Search code examples
cudagraph-theorynpp

NPP graph cut for non-integer edge capacities

The nppiGraphcut_32s8u function works on 32 bit signed integer edge capacities. In the grabCut example included with the CUDA 6.5 toolkit, they simply cast Npp32f to Npp32s. For example, from GrabcutUtil.cu in the grabcutNPP example folder:

// Top/Bottom
Npp32s bottom;
bottom = _FIXED(edge_weight(center, tex2D(imageTex, x + 0.5f ,y + 1.5f), alpha, beta, 1.0f));

where edge_weight returns a Npp32f. Did they scale up their numbers so that the edge capacities would be large values and thus avoid rounding off error?

Solution

  • It is not a 'cast'. Look at the implementation: _FIXED first multiplies the number by 10 and then calls a rounding routine.

    So yes, the number is scaled up, although only by one digit.