Improve SIFT for arm64/Apple silicon

[Developer-Ecosystem-Engineering]

Reduce branch density by collapsing compares.

Performance improvements from 1.03 to 1.53 with existing tests

% ./modules/ts/misc/summary.py -u us -m median ./build-{3.4,improv-sift}/*.xml | grep SIFT
(x-factor)
detect::OCL_feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 50256.333 48789.417 1.03
detect::OCL_feature2d::(SIFT_DEFAULT, "stitching/a3.png") 64591.083 52250.542 1.24
detect::OCL_feature2d::(SIFT_DEFAULT, "stitching/s2.jpg") 132094.667 99417.667 1.33
detect::feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 77521.167 60114.083 1.29
detect::feature2d::(SIFT_DEFAULT, "stitching/a3.png") 70179.083 55914.917 1.26
detect::feature2d::(SIFT_DEFAULT, "stitching/s2.jpg") 137359.208 101112.250 1.36
detectAndExtract::OCL_feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 99694.083 67476.833 1.48
detectAndExtract::OCL_feature2d::(SIFT_DEFAULT, "stitching/a3.png") 101392.958 71199.708 1.42
detectAndExtract::OCL_feature2d::(SIFT_DEFAULT, "stitching/s2.jpg") 203133.500 133395.083 1.52
detectAndExtract::feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 98465.083 66492.083 1.48
detectAndExtract::feature2d::(SIFT_DEFAULT, "stitching/a3.png") 100793.042 68811.333 1.46
detectAndExtract::feature2d::(SIFT_DEFAULT, "stitching/s2.jpg") 203328.083 132630.917 1.53
extract::OCL_feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 26723.042 19980.250 1.34
extract::OCL_feature2d::(SIFT_DEFAULT, "stitching/a3.png") 19655.625 15591.500 1.26
extract::OCL_feature2d::(SIFT_DEFAULT, "stitching/s2.jpg") 59479.917 44881.208 1.33
extract::feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 27984.542 19793.625 1.41
extract::feature2d::(SIFT_DEFAULT, "stitching/a3.png") 21164.250 16235.250 1.30
extract::feature2d::(SIFT_DEFAULT, "stitching/s2.jpg")
extract::feature2d::(SIFT_DEFAULT, "stitching/s2.jpg") 58240.000 44191.833 1.32

Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

• I agree to contribute to the project under Apache 2 License.
• To the best of my knowledge, the proposed patch is not based on a code under GPL or other license that is incompatible with OpenCV
• The PR is proposed to proper branch
• There is accuracy test, performance test and test data in opencv_extra repository, if applicable
  Patch to opencv_extra has the same branch name.
• The feature is well documented and sample code can be built with the project CMake

force_builders=ARMv7,ARMv8,Custom
buildworker:Custom=linux-3
build_image:Custom=ubuntu:18.04
CPU_BASELINE:Custom=AVX512_SKX
disable_ipp=ON

[Developer-Ecosystem-Engineering]

Investigating build failures

[tomoaki0705]

TIPS 1 : rather filtering with grep, I recommend to filter using --gtest_filter option that you pass to the gtest executable file

./opnecv_test_features2d --gtest_filter=*SIFT*

TIPS 2: use -o markdown option of summary.py

summary.py -o markdown -m median build-before.xml build-after.xml

Will give you something like this

Median
|Name of Test|build-before|build-after|build-before vs build-after(x-factor)|
|---|:-:|:-:|:-:|
|detect::OCL_feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png")|50.256 ms|48.789417 ms|1.03|
|detect::OCL_feature2d::(SIFT_DEFAULT, "stitching/a3.png")|64.591083ms| 52.250542ms | 1.24 |

Which is WAY MORE READABLE on Github

Median

Name of Test	build-before	build-after	build-before vs build-after(x-factor)
detect::OCL_feature2d::(SIFT_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png")	50.256 ms	48.789 ms	1.03
detect::OCL_feature2d::(SIFT_DEFAULT, "stitching/a3.png")	64.591ms	52.250ms	1.24

Also, the commit message

only enable for NEON
is a red flag for me.

[Developer-Ecosystem-Engineering]

Thanks for the tips, will reformat.

only enable for NEON
is a red flag for me.

We tested full vector on coffeelake, marginal up (and down).
Tested 128 bit vector and saw even lower benefits.

Since the benefit was clear on NEON, thats where we landed.

[vpisarev]

@Developer-Ecosystem-Engineering, thank you for the contribution! This work on improving OpenCV@M1 performance is brilliant!

Note, however, that except for the kernels in DNN module, which are few and really critical, we do not accept native optimizations any longer. It would be just impossible for our tiny team to maintain all those branches. Please, rewrite the native NEON code using our universal intrinsics.

[Developer-Ecosystem-Engineering]

@Developer-Ecosystem-Engineering, thank you for the contribution! This work on improving OpenCV@M1 performance is brilliant!

Note, however, that except for the kernels in DNN module, which are few and really critical, we do not accept native optimizations any longer. It would be just impossible for our tiny team to maintain all those branches. Please, rewrite the native NEON code using our universal intrinsics.

An updated patch is available with it rewritten.

[vpisarev]

@Developer-Ecosystem-Engineering, thank you! the patch is almost ready to be merged. Please, fix the compile warnings on Windows (see pullrequest.opencv.org) and squash commits into one.

[Developer-Ecosystem-Engineering]

@Developer-Ecosystem-Engineering, thank you! the patch is almost ready to be merged. Please, fix the compile warnings on Windows (see pullrequest.opencv.org) and squash commits into one.

• Squashed into a single commit
• Fix windows build errors
• Use OpenCV universal intrinsics
• Remove CV_NEON/CV_SIMD guards

[alalek]

Thank you for contribution!

[Developer-Ecosystem-Engineering]

Modifications requested by @alalek have been integrated and re-squashed.

[alalek]

Looks good to me 👍

/cc @vpisarev