purpose of tile-based rendering
- Thread starter mdhome
- Start date
Low latency/high bandwith access to the framebuffer is the most usefull result of tiling IMO (it can also reduce the storage needed per pixel, but nowaday's thats less of an issue).
For instance it reduces normal framebuffer bandwith, but it also allows you to do things which could overwhelm external memory bandwith. Deferred shading is very hard (to do efficiently) without tiling for instance, because in some way the shading parameters for all the pixels have to be stored ... you can do it by laying down the Z-buffer first, and having the Z values together with the original geometry encode the shading parameters per pixel (a bit of a stretch to call it an encoding there, but Im trying to make a point
but you have to transform everything twice so thats hardly an enticing option. Or you can do it by storing a post transform display list and a reference per pixel, or you can even generate all the shading parameters during HSR and store them per pixel (without shading it) ... that last option haa large storage requirements, and by association also high bandwith requirements, but is not a huge problem for tiling.
Multisampling is another example, again generating very high storage/bandwith requirements per pixel ... again much less of an issue for tiling.
For instance it reduces normal framebuffer bandwith, but it also allows you to do things which could overwhelm external memory bandwith. Deferred shading is very hard (to do efficiently) without tiling for instance, because in some way the shading parameters for all the pixels have to be stored ... you can do it by laying down the Z-buffer first, and having the Z values together with the original geometry encode the shading parameters per pixel (a bit of a stretch to call it an encoding there, but Im trying to make a point
but you have to transform everything twice so thats hardly an enticing option. Or you can do it by storing a post transform display list and a reference per pixel, or you can even generate all the shading parameters during HSR and store them per pixel (without shading it) ... that last option haa large storage requirements, and by association also high bandwith requirements, but is not a huge problem for tiling.Multisampling is another example, again generating very high storage/bandwith requirements per pixel ... again much less of an issue for tiling.
mdhome said:You will have to excuse me and I know that this may seem like a silly and simplistic question but what is the purpose of tile-based rendering? Is it to increase fillrate or bandwidth?
Regards Michael
The idea behind TBR is to optimise bandwidth and to use fill-rate more effectively.
As far as the Kyro series does it, first off there's bandwidth enhancement in the form of near-perfect HSR. Pretty much everything not visible is culled by the chip before texturing begins - so a lot of texture accesses are stopped before they start. Also, this means that all the culled pixels aren't processed, so the fill-rate is saved as well.
Also, bandwidth is saved by all Z-buffering being done on a medium-size on-chip buffer - no external bandwidth necessary.
[/i]...That "medium-size" in Kyro being a grand total of 32x16 pixels = one tile.
Generally, one could also say that "the purpose of TBR" is simply to do it smarter. Divide the screen into tiles, get the polygons, send them to proper tile bins, then do one bin (and tile) at a time, and voila! now it's easy to check first what to render and what not, because it's only a small binful of the stuff (polygons) at a time; and when you render, the Z-buffer for visibility testing and the back-buffer where you render the image can be only medium-size so you can make them internal where you have bandwidth out the wazoo, as Tagrineth already explained, so this post is kinda useless
Generally, one could also say that "the purpose of TBR" is simply to do it smarter. Divide the screen into tiles, get the polygons, send them to proper tile bins, then do one bin (and tile) at a time, and voila! now it's easy to check first what to render and what not, because it's only a small binful of the stuff (polygons) at a time; and when you render, the Z-buffer for visibility testing and the back-buffer where you render the image can be only medium-size
so you can make them internal where you have bandwidth out the wazoo, as Tagrineth already explained, so this post is kinda useless 
mdhome,
Perhaps you'd like to read up on it a little:
PowerVR Tile Based Rendering Explained.
Our reviews also contain information about the architecture:
Neon250
KYRO Vivid!
KYROII Vivid!XS
Hope they're of help.
Perhaps you'd like to read up on it a little:
PowerVR Tile Based Rendering Explained.
Our reviews also contain information about the architecture:
Neon250
KYRO Vivid!
KYROII Vivid!XS
Hope they're of help.
mdhome said:
There may be a few people frequenting B3D who do.
EDIT: Oh, BTW, Dave mentioned Kristof's article: PowerVR Tile Based Rendering Explained.
I'll warn you that it's not 100% correct (it was written before K started his current job) , but it's a reasonable starting point.
king_iron_fist
Newcomer
In my tiny mind there are quite a few benefits for tilers, in simple terms :
Architecture means that only a fraction of the bandwidth is needed. Something which would require 8gb/s on an IMR may well only need under 1Gb/s on a tiler.
This means that expensive memory isn't needed therefore Tilers can be much cheaper than everyone else.
