cv::magnitudeSqr() #15683

chacha21 · 2019-10-10T11:57:05Z

TODO: ippicv must implement ippicvsPowerSpectr_32f/ippicvsPowerSpectr_64f

accuracy tests
performance tests

TODO: ippicv must implement ippicvsPowerSpectr_32f/ippicvsPowerSpectr_64f

alalek · 2019-10-10T16:24:11Z

modules/core/perf/opencl/perf_arithm.cpp

+
+    OCL_TEST_CYCLE() cv::magnitudeSqr(src1, src2, dst);
+
+    SANITY_CHECK(dst, 1e-6);


Use SANITY_CHECK_NOTHING(); here.

chacha21 · 2019-10-10T17:09:13Z

Where can I ask to get ippicvsPowerSpectr_32f/ippicvsPowerSpectr_64f available ?

alalek · 2019-10-10T18:24:02Z

In the nearest future no updates of IPPICV are planned.
You can guard these calls by #ifndef HAVE_IPP_ICV (use standalone IPP calls if you have it). Need to measure performance benefits of separate IPP implementations of these calls first.

asmorkalov · 2019-10-21T06:39:51Z

@chacha21 Do you have any progress on the patch?

chacha21 · 2019-10-21T08:32:56Z

I do not know how to add accuracy/performance tests. I always had trouble with that ( #13879).
So I do not know how to go further.

chacha21 · 2020-04-16T08:51:24Z

IPPICV has been updated for OpenCV 4.3.0, but ippicvsPowerSpectr_32f/ippicvsPowerSpectr_64f are still not available.
Not a big issue, just an update about that topic.

asenyaev · 2021-04-07T20:10:56Z

jenkins cn please retry a build

asmorkalov · 2023-07-04T13:24:55Z

@chacha21 Is the PR still relevant? Do you plan to work on it?

chacha21 · 2023-07-04T13:40:20Z

@chacha21 Is the PR still relevant? Do you plan to work on it?

If there is no hope to benefit from ippicvsPowerSpectr_32f/ippicvsPowerSpectr_64f , the risk is that magnitudeSqr() could be slower than magnitude(m)^2 in the IPP case. The IPP case could be removed from magnitudeSqr() and only rely on the hal implementation.
Since I have no benchmark procedure for different machine configurations, I can't tell what's best.

since ippsPowerSpectr_32f/64f isnot available, still rely on ippsMagnitude_32f/64f followed by a new square function, unfortunately with no available IPP backend for the 64f version (but vectorized with hal, though)

chacha21 · 2023-07-04T15:21:46Z

ASAP, I will add validity Gtests, but I am unable to provide perf tests

trailing spaces

rely on HAL rather than two successive calls to IPP functions

…o magnitudeSqr

mshabunin · 2023-11-21T20:36:51Z

modules/core/src/mathfuncs.cpp

+    CV_INSTRUMENT_REGION();
+
+    int type = src1.type(), depth = src1.depth(), cn = src1.channels();
+    CV_Assert( src1.size() == src2.size() && type == src2.type() && (depth == CV_32F || depth == CV_64F));


Shouldn't we also check all arrays for continuity? hal::magnitudeSqr* functions work with 1D arrays and do not know about row step.

continuity should not be a problem thanks to the NAryMatIterator that split data into continuous "planes" (which happen to be rows in this case)

mshabunin

I tried to compare performance of the new function magnitudeSqr and magnitude+multiply/sqr and it seems that fused operation is faster (x86_64).

// magnitude(x, y, dst); multiply(dst, dst, dst);
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(640x480, 32FC1)   0.132 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(640x480, 32FC4)   0.642 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(1280x720, 32FC1)  0.436 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(1280x720, 32FC4)  2.459 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(1920x1080, 32FC1) 1.236 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(1920x1080, 32FC4) 5.845 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(3840x2160, 32FC1) 5.861 
MagnitudeAndSqr::OCL_MagnitudeSqrFixture::(3840x2160, 32FC4) 24.528

// magnitudeSqr(x, y, dst);
MagnitudeSqr::OCL_MagnitudeSqrFixture::(640x480, 32FC1)      0.076 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(640x480, 32FC4)      0.417 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(1280x720, 32FC1)     0.263 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(1280x720, 32FC4)     1.673 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(1920x1080, 32FC1)    0.858 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(1920x1080, 32FC4)    3.995 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(3840x2160, 32FC1)    3.995 
MagnitudeSqr::OCL_MagnitudeSqrFixture::(3840x2160, 32FC4)    16.533

I've updated intrinsics to the modern scalable format and enabled these blocks in scalable mode.

