tinyFPS
use FPGA to accelerate the Farthest Point Sampling operation
what is FPS
Farthest Point Sampling is a important operation using in PointCloud Sampling. For a PointCloud P, this operation downsample a small pointset S.
Why we accelerate the FPS
FPS is a normal operation for PointCloud. In recent study, we find this operation has long running time.
e.g.
| num_point | num_sample | CPU(ms) | NVIDIA 1080Ti runtime(ms) |
|---|---|---|---|
| 1024 | 32 | 4.862 | 0.178 |
| 1024 | 64 | 7.259 | 0.227 |
| 1024 | 128 | 11.010 | 0.422 |
| 1024 | 256 | 18.134 | 0.719 |
How to accelerate
we will design a parallel architecture to excute it more efficiently by data reuse and overlapping the transformation and calculation stage, and using the Xilinx HLS tool to implement this architecture.
HLS
use Vitis_hls 2020.1, Compared to the Intel CPU, we achieves on 7.04x speedup when the 1024 points sample 32 points
solution
there are two solution: 'basic' and 'opt', the 'basic' solution has no directive, and in 'opt' solution, we try to optimize the bottleneck by use 'ARRAY_PARTITION', 'UNROLL', 'PIPELINE' directive, and get a little optimization.
the Performance Estimates:
the Utilization Estimates:
