LiSPPSVSM (Realtime Soft Shadows)

As some of you may already noticed, I am again trying to get some usable shadowmapping for big outdoor environments to work.
I think the right way to put the names of different techniques together is what you see in the Subject: LiSPPSVSM
The V will probably get a PCS but for now it is a V.
Those letters stand for "Light Space Perspective Parallel Split Variance Shadow Mapping", PCS would make the "Variance" to "Percentage Closer Soft" .

The techniques I am using are as the name states:
1) Light Space Perspective Shadow Mapping: http://www.cg.tuwien.ac.at/research/vr/lispsm/index.html
2) Parallel Split Shadow Mapping: http://hax.fi/asko/PSSM.html
3a) Variance Shadow Mapping: http://www.punkuser.net/vsm/
3b) Percentage Closer Filtering / Percentage Closer Soft Shadows: http://www.cs.utah.edu/classes/cs5610/projects-2005/lha/

This is btw basicly the same as what has just been integrated into Ogre.

For those that don´t understand anything so far and aren´t interested in those articles, this is about realtime soft shadows for large outdoor levels.

So far I am done with the Parallel Split and the Variance Shadow Mapping and will hopefully get the Light Space Projective part done during the next hours, which means just a few lines more as everything is already prepared to work with it. That should again increase the shadowresolution a lot.

This is the very first screenshot I took of my shadows. It shows the first optimisation of the shadow resolution, which is the first part of what is explained in the LiSPSM Paper:


This is the same shadow as above but blurred using some kind of "Plug and Play" Percantage Closer Filtering from NVIDIA. The problem is that they forgot to attach the dx9 version to the paper (which btw seems to have disappeared) and some variables had wrong names and thus I had to convert the dx10 version to dx9 and fix the wrong names, which causes this due to a lack of optimisations to be very slow with this pretty high quality:


This shows the same technique as the first screenshot but already with a little variance shadow mapping and I am now using two depthmaps. This was the first breakthrough to parallel split shadow mapping:


This next screenshot uses three depthmaps with each a resolution of 1024*768 (which I did not want, but currently the results are strange when using 1024*1024).
This screenshot shows the best the current state of the shadows. As you see, I did not yet work on correct selfshadowing, very soft shadows and the right brightness, but just look at the shadowresolution :


This screenshot is a little bit faked in terms of shadowresolution because the first depthmap covers only a VERY small area which makes already the trunk to show pretty strong aliasing, but improves the Shadowresolution of the rest of the shadow a lot. But what you should look at here is the fps in comparison to the screenshot before. The only big difference is that I rewrote the Lite-C code and had due to a bug to change to the latest beta, which seems to have improved speed greatly:


I also hope to improve the speed a lot as posted here: http://www.coniserver.net/ubb7/ubbthreads.php?ubb=showflat&Number=223541#Post223541

At the moment I am using a dll for the math but I plan to "port" it to Lite-C.

I hope you like it so far and would of course like some kind of feedback and especially if someone is strong at matrix math, it would be great if someone could may contact me and help me with getting the lispsm to work, I will post an update as soon as it works as I am of course trying to get it to work on my own if nobody can help me :P.

I really like your work and I see you have got already sucess with it. Keep up the good work and let's hope to get a good and fast LiSPPSVSM

Scorpion

Looks , good , didnt understand much though ,lolz

i had the same problem. you have to use bmap_zbuffer() and the latest beta to fix this.

how does this work? how can it be enabled?

Can't help you, but wish you success! It looks promising!

Looks very cool.

Nice work Nils.

Thanks for your nice words

ventilator, bmap_zbuffer somehow doesn´t really change the problem, it allows for higher resolution rendertargets but things are somehow still streched-.- although 512*512 works fine... I have to look some more into that...
With "Plug and Play" I meant that there was a paper with code attached and the paper described how to use the code. It basicly only means to copy that code on top of your shader and change the depthtest to their function with the parameters you would usually use for the depthtest. It then returns the greyscale shadow.
But there is also Hardware Shadowmapping on NVIDIA cards which seems to be that the graphicscard somehow notices that you want to create a shadow and also blurs it a little somehow... And I read that there is something similar for ATI cards or it is going to be there, don´t know :P

Looks really good Slin! Glad to see you working on this again.

i think i read somewhere about some kind of percentage closer filtering nvidia cards have implemented in hardware and when i read about your plug&play thing this came to my mind.

i am looking forward to see a working demo of your shadows.

i still only have simple pssm without any filtering. i didn't have time to continue working on it.

