According to many papers, about 70%-80% of rendering time is spent on intersections. So the first target to optimize for performance is to accelerate intersections. Due to the parallelism essence of rays, GPU is of course the best choice.
After copying all geometry and lights information to GPU (CUDA compiler has so many constrains so it is not a easy task..), together with an iterative-style kd-tree on GPU(the same style as the recursive ray-tracing actions on GPU), I got a significant performance speed-up:
Testing Environment:
CPU: Intel Core(TM)2 Quad CPU Q9400 @ 2.66GHz 2.67GHz
GPU: NVIDIA GeForce 310
[Venus.obj: 1418 Triangles
] (values in unit of second):
| KD-tree Depth | 2 | 6 |
| GPU | 11.99 | 21.71 |
| CPU | 27.52 | 31.87 |
| Speed-up | 129.52% |
46.80% |
As above, it is clear that kd-tree’s operation cost will increase rapidly on GPU because of the large amount of “if-else
” executions, which is one of the main killers against GPU execution. It also happens on CPU.
Another object: [castle.obj, 13114 triangles
]
| KD-tree Depth | 2 | 6 |
| GPU | 90.06 | 107.99 |
| CPU | 337.11 | 340.51 |
| Speed-up | 274.32% |
215.32% |
More triangles can mask the kd-tree operation cost efficiently so this object shows the result better.