[GSOC] Speeding-up AKAZE, part #2 (#8951)
* feature2d: instrument more functions used in AKAZE
* rework Compute_Determinant_Hessian_Response
* this takes 84% of time of Feature_Detection
* run everything in parallel
* compute Scharr kernels just once
* compute sigma more efficiently
* allocate all matrices in evolution without zeroing
* features2d: add one bigger image to tests
* now test have images: 600x768, 900x600 and 1385x700 to cover different resolutions
* explicitly zero Lx and Ly
* add Lflow and Lstep to evolution as in original AKAZE code
* reworked computing keypoints orientation
integrated faster function from https://github.com/h2suzuki/fast_akaze
* use standard fastAtan2 instead of getAngle
* compute keypoints orientation in parallel
* fix visual studio warnings
* replace some wrapped functions with direct calls to OpenCV functions
* improved readability for people familiar with opencv
* do not same image twice in base level
* rework diffusity stencil
* use one pass stencil for diffusity from https://github.com/h2suzuki/fast_akaze
* improve locality in Create_Scale_Space
* always compute determinat od hessian and spacial derivatives
* this needs to be computed always as we need derivatives while computing descriptors
* fixed tests of AKAZE with KAZE descriptors which have been affected by this
Currently it computes all first and second order derivatives together and the determiant of the hessian. For descriptors it would be enough to compute just first order derivates, but it is not probably worth it optimize for scenario where descriptors and keypoints are computed separately, since it is already very inefficient. When computing keypoint and descriptors together it is faster to do it the current way (preserves locality).
* parallelize non linear diffusion computation
* do multiplication right in the nlp diffusity kernel
* rework kfactor computation
* get rid of sharing buffers when creating scale space pyramid, the performace impact is neglegible
* features2d: initialize TBB scheduler in perf tests
* ensures more stable output
* more reasonable profiles, since the first call of parallel_for_ is not getting big performace hit
* compute_kfactor: interleave finding of maximum and computing distance
* no need to go twice through the data
* start to use UMats in AKAZE to leverage OpenCl in the future
* fixed bug that prevented computing determinant for scale pyramid of size 1 (just the base image)
* all descriptors now support writing to uninitialized memory
* use InputArray and OutputArray for input image and descriptors, allows to make use UMAt that user passes to us
* enable use of all existing ocl paths in AKAZE
* all parts that uses ocl-enabled functions should use ocl by now
* imgproc: fix dispatching of IPP version when OCL is disabled
* when OCL is disabled IPP version should be always prefered (even when the dst is UMat)
* get rid of copy in DeterminantHessian response
* this slows CPU version considerably
* do no run in parallel when running with OCL
* store derivations as UMat in pyramid
* enables OCL path computing of determint hessian
* will allow to compute descriptors on GPU in the future
* port diffusivity to OCL
* diffusivity itself is not a blocker, but this saves us downloading and uploading derivations
* implement kernel for nonlinear scalar diffusion step
* download the pyramid from GPU just once
we don't want to downlaod matrices ad hoc from gpu when the function in AKAZE needs it. There is a HUGE mapping overhead and without shared memory support a LOT of unnecessary transfers.
This maps/downloads matrices just once.
* fix bug with uninitialized values in non linear diffusion
* this was causing spurious segfaults in stitching tests due to propagation of NaNs
* added new test, which checks for NaNs (added new debug asserts for NaNs)
* valgrind now says everything is ok
* add nonlinear diffusion step OCL implementation
* Lt in pyramid changed to UMat, it will be downlaoded from GPU along with Lx, Ly
* fix bug in pm_g2 kernel. OpenCV mangles dimensions passed to OpenCL, so we need to check for boundaries in each OCL kernel.
* port computing of determinant to OCL
* computing of determinant is not a blocker, but with this change we don't need to download all spatial derivatives to CPU, we only download determinant
* make Ldet in the pyramid UMat, download it from CPU together with the other parts of the pyramid
* add profiling macros
* fix visual studio warning
* instrument non_linear_diffusion
* remove changes I have made to TEvolution
* TEvolution is used only in KAZE now
* Revert "features2d: initialize TBB scheduler in perf tests"
This reverts commit ba81e2a711.