Already told you on ICQ how awesome this is

Good luck with this one. You'll need it
And please keep us updated, the community is thirsting for a nice shadowmapping shader i'd say

agree here

BoH_Havoc, you know very well that the hard part isn´t really the shader but the math, especially for the matrices calculated on the cpu :P
And thank you for getting me started with shadowmapping in the end of the last year Although I sometimes regret that I had you helping me... It could have saved me a lot of time and frustration if I had given up on this :P

Thanks to you and to Wiliam for commenting

I think this time I will wait until giving away a demo but I hope to be that far until sunday, because afterwards holidays are over and I´ll have a little less time... .

Here is a new screenshot I just took

I finally found out last night at about 3am that the example LiSPSM code from the paper multplicated two matrices the wrong way around, which caused it not to work...
Now it does work and I also already implemented with some help from Error014 a dynamic changing value for the lispsm quality to make sure that the LiSPSM quality stays nearly constant even with "bad" camera angles.

On the screenshot the quality looks worse than before, but the distance of visible shadows increased with a factor of about 3 and the better usage of the depthmaps causes the VSM to work worse than before, which made them look much better than they really were.

A complex name for a complex aspect, makes sense ^^
If you succeed in this, this will become one of greatest UCs ever
You're so close to victory Slin, don't let unexpected events turn down your focus on SM

Cheers

Frazzle

I can't wait for demo..

Que?

Looking good though, keep on it, I hope you can crack the puzzle and achieve some fast soft shadows.

Wow it looks realy cool!
I hope that you'll suceed on this and Conitec will buy it from you

Dark_Samurai

I am currently improving the shaders to make them not only faster but also to get rid of some aliasing and movement caused in the shader.

This is how it looks with three 256*256 depthmaps:

Nice effect, maybe even better than smoothing the shadow!

Nope it is not, because this is already smoothed through vsm (which gets stronger with lower resolution ;)), it would look just ugly without :P. Only thing missing is prefiltering to make it more smooth.
The problem with this low depthmap resolution is that the shadow is changing a lot which makes it look a little ugly when you see it in realtime.

What is the difference in fps between this low resolution and the next higher one?

Currently, in my testlevel about 10fps.
This would be the next one (512*512):

That's not too much. At least without any motion the shadow looks great!

Fear411 offered me to build a little level for the shadows, these are the results with a shadowmap resolution of 1024*768:



There are still A LOT of issues I have to fix but I am still very happy with my results and am optimistic that I am able to fix the problem :), at least with the help of some papers and some "controlled" copy and paste :P.
And I think that there is definatly some Ambient Occlusion missing... :P

Man,this is nice..so cool..

Yeah, the new screens look very promising!

wow, looking great:)

totally awesome Slin!
Great Job!

Looks like a Crysis level
Really cool!!! Are you planning to add more lights (for indoor levels)?

Dark_Samurai

Looks really good!
Please don't exaggerate the list of features!
Ambient Occlusion can be baked, at least, as long the things are not movable.

If this works with one light and many models with a good fps rate, you already achieved a lot!

Lol, no it doesn't, but it's still looking very very good.

do you still do some stuff in a plugin? what exactly do you need the plugin for?

i started to work a bit on my pssm project again. i try to do everything without any plugin but i continuously run into problems where i don't know if it's my fault or not. shader projects can be such a pain to debug.

Thanks for all your comments

Dark_samurai, I don´t plan to do indoor shadows at the moment. There are different problems and different solutions, thus it is a completely different project. I will may give it a try as well, but that has some time.

Pappenheimer, I ofcourse don´t add Ambient Occlusion before I am done with the shadow and probably not even then. I just think that there is some depth missing when looking at those screenshots.

ventilator, I am still using my plugin but I don´t think that it is needed. I am only calculating my matrices in there and am passing them to my material matrix and I also have a simple function in there to pass an "endless" number of vectors and matrices to a shader.
What I am curious about is how you are choosing on which depthmap to do the lookup. Are you using several passes or a one pass approach like this: http://blog.makingartstudios.com/?p=8
I am currently trying to get it done in one which should work quite well with a depthmap, I just had to do a lot of other thing so I didn´t really succeed yet-.- I did it with if else previously and got rid of them now, but at the moment it causes the wrong depthmap to be used when looking down... The other four things I am still missing are some kind of "deep slope bias", to get rid of the vsm artefacts on some edges, to get the sky to be completely not shadowed and some culling for better fps, with which jcl has to help me...
And I still can´t get rid of the problem with 1024*1024 textures... 512*512 and 1600*1200 are btw working fine-.-

with your 1024x1024 problem it might be the same type of issue that happens with the a6 comm or< r2t.dll, if the render target is not either smaller than, or in the same ratio as the current resolution it will cause an error. not sure if thats it but just a thought.

