, not having a degree creates a somewhat volatile career situation so I had to stick that on the backburners yet again.

In short, I'm still interested and excited to work on this, I just don't have the time for it at the moment.

[o40]

I was messing around yesterday with using SIMD (_mm_dp_ps) for replacing the dot product calculations but I did not get any performance boost.

Maybe I can improve the AddVector functionality instead.

EDIT:

The main reason why I want shorter compile times is because of large generated maps such as this one:

http://ws.q3df.org/map/mineqraft/

VIS took 6 hours more or less.

[^misantropia^]

@o40: check gcc's assembly output (compile with -S), it might already be generating SIMD instructions.

[o40]

Good tips! Will do when I feel the urge.