RGB2Lab_f added, bugs fixed, moved to float
several bugs fixed
LUT fixed, no switch in tetraInterpolate()
temporary code; to be removed and rewritten
before refactoring
extra interpolations removed, some things to do left
added Lab2RGB_b +XYZ version, etc.
basic version is done, to be sped up
tetra refactored
interpolations: LUT for weights, refactor., etc.
address arithm optimized
initial version of vectorized code added (not compiling now)
compilation fixed, now segfaults
a lot of fixes, vectorization temp. disabled
fixed trilinear shift size, max error dropped from 19 to 10
fixed several bugs (255 vs 256, signed vs unsigned, bIdx)
minor changes
packed: address arithmetics fixed
shorter code
experiments with pure integer calculations
Lab2RGB max error decreased to 2; need to clean the code
ready for vectorization; need cleaning
vectorized, to be debugged
precision fixed, max error is 2
Lab->XYZ shortened
minor fixes
Lab2RGB_f version fixed, to be completely rewritten using _b code
RGB2Lab_f vectorized
minors
moved to separate file
refactored Lab2RGB to float and int versions
minor fix
Lab2RGB_f vectorized
minor refactoring
Lab2RGBint refactored: process methods, vectorize by 4 pix
Lab2RGB_f int version is done
cleanup extra code
code copied to color.cpp
fixed blue idx bug
optimizations enabled when testing; mulFracConst introduced
divConst -> mulFracConst
calc min time in perf instead of avg
minors
process() slightly sped up
Lab2RGB_f: disabled int version
reinterpret added, minor fixes in names
some warnings fixed
changes transferred to color.cpp
RGB2Lab_f code (and trilinear interpolation code) moved to rgb2lab_faster
whitespace
shift negative fixed
more warnings fixed
"constant condition" warnings fixed, little speed up
minor changes
test_photo decolor fixed
changes copied to test_lab.cpp
idx bounds checking in LUT init
several fixes
WIP: softfloat almost integrated
test_lab partially rewritten to SoftFloat
color.cpp rewritten to SoftFloat
test_lab.cpp: accuracy code added
several fixes
RGB2Lab_b testing fixed
splineBuild() rewritten to SoftFloat
accuracy control improved
rounding fixed
Luv <=> RGB: rewritten to SoftFloat
OCL cvtColor Lab and Lut rewritten to SoftFloat
minor fixes
refactored to new SoftFloat interface
round() -> cvRound, etc.
fixed OCL tests
softfloat.cpp: internal functions made static, unused ones removed
meaningful constants
extra lines removed
unused function removed
unfinished work
it works, need to fix TODOs
refactoring; more calls rewritten
mulFracConst removed
constants made bit exact; minors
changes moved to color.cpp
fixed 1 bug and 4 warnings
OCL: fixed constants
pow(x, _1_3f) replaced by cubeRoot(x)
fixed compilation on MSVC32
magic constants explained
file with internal accuracy&speed tests moved to lab_tetra branch
merge_histogram kernel only need "BINS" theads to accumulate the
histgrams, it is not efficient to directly use maxGroupSize as
local size if maxGroupSize is far greater then BINS.
Remove unnecessary Non-ASCII characters from source code (#9075)
* Remove unnecessary Non-ASCII characters from source code
Remove unnecessary Non-ASCII characters and replace them with ASCII
characters
* Remove dashes in the @param statement
Remove dashes and place single space in the @param statement to keep
coding style
* misc: more fixes for non-ASCII symbols
* misc: fix non-ASCII symbol in CMake file
* another round of dnn optimization:
* increased malloc alignment across OpenCV from 16 to 64 bytes to make it AVX2 and even AVX-512 friendly
* improved SIMD optimization of pooling layer, optimized average pooling
* cleaned up convolution layer implementation
* made activation layer "attacheable" to all other layers, including fully connected and addition layer.
* fixed bug in the fusion algorithm: "LayerData::consumers" should not be cleared, because it desctibes the topology.
