vertex colors are not exported

[corentinjaffre]

They are by the autodesk 3dsMax collada exporter, but not at all by openCOLLADA.

[bcalidas]

We would strongly prefer option 1. Under option 2, we are forced to change our implementation to calculate values at a higher precision. This leads to loss of performance which defeats the purpose of using half-floats in the first place. When an application chooses to use fp16 , it should also be prepared to handle/accept any underflow/overflow issues from the lower precision level.

[bashbaug]

FWIW, we currently implement dot with pure fp16:

https://github.com/intel/intel-graphics-compiler/blob/master/IGC/BiFModule/Implementation/Geometric/dot.cl#L69

We implement cross though by upconverting to fp32 - not sure why:

https://github.com/intel/intel-graphics-compiler/blob/master/IGC/BiFModule/Implementation/Geometric/cross.cl#L53

I agree it would be nice to implement these functions with fp16 to improve performance.

The one part I don't quite understand just yet is how we're passing the new CTS tests when we implement dot with fp16. We are implementing this built-in with a sequence of fmas, rather than multiplies and adds. Could that have something to do with it?

If I get a chance I'll try a pure fp16 cross also to see what happens.

[bashbaug]

I checked with a pure fp16 cross and we fail the new test with this implementation also:

half3 __attribute__((overloadable)) my_cross(half3 p0, half3 p1) {
    half3 result;
    result.x = fma(p0.y, p1.z, -p0.z * p1.y);
    result.y = fma(p0.z, p1.x, -p0.x * p1.z);
    result.z = fma(p0.x, p1.y, -p0.y * p1.x);
    return result;
}

The failure is (I haven't dug deeper):

ERROR: Data sample 124 does not validate! Expected (inf,0x1.bdcp+15,inf,0x0p+0), got (0x1.bdcp+15,inf,0x0p+0,inf)
    Input: (0x1.07p+8 0x1.16cp+7 -0x1.0a4p+8) and (-0x1.e44p+8 0x1.dap+7 0x1.114p+8)
    Errors: (nan out of 0x1.b57e58p+7), (inf out of 0x1.5780c6p+9), (nan out of 0x1.5780c6p+9)
     ulp -3831.000000

I'll comment on the CTS PR also. It would be great to get this clarified in the spec (and then fix the test to match).

[bcalidas]

Do we have agreement that it is conformant to implement dot and cross products with fp16 precision.