Significant regression when moving from ippiErode_8u_C1R (8.1) to ippiErodeBorder_8u_C1R (9.0.1) - about 14x for 640x480 unit8

[Ashish_U_]

As part of our IPP 9 upgrade, we are transitioning from ippiErode_8u_C1R to ippiErodeBorder_8u_C1R. I see significant performance regression (glnxa64, avx2) with this transition. 

I have attached some whittled down standalone repro steps for 8.1 and 9.0.1 on glnxa64 platform. 8.1 takes 0.33ms (fastest from multiple runs) and 9.0.1 takes 4.66 ms (fastest time from multiple runs). While this is for 640x480 images, I notice this pattern for a full range of images from 10x10 to 2kx2k and higher.

I double checked the numerics, and they match, so I assume we have the calling syntax correct. However, this regression is very significant - would you please cross check with your internal performance tests?

[Andrey_B_Intel]

Hi Ashish.

Thanks for your feedback. You are right with this issue and we already have prepared performance fix for Erode/Dilate as part our nearest release of IPP.

[Zhen_Z_Intel]

Hi Ashish,

There's indeed a performance regression for Erode function. I already submitted CQ asking developer team to fix this problem, I will inform you once this problem has been fixed. Thank you.

Best regards,
Fiona