* greatly optimized permutation layer, which improved SSD performance
* parallelized element-wise binary/ternary/... ops (sum, prod, max)
* also, added missing copyrights to many of the layer implementation files
* temporarily disabled (again) the check for intermediate blobs consistency; fixed warnings from various builders
Parallelize Canny with custom gradient (#8694)
* New Canny implementation. Restructuring code in parallelCanny class. Align mag buffer and map.
* Fix warnings.
* Missing SIMD check added.
* Replaced local trailingZeros in contours.cpp. Use alignSize in canny.cpp
* Fix warnings in alignSize and allocate just minimum extra columns.
* Fix another warning in map.create.
* Exchange for loop by do loop to avoid double check at the beginning.
Define extra SIMD CANNY_CHECK to avoid unnecessary continue.
Updated integrations for:
cv::split
cv::merge
cv::insertChannel
cv::extractChannel
cv::Mat::convertTo - now with scaled conversions support
cv::LUT - disabled due to performance issues
Mat::copyTo
Mat::setTo
cv::flip
cv::copyMakeBorder - currently disabled
cv::polarToCart
cv::pow - ipp pow function was removed due to performance issues
cv::hal::magnitude32f/64f - disabled for <= SSE42, poor performance
cv::countNonZero
cv::minMaxIdx
cv::norm
cv::canny - new integration. Disabled for threaded;
cv::cornerHarris
cv::boxFilter
cv::bilateralFilter
cv::integral
Added assertios to remap and warpAffine functions
As @mshabunin said, remap and warpAffine functions do not support more than 4 channels in
Bicubic and Lanczos4 interpolation modes. Assertions were added. Appropriate test was chenged.
resolves#8272
Warping a matrix with more than 4 channels using BORDER_CONSTANT and
INTER_NEAREST, INTER_CUBIC or INTER_LANCZOS4 interpolation led to
undefined behaviour. This commit changes the behavior of these methods
to be similar to that of INTER_LINEAR. Changed the scope of some of the
variables to more local. Modified some tests to be able to detect the
error described.
added 64b optimization for 3 channels case
not added 64b optimization for 4 channels case since timings did not
show any improvement
split ICV_HLINE cases into inline functions instead of macro for code
size reduction, without significand speed drawback at first sight
medianBlur called with "empty" source and ksize >= 7 crashes application with accessviolation. With this extra assert this is avoided and the application may normally catch the thrown exception.
- don't use undefined flag=0. It should be CONSTANT instead.
- don't allow 'UMat* m=NULL' argument (except LOCAL/CONSTANT flags).
This case is not handled well to provide NULL __global pointers.
It is better to use '-D' macro defines instead (at least for performance)
* OpenVX HAL updated to use generic OpenVX wrappers
* vxErr class from OpenVX HAL replaced with ivx::WrapperError
* reduced usage of vxImage class from OpenVX HAL replaced with ivx::Image
* vxImage class rewritten as ivx::Image subclass that calls swapHandle prior release
* Fix OpenVX HAL build
* Fix for review comments
OpenVX pyrDown wrappers (#7793)
* wrappers for vx_pyramid added
* initial version of pyrDown() wrapper added
* disabled for Khronos
* rewritten for new macro use; border mode added to node
Add new 5x5 gaussian blur kernel for CV_8UC1 format,
it is 50% ~ 70% faster than current ocl kernel in the perf test.
Signed-off-by: Li Peng <peng.li@intel.com>
Add new OpenCL kernels for bicubic interploation, it is 20% faster
than current warp image kernel with bicubic interploation.
Signed-off-by: Li Peng <peng.li@intel.com>
Add new ocl kernels for warpAffine and warpPerspective,
The average performance improvemnt is about 30%. The new
ocl kernels require CV_8UC1 format and support nearest
neighbor and bilinear interpolation.
Signed-off-by: Li Peng <peng.li@intel.com>
This ocl kernel is 46%~171% faster than current laplacian 3x3
ocl kernel in the perf test, with image format "CV_8UC1".
Signed-off-by: Li Peng <peng.li@intel.com>
Change contour test images to be very wide (#7464)
* Change contour test images to be very wide (#7409, #7458)
Unfortunately, slows down the tests.
