OGL 3.0: Programmable filtering, blending, stencil, etc. - Sooner than we think?

[ShootMyMonkey]

from memory ~1-2 years ago with nvidia drivers (donno 50.xx-60.xx)
an average vs + fs program ~100 instructions were taking 600-700msec
now with 90.xx thats down to ~50msec on the same machine

With MS' HLSL compiler, it's gone in the opposite direction, FWIW. It used to be around 400-600 msec a year ago for a vs+ps of that size, but now that's around 650-950 msecs. Though flow control really throws it into hell -- loops bring it from ms to secs. Though I suppose that any case where you feel the need for flow control, you also have something pretty large (and I do have to deal with generated shaders that amount to a few thousand instructions). All the same, though, the end results are better than they were back then. The 2.0 targets were pretty darn good a year ago, but the 3.0 targets produced some really laughable excuses for optimization. Nowadays, they're about even.

 

[mczak]

from memory ~1-2 years ago with nvidia drivers (donno 50.xx-60.xx)
an average vs + fs program ~100 instructions were taking 600-700msec
now with 90.xx thats down to ~50msec on the same machine

almost a second for 100 instructions? That seems really terribly slow - if c compilers were that slow I'd still be compiling linux kernel 1.0... 50msec sounds more reasonable, though it doesn't look terribly fast to me neither.

Just to give some rough numbers, we needed 8 hours to compile about 60k VS + 60k PS on a modern CPU.

That's a hell of a lot of shaders (well compared to doom3 at least )! Ok I see now why offline-compiles would be useful... However, if you want to recompile if there's a new compiler, you'd still have a problem.

 

[ERP]

almost a second for 100 instructions? That seems really terribly slow - if c compilers were that slow I'd still be compiling linux kernel 1.0... 50msec sounds more reasonable, though it doesn't look terribly fast to me neither.

Often the output of the compiler bears little resemblence to the input.
The shader assemblers IMO do some extremly impressive transforms on the specified data during optimisation. My guess would be there are several bruteforce algorithms involved.

 

[Farid]

Knowing the ARB if they were to release it they'd probably call it OpenGl 2.1, just to emphasise the anticlimatic buzz they build along the year around their API.

Let's just hope that this division happens fast and that they get to work on the new extensions and the GLSL improvements.

 

[nAo]

That's a hell of a lot of shaders (well compared to doom3 at least )! Ok I see now why offline-compiles would be useful... However, if you want to recompile if there's a new compiler, you'd still have a problem.

new compilers arrive all the time...but now we are not as dumb as we used to be: we support much more shaders (I did not compute the right number but it's in the order of milions) but they are generated per level using a system which queries our database and generates only shaders we need for that level. This make things much more quickier a what was taking hours now takes minutes.

 

[Reverend]

new compilers arrive all the time...but now we are not as dumb as we used to be: we support much more shaders ..[...]

Honestly, I am not sure what you mean by the word "we" in the 3 instances it was used.

 

[nAo]

Honestly, I am not sure what you mean by the word "we" in the 3 instances it was used.

we as "the team of very talented people I'm honoured to work with"

 

[DmitryKo]

[thread=32539]OpenGL ARB to Pass Control of OpenGL Specification to Khronos Group[/thread]

SIGGRAPH 2006 presentation, ATI - OpenGL 3.0 Directions

 

[nAo]

That's a bit more interesting imho: http://www.khronos.org/developers/l...nGL_BOF/NVIDIA_-_OpenGL_upcoming_features.ppt

 

[ShootMyMonkey]

we as "the team of very talented people I'm honoured to work with"

Interesting that you speak of them as previously having been dumb... or at least dumber relative to now. I'm sure there are a lot of people who'd be happy to hear that.

 

[trinibwoy]

That's a bit more interesting imho: http://www.khronos.org/developers/l...nGL_BOF/NVIDIA_-_OpenGL_upcoming_features.ppt

Varying variable interpolation modifiers
-flat, no-perspective, centroid
-Full shader control

Does programmable interpolation tell us anything about G80 pipeline structure ?

 

[nAo]

Interesting that you speak of them as previously having been dumb... or at least dumber relative to now. I'm sure there are a lot of people who'd be happy to hear that.

Umh...excuse me?!?! what are you referring to?

 

[ShootMyMonkey]

Umh...excuse me?!?! what are you referring to?

Joke. But, while you did include yourself, you did say "now we are not as dumb as we used to be."

 

[nAo]

Joke. But, while you did include yourself, you did say "now we are not as dumb as we used to be."

LOL, I was obviously referring to the fact that we were using a dumb/simple approach