Here is a breakdown of some of the Neon intrinsics that were used to optimize the AABB collision detection in the function called NeonAABBObjCollisionDetectionUnrolled. It can be found at line 718 in CollisionCalculationScript.cs.
NeonAABBObjCollisionDetectionUnrolled performs the collision detection between the characters and the walls. The outer loop iterates through all of the characters, while the inner loop iterates through the walls. The result is an array of boolean values (true denotes a collision has occured) which tells us which characters have collided with which walls.
Unity.Burst.Intrinsics.v64 (loading data into a vector register)
Line 721: var tblindex1 = new Unity.Burst.Intrinsics.v64((byte)0, 4, 8, 12, 255, 255, 255, 255)
Create a 64-bit vector with 8 8-bit elements with values (0, 4, 8, 12, 255, 255, 255, 255). This is used as a lookup table on line 741.
vdupq_n_f32
Line 728: charMaxXs = vdupq_n_f32(*(characters + c))
Duplicate floating point values into all 4 lanes of the 128-bit returned vector. The returned vector will contain 4 copies of a single Max X value (of character bounds).
vld1q_f32
Line 736: wallMinXs = vld1q_f32(walls + w)
Load multiple floating point values from memory into a single vector register. The returned vector will contain Min X values from 4 different walls.
vcgeq_f32
Line 741: vcgeq_f32(wallMinXs, charMaxXs)
Floating point comparisons (greater-than or equal). It compares 4 walls at once with a character’s Max X. Each of the four results will either be all ones (true) or all zeros (false).
vorrq_u32
Line 741: vorrq_u32(vcgeq_f32(wallMinXs, charMaxXs), vcgeq_f32(wallMinYs, charMaxYs))
Bitwise inclusive OR. The nested calls to vcgeq\_f32 are comparing the walls (Min X and Min Y) against the characters’ Max X and Max Y. The four comparison results are combined with a bitwise OR.
vqtbl1_u8
Line 741: results = vqtbl1_u8(_result of ORs_, tblindex1)
Table lookup function that selects elements from an array based on the indices provided. The result of the OR operations will be treated as an array of 8-bit values. The values from tblindex1 (0, 4, 8 and 12) ensure that we select the most significant bytes from each u32 OR result. So 4 character-wall comparisons are being merged into one 128-bit vector along with 4 dummies (because of the out of range values in tblindex1) that will be replaced later with the 64-bit value of the next 4 wall comparisons (from wmvn_u8).
vqtbx1_u8
Line 751: vqtbx1_u8(results, …)
Table lookup function except when an index is out of range as it leaves the existing data alone. This has the effect of selecting 4 bytes (using indices from tblindex2) from the results of the last 4 comparisons and combines them with the previous 4 results. results will now contain the results of 8 wall-character comparisons.
vmvn_u8
Line 751: results = vmvn_u8(...)
Bitwise NOT operation. This negates each of the 8 character-wall comparisons. It is effectively the ! (NOT) in our AABB intersection function except that it is working on 8 results instead of 1.
Unity.Burst.Intrinsics.v64 (storing to memory)
Line 755: *(Unity.Burst.Intrinsics.v64*)(collisions + (c * numWalls + w - 4)) = results;
This is a vectorized form of writing out the 8 results to memory all at once. At the time of writing, Burst doesn’t include the VSTR Neon intrinsic that would do this so this is the most efficient Burst way to get data out of Neon registers.