* Decrease the number of contour test cases, in order to (at least partially) offset the test run duration increase caused by making the test images wider
* Don't test with very wide images on 32-bit architectures
Maximum depth limit var was added to the instrumentation structure;
Trace names output console output fix: improper tree formatting could happen;
Output in case of error was added;
Custom regions improvements;
Improved timing and weight calculation for parallel regions; New TC (threads counter) value to indicate how many different threads accessed particular node;
parallel_for, warnings fixes and ReturnAddress code from Alexander Alekhin;
This ocl kernel is for 3x3 kernel size and CV_8UC1 format
It is 115% ~ 300% faster than current ocl path in perf test
python ./modules/ts/misc/run.py -t imgproc --gtest_filter=OCL_GaussianBlurFixture*
Signed-off-by: Li Peng <peng.li@intel.com>
This kernel is for CV_8UC1 format and 3x3 kernel size,
It is about 33% ~ 55% faster than current ocl kernel with below perf test
python ./modules/ts/misc/run.py -t imgproc --gtest_filter=OCL_ErodeFixture*
python ./modules/ts/misc/run.py -t imgproc --gtest_filter=OCL_DilateFixture*
Also add accuracy test cases for this kernel, the test command is
./bin/opencv_test_imgproc --gtest_filter=OCL_Filter/MorphFilter3x3*
Signed-off-by: Li Peng <peng.li@intel.com>
* use hasSIMD128 rather than calling checkHardwareSupport
* add SIMD check in spartialgradient.cpp
* add SIMD check in stereosgbm.cpp
* add SIMD check in canny.cpp
The optimization is for CV_8UC1 format and 3x3 box filter,
it is 15%~87% faster than current ocl kernel with below perf test
./modules/ts/misc/run.py -t imgproc --gtest_filter=OCL_BlurFixture*
Also add test cases for this ocl kernel.
Signed-off-by: Li Peng <peng.li@intel.com>
Fix findContours crash for very large images (#7451)
* Cast step to size_t in order to avoid integer overflow when processing very large images
* Change assert to CV_Assert
* seriously improved performance of blur function, especially 3x3 and 5x5 cases
* trying to fix warnings and test failures
* replaced #if 0 with #if IPP_DISABLE_BLOCK
* Improve Canny by using _mm_movemask_epi8 to find next pixel magnitude greater than lower threshold. Added parallelized finalPass to Canny with variable gradients. Little changes in finalPass.
* Some things fixed
* use universal intrinsic for accumulate series using float/double
* accumulate, accumulateSquare, accumulateProduct and accumulateWeighted
* add v_cvt_f64_high in both SSE/NEON
* add test for conversion v_cvt_f64_high in test_intrin.cpp
* improve some existing universal intrinsic by using new instructions in Aarch64
* add workaround for Android build in intrin_neon.hpp
* Add Grana's connected components algorithm for 8-way connectivity. That algorithm is faster than Wu's one (currently implemented in opencv). For more details see https://github.com/prittt/YACCLAB.
* New functions signature and distance transform compatibility
* Add tests to imgproc/test/test_connectedcomponents.cpp
* Change of test_connectedcomponents.cpp for c++98 support
There is an issue with processing of abs(short) function for
negative argument.
Affected OpenCL devices:
- iGPU: Intel(R) HD Graphics 520 (OpenCL 2.0 )
- CPU: Intel(R) Core(TM) i5-6300U CPU @ 2.40GHz (OpenCL 2.0 (Build 10094))
* Common Canny parallelization added. TBB and single thread code removed. Final pass vectorized with SSE2 intrinsics.
* wrong #ifdef replaced with #if
* Merged to actual Canny version
* Merged common parallelized Canny with actual Canny implementation
* Remove 'Mutex *mutex' and pass 'Mutex mutex' from outside to parallelCanny
* Replaced extern Mutex with intern mutable Mutex.
When using OCL, the results of goodFeaturesToTrack() vary slightly from
run to run. This appears to be because the order of the results from
the findCorners kernel depends on thread execution and the sorting
function that is used at the end to rank the features only enforces are
partial sort order.
This does not materially impact the quality of the results, but it
makes it hard to build regression tests and generally introduces noise
into the system that should be avoided.
An easy fix is to change the sort function to enforce a total sort on
the features, even in cases where the match quality is exactly the same
for two features.
