Welcome to Fractal Forums

Greetings,
I bring you this report summary of VTune Results; Compute vs. CVector3::IsNotANumber.
The data focuses on CVector3::IsNotANumber.
(http://i.imgur.com/ujF1izC.png)
 

This case is interesting because we are spending much time in Compute function itself.
Within the scope of this Compute function, we use CVector3::IsNotANumber where the majority instructions retire.
Also within the scope of this Compute function call, other operations retire a large share of instructions to self.
The Bottom-up view shows that we retire most instructions in CVector3::IsNotANumber.
(http://i.imgur.com/pusfpiw.png)
 

The Caller/Callee view reinforces the separate and nested scopes where the majority of CPU instructions takes place.
It shows that we spend 68.2% of total time in Compute function.
It further shows that we spend 22.5% of total time in CVector3::IsNotANumber function.
However, the report clearly states that CVector3::IsNotANumber retired nearly double the CPU instructions to self as opposed to the “runner-up” Compute function.
(http://i.imgur.com/KcknmZz.png)
 

There are instances of Compute function throughout the RayMarcher, ObjectShader, CalculateNormals, etc…
(http://i.imgur.com/hW8oFZb.png)
 

This source code view below shows one call stack instance of Compute function.
Here we call CVector3::IsNotANumber()
This instance alone accounts for 11.2% of total CPU time.
(http://i.imgur.com/gNVoT7b.png)
 

This initial source view of algebra.hpp pinpoints the critical section.
(http://i.imgur.com/RQ8c1jA.png)
 

This secondary source view of algebra.hpp breaks down the critical section.
(http://i.imgur.com/jSIbq4F.png)
 

When testing for NaN, any ideas to optimize?

Thanks,
coast