nVidia going for 16 sample FSAA?
- Thread starter LeStoffer
- Start date
But, given that flipquad is only using 2 samples per pixel, it's not bad!OpenGL guy said:I like them for the cleverness of the idea, but they cause too much blurring IMO. The samples pictures I've seen look "ok" until you take a closer look and realize that the blur is overdone.davepermen said:i like flipquad and fliptri samplings..
Ailuros said:
I still prefer the "many samples" part of that pic (presupposition that it's actually usable)
Usable? 10000 samples per pixel? Hmmm...
Looks like it just needs some more bandwidth for HD contents. Current DVD bandwidth (9.8Mbps) is certainly not enough. MPEG-4 helps, but mostly done with a sophisticated post filter (the Microsoft HDTV samples look good, though).
I did some search for BS Digital, and I found that it uses 28.86 Mbps for one channel (MPEG-2 video + MPEG-2 AAC). That's three times of the normal DVD bandwidth.
BS Digital is very popular in Japan. Last time I visited Tokyo, almost evey household has a small satellite dish. That's impressive.
I did some search for BS Digital, and I found that it uses 28.86 Mbps for one channel (MPEG-2 video + MPEG-2 AAC). That's three times of the normal DVD bandwidth.
BS Digital is very popular in Japan. Last time I visited Tokyo, almost evey household has a small satellite dish. That's impressive.
davepermen
Regular
OpenGL guy said:I like them for the cleverness of the idea, but they cause too much blurring IMO. The samples pictures I've seen look "ok" until you take a closer look and realize that the blur is overdone.davepermen said:i like flipquad and fliptri samplings..
i know it can't beat 16sample stochastic aa. but for the cost it has, it is by far the best result you can get. it has about the quality of the quinquirks samplers from nvidia, but at half the cost (and less). and actually, quality is bether..
i'm merely interested into finding bether lowcost algorithms. and then, we can always bruteforce them...
I did a google search on fliptri and flipquad sampling and found the two Akenine-Möller pdfs. While I think flipquad beats quincunx in every aspect, I still think it's bad because of the texture blurring. Even more so with fliptri, which has the additional problem of how to distribute colors to 1.25 samples per pixel
Youre spot on, DC.DemoCoder said:
Its just that at this time and age, i think for faster advancement of perceived image quality on consumer realtime 3D graphics, general advancement in global lighting solutions, actually useable temporal and spatial anti-aliasing solutions, shadowing techniques, procedural detail generation, would give much better bang for the buck than brute pixel churning power.
"Bang" meaning perceived image quality ( which is obviously very hard to quantify ). "Buck" meaning silicon space, invested R&D effort ( which loosely translate to consumer paid $$$ ).
Of course, this one pdf isnt exactly proof of NV focussing on brute force pixel crunching power only. They'd be stupid if they did, and after all NV was spearheading the move to more programmable hardware just a short time ago, i dont think they have completely lost their edge just yet.
Oh, and btw, i think the next really significant paradigm shift in consumer 3D will not come from hardware side ( unless Volari or R420 implements photon mapping in HW ) but from really cool content generation ( think speedTrees and darkTrees )
davepermen
Regular
Xmas said:I did a google search on fliptri and flipquad sampling and found the two Akenine-Möller pdfs. While I think flipquad beats quincunx in every aspect, I still think it's bad because of the texture blurring. Even more so with fliptri, which has the additional problem of how to distribute colors to 1.25 samples per pixel
you don't have to use it for the texture samples.. no gpu uses their aa for texturesamples, as well.. (and the whole pixelshaders don't work on the samples as well
). with the known advantages and disadvantages..What I meant is, you need a color for each sample. The usual multisampling approach is, you calculate one color value per pixel and assign that to all samples that belong to that pixel and are covered by the current triangle. With fliptri and flipquad, you don't have that "sample belongs to pixel" relation. Using flipquad, you can calculate one color value per pixel and assign that to two samples. But with fliptri that's more difficult. You could turn it into supersampling and calculate one color value per sample. Or you take one color(texture) value in the upper left corner of each pixel and, depending on whether this corner is covered by one sample position or surrounded by four, use it for those samples. But I expect it to blur textures a lot.davepermen said:you don't have to use it for the texture samples.. no gpu uses their aa for texturesamples, as well.. (and the whole pixelshaders don't work on the samples as well