this article is baseless!

Broken Link.

The authors agree with you. they pulled it.

Luckily, I found it.

I'll bite. What's baseless about it?

http://www.comp.nus.edu.sg/~tants/softShadow.html

Abt the "baseless" issue:
As far as i realize, algorithms designed for the GPU should be 100% parallelizable. Their "algorithm" actually work by processing pixels where there will be NO penumbra classification for that frame. These intermediate values are dependant for the next rendering pass. So it is still propagative (meaning not fully parallelizable).

Secondly, their "algorithm" is not 100% accurate. It can process wrong intermediate values at certain times.

Thirdly, they claimed that there are no proper standards in the design of GPU based algorithms, which I think is part of their ignorance. There are in fact, some articles written by IHVs (NVidia and ATI) that the authors did not read.

There are alot more issues, but for now, i let u pple voice your views.

Quote:

Original post by edwinnie Abt the "baseless" issue: As far as i realize, algorithms designed for the GPU should be 100% parallelizable. Their "algorithm" actually work by processing pixels where there will be NO penumbra classification for that frame. These intermediate values are dependant for the next rendering pass. So it is still propagative (meaning not fully parallelizable). There are alot more issues, but for now, i let u pple voice your views.

I'm not sure I understand what you're implying, re: propagative.

Because the technique uses multiple taps, it's not GPGPU?

Quote:

Original post by edwinnie Abt the "baseless" issue: As far as i realize, algorithms designed for the GPU should be 100% parallelizable. Their "algorithm" actually work by processing pixels where there will be NO penumbra classification for that frame. These intermediate values are dependant for the next rendering pass. So it is still propagative (meaning not fully parallelizable).

no, that is not a requirement. Would be damn near impossible to achieve.

Quote:

Secondly, their "algorithm" is not 100% accurate. It can process wrong intermediate values at certain times.

there is no 100% accurate lighting and shading algorithm outside of ray-tracign. They all operate on simplified models, some of which are down-right unscientific but just look good.

Quote:

Thirdly, they claimed that there are no proper standards in the design of GPU based algorithms, which I think is part of their ignorance. There are in fact, some articles written by IHVs (NVidia and ATI) that the authors did not read.

an article from an IHV does not a standard make.

hmmm.... can someone explain to me what all the fuss is about in non-programmer terms?

Quote:

Original post by Jarrod1937 hmmm.... can someone explain to me what all the fuss is about in non-programmer terms?

The algorithm the article introduces has several propagative steps, that is, each step is dependant on the previous one. Because of this dependancy, it would be very difficult to achieve parallelism (to make it run better with multiple cores).

I guess, I'm not really sure myself :]

Quote:

Original post by Mushu Quote: Original post by Jarrod1937 hmmm.... can someone explain to me what all the fuss is about in non-programmer terms? The algorithm the article introduces has several propagative steps, that is, each step is dependant on the previous one. Because of this dependancy, it would be very difficult to achieve parallelism (to make it run better with multiple cores). I guess, I'm not really sure myself :]

ok, if that is what it is then i understand.

Hi capn_midnight and all!

It is actually possible to design a 100% parallelizable algorithm for this matter. What is important to understand which algorithm the parallelization is for. I am not referring to their "voronoi-based algorithm", but to the soft shadow algorithm by Arvo, in which their article cited as their main soft shadow reference, which relies on a propagative way to generate penumbra.

I am not refering to the lighting models and penumbra accuracy with respect to low polygonal models. I am refering to the fact that their "voronoi-based algorithm" can compute some wrong intermediate values. Their algorithm rely on single intemediate values, and if one of these is wrong, you can potentially see "bad spots" on the final image. Do not be fooled by the images you see on their website.

I would prefer to bank on IHV standards in order not to contravene the way the graphics hardware is to be utilized. Deviation from the norm will have future consequences with respect to this matter. For example, I can think of a future upgrade to localise computations at a coarse level. The fact that their "voronoi-based algorithm" depends on intermediate values that are to be calculated on pixels that will NOT be classified as penumbra, immediately implies that their algorithm does not take into account of "ending optimizations".

The authors appeared to just want to focus on the penumbra intensity with no regard to IHV standards. As far as I realize, any beginner in GPGPU can EASILY think of a 100% parallelziable algorithm to parallelize Arvo's soft shadow algorithm. In fact, I do can think of two, and I bet anyone can easily do the same too.

thx fer letting me to ranter my frustrations...

Quote:

In fact, I do can think of two, and I bet anyone can easily do the same too.

This being the case, perhaps the authors did not feel that this was an important matter to discuss in the paper. It is not uncommon for academic papers in computing to present algorithms that are non-hardware specific- The focus being on computing the solution, not implimenting it. An unoptimized rendition still shows that the solution can be computed, and the best implimentation for GPU hardware can be the subject of another paper.

Edit: On the other hand, they do claim an increase in speed over previous algorithms, which would indicate some consideration was intended in the implimentation.