Coder Corner

Part 10 – integer SIMD

Interlude:

Similar to what we did in version 7, and in sake of completeness, I tried using integer comparisons in the SIMD version as well.

The changes are straightforward: encode floats as integers like we did in version 7, replace SIMD floating-point intrinsics with SIMD integer intrinsics.

There are no special traps here, and not much to report, because at the end of the day the results are slower:

Home PC:

Complete test (brute force): found 11811 intersections in 781780 K-cycles.
17454 K-cycles.
16681 K-cycles.
16669 K-cycles.
17145 K-cycles.
16703 K-cycles.
16683 K-cycles.
16668 K-cycles.
16667 K-cycles.
16862 K-cycles.
16707 K-cycles.
16670 K-cycles.
16689 K-cycles.
16668 K-cycles.
16949 K-cycles.
16710 K-cycles.
16667 K-cycles.
Complete test (box pruning): found 11715 intersections in 16667 K-cycles.

Office PC:

Complete test (brute force): found 11811 intersections in 810906 K-cycles.
16607 K-cycles.
15927 K-cycles.
15648 K-cycles.
15971 K-cycles.
15648 K-cycles.
15960 K-cycles.
15742 K-cycles.
15990 K-cycles.
15837 K-cycles.
15741 K-cycles.
15970 K-cycles.
15651 K-cycles.
16247 K-cycles.
15649 K-cycles.
15834 K-cycles.
15738 K-cycles.
Complete test (box pruning): found 11715 intersections in 15648 K-cycles.

The gains are summarized here:

What we learnt:

Don’t bother with integer comparisons anymore.

We are still using integer SIMD in PhysX for this, so it appears that the PhysX version is sub-optimal. Expect some performance gains in PhysX 3.5.

GitHub code for part 10

This entry was posted on Tuesday, February 14th, 2017 at 7:22 am and is filed under Physics, Programming. You can follow any responses to this entry through the RSS 2.0 feed. Both comments and pings are currently closed.