this is something I would pay money for if you manage to get the frame up a bit.
Very nice work.

yes, i use one pass with ifs. i have to look if this can be done in a better way. thanks for your link! it seems to be interesting.

1024*1024 works for me if i set up a new depth buffer with bmap_zbuffer(). strange that this doesn't work for you. what graphics card do you use?

your project is more advanced than mine. i don't know much about vsm or lispsm yet. at the moment i just want to get my simple pssm a bit more optimized and then experiment with different filtering methods.

will this be for a6 or a7 or both?

It will be Lite-C only and should be used with A7.5 and above.

ventilator, sorry that I didn´t answer earlier... I´ve got a 8800gts 640mb.

I am of course still working on this, but I have currently less time and am still trying to get rid of artefacts and to blurr the shadows. Blurring turned out to be a lot harder than I expected, I can blur the depthmaps for VSM, but the performence gets somehow very bad then. Some more advanced PCF in one pass is even slower and both techniques especially PCF have problems with "surface acne". Sure it can be tweaked with a bias, but that is not really perfect.

NOT that i'm a shader/engine guy who knows how everything works, but would it maybe help taking some ideas of how CryTek did it for crysis? :P

please exaggerate (whatever this means ) !
baking is not cool and the ssao you showed in your shader thread looks very cool ..

Sweet!

cool

Looks mighty accurate to me, great work!

I got a huge erection.
Nice work, mate!

well, it's not, the right part of the fence doesn't receive the shadow propery. but who cares, i don't. looks really cool.

great! looks very satisfying to me

they said it all, what else to add?

very nice work

... cool models as well :P, and cool grass. but I guess you're not showcasing those ^^

Due to currently everyone showing around their pointlight shadowmapping approaches, I decided to dig out this thread again and show you one of my latest screenshots of this:

REAL pro work!

When are you planning to finish this? What do you still have to do until this is finshed?

looks great slin, now we just need you to get some ssao and you'll be a competitor soon :P I just love all these new dynamic shading features people are creating. it means a7 is finally back up to par with many other engines.

Competitor for what?^^
As far as I know there is no one with a comparable outdoor shadow technique here :P

Did you by the way notice that these are dynamic soft shadows?


The Screenshot is taken out of my Project called "Demoscene" with which I took part in the conitec contest last winter/spring/whatever...

for the top shader collection of course :P (looking at you, boh_havoc, chrisB, and foxfire)

granted its no actual competition but still great to see so many incredible projects.

Looks great!

yeah, what if we all worked on the same project? Couldn't we get better results (and faster)? Anyway, good work, let me know if you want to work together (anyone).
-Mike-

This evening I also gave pointlightshadows a try. So far it was just about three hours of work and thus you shouldn´t expect to much, anyways I uploaded a first demo which can be downloaded here: http://files.filefront.com/VSM+maincdrar/;12576047;/fileinfo.html
I am pretty sure that it will work with a geforce 8 and newer nvidia chips, but it mostprobably won´t work with older nvidia or any ATI cards. It may does but I just don´t know that and I just made some bad experiences with shadows and ati...

Sieht ganz gut aus... FPS liegt bei mir zwishen 80 und 120... Hab ne Geforce 8600GT...

Aber was soll der Shadow da sein?(siehe Bild)


cYa Sebastian

There is hole in the terrain which throws that shadow, but yes it really is a little ugly :P

Anyways, I just got it to work with 4 Lightsources:


(Thanks to Machinery_Frank for allowing me to use that roommodel and those textures, if you like them just check out dexsoft)

I improved some things and I´d say that it looks a lot better now

Dude, you rule.

It looks great.

nice work Slin .

definitely great work

Thanks for your comments

Wohoow! More please!

I think in this scene the shadows still could fit if they were a lot darker than that.

looking better than in Oblivion :)))

now to soften the edges :P

I already am using variance shadow mapping, but I will have to blurr the depthmap to make them softer, as it currently is too little.

