07-11-2018 07:33 PM
I can run CHaiDNN successfully on ZCU102.
However, I don't know how make CHaiDNN run faster. Do you have a plan to make it come true?
thanks in advance,
07-12-2018 12:42 AM
What scale of improvement do you need? I was able to make it faster by pipelining multiple images. I obtained 10% increase in performance (images per second) by pipelining two images. In original CHaiDNN source code, it executes one image at a time. You have to edit the xi_scheduler.cpp and <example>_ex.cpp(googlenet_ex.cpp) codes to do that.
07-12-2018 05:08 AM
Thanks your advise,
Today, I notice the number of images just 1 in hw_setting.h and xi_scheduler.cpp
//# Number of images in the Queue
#define NUM_IMG 1
int totalImages = NUM_IMG;//total_layers;
So I have the same idea as you. Two images is not enough this issue , maybe wo can try pipelining 10 or 100 images to show how effective this method is.
Perhaps there are different ways to make it faster, like using more DSP or reduce weights loading time. If you have more methods, please leave a message here, I will reply as faster as I can.
Thanks a lot.