Overall PR looks good to me.

mshabunin · 2023-11-22T16:11:52Z

modules/core/include/opencv2/core/hal/hal.hpp

 CV_EXPORTS void invSqrt32f(const float* src, float* dst, int len);
 CV_EXPORTS void invSqrt64f(const double* src, double* dst, int len);

+CV_EXPORTS void sqr64f(const double* src, double* dst, int len);


This function is not used because ipp branch is commented, maybe we can remove it? Or add sqr32f and cv::sqr() for completeness?

mshabunin · 2023-11-22T16:13:38Z

modules/core/src/mathfuncs_core.simd.hpp

+        v_float32 x0 = vx_load(x + i), x1 = vx_load(x + i + VECSZ);
+        v_float32 y0 = vx_load(y + i), y1 = vx_load(y + i + VECSZ);
+        x0 = v_muladd(x0, x0, v_mul(y0, y0));
+        x1 = v_muladd(x1, x1, v_mul(y1, y1));


I tried to use v_sqr_magnitude intrinsic and it works well too. Performance on x86_64 is the same at least.

mshabunin · 2023-11-22T16:16:23Z

modules/core/src/mathfuncs_core.simd.hpp

+        v_store(mag + i, x0);
+        v_store(mag + i + VECSZ, x1);
+    }
+    vx_cleanup();


There is new approach to using vx_cleanup, see #23098 (comment)

chacha21 added 3 commits October 10, 2019 13:55

cv::magnitudeSqr()

eb3b7bc

TODO: ippicv must implement ippicvsPowerSpectr_32f/ippicvsPowerSpectr_64f

Merge remote-tracking branch 'upstream/master' into magnitudeSqr

f72ac09

updated doc and tests

b1d9f18

alalek reviewed Oct 10, 2019

View reviewed changes

chacha21 added 3 commits October 11, 2019 09:40

fixed test

66f3800

fixed test

7e397ad

IPP implementation

1a75c6b

asmorkalov added test and removed test labels Jan 15, 2020

asmorkalov assigned VadimLevin Mar 20, 2020

chacha21 added 2 commits July 4, 2023 16:48

Merge branch '4.x' into magnitudeSqr

a3cab0c

use fallback for IPP

a4c45f4

since ippsPowerSpectr_32f/64f isnot available, still rely on ippsMagnitude_32f/64f followed by a new square function, unfortunately with no available IPP backend for the 64f version (but vectorized with hal, though)

chacha21 added 5 commits July 4, 2023 18:45

fixed code stye

3d662e3

trailing spaces

disable magnitudeSqr through IPP

586a433

rely on HAL rather than two successive calls to IPP functions

added accuracy tests

fae0c68

Merge branch 'magnitudeSqr' of https://github.com/chacha21/opencv int…

230f91c

…o magnitudeSqr

Merge branch '4.x' into magnitudeSqr

bcf327b

mshabunin reviewed Nov 21, 2023

View reviewed changes

mshabunin added 2 commits November 22, 2023 16:34

Merge branch '4.x' into pr15683

dbcb75c

Fix scalable intrinsics for magnitudeSqr

5965e16

mshabunin reviewed Nov 22, 2023

View reviewed changes


		OCL_TEST_CYCLE() cv::magnitudeSqr(src1, src2, dst);

		SANITY_CHECK(dst, 1e-6);

Uh oh!

cv::magnitudeSqr() #15683

Are you sure you want to change the base?

cv::magnitudeSqr() #15683

Uh oh!

Conversation

chacha21 commented Oct 10, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alalek Oct 10, 2019

Choose a reason for hiding this comment

Uh oh!

chacha21 commented Oct 10, 2019

Uh oh!

alalek commented Oct 10, 2019

Uh oh!

asmorkalov commented Oct 21, 2019

Uh oh!

chacha21 commented Oct 21, 2019

Uh oh!

chacha21 commented Apr 16, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

asenyaev commented Apr 7, 2021

Uh oh!

asmorkalov commented Jul 4, 2023

Uh oh!

chacha21 commented Jul 4, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chacha21 commented Jul 4, 2023

Uh oh!

mshabunin Nov 21, 2023

Choose a reason for hiding this comment

Uh oh!

chacha21 Nov 22, 2023

Choose a reason for hiding this comment

Uh oh!

mshabunin left a comment

Choose a reason for hiding this comment

Uh oh!

mshabunin Nov 22, 2023

Choose a reason for hiding this comment

Uh oh!

mshabunin Nov 22, 2023

Choose a reason for hiding this comment

Uh oh!

mshabunin Nov 22, 2023

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

chacha21 commented Oct 10, 2019 •

edited

Loading

chacha21 commented Apr 16, 2020 •

edited

Loading

chacha21 commented Jul 4, 2023 •

edited

Loading