Gamefest content is up
- Thread starter pipo
- Start date
Thanks for the link...
TurnDragoZeroV2G
Regular
Lots of great content. It's unfortunate so few people care/are paying any attention.
*Seems that the PIX performance tools continue to grow, and are becoming much more useful as time goes on. It's also interesting that MS didn't really get great documentation on the performance counters in the GPU and instead have had to play around with them to find out which ones are reliable, what they do, and so on. Rather than have a situation where MS went "Hey, build us some performance counters into these chips...." it was more like "you guys use these counters when you're engineering these chips, so leave them in and we'll make use of them too. And add a couple extra ones in too, please." Also, even though there are many PerfCounters, they can't all be used at once. Then-recent updates allow the tool to run through multiple times to get all the data available.
*Skinning: Wow. It's impressive how much of a win Memexport is (and stream-out on D3D10 should be). Skinning the characters once and then outputting the results to memory is significantly faster than the other methods tested once several attributes are being used. Much better than vertex fetch through either the texture cache or vertex cache, as well as the constant-buffers added to the system.
*Lots of updates on the compiler and API side of things. Several things added to help better manage output code, and tweak code to prevent z-fighting issues from occurring when you're doing something like z-prepassing and you don't necessarily want to run the same complicated vertex shaders in all your passes. Updates to other areas like predicated tiling.
Oh, and not many ******s on the 360 side of the fence should go about calling Heavenly Sword's HDR a hack anymore (though there really aren't any here....). In any case, Bungie apparently wasn't willing to give up blending support (light accumulation), and so passed on FP16 in favor of rendering to a LDR fp-10 (7e3) buffer and then also outputting another rendertarget in an 8-bit gamma 2.0 format for bloom, then combining the two after and through some postprocessing (including an unrealistic part) to achieve the final result. They use an 8-bit format RGB + multiplier for HDR cubemaps, though, so they don't need blending there/at that stage apparently.
Edit: Well, I suppose I have had quite a bit of free time today
*Seems that the PIX performance tools continue to grow, and are becoming much more useful as time goes on. It's also interesting that MS didn't really get great documentation on the performance counters in the GPU and instead have had to play around with them to find out which ones are reliable, what they do, and so on. Rather than have a situation where MS went "Hey, build us some performance counters into these chips...." it was more like "you guys use these counters when you're engineering these chips, so leave them in and we'll make use of them too. And add a couple extra ones in too, please." Also, even though there are many PerfCounters, they can't all be used at once. Then-recent updates allow the tool to run through multiple times to get all the data available.
*Skinning: Wow. It's impressive how much of a win Memexport is (and stream-out on D3D10 should be). Skinning the characters once and then outputting the results to memory is significantly faster than the other methods tested once several attributes are being used. Much better than vertex fetch through either the texture cache or vertex cache, as well as the constant-buffers added to the system.
*Lots of updates on the compiler and API side of things. Several things added to help better manage output code, and tweak code to prevent z-fighting issues from occurring when you're doing something like z-prepassing and you don't necessarily want to run the same complicated vertex shaders in all your passes. Updates to other areas like predicated tiling.
Oh, and not many ******s on the 360 side of the fence should go about calling Heavenly Sword's HDR a hack anymore (though there really aren't any here....). In any case, Bungie apparently wasn't willing to give up blending support (light accumulation), and so passed on FP16 in favor of rendering to a LDR fp-10 (7e3) buffer and then also outputting another rendertarget in an 8-bit gamma 2.0 format for bloom, then combining the two after and through some postprocessing (including an unrealistic part) to achieve the final result. They use an 8-bit format RGB + multiplier for HDR cubemaps, though, so they don't need blending there/at that stage apparently.
Edit: Well, I suppose I have had quite a bit of free time today
Last edited by a moderator:
The Heavenly Sword HDR is a hack. A beautiful, wish-I-would-have-thought-of-it, just-looking-for-an-excuse-to-use-it hack. Kudos to NAO.
Sorry but that's so wrong. It provides a floating-point high dynamic range intensity format. That's what HDR is (doesn't need to be FP either). A lot of people seem to get confused between data types and data ranges.
Or rather, other sections need the sort of attention to news items and passion for exotic technologies that the console crowd has!