Tilers have virtually zero overdraw. Where an IMR may render at least 3 times the polys/texels actually on the scene a tiler will only render what's there. This means a relatively low fillrate can comptete with a much higher theoretical fill rate on an IMR. Also the Peak fillrate of a tiler is almost always reached e.g Kyro II theoretical fillrate of 350 mpixels/s, in 3d mark test it gets 340+ sustained.
Certain features can be implemented on tilers for virtually no performance hit. Things like 32 bit and higer colour depth, Free AA and 8+ sample multitexturing can all be implemented for a negligble performance drop.
Architecture means that only a fraction of the bandwidth is needed. Something which would require 8gb/s on an IMR may well only need under 1Gb/s on a tiler.
This means that expensive memory isn't needed therefore Tilers can be much cheaper than everyone else.
Tilers have virtually zero overdraw. Where an IMR may render at least 3 times the polys/texels actually on the scene a tiler will only render what's there. This means a relatively low fillrate can comptete with a much higher theoretical fill rate on an IMR. Also the Peak fillrate of a tiler is almost always reached e.g Kyro II theoretical fillrate of 350 mpixels/s, in 3d mark test it gets 340+ sustained.
Certain features can be implemented on tilers for virtually no performance hit. Things like 32 bit and higer colour depth, Free AA and 8+ sample multitexturing can all be implemented for a negligble performance drop.
I find it really unforunate that PVR seems to be in such a weird postion right now.
It would so kick ass to see a PVR card with a solid TnL unit and DX8/9 class shaders. I bet with about 8 gigs of memory bandwidth, it would totally whomp on just about everything if the drivers were decent.
It would so kick ass to see a PVR card with a solid TnL unit and DX8/9 class shaders. I bet with about 8 gigs of memory bandwidth, it would totally whomp on just about everything if the drivers were decent.
Typedef Enum
Regular
Unfortunately, we've had to say this for a good number of years now...and when they put something out, it almost always falls short of these kind of expectations...
Would it be kool if it were to happen? Of course. Am I expecting it to happen anytime soon? I've not heard or seen anything that would cause me to have optimism.
Joe DeFuria
Legend
SHHHHH!! You'll wake up Teasy.
Do I count Joe?
Typedef,
I wasn´t expecting that much from either 3DLabs or Matrox prior to their official announcements of their upcoming offerings. Just because I hadn´t seen anything worth mentioning for years in the PC desktop space.
All PowerVR would need is a partner that is willing to push the envelope and won´t back out with the first best chance. Frankly what did ST Micro expect from the KYRO line? To sell dozens of million units at first attempt? It´s a tough market and you don´t get recognition from one day to the other that easily. If they weren´t realizing what they were dealing with, then they shouldn´t ever had attempted to enter the graphics market.
Wasn´t it apparent from the beginning that even the KYRO II would have had a hard time to penetrate the OEM market? Just about when the offering that had the potential at least to get a far more significant OEM share was close to being ready they decided to abandon effort.
3DLabs is part of Creative right now and Matrox is not exactly a small company lacking of resources to manufacture or distribute. I don´t see how an IP selling company with the size and resources like ImgTech could stand up and do it all by itself out of the blue. Do you really think they´re "that" incapable of delivering?
Typedef,
I wasn´t expecting that much from either 3DLabs or Matrox prior to their official announcements of their upcoming offerings. Just because I hadn´t seen anything worth mentioning for years in the PC desktop space.
All PowerVR would need is a partner that is willing to push the envelope and won´t back out with the first best chance. Frankly what did ST Micro expect from the KYRO line? To sell dozens of million units at first attempt? It´s a tough market and you don´t get recognition from one day to the other that easily. If they weren´t realizing what they were dealing with, then they shouldn´t ever had attempted to enter the graphics market.
Wasn´t it apparent from the beginning that even the KYRO II would have had a hard time to penetrate the OEM market? Just about when the offering that had the potential at least to get a far more significant OEM share was close to being ready they decided to abandon effort.
3DLabs is part of Creative right now and Matrox is not exactly a small company lacking of resources to manufacture or distribute. I don´t see how an IP selling company with the size and resources like ImgTech could stand up and do it all by itself out of the blue. Do you really think they´re "that" incapable of delivering?
MfA said:
Probably not untrue. I guess IP selling companies aren't the best business scheme for the PC market after all. How successful was Gigapixel, before 3dfx bought them? No product ever released and a blown Xbox deal.
Quite frankly if I'd be in the shoes of a vendor I'd prefer the sollution of a small company buyout, instead of licensing constantly and paying royalties to 3rd parties.
I'd dare to guestimate that ImgTech is probably doing well in all the other markets it addresses, because I haven't seen any major success in the PC market so far.
Don't know what Kristof would have to say about abusive boyfriends though
