Skip to content

Change fast corner flags in HAL version from char array to single int#15357

Merged
opencv-pushbot merged 1 commit intoopencv:3.4from
ChipKerchner:fastCorner
Aug 29, 2019
Merged

Change fast corner flags in HAL version from char array to single int#15357
opencv-pushbot merged 1 commit intoopencv:3.4from
ChipKerchner:fastCorner

Conversation

@ChipKerchner
Copy link
Copy Markdown
Contributor

@ChipKerchner ChipKerchner commented Aug 20, 2019

Change fast corner flags in HAL version from char array to single int

@alalek
Copy link
Copy Markdown
Member

alalek commented Aug 21, 2019

relates #14916

@ChipKerchner
Copy link
Copy Markdown
Contributor Author

ChipKerchner commented Aug 21, 2019 •
edited
Loading

I'm seeing slightly better performance with v_signmask than the array of chars. If you disagree we can cancel this pull request - I didn't realize that I'm undoing a previous check-in. You would think a single variable would be better than multiple memory accesses.

@alalek
Copy link
Copy Markdown
Member

alalek commented Aug 21, 2019

This code is under CV_SIMD128, so v_signmask() is probably fine there.
@terfendail Could you please check performance changes?

@terfendail
Copy link
Copy Markdown
Contributor

terfendail commented Aug 28, 2019

Performance for SSE2 baseline
Performance test Reference time PR time Speedup
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.241 1.307 0.95
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/a3.png") 0.718 0.752 0.96
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/s2.jpg") 4.954 5.014 0.99
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 3.916 3.949 0.99
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/a3.png") 2.195 2.233 0.98
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/s2.jpg") 14.500 14.498 1.00
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.077 1.039 1.04
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/a3.png") 0.731 0.715 1.02
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/s2.jpg") 3.973 3.807 1.04
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.311 1.313 1.00
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/a3.png") 0.752 0.753 1.00
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/s2.jpg") 5.113 5.169 0.99
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 4.122 4.228 0.97
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/a3.png") 2.397 2.456 0.98
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/s2.jpg") 15.509 16.036 0.97
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.374 1.394 0.99
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/a3.png") 1.024 1.050 0.98
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/s2.jpg") 5.815 5.976 0.97
detect::feature2d::(FAST_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 2.569 2.427 1.06
detect::feature2d::(FAST_DEFAULT, "stitching/a3.png") 2.076 2.056 1.01
detect::feature2d::(FAST_DEFAULT, "stitching/s2.jpg") 9.493 9.404 1.01
Performance for SSE3 baseline
Performance test Reference time PR time Speedup
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.287 1.300 0.99
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/a3.png") 0.749 0.753 0.99
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/s2.jpg") 4.953 4.951 1.00
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 3.842 3.770 1.02
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/a3.png") 2.198 2.171 1.01
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/s2.jpg") 14.532 14.360 1.01
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.068 1.015 1.05
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/a3.png") 0.726 0.692 1.05
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/s2.jpg") 3.965 3.746 1.06
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.297 1.252 1.04
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/a3.png") 0.749 0.718 1.04
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/s2.jpg") 5.104 4.930 1.04
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 4.070 3.893 1.05
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/a3.png") 2.332 2.370 0.98
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/s2.jpg") 15.205 15.809 0.96
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.438 1.397 1.03
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/a3.png") 1.084 1.053 1.03
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/s2.jpg") 6.188 5.981 1.03
detect::feature2d::(FAST_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 2.550 2.427 1.05
detect::feature2d::(FAST_DEFAULT, "stitching/a3.png") 2.143 2.000 1.07
detect::feature2d::(FAST_DEFAULT, "stitching/s2.jpg") 9.804 8.987 1.09
Performance for SSE4_2 baseline
Performance test Reference time PR time Speedup
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.276 1.262 1.01
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/a3.png") 0.731 0.736 0.99
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/s2.jpg") 4.814 4.869 0.99
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 3.779 3.893 0.97
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/a3.png") 2.152 2.221 0.97
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/s2.jpg") 13.805 14.470 0.95
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.026 1.053 0.97
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/a3.png") 0.700 0.709 0.99
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/s2.jpg") 3.962 3.841 1.03
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.224 1.271 0.96
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/a3.png") 0.706 0.732 0.96
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/s2.jpg") 4.817 5.005 0.96
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 4.163 4.111 1.01
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/a3.png") 2.404 2.415 1.00
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/s2.jpg") 15.977 16.113 0.99
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.454 1.398 1.04
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/a3.png") 1.093 1.049 1.04
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/s2.jpg") 6.262 5.947 1.05
detect::feature2d::(FAST_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 2.646 2.542 1.04
detect::feature2d::(FAST_DEFAULT, "stitching/a3.png") 2.135 2.045 1.04
detect::feature2d::(FAST_DEFAULT, "stitching/s2.jpg") 9.831 9.366 1.05
Performance for AVX2 baseline
Performance test Reference time PR time Speedup
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.338 1.296 1.03
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/a3.png") 0.777 0.757 1.03
detect::feature2d::(FAST_20_FALSE_TYPE5_8, "stitching/s2.jpg") 5.079 4.935 1.03
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 3.854 3.863 1.00
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/a3.png") 2.215 2.203 1.01
detect::feature2d::(FAST_20_FALSE_TYPE7_12, "stitching/s2.jpg") 14.048 14.471 0.97
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 0.800 0.837 0.96
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/a3.png") 0.547 0.568 0.96
detect::feature2d::(FAST_20_FALSE_TYPE9_16, "stitching/s2.jpg") 2.770 2.906 0.95
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.346 1.313 1.03
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/a3.png") 0.780 0.758 1.03
detect::feature2d::(FAST_20_TRUE_TYPE5_8, "stitching/s2.jpg") 5.057 5.224 0.97
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 3.923 4.106 0.96
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/a3.png") 2.311 2.301 1.00
detect::feature2d::(FAST_20_TRUE_TYPE7_12, "stitching/s2.jpg") 15.796 15.269 1.03
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 1.184 1.143 1.04
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/a3.png") 0.908 0.869 1.05
detect::feature2d::(FAST_20_TRUE_TYPE9_16, "stitching/s2.jpg") 5.027 4.801 1.05
detect::feature2d::(FAST_DEFAULT, "cv/detectors_descriptors_evaluation/images_datasets/leuven/img1.png") 2.110 2.111 1.00
detect::feature2d::(FAST_DEFAULT, "stitching/a3.png") 1.740 1.675 1.04
detect::feature2d::(FAST_DEFAULT, "stitching/s2.jpg") 8.354 7.954 1.05

@terfendail
Copy link
Copy Markdown
Contributor

terfendail commented Aug 28, 2019

It looks like the change makes performance better a bit. Actually the change to array was a part of avoiding of non-universal v_signmask, but I agree that for SIMD128 it's not a big issue to use it.

opencv-pushbot pushed a commit that referenced this pull request Aug 29, 2019
@alalek
Merge pull request #15357 from ChipKerchner:fastCorner
8d0b3da
@opencv-pushbot opencv-pushbot merged commit 51ceabb into opencv:3.4 Aug 29, 2019
@alalek alalek mentioned this pull request Aug 30, 2019
Merged
@ChipKerchner ChipKerchner deleted the fastCorner branch September 3, 2019 13:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@terfendail terfendail terfendail approved these changes

Assignees

@terfendail terfendail

Labels

category: features2d optimization

Projects

None yet

Milestone

3.4.8

Development

Successfully merging this pull request may close these issues.

4 participants

@ChipKerchner @alalek @terfendail @opencv-pushbot