OCL: speeded up ocl::distanceToCenters

[ilya-lavrenov]

merge with opencv/opencv_extra#92

[hgaspar]

Guys, what is the current performance? Better than 20x slower than cpu i hope? was this because something got broken at some point?

[ilya-lavrenov]

@hgaspar, previous performance:

distanceToCenters::distanceToCentersFixture::(256x256, NORM_L1)      2.244 ms    34.951 ms   0.06
distanceToCenters::distanceToCentersFixture::(256x256, NORM_L2SQR)   3.018 ms    35.044 ms   0.09
distanceToCenters::distanceToCentersFixture::(512x512, NORM_L1)      7.505 ms    155.863 ms  0.05
distanceToCenters::distanceToCentersFixture::(512x512, NORM_L2SQR)   7.150 ms    155.907 ms  0.05

current performance:

distanceToCenters::distanceToCentersFixture::(256x256, NORM_L1)      1.961 ms    1.903 ms    1.03
distanceToCenters::distanceToCentersFixture::(256x256, NORM_L2SQR)   3.163 ms    1.742 ms    1.82
distanceToCenters::distanceToCentersFixture::(512x512, NORM_L1)      7.757 ms    10.135 ms   0.77
distanceToCenters::distanceToCentersFixture::(512x512, NORM_L2SQR)   7.152 ms    10.239 ms   0.70

Cycle, that performs calculation of vector distance (L1 or L2SQR), has wrong condition and as a result only 1/4 of all the calculations were vectorized. And I change calculation scheme: previously work-item calculates distance for each row of centers, e.g. if centers contains n rows, then each work-item calculates n distances. I think it's too much work for each work-item and currently OpenCL kernel contains only one distance calculation.
These fixes and optimizations give resultant acceleration - faster 15-20x then previous, but CPU version is also very fast because it's optimized using TBB and SSE2.

[apavlenko]

👍

[pengx17]

Oh, I did not realize the bug in previous dist implementation ... and it shocked me that find the best match is much faster on CPU.
Thanks for the in sights 👍

[opoplawski]

It appears that this broke API and ABI compatibility for distanceToCenters. Isn't that a bad thing to do in a stable release series?

[apavlenko]

2.4.x releases kept backward binary compatibility, but this was not true for ocl module that was in active development stage in 2013 (that was declared from the very beginning).