"legit" 3dmark05 or whatever it'll be called pic
- Thread starter tEd
- Start date
tEd said:
How many samples you need to smooth the shadows depends on how smooth you want your shadows, not whether you use PSM or plain shadowmaps.
The overhead of PSMs over plain shadowmaps is usually a tiny bit of code tweaking the matrices (negligible), and potentially an enumeration of potential shadowcaster objects (usually also negligible, and much less than what you'd do with them with stencil shadows).
The benefit of PSMs is that you get less pixelization, as you use the shadowmap resolution much better; thus, if you want sharp shadows, you may completely skip blurring the shadows via multisampling. If you want smooth shadows, however, you need to multisample even with plain, jagged shadowmaps.
The real drawback of PSMs isn't in their visual quality or the rendering overhead - it is in the complexity of implementation, and the many special cases you need to handle.
Luminescent
Veteran
I guess I've attained a little insight as to those words from being on helpful sites such as Beyond 3D for the past couple of years. I did percieve the term PSM incorrectly, though, and interpreted it as shadow buffers for some reason. In the Quakecon 04 presentation Carmack stated something that approximated: for the shadow technique in my follow up to Doom 3, hardware is not yet optimized and performs less than spectacularly atm, which is why I made the reference to poor performance.assen said:
Luminescent
Veteran
Interesting snippet from the interview.Patric said:
Luminescent said:
You'll want to check out what Deano C had to say about it also (in case you haven't already):
http://www.beyond3d.com/forum/viewtopic.php?t=14361
Luminescent said:I guess I've attained a little insight as to those words from being on helpful sites such as Beyond 3D for the past couple of years. I did percieve the term PSM incorrectly, though, and interpreted it as shadow buffers for some reason. In the Quakecon 04 presentation Carmack stated something that approximated: for the shadow technique in my follow up to Doom 3, hardware is not yet optimized and performs less than spectacularly atm, which is why I made the reference to poor performance.
Shadow buffers and shadow maps are related terms and usually used to mean one and the same things. The "perspective" part is a slight tweaking of the math involved which sometimes achieves results which are drastically better, but performance-wise is the same as "normal" shadowmaps.
Generally, shadow mapping algorithms work in two parts:
1. rendering a "shadow map" - a z-only pass from the point of view of the light. Recent NVIDIA GPUs are good at that thanks to their double z-only fillrate. Basically, you store a z value into a texture.
2. normal rendering of the scene, which includes a comparison of one sample (or several samples, if you want smooth shadows) of the shadowmap with the current, normal z-value of the pixel. NVIDIA GPUs (since Geforce2 IIRC) can do this natively, while on ATI you need to perform pixel shader trickery to pack z values into color channels, and unpack them back. This is probably what Carmack was referring to.
I doubt Futuremark would have chosen PSM, if they couldn't get it to run decently even on previous-generation ATI hardware (9500-9800).
assen said:
Not entirely correct, I think.
Firstly, GeForce3 was the first with hardware shadowmap comparisons, I think...
Secondly, the 'pixelshader trickery' on ATi is not related to packing z values in colour channels. You can simply use a float texture with a 16 or 32 bit float channel (even 8 bit integer will work in most cases, that is not the biggest problem really).
Rather, the hardware comparisons on GeForces use the so-called Percentage Closer Filtering (PCF) technique. Think of it as a 2x2 bilinear filter, which is applied AFTER the depth-comparisons, rather than before, as a normal texture filter would. This will give you an estimate of the percentage of surface that is closer than the current pixel, and this percentage can be used to attenuate the light to get soft shadow edges. While not correct, it does give a nice and smooth image.
On an ATi card (or any other brand, really), there is no special function for sampling a texture that is used as a shadowmap, so you can only get the same performance as the GeForces if you do not apply a filter at all. This doesn't look good ofcourse, so generally people will opt for manually filtering in a pixelshader. Ofcourse this requires 4 separate texture fetches for a 2x2 filter, which means you get a considerable performance hit.
This is what Carmack was referring to.
3DMark03 ran horribly even on the current generation NVIDIA hardware at the time (GF4), and one test didn't work at all, and that didn't stop them either.
Those that have a clue really couldn't answer that one either way right now, that's a pretty big bit of the NDA thingy I'd imagine. :?Sxotty said:
Oh, but I can tell you it'll be called 3dm2k5. 8)
Well all I can say then is, "WOOOOOO-HOOOOOOO!!!!!".worm[Futuremark said:
I thought the extremely low CPU dependence of 3DMark03 was quite interesting actually. It showed everyone just how far you could go with 3d acceleration, and how much CPU you could spare for other game-logic.
Makes me wonder exactly why 3DMark05 will be more CPU-dependent...
If it's just for trivial reasons, eg because the scenes are larger/more animated objects etc, that makes sense...
If not, I wonder what those CPU-cycles are being used for
Makes me wonder exactly why 3DMark05 will be more CPU-dependent...
If it's just for trivial reasons, eg because the scenes are larger/more animated objects etc, that makes sense...
If not, I wonder what those CPU-cycles are being used for
Not actually sure with which card or system it was taken. All I know is that there is a 2048x1536 version of it too, which was a tad too big to be posted.MuFu said:Looks very sweet indeed.
Wouldn't happen to be captured on an ATi board (w/6xFSAA, no doubt), would it? Main reason I ask is because of the aliasing on the floor textures, a la D3.
+ a free, feature-limited version, right?
Yes there will of course be the free version available! The thing is that this time around we really have tried to focus on the Pro version, and to offer the Pro users more features, options etc. than before. I'd say that the free version is only a crumb from a cookie in comparison to Pro.
