cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
buaaneo
Observer
Observer
229 Views
Registered: ‎01-27-2021

DPU test_performance

According to the custom platform flow, I tested the image processing speed of one DPU and two DPUs, and found that the image processing speed is the same. After I changed the frequency of the DPU, I found that the image processing speed was still the same. But one DPU test_performance's FPS is relatively low, and after reducing the frequency, test_performance's FPS will also be reduced.

That is to say, changing the number and frequency of DPUs will not affect the image processing speed, but will only affect the test_performance's FPS.

I don't know how to make full use of DPU's resources when batch processing images? How to write c++ postprocessing code to approach test_performance's FPS?

0 Kudos
2 Replies
chaoz
Xilinx Employee
Xilinx Employee
179 Views
Registered: ‎09-14-2018

@buaaneo 

From my knowledge edge DPU don't support batch (Alveo DPU do support batch). 
If you implement 2 DPUs, they can both process one image simultaneously, that will enhance the performance but still the batch size is one. 
Hope this can explain it a bit.

Chao

----------------------------------------------------------------------------------------------
Don't forget to "Accept as solution" or "Kudo" if it helps. Thanks!
----------------------------------------------------------------------------------------------

Chao
----------------------------------------------------------------------------------------------
如果帖子有帮助,别忘“接受为解决方案”或“奖励”。谢谢!
Don't forget to "Accept as solution" or "Kudo" if it helps. Thanks!
----------------------------------------------------------------------------------------------
buaaneo
Observer
Observer
165 Views
Registered: ‎01-27-2021

Thank you for your reply and attention.

My task is to deal with a large number of images continuously. Follow the Test_ Jpeg processing method, I process one picture at a time, found that two DPUs can not improve the processing speed, and the time spent is the same as one DPU. Sequential processing can not reach the level of Test_ Performance.

I found that Test_ Performance is a test method using multithreading. Can multithreading improve the processing speed of my tasks? Where to use multithreading?

0 Kudos