Added flag to GaussianBlur for faster but not bit-exact implementation #25792

asmorkalov · 2024-06-20T09:25:35Z

Rationale:
Current implementation of GaussianBlur is almost always bit-exact. It helps to get predictable results according platforms, but prohibits most of approximations and optimization tricks.

The patch converts borderType parameter to more generic flags and introduces GAUSS_ALLOW_APPROXIMATIONS flag to allow not bit-exact implementation. With the flag IPP and generic HAL implementation are called first. The flag naming and location is a subject for discussion.

Replaces #22073
Possibly related issue: #24135

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 another license that is incompatible with OpenCV
The PR is proposed to the proper branch
There is a reference to the original bug report and related work
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

modules/imgproc/include/opencv2/imgproc.hpp

asmorkalov · 2024-06-21T07:07:05Z

Discussion result:

move enum to core
dedicated parameter for perf hint, do not mix with border type.
check python and java bindings

asmorkalov · 2024-06-26T08:45:13Z

@mshabunin @opencv-alalek @vpisarev I reworked interface as discussed offline and added accuracy tests for IPP branch. Could you take a look?

modules/imgproc/src/smooth.dispatch.cpp

modules/core/include/opencv2/core.hpp

modules/imgproc/include/opencv2/imgproc.hpp

…mentation.

CMakeLists.txt

modules/imgproc/include/opencv2/imgproc.hpp

opencv-alalek · 2024-07-11T07:27:26Z

modules/imgproc/test/test_smooth_bitexact.cpp

+    cv::absdiff(dst, gt, diff);
+    cv::Mat flatten_diff = diff.reshape(1, diff.rows);
+
+    int nz = countNonZero(flatten_diff);


norm is faster than countNonZero approach.

Use relative NORM_L1/L2 and NORM_INF instead.

I intentionally split the check on min-max deviation and amount of different pixels.

EXPECT_LE(max_val, 2); // expectes results floating +-1

comment doesn't follow to the check anyway.

NORM_INF <=1 works perfect.

With 1-limited NORM_INF, we could use NORM_L1 + RELATIVE to define the part of pixels of different values.

modules/core/include/opencv2/core.hpp

asmorkalov · 2024-07-11T08:38:31Z

@opencv-alalek I fixed your review notes. Please take a look again.

opencv-alalek · 2024-07-14T23:31:23Z

modules/core/include/opencv2/core.hpp

+    ALGO_APPROX = 2, //!< Allow alternative approximations to get faster implementation. Behaviour and result depends on a platform
+};
+
+/*! @brief Returns ImplementationHint selected by default, a.k.a. `IMPL_DEFAULT` defined during OpenCV compilation.


ImplementationHint

not renamed

IMPL_DEFAULT

What is that?

opencv-alalek · 2024-07-14T23:31:37Z

modules/core/include/opencv2/core.hpp

+*/
+enum AlgorithmHint {
+    ALGO_DEFAULT = 0, //!< Default algorithm behaviour defined during OpenCV build
+    ALGO_ACCURATE = 1, //!< Use generic portable implementation


ALGO_HINT_ then.

opencv-alalek · 2024-07-14T23:32:43Z

modules/core/src/system.cpp

@@ -46,6 +46,7 @@
 #include <iostream>
 #include <ostream>

+#include <opencv2/core.hpp>


To be removed.

opencv-alalek · 2024-07-14T23:35:25Z

modules/core/include/opencv2/core.hpp

+
+/*! @brief Returns ImplementationHint selected by default, a.k.a. `IMPL_DEFAULT` defined during OpenCV compilation.
+ */
+CV_EXPORTS_W AlgorithmHint getDefaultAlgorithmHint();


Should go to utility.hpp somewhere near setUseOptimized()

setUseOptimized() should also control behavior of that:

setUseOptimized(false) disables these hints and use accurate versions.

opencv-alalek · 2024-07-14T23:40:06Z

modules/imgproc/test/test_smooth_bitexact.cpp

+    cv::absdiff(dst, gt, diff);
+    cv::Mat flatten_diff = diff.reshape(1, diff.rows);
+
+    int nz = countNonZero(flatten_diff);


EXPECT_LE(max_val, 2); // expectes results floating +-1

comment doesn't follow to the check anyway.

NORM_INF <=1 works perfect.

With 1-limited NORM_INF, we could use NORM_L1 + RELATIVE to define the part of pixels of different values.

Added xxxApprox overloads for YUV color conversions in HAL and AlgorithmHint to cvtColor #25932 The xxxApprox to implement HAL functions with less bits for arithmetic of FP. The hint was introduced in #25792 and #25911 ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

Added flag to GaussianBlur for faster but not bit-exact implementation opencv#25792 Rationale: Current implementation of GaussianBlur is almost always bit-exact. It helps to get predictable results according platforms, but prohibits most of approximations and optimization tricks. The patch converts `borderType` parameter to more generic `flags` and introduces `GAUSS_ALLOW_APPROXIMATIONS` flag to allow not bit-exact implementation. With the flag IPP and generic HAL implementation are called first. The flag naming and location is a subject for discussion. Replaces opencv#22073 Possibly related issue: opencv#24135 ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [ ] There is a reference to the original bug report and related work - [x] 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

Added xxxApprox overloads for YUV color conversions in HAL and AlgorithmHint to cvtColor opencv#25932 The xxxApprox to implement HAL functions with less bits for arithmetic of FP. The hint was introduced in opencv#25792 and opencv#25911 ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

