Turning a Photoshop action into a pixel bender script

Question

We have a photoshop action that takes images and 'plasticifies' them. Specifically the action performs several filters that we're looking to convert into pixel bender scripts so that the action can be performed in Flash. I'm trying to discover if there are existing examples to perform these actions, and if not if it would be possible to create them. The filters that are being used are:
surface blur
diffuse
layer copy
high pass filter
layer overlay
hue / saturation
shadow / highlight
unsharp mask
I've attached the script to the post for reference and provided an example below. If this isn't something that we could combine several pre-existing filters together, would someone be up for working with me to create this or provide some guidance on how to move forward?
Original Image
Plasticify Effect
File: autoPlastic_test.atn (created in Photoshop CS4).
URL: http://drop.io/t7gez0g
Thanks,
Roy

Answer 1

AIF Kevin wrote:
The Bilateral Blur in After Effects is implemented in Pixel Bender. It  was tricky to get the best possible performance for it. Pixel Bender  works pretty well for Anistropic filtering as well, we are doing that  now for some new stuff.
Royi A wrote:
 I think in the GPU each basic variable (Pixel) is a vector of 4. Applying something on 4 of items in the vector or only 1 would "cost" the same.Someone of the AIF could confirm it?
This is sort of true. We do take advantage of the vector types to generate SSE code, but we do it not only within a pixel, but also across multiple pixels so a pixel1 value within your kernel may be computed with pixel1 values from neighboring pixels to improve performance. In general, it is best to write your code naturally and then only look to optimizing at this level if performance is unacceptable and the more obvious problems (minimizing conditional branching, for example) do not solve your problem.
    Kevin
Kevin, how do you iterate in the Anisotropic Filter?
One the things I miss is to loop a graph or something like that to iterate over the image.
Regarding the Bilateral Filter, My first PB kernel was a Gaussian Blur, just to get the feeling.
At first I didn't use its separability property.
It was slow, very slow!
For the Bilateral Filter you have to do even more operations. Unless you implemented its "Separable Bilateral" approximation, Or any other known optimization (Grids, Linear Approximation, etc...).
What about the GPU case regarding what I wrote?
Let's set I convert the image into LAB space and perform Gaussian Blur and each time only operate one CurrentPixel.r (The first item in the "Pixel" vector), will that make any performance difference?
Or even make it completely different. Creating a node which take as input image4 and output of image1 (L Channel). Then Gaussian Blur (2 Nodes) on image1, would that be faster (Again, GPU wise)?
Thanks.