Add a Approximate Distance Field computation extension to irr::ext

[devshgraphicsprogramming]

Result will be used for Sphere Tracing (SDF Shadows) - but those uses are outside the scope of the extension.
http://jamie-wong.com/2016/07/15/ray-marching-signed-distance-functions/

Use Compute Shader, but make the extension a CGLSLFunctionGenerator function only for processing input that is already a distance field along an axis.

Marching parabolas algorithm can be implemented in shared memory, much like the Blur function.
https://prideout.net/blog/distance_fields/

Flow:

1. Input is 2D or 3D texture that has already been swept in the X direction
2. Sweep along Y with extension compute shader
3. Memory Barrier
4. Sweep along Z
5. Done

The distance field should accept a quantization factor (rescaling factor) so that the EDF does not need floating point storage.

[devshgraphicsprogramming]

https://www.semanticscholar.org/paper/Almost-optimal-column-wise-prefix-sum-computation-Tokura-Fujita/4c60f16b9998eca0ac4dadac14b7964f8ab90c2c

[devshgraphicsprogramming]

Pseudo code for a non optimal version (apparently one can use a prefix sum for this)

uint _out = uint(input*QUANITZATION_FACTOR); // values > EDF_RANGE*EDF_RANGE indicate "infinity"
sharedMem[rowInvocationID] = _out ;
BARRIER();
for (int i=-EDF_RANGE; i<EDF_RANGE; i++) // could optimize by splitting into two loops and not processing i==0
{
    int newCoordinate = coordinateWrapAround(rowInvocationID+i); // usually just a x%ROW_LENGTH
    int distanceBetween = rowInvocationID-newCoordinate;
    _out = min(_out ,sharedMem[newCoordinate]+distanceBetween*distanceBetween );
}
return float(_out )*QUANITZATION_FACTOR_RCP;

EDF_RANGE is max quantizable distance, if storing to a UNORM_R8 texture then it will be 255, if storing to SNORM_R8 then it will be 127.
Naturally the intermediate result texture will have to have 2ce the bits.

[devshgraphicsprogramming]

For rows longer than LOCAL_WORKGROUP_SIZE_X (which should be 256 because of mobile devices), then sharedMem would have to be pre-loaded with LOCAL_WORKGROUP_SIZE_X+2*EDF_RANGE samples.

If performance for example 13.SphereTracing is acceptable then the non-optimal (no prefix-sum/scan version can stay)

[devshgraphicsprogramming]

Perf bound on a 320GB/s GDDR5 desktop GPU will be around 6ms for SDF computation.

I expect the non-optimal (non-scan) algorithm to take between 1000-500ms assuming 2TB/s shared memory speed and EDF_RANGE=255.

So to optimize either use prefix-sum/scan or implement the exchange of shared memory within a warp via NV_shuffle (and similar)

[devshgraphicsprogramming]

Reading Resources about rendering those
http://advances.realtimerendering.com/s2015/AlexEvans_SIGGRAPH-2015-sml.pdf