Go Go Slin!!! This is awesome!!!

Yes, of course this looks better than Oblivion and better than Fallout 3 because these 2 don't use shadows for their environments.

Slin did a great job here!

Really pro work!! You're a real shadow god ^^

I agree with Pappenheimer, the shadows should be darker.

Just saw this, really nice work!

any progress with this great! project Slin?

Maybe

hehe, no really.
do you have plans to finish it soon or it's still in early stage of development?

Let's just hope that this years Christmas Contribution by Slin will have something to do with LiSPPSVSM

The shader-gift you provided last year at christmas was amazing Slin!

that would be awesome!

This is really great work..It look's awesome!

Rayn...
Why did you have to tell them-.- :P

http://www.coniserver.net/ubb7/ubbthreads.php?ubb=showflat&Number=242546#Post242546

Dear Slin

What can I say - I'm sincerely very sorry! That was just very clumsy and stupid of me! My hopes for your shadow-map-project boiled over I guess - and made me thoughtless. I wouldn't want to bite the hand that feds me! Again, sorry! Thank you so much for your Christmas present!

Best of luck to you!

Sincerely

I just want to show you some unblurred shadows

I rewrote the shadercode, which allows me now to do an a lot better looking blur with better performence

Now that looks awsome!

Seems that you got rid of the ugly stripes with this. Great.

Here is another screenshot...

Looking good, as always

Any info on what you are using? I guess PSSM and some nifty blurring? Did you try out the ESM thingy?

At the moment it is just some kind of PSSM and PCF.
It looks also quite well in motion and is very light/shadow bleeding free
The only disadvantage at the moment is the speed or to be more precise, there are just sometimes some moments where the ref time goes up a lot, while most of the time it runs quite okay...
I will may look into ESM somewhen later, but for now, this seems to be the most stable way to blur the shadows.

The last screenshot looks great!!

awesome!
Cant wait to see the final version ^^
nice job slin

For now, I disabled the blur (to be honest, it is still calculated, just not visible...), added a fourth depthmap and did some other minor changes.
This is what it looks like:


There are still some artifacts, but most should disappear after some blurring and the rest gets hopefully hidden by the shading.
I also uploaded a published version of the shadows and would like to get some feedback on strange behaviour like shadows nearly everywhere and something like that. If you encounter any other problems than a bad fps I would like you to post a screenshot of the problem and some information on what graphics card you are using. Thanks.
Download it here: [url=www.slindev.com/files/projects/SunShadowsV1.7z]www.slindev.com/files/projects/SunShadowsV1.7z[/url]

For some reasons the camera starts to fly lightning fast to the front when I press "0". Could you please fix that in your published version and can you please add a toggle key for the blur? I'm curious how it looks like!

That is very strange, as everything works fine for me and also for others ´ve sent it to through an IM...
The blur isn´t there at the moment, as I am not very happy with it and want to rework it.

when I set a higher resolution, the shadows "Move" when I change the camera tilt

Works well here. small artifacts (especially on the tree trunks, it seems), but nothing horrible. It's interesting how the sprites seem to throw shadows onto other sprites...?
so why is the blur disabled? How do i reactivate it? I'd like to see it without those artifacts :=

It would be great if someone with a quite modern ATI graphics card could give me some feedback on the already posted published version (www.slindev.com/files/projects/SunShadowsV1.7z), as I currently can´t find anyone with a real AMD/ATI card...

I've got the Ati X300, and there are no satisfying news from that card!
I don't want to frustrate you.

Edit:
It has no Shader 3.0

Hey Slin works quite well here on my ATI HD4870 X2 with 75 FPS! no Bugs or anything bad atm great Job

grz

Core

4870X2 at 75FPS????

At what resolution? Those cards should eat anything gamestudio throws at them!

Thanks _Core_, that is all I wanted to know for now

And yes, shadermodel 3.0 is needed for those shadows. But you anyways wouldn´t get an acceptable performence with an graphics card without shadermodel 3.0 support.

Edit: Tai, those are shadows. highquality shadowmapping is just VERY demanding for hardware. In any modern game, you will notice a big fps difference between enabled and disabled shadows. That has nothing to do with gamestudio (at least not that much).

So dynamic shadowmapping has that big of an FPS hit? Ouch!

I'm getting 85fps when zooming out and having everything visible on a HD4870. Shadows do work here. BTW is it intentional that the shadows blend out at a certain distance?

Enjoy your meal
Toast