Am 20.01.2011 10:20, schrieb anupama:
> Hi,
>     I am working on povray and the objective is to test if povray is build
> correctly across compilers for different optimizations. I have choosen the CPU
> benchmark provided at http://www.povray.org/download/benchmark.php for this
> exercise. I have enabled "All_File=povray.log" in the benchmark.ini to capture
> the povray log.
>     But I see that the povray log file generated has different output for
> different compiler but the image file size remains same. A sample of the
> parameters that are differing are given below:
> ----------------------------------------------------------------------------
> Ray->Shape Intersection          Tests       Succeeded  Percentage
> ----------------------------------------------------------------------------
> Box                          different        different      different
> Cone/Cylinder                different        different      different
> CSG Intersection             different        different      different
> CSG Merge                    different        different      different
> Fractal                      different        different      different
> Height Field                 different        different      different
> Height Field Box             different        different      different
> Height Field Triangle        different        different      different
> ----------------------------------------------------------------------------
>
> where "different" implies the number of "Tests" "Succeeded" and "Percentage" is
> different for different compilers.
>
> Can anyone please help me to understand how to verify if the rendering is done
> correctly using the log file? Or any other method.

Try using "+WT1" to use just a single render thread. With multiple 
threads, OS scheduling adds a random element to the render, leading to 
some differences.

You may also try "+HR" which should increase reproducability, but 
currently it's not a cure-all.

Also note that different platforms might be using different rounding 
modes for floating point arithmetic, which may affect the result as well.

The effects on rendered output should be very subtle, but render 
statistics will differ nonetheless, so I'd suggest comparing the output 
images instead.

Post a reply to this message