* Vectorize calculating integral for line for single and multiple channels
* Single vector processing for 4-channels - 25-30% faster
* Single vector processing for 4-channels - 25-30% faster
* Fixed AVX512 code for 4 channels
* Disable 3 channel 8UC1 to 32S for SSE2 and SSE3 (slower). Use new version of 8UC1 to 64F for AVX512.
fixed the ordering of contour convex hull points
* partially fixed the issue #4539
* fixed warnings and test failures
* fixed integer overflow (issue #14521)
* added comment to force buildbot to re-run
* extended the test for the issue 4539. Check the expected behaviour on the original contour as well
* added comment; fixed typo, renamed another variable for a little better clarity
* added yet another part to the test for issue #4539, where we run convexHull and convexityDetects on the original contour, without any manipulations. the rest of the test stays the same
* fixed several problems when running tests on Mac:
* OCL_pyrUp
* OCL_flip
* some basic UMat tests
* histogram badarg test (out of range access)
* retained the storepix fix in ocl_flip only for 16U/16S datatype, where the OpenCL compiler on Mac generates incorrect code
* moved deletion of ACCESS_FAST flag to non-SVM branch (where SVM is shared virtual memory (in OpenCL 2.x), not support vector machine)
* force OpenCL to use read/write for GPU<=>CPU memory transfers on machines with discrete video only on Macs. On Windows/Linux the drivers are seemingly smart enough to implement map/unmap properly (and maybe more efficiently than explicit read/write)
* Fix NN resize with dimentions > 4
* add test check for nn resize with channels > 4
* Change types from float to double
* Del unnecessary test file. Move nn test to test_imgwarp. Add 5 channels test only.
Actually, we can do this in constant time. xofs always
contains same or increasing offset values. We can instead
find the most extreme value used and never attempt to load it.
Similarly, we can note for all dx >= 0 and dx < (dwidth - cn)
where xofs[dx] + cn < xofs[dwidth-cn] implies dx < (dwidth - cn).
Thus, we can use this to control our loop termination optimally.
This fixes#16137 with little or no performance impact. I have
also added a debug check as a sanity check.
* Handle det == 0 in findCircle3pts.
Issue 16051 shows a case where findCircle3pts returns NaN for the
center coordinates and radius due to dividing by a determinant of 0. In
this case, the points are colinear, so the longest distance between any
2 points is the diameter of the minimum enclosing circle.
* imgproc(test): update test checks for minEnclosingCircle()
* imgproc: fix handling of special cases in minEnclosingCircle()
* imgproc: Prevent 1B overrun of 8C3 SIMD optimization
The fourth value read via v_load_q is essentially ignored,
but can cause trouble if it happens to cross page boundaries.
The final few iterations may attempt to read the most extreme
elements of S, which will read 1B beyond the array in most
aligment cases. Dynamically compute the stop. This could be
hoised from the loop, but will require a more extensive change.
Likewise, cleanup the iteration increment statements to make
it more obvious they do channel count (3) elements per pass.
This should resolve#16137
* imgproc(resize): extra check
* resize: HResizeLinear reduce duplicate work
There appears to be a 2x unroll of the HResizeLinear against k,
however the k value is only incremented by 1 during the unroll. This
results in k - 1 duplicate passes when k > 1.
Likewise, the final pass may not respect the work done by the vector
loop. Start it with the offset returned by the vector op if
implemented. Note, no vector ops are implemented today.
The performance is most noticable on a linear downscale. A set of
performance tests are added to characterize this. The performance
improvement is 10-50% depending on the scaling.
* imgproc: vectorize HResizeLinear
Performance is mostly gated by the gather operations
for x inputs.
Likewise, provide a 2x unroll against k, this reduces the
number of alpha gathers by 1/2 for larger k.
While not a 4x improvement, it still performs substantially
better under P9 for a 1.4x improvement. P8 baseline is
1.05-1.10x due to reduced VSX instruction set.
For float types, this results in a more modest
1.2x improvement.
* Update U8 processing for non-bitexact linear resize
* core: hal: vsx: improve v_load_expand_q
With a little help, we can do this quickly without gprs on
all VSX enabled targets.
* resize: Fix cn == 3 step per feedback
Per feedback, ensure we don't overrun. This was caught via the
failure observed in Test_TensorFlow.inception_accuracy.
Improving VSX performance of integral function
* Adding support for vector get function on VSX datatypes so the
integral function gains a bit of performance.
* Removing get as a datatype member function and implementing a new HAL
instruction v_extract_n to get the n-th element of a vector register.
* Adding SSE/NEON/AVX intrinsics.
* Implement new HAL instruction v_broadcast_element on VSX/AVX/NEON/SSE.
* core(simd): add tests for v_extract_n/v_broadcast_element
- updated docs
- commented out code to repair compilation
- added WASM and MSA default implementations
* core(simd): fix compilation
- x86: avoid _mm256_extract_epi64/32/16/8 with MSVS 2015
- x86: _mm_extract_epi64 is 64-bit only
* cleanup
* Convert moments in tile algorithms to HAL (1.3x faster for VSX).
