Timing performance of vpiSubmitOpticalFlowPyrLK() with different backend flags

[Appletree]

Hi,

I use the attached function to find the corresponding points between two frames on Orin.
tmp.txt (5.1 KB)

When monitoring the used time during calling NvVpiPyrLKOpticalFlow::sparse(), I found the difference between VPI_BACKEND_CPU and VPI_BACKEND_CUDA is small in both the following two scenarios. Is this normal? I also found that there are a few milliseconds difference by setting flags=0 and flags = vpiBackend_, its hard to understand also. The third thing confusing me is that I still saw some GPU usage activities when
backend flag is VPI_BACKEND_CPU via jtop.

How to explain these behaviors?

scenario 1:
line 264 const uint64_t flags = vpiBackend_;
backendFlag=0 [VPI_BACKEND_CPU], w=1920, h=1080
frameCnt=200, perFrameTimeUsage(mu=24020.3, std=4153.9)(macro sec)
backendFlag=1 [VPI_BACKEND_CUDA], w=1920, h=1080
frameCnt=200, perFrameTimeUsage(mu=21908.3, std=4589.6)(macro sec)

scenario 2:
line 264 const uint64_t flags = 0;
backendFlag=0 [VPI_BACKEND_CPU], w=1920, h=1080
frameCnt=200, perFrameTimeUsage(mu=30842.4, std=3925.4)(macro sec)

backendFlag=1 [VPI_BACKEND_CUDA], w=1920, h=1080
frameCnt=200, perFrameTimeUsage(mu=24420.1, std=4000.2)(macro sec)

[AastaLLL]

Hi,

When you test the performance, have you maximized the device clock?
https://docs.nvidia.com/vpi/algo_performance.html#maxout_clocks

We do list the performance of the LK tracker here:
For a 1920x1080 image, CPU takes 0.412±0.002 ms and CUDA takes 0.0586±0.0002 ms.
https://docs.nvidia.com/vpi/algo_optflow_lk.html

Tracking depends on the number of detected objects.
So it’s possible to vary for different scenarios.

Thanks.

[Appletree]

@AastaLLL

maximized the device clock

Thank you, when the clock is maximized, I do see a factor of 2 improvement. But it’s not clear where are the start and end points to measure the elapsed time in the performance table mentioned in the above link.

[AastaLLL]

Hi,

The benchmark is tested similarly to our 05_benchmark example.
It contains the warm-up and benchmark algorithm in a batch manner.

https://docs.nvidia.com/vpi/sample_benchmark.html

For more details, please find the “Benchmarking Method” info below:
https://docs.nvidia.com/vpi/algo_performance.html#benchmark

Thanks.