Added flag to GaussianBlur for faster but not bit-exact implementation opencv#25792 Rationale: Current implementation of GaussianBlur is almost always bit-exact. It helps to get predictable results according platforms, but prohibits most of approximations and optimization tricks. The patch converts `borderType` parameter to more generic `flags` and introduces `GAUSS_ALLOW_APPROXIMATIONS` flag to allow not bit-exact implementation. With the flag IPP and generic HAL implementation are called first. The flag naming and location is a subject for discussion. Replaces opencv#22073 Possibly related issue: opencv#24135 See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [ ] There is a reference to the original bug report and related work - [x] 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

Added flag to GaussianBlur for faster but not bit-exact implementation opencv#25792 Rationale: Current implementation of GaussianBlur is almost always bit-exact. It helps to get predictable results according platforms, but prohibits most of approximations and optimization tricks. The patch converts `borderType` parameter to more generic `flags` and introduces `GAUSS_ALLOW_APPROXIMATIONS` flag to allow not bit-exact implementation. With the flag IPP and generic HAL implementation are called first. The flag naming and location is a subject for discussion. Replaces opencv#22073 Possibly related issue: opencv#24135 ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [ ] There is a reference to the original bug report and related work - [x] 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

Added xxxApprox overloads for YUV color conversions in HAL and AlgorithmHint to cvtColor opencv#25932 The xxxApprox to implement HAL functions with less bits for arithmetic of FP. The hint was introduced in opencv#25792 and opencv#25911 ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

asmorkalov added optimization category: imgproc RFC labels Jun 20, 2024

asmorkalov added this to the 4.11.0 milestone Jun 20, 2024

asmorkalov requested review from vpisarev, mshabunin and opencv-alalek June 20, 2024 09:25

asmorkalov force-pushed the as/HAL_fast_GaussianBlur branch 2 times, most recently from 1c62df7 to b5df286 Compare June 20, 2024 10:16

asmorkalov added the pr: Discussion Required label Jun 20, 2024

asmorkalov changed the title ~~Added flag to GaussianBlur for faster but not bit-exact implementation.~~ Added flag to GaussianBlur for faster but not bit-exact implementation Jun 20, 2024

asmorkalov force-pushed the as/HAL_fast_GaussianBlur branch 2 times, most recently from 671ce14 to ea56bd9 Compare June 20, 2024 12:37

asmorkalov added the pr: needs test New functionality requires minimal tests set label Jun 20, 2024

vpisarev reviewed Jun 21, 2024

View reviewed changes

modules/imgproc/include/opencv2/imgproc.hpp Outdated Show resolved Hide resolved

asmorkalov added category: core test and removed pr: Discussion Required pr: needs test New functionality requires minimal tests set labels Jun 25, 2024

mshabunin reviewed Jun 26, 2024

View reviewed changes

modules/imgproc/src/smooth.dispatch.cpp Outdated Show resolved Hide resolved

vpisarev reviewed Jun 28, 2024

View reviewed changes

modules/core/include/opencv2/core.hpp Show resolved Hide resolved

vpisarev reviewed Jun 28, 2024

View reviewed changes

modules/imgproc/include/opencv2/imgproc.hpp Outdated Show resolved Hide resolved

opencv-alalek reviewed Jun 28, 2024

View reviewed changes

modules/imgproc/include/opencv2/imgproc.hpp Outdated Show resolved Hide resolved

asmorkalov added 3 commits July 10, 2024 13:39

Added flag to GaussianBlur for faster but not bit-exact implementation.

f702da6

Implemented alternative interface for implementation hints.

12a5914

Accuracy test for GaussianBlur with IMPL_ALLOW_APPROXIMATION.

771034c

Added CMake option to define Implementation hint with command line.

8fa2e47

asmorkalov force-pushed the as/HAL_fast_GaussianBlur branch from a753443 to 8fa2e47 Compare July 10, 2024 11:49

asmorkalov added 3 commits July 10, 2024 15:20

Added Implementation hint to test system reports and CMake flags docu…

a7ce249

…mentation.

Code review fixes.

bf8a99d

Code review fixes.

a677218

vpisarev self-requested a review July 10, 2024 19:22

vpisarev approved these changes Jul 10, 2024

View reviewed changes

opencv-alalek reviewed Jul 11, 2024

View reviewed changes

asmorkalov added feature and removed RFC labels Jul 11, 2024

asmorkalov force-pushed the as/HAL_fast_GaussianBlur branch from a47f442 to ee840b5 Compare July 11, 2024 09:48

Code review fixes.

13b6caa

asmorkalov force-pushed the as/HAL_fast_GaussianBlur branch from ee840b5 to 13b6caa Compare July 11, 2024 11:04

asmorkalov merged commit 15783d6 into opencv:4.x Jul 12, 2024
27 of 30 checks passed

Kumataro mentioned this pull request Jul 13, 2024

core: hal: avoid to use _tzcnt_u32 for ARM64EC #25903

Merged

6 tasks

opencv-alalek reviewed Jul 14, 2024

View reviewed changes

savuor mentioned this pull request Nov 1, 2024

Recent HAL changes ported to 4.9 #26395

Closed

6 tasks

Uh oh!

Added flag to GaussianBlur for faster but not bit-exact implementation #25792

Added flag to GaussianBlur for faster but not bit-exact implementation #25792

Uh oh!

Conversation

asmorkalov commented Jun 20, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Readiness Checklist

Uh oh!

Uh oh!

asmorkalov commented Jun 21, 2024

Uh oh!

asmorkalov commented Jun 26, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

asmorkalov commented Jul 11, 2024

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

asmorkalov commented Jun 20, 2024 •

edited

Loading