Benchmark

Datasets

Dataset is selected based on maximum resolution of the image to be processed, number of pixels processed per clock cycle (NPPC) and few other function specific parameters.

Performance

Resource utilization numbers along with the achieved Frames Per Second (FPS) are represented for standalone kernel for both Full HD(1920x1080) and 4K(3840x2160) image resolutions. CPU numbers are an average of 100 runs and are calculated on Intel(R) Xeon(R) CPU E5-2680 @ 2.70GHz.

Note

Some of the reference functions against which benchmark is done, although follow similar algorithm as Vitis Vision function, doesnot match functionally with the corresponding Vitis Vision function because of different output format. In such cases there are two reference functions, one, a standard function, used to calculate the performance which can be enabled or disabled through the macro “__XF_BENCHMARK” defined in the config file. The other is internal reference function used for verifying functional correctness against the Vitis Vision function.

Table 1 Performance on FPGA

Algorithm	Dataset			FPS(CPU)	FPS(FPGA)	LUT	BRAM	FF	DSP
	Resolution	NPPC	other params
Canny Edge tracing	4K	8	L1 Norm, Filter - 3x3	9	95	31408	132	19148	96
	FHD	8	L1 Norm, Filter - 3x3	25	333	17451	65	11256	63
Harris Corner Detection	4K	8	L1 Norm, Filter - 3x3	22	289	40460	227	27236	208
	FHD	8	L1 Norm, Filter - 3x3	62	1100	22478	113	16021	138
Fast Corner Detection	4K	8	NA	79	289	21171	10	13396	0
	FHD	8	NA	186	1100	20437	10	14322	0
Kalman filter	NA	1	SV - 16x16x16	6.57ms (latency)	0.55 ms (latency)	59342	98	90762	361
Lk Dense Pyramidal Optical Flow(2 frames)	4K	1	5 iterations, 5 levels	0.15	3	30781	182	26169	83
	FHD	1	5 iterations, 5 levels	0.63	12	30839	107	25714	83
Corner Tracker (2 frames)	4K	1	MAXCORNERS-10K	0.53	3	37547	225	35089	115
	FHD	1	MAXCORNERS-10K	2	11	34624	129	34198	115
Colordetect	4K	1	Filter - 3x3	15	289	16523	176	9145	3
	FHD	1	Filter - 3x3	28	1100	9961	91	5432	1
Gaussian difference	4K	1	Filter - 3x3	126	289	27235	35	37426	281
	FHD	1	Filter - 3x3	500	1100	15146	19	22032	193
Bilateral filter	4K	8	Filter - 3x3	72	289	22523	35	22862	65
	FHD	8	Filter - 3x3	200	1100	12513	18	13451	44
Stereo LBM	4K	1	PARALLEL Units - 32 Disparity - 32	13	34	19005	26	21113	7
	FHD	1	PARALLEL Units - 32 Disparity - 32	35	135	19380	13	20670	7
ISP Pipeline	4K	4	16-bit	0.11	135	18987	24	17713	91
	FHD	4	16-bit	0.44	520	19254	19	17534	88

Below are the links to the individual benchmark result files.

• Canny Edge Detection
• Harris Corner Detection
• FAST Corner Detection
• Kalman Filter
• Dense Pyramidal LK Optical Flow
• Corner Tracker
• Color Detect
• Gaussian Difference
• Bilateral Filter
• Stereo Local Block Matching
• Image Sensor Processing (ISP) Pipeline

Vitis Vision Library

• Download code

These Vitis Vision benchmarks can be downloaded from vitis libraries master branch.

git clone https://github.com/Xilinx/Vitis_Libraries.git
cd Vitis_Libraries
git checkout master
cd vision

• Setup environment

Specify the corresponding Vitis, XRT, FPGA device, and path to the OpenCV libs by running following commands.

source <intstall_path>/installs/lin64/Vitis/2021.1_released/settings64.sh
source <path-to-xrt-installation>/setup.sh
export DEVICE=< path-to-platform-directory >/< platform >.xpfm
export OPENCV_INCLUDE=< path-to-opencv-include-folder >
export OPENCV_LIB=< path-to-opencv-lib-folder >
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:< path-to-opencv-lib-folder >