source: GTP/trunk/App/Demos/Vis/FriendlyCulling/toto.txt @ 3203

1) change shader system: look into fx files
2) tone mapping
   bloom (gaussian blur)
   use downsampling: first 4x4 to luminance then 3 times further to 1x1
   use 3 new render render targets for this

3) mrt channels:

2 mrts:

mrt 1: rgba color + emmissive / luminance
mrt 2: world pos + projected depth
mrt 3: normals + w coordinate
mrt 4: flip-flop for mrt1


4) idea: donwsample color texture for ao

5) use depth formula on projected depth (or use linear depth)
   in order to combine ao + diffuse lookup into on texture lookup


reorder mrts:

mrt 1: rgba color + projected depth
mrt 2: luminance
mrt 3: normals + w coordinate
mrt 4: flip-flop with mrt1 / store luminance after ssao


6) render sun disc



ssao:

1) implement bilateral filtering
2) check physical properties 

paper of 2007 arikan formula:

SW(P, C, r) = 2 * pi * (1 - cos(asin (r / |PC|)))

but not working properly!

3) normal mapping


problems:

4) update of converged regions

something could become visible 

a) from outside
b) from previously occluded regions

a) could use ratio of samples outside current frame / last frame
   but slows donw code

5) dynamic objects:

a) make ao stick on object: this should
   be possible somehow, as the information is still available!!
 
   tried to use difference of ao intensity between previous and
   current frame to find out if pixel ao is not valid anymore.
   but problems as some flickering was introduced while update of 
   dynamic objects was not fast enough (annoying grey fade effect)

b) fix the contact shadow on the floor:

check when pixel not valid anymore: do that by checking for each
sample if they were invalidated recently. if so, then invalidate
current pixel ao. for each point, theoretically you don't have
to compare the depth of the current pixel, but the depth of the
samples taken for ao


idea for incorporation of dynamic objects:

store object id with render target
for each object we know the trafo
when doing the back projection =>

as usual:
we have the world space position of the current pixel
find pixel from last frame using the old projection view trafo

but now we first apply the inverse transformation that brought the
last pixel to the current pixel and then the old projection view!!

=> now we do the equality comparison as usual


for reducing flickering:


keep chain of kernels constant based on the state of convergence=>

only use a single kernel, rotate based on noise texture
we use fixed offset into the noise texture based on the #frames samples were accumulated

problem: 



=========================

1) ssao filter kernel: grows smaller too fast => flickering ... use
adaptive number of samples instead of size?
2) fix color bleeding
3) fix ssao2 so that ssao contribution stays constant
4) find physical expressions for constants in ssao
5) retry normal discontinuity for ssao filter
6) fix adaptive sampling for ssao and ssao2
   a) sample adaptive if found invalid sample
   b) sample adaptive if number of valid samples small?
   
   q: is it good to choose new weight as we do? (approx 4 frames)
   or shouldn't we just completely reset sample?