* Adding NEON code back in for non 64-bit platforms.
* Remove floats from post processing.
- move TLS & instrumentation code out of core/utility.hpp
- (*) TLSData lost .gather() method (to dispose thread data on thread termination)
- use TLSDataAccumulator for reliable collecting of thread data
- prefer using of .detachData() + .cleanupDetachedData() instead of .gather() method
(*) API is broken: replace TLSData => TLSDataAccumulator if gather required
(objects disposal on threads termination is not available in accumulator mode)
* Adding support for vectorized masking for uchar/ushort.
* Fixing bug where mask was zeroing the dst. Improved the way to calculate
the mask and tweaked for further performance improvements.
* Fixing mask comparison test.
* Restricting to one channel.
* Adding support for 3 channels, switch old approach to start using HAL's
v_select.
* Adding all possible data type interactions to the perf tests since some
use SIMD acceleration and others do not.
* Disabling full tests by default.
* Giving proper names, removing magic numbers and sanity checks of new
performance tests for the integral function.
* Giving proper names, making array static.
* Convert ImgWarp from SSE SIMD to HAL - 2.8x faster on Power (VSX) and 15% speedup on x86
* Change compile flag from CV_SIMD128 to CV_SIMD128_64F for use of v_float64x2 type
* Changing WarpPerspectiveLine from class functions and dispatching to static functions.
* Re-add dynamic runtime and dispatch execution.
* RRestore SSE4_1 optimizations inside opt_SSE4_1 namespace
Crosscorr cleanup (#14936)
* Simplify code for convolution destination type/size
For the 2d filter code, destination size equals source size, and the
crossCorr function even (re-)creates the output matrix with the given size.
The number of channels also have to match. The destination type() is the
one used to create the output matrix, so we can use its type() here.
This is a preparatory patch.
Signed-off-by: Stefan Brüns <stefan.bruens@rwth-aachen.de>
* Remove redundant destination size and type parameters from crossCorr
All calling sites of crossCorr already use (...,
mat, mat.size(), mat.type(), ...), so the parameters are redundant.
Signed-off-by: Stefan Brüns <stefan.bruens@rwth-aachen.de>
Lab, Luv and XYZ conversions rewritten to wide intrinsics (#14106)
* rgb2xyz<float> re-vectorized
* rgb2xyz_i vectorized for ushort and uchar
* xyz2rgb<float> vectorized
* xyz2rgb_i vectorized for both uchar and ushort
* intermediate conversions (int->float) rewritten
* packed rgb2luv rewritten
* (some) float conversions rewritten
* burnt volatile int _3 and similar
* RGB2Lab_b rewritten
* tests: logging made better
* RGB2Lab_f (LRGB path) rewritten
* Lab2RGBfloat rewritten
* Lab2RGBinteger and Lab2RGB_b rewritten to wide universal intrinsics
* Luv2RGBinteger wide vectorized
* RGB2Lab_b fixed: v_sub_wrap instead of saturated sub
* warnings fixed
* trying to fix compilation on older compilers
* using 16x8 registers for 8-element dot product
* cleanup added
* splineInterpolate: loop unrolled, perf fix for f32x4
* Lab2RGBfloat: grab 2x more data to process on f32x4
* nrepeats for Luv2RGBfloat, +20% perf
* minor
* nrepeats to RGB2Lab_f
* Lab2RGBinteger: no tab for linear BGR
* nrepeats for RGB2Luvfloat
* Luv2RGBinteger: no tab for linear RGB
* +10% more to perf of Luv2RGBfloat
* nrepeats for 256-simd for Lab2RGBfloat
* less warnings
* BOM removed
* CV_SIMD_WIDTH used for lanes number checking
* trilinearPackedInterpolate: 128-bit specialization added
* fix build; no vx_cleanup(), instrumentation instead
Lab/XYZ modes have been postponed (color_lab.cpp):
- need to split code for tables initialization and for pixels processing first
- no significant performance improvements for switching between SSE42 / AVX2 code generation
Resize reworked using wide universal intrinsics (#13781)
* Added wide universal intrinsics optimized implementation for 3 channel bit-exact linear resize
* Reworked linear resize using new wide LUT intrinsics
* Fix for VSX intrinsics
Due to size limit of shared memory, histogram is built on
the global memory for CV_16UC1 case.
The amount of memory needed for building histogram is:
65536 * 4byte = 256KB
and shared memory limit is 48KB typically.
Added test cases for CV_16UC1 and various clip limits.
Added perf tests for CV_16UC1 on both CPU and CUDA code.
There was also a bug in CV_8UC1 case when redistributing
"residual" clipped pixels. Adding the test case where clip
limit is 5.0 exposes this bug.
PyrDown: Fix bug #12961 (#13672)
* Force unaligned pointer and create test
* More cross-platform solution
* MSVC expects a proper order
* Remove useless clang macro
