Re: [filters] Request: Support for convolution in CSS Shaders

Roy,

I'd also encourage you to take a look at Eric Dahlstrom's implementation of
a ripple effect using a SMIL-animated feDisplacementMap.

http://svg-wow.org/filterEffects/ripple.svg

-- michael

On Thu, Oct 11, 2012 at 6:26 PM, Roy Williams <royw@fb.com> wrote:

>  Lots of effects like reflections or distortions, etc... require
> sampling from different locations in a texture to get the desired effect.
> Effects like Blurs or edge detection require convolution.  When I first
> heard about CSS Filter Effects and sat down with the code, I tried to
> implement a water ripple originating from where the user clicked.  I found
> I was unable to do so with the current spec, I wasn't able to distort the
> element I was styling.
>
>  Convolution is a superset of modifying the sample location, so I would
> propose adding two (maybe 3?) new output parameters from the fragment
> shader,
>
>  vec2[KERNEL_SIZE] css_SampleOffsets (css_SampleLocations?)
> and
> float[KERNEL_SIZE] css_SampleWeight
> (maybe mat4[KERNEL_SIZE] css_SampleColorMatrix as well?)
>
>  The first, css_SampleOffsets, would contain the offsets (or absolute
> locations?) into the texture to sample from. css_SampleColorMatrix would
> have the same functionality as css_ColorMatrix, but applied to each
> sample. css_SampleWeights would contain the weight with which each sample
> should contribute to the final color. Finally, css_SampleWeight would be
> used to weigh each sample.
>
>  The actual fragment shader executing would look something like:
>
>  void main() {
>   shaderMain();
>   float totalWeight = 0;
>   vec4 sum = vec4(0.);
>   for (int i = 0; i < KERNEL_SIZE; ++i) {
>     sum += css_SampleColorMatrix[i] *
>     texture2D(tex, css_SampleLocations[i]) * css_SampleWeight[i];
>     totalWeight += css_SampleWeight[i];
>   }
>   multColor = css_ColorMatrix * (sum / max(totalWeight, 1.));
> }
>
>  This doesn't seem like it would disclose any additional information nor
> does it seem like it could be exploited by a timing attack, but it would
> greatly expand the kinds of effects you can achieve with the shader.
>
>  I'd make css_SampleColorMatrix optional. There are a large number of
> effects that would default to the identity matrix, and it would be
> wasteful to use all of that additional memory.  If there aren't an effects
> that would need this kind of functionality, I wouldn't bother with it.
>
>  Thanks,
>
>  Roy Williams
> Software Engineer, Facebook
> royw@fb.com
>