Benchmarking with TI DSP
Introduction:
Embedded vision is a method of implementing computer vision algorithms on embedded systems. Embedded systems are specific purpose microprocessor based computer which can be found in many daily appliances like automobiles, surveillance systems, PDAs and medical equipment.
Today, there are many powerful, low-cost, and energy-efficient processors so that it has become possible to use many computationally extensive computer vision algorithms in real word in a wide range of embedded systems.
One of the current applicable processors used in large scale these days is specialized Digital Signal Processors (DSPs). DSPs provide better performance, lower latency and lower power consumption rather than general purpose CPUs. For example Texas Instruments TMS320 family were created, allowing for a more efficient execution of signal processing code compared to general purpose CPU.
In this research, we explore how computer vision algorithms can be optimized and parallelized for this type of embedded systems and evaluating benchmarking some various implementation strategies (using SIMD, EDMA and Caching).
| Algorithm | Elapsed time | Size of Image |
|---|---|---|
| Median | 8.6 ms | 480 X 640 |
| Median | 2.2 ms | 240 X 320 |
| Median | 0.52 ms | 120 X 160 |
| Fast9 | 6.4 ms | 480 X 640 |
| Fast9 | 1.6 ms | 240 X 320 |
| Fast9 | 0.38 ms | 120 X 160 |
| 8-bit Histogram | 2.22 ms | 480 X 640 |
| GFTT | 187.82 ms | 480 X 640 |
| Gaussian 7*7 | 36.35 ms | 480 X 640 |
| Harris without Non Max Suppression | 213.33 ms | 480 X 640 |
| Harris with Non Max Suppression | 214.77 ms | 480 X 640 |
| Filter 2D | 36.33 ms | 480 X 640 |
| Sobel (3*3) | 46.04 ms | 480 X 640 |
| Sobel (5*5) | 104 ms | 480 X 640 |
| RGB to Grayscale | 46.04 ms | 480 X 640 |
| RGB to YUV | 63.59 ms | 480 X 640 |