sayan1an / chakumrender Goto Github PK

At eb35b96.
Notes to self:
Dummy is very sensitive to the sequence of points. Performance of dummy is highest when points are arranged as per their closeness to each other.

Bvh right now is more time consuming than Dummy with SPLIT_HALF formulation. Performance for BVH increases with increasing no. of primitives in leaf-node. So, there is significant overhead
in traversal code and needs further optimizations. Ray - Box intersections needs more optimization.

Test conditions:
Using A8-4500M on all cores. Size of image 50x50, X, Y, Z = 45, 0, -20
mat3 M = scale(27, 27, 27).mul(rotateY(PI/4 + PI));
Object:Aventador1.obj

Timing dummy:
8s - with all primitives arranged as per their closeness to each other.
21s- with primitives reshuffled.

Timing Bvh:
45s- Using no. of primitives in leaf node = 1.
25s - Using no. of primitives in leaf node = 1000.

There is some serious bug in bvh code. Each ray should intersect 2log(n) nodes but actually it is way more.