Precision or mathematic problem

Polar01 — Tue, 05 Apr 2011 08:51:53 GMT

Hi,

I'm working on a rendering software (http://spectralbattle.wordpress.com/) and use several SDK to test my code. Using several SDK allow me to check if the errors are in my code... or in the SDK.

When I use the Intel SDK I have some problems :

Here is a rendering with AMD SDK :https://picasaweb.google.com/polarlights01/CornellAMD#5592014900335257794

Here is the same rendering with Intel SDK :

https://picasaweb.google.com/polarlights01/CornellIntel#5592017101324392082

The problem can be seen on the top of the box, a "normal" sounds to be inversed !

The normal orientation is simply computed with a dot product like this :

int isInside = (dot(Ng, I) < 0);

N = isInside ? -Ng : Ng;

I have also try to compute directly the dot product (Ng.x*I.x+Ng.y*I.y+Ng.z*I.z) , but I have the same effect !

Maybe the error is somewhere, but what I see is that I don't have this problem with AMD SDK !

Precision or mathematic problem

Polar01 — Tue, 05 Apr 2011 11:49:58 GMT

Finally,

I have found a way to fix it !

But really it is not related to my code ! I have just replace some operation with corresponding float3 operations and it works !

So, there is a bug somewhere :-P

Thanks

topic Precision or mathematic problem in OpenCL* for CPU

Precision or mathematic problem

Precision or mathematic problem