NATIVE mode - This made the library use BYOND's native pixel movement feature instead of its own. This improved performance but limited the features.
LIBRARY mode - This was the old implementation that was slower, but had support for additional features (3D movement, ramps, etc.)
The v4.1 update adds a new mode - HYBRID mode. This new mode uses some built-in procs (like obounds) to improve the performance of library mode. This means that hybrid mode's performance is on par with native mode and it supports features that native mode doesn't.
The reason why library mode was slow was because every time you moved, the library had to look for nearby obstacles that you might hit. It checked your move against all atoms in oview(1). Most of those atoms you were certainly not going to hit but it checked anyway. Hybrid mode doesn't check all atoms in oview(1), it uses obounds(src) to check your move against atoms in a much smaller area.
Hybrid mode's performance is on par with native mode for 2D movement (because native movement only supports 2D). To compare performance I ran the benchmarks\demo-1 demo five times in each mode until pixel_move was called 100,000 times. I dropped the best and worst runs for each mode, then averaged the remaining three trials. I also ran it in library mode to show how much better the other two modes are:
NATIVE proc Time (s) Calls Avg pixel_move 22.970 302277 76.0 set_flags 17.571 302277 58.1 set_pos 9.941 302277 32.9 TOTAL 167.0 HYBRID proc Time (s) Calls Avg pixel_move 18.279 302073 60.5 set_flags 12.037 302073 39.8 set_pos 10.430 302073 34.9 TOTAL 134.9 LIBRARY proc Time (s) Calls Avg pixel_move 55.521 301180 184.3 set_flags 30.207 301180 100.3 set_pos 32.841 301180 109.0 TOTAL 393.7The "Time" column is the value shown in the profiler for "Real Time". The "Calls" column is the number of total calls to each proc. The "Avg" column is the average time per call in microseconds (millionths of a second).
These are the main pixel movement procs and both modes are significantly better than library mode. Hybrid mode appears to be the best, but the results are a little deceiving. The results fluctuated greatly.
Hybrid mode had four decent trials and one good one. Native mode had three good trials and two bad ones. I dropped the best and worst to help balance it out. Native mode's total value (the last row in the table) was around 130 for each trial, where hybrid mode was around 140. Native mode had a trial with a total of 247 which pulled its average up. I tried to keep the fluctuations out of the results but that's part of the performance.