Anyway I do agree though that this thread is under-appreciated.
Jawed
Legend
Memexport is very impressive.
Sadly the presentation, "Seven Ways to Skin a Mesh", indicates that D3D10 streamout won't perform so well
"likely winner will be tfetch".But this is based on one example, so it's early days...
Jawed
Or rather, other sections need the sort of attention to news items and passion for exotic technologies that the console crowd has!
Anyway I do agree though that this thread is under-appreciated.
Its appreciated by some of us...I know which ones I am going after
Ooh, tasty anti-aliasing coming to D3D10?
from "Introduction to D3D 10.ppt" in the Graphics track.
Jawed
p.s. Gamefest isn't just about XB360, thread needs moving...
Oooh. Where's 2x and 6x?
heliosphere
Newcomer
I think he means tfetch is the likely winner over vfetch (using an extra vertex stream for your skinning matrices). That's how I understood it from the slides and audio anyway. He also says he hasn't had access to any DX10 hardware yet so any performance estimates are speculative.Memexport is very impressive.
Sadly the presentation, "Seven Ways to Skin a Mesh", indicates that D3D10 streamout won't perform so well
I think it's pretty interesting that they posted all the 360 slides and audio uncensored, for anyone who's interested you can find out a lot of detail about the 360 hardware that hasn't been publicly available before.
Jawed
Legend
Having gone through all the graphics track presentations (mostly I've ignored the audio), the real highlight is "Advanced Lighting and More From Microsoft".
Real-time Soft Shadows in Dynamic Scenes using Spherical Harmonic Exponentiation
Video
Paper
Appearance-space texture synthesis
Video
Paper
Perfect spatial hashing
Video
Paper
At the very least, watch the videos to see these things in action
These are the home pages for the research I've linked. Huge amounts of fascinating stuff
http://research.microsoft.com/users/kunzhou/
http://research.microsoft.com/~hoppe/
Jawed
Real-time Soft Shadows in Dynamic Scenes using Spherical Harmonic Exponentiation
Video
Paper
Appearance-space texture synthesis
Video
Paper
Perfect spatial hashing
Video
Paper
At the very least, watch the videos to see these things in action
These are the home pages for the research I've linked. Huge amounts of fascinating stuff
http://research.microsoft.com/users/kunzhou/
http://research.microsoft.com/~hoppe/
Jawed
Having gone through all the graphics track presentations (mostly I've ignored the audio), the real highlight is "Advanced Lighting and More From Microsoft".
Real-time Soft Shadows in Dynamic Scenes using Spherical Harmonic Exponentiation
Video
Paper
Appearance-space texture synthesis
Video
Paper
Perfect spatial hashing
Video
Paper
At the very least, watch the videos to see these things in action
These are the home pages for the research I've linked. Huge amounts of fascinating stuff
http://research.microsoft.com/users/kunzhou/
http://research.microsoft.com/~hoppe/
Jawed
I take it that most of these exotic approaches are only viable on true DX10 PC hardware?
heliosphere
Newcomer
Actually, they were all demoed on DX9 hardware, none of the techniques require DX10 features, though the soft shadow stuff might be a bit expensive to be practical for most games on DX9 hardware.
LightHeaven
Regular
Ooh, tasty anti-aliasing coming to D3D10?
from "Introduction to D3D 10.ppt" in the Graphics track.
Jawed
p.s. Gamefest isn't just about XB360, thread needs moving...
That means Dx10 hardfware will be able to multi sample even when deffered techs are used?
Jawed
Legend
It appears so. I've heard hints about this being a possibility before, but I haven't got my head round how it would work...
Jawed
Jawed
Legend
Thinking a bit more, it may be that the programmer has to maintain their own "coverage mask" for every pixel containing multiple Z values (i.e. edge pixels).
Hmm, time to start a new thread...
http://www.beyond3d.com/forum/showthread.php?t=33678
Jawed
Hmm, time to start a new thread...
http://www.beyond3d.com/forum/showthread.php?t=33678
Jawed
LightHeaven
Regular
Really sweet if true. I got the impression from that slide that the multi sample could be achieved by a shader... if thats the case, any chance that the pipeline that allow this feat could be implemented on a non Dx10 hardware (asking directly: any chance of seeing something like this on next gen consoles?
)