Optimized sample generation in Sobol sequence code

[sshudler]

This code does two things:

• Sets the denormal mode and flush to zero mode (intrinsics API) to improve the performance of Embree (as per Embree recommendation).
• Optimizes the generation of samples in Sobol sequence code. The original code compiled with MSVC++ had an extra load instruction in the loop, which affected branch prediction performance and lowered the IPC. Also, the loop was manually unrolled.

We noticed a performance improvement of about 6%-15% in some scenarios on some systems.

@csakthiv-intel

[BeheadedKamikaze]

There is no longer a bounds check on i for 3 of the 4 indexes into directions before the loop repeats, how do you ensure the offset is always a valid index?

[sshudler]

Before the change and after the change i wasn't used as an index into directions. Whenever the loop was hitting a 1 bit in i it would index into directions by that bit's offset. The way directions defined, starting from the offset it has SOBOL_BITS range for access, so in the new code the access to directions is valid and when i becomes 0 none of these directions values will be used in the result because of & with (0 & 1). Also, SOBOL_BITS == sizeof(i).