Thank you for the contribution!

May I ask you to share performance numbers for AVX2 vs AVX512.

Something like these:

• setup env: OPENCV_TEST_DATA_PATH=<opencv_extra>/testdata
• setup env: OPENCV_DNN_TEST_DATA_PATH=<some_dir_with_dnn_subfolder>
  use this script to download test DNN models (into "dnn" subfolder, size ~2Gb): https://github.com/opencv/opencv_extra/blob/master/testdata/dnn/download_models.py
• compile baseline code (without patch) and run:

  ./bin/opencv_perf_dnn --gtest_output=xml:base.xml
  

• compile patched code and run:

  ./bin/opencv_perf_dnn --gtest_output=xml:optimized.xml
  

• generate report:

  python <opencv_src>/modules/ts/misc/summary.py base.xml optimized.xml
  

• post results here (use "-o markdown" to generate report compatible with GitHub comments)

(somewhat cleaned up and simplified patch updated, which shows this data)

Geometric mean

Thank you for the update!

It seems there is old design problem with AVX512 - see this issue: #8974
Problem here is that there are many "independent" AVX512 instruction sets (you use/check for AVX512DQ in this patch).

I'm pretty sure I accidentally fixed that other issue in the first patch of this PR;
AVX512 works with this PR.

yes AVX512 is a family not a single point; generally I check DQ since that is what is actually shipping as a superset of the base.. (e.g. Core i9 etc go beyond the 512F base)

(or in other words, 512DQ is a reasonable line to put in the sand for "what is AVX 512")

Current CMake scripts are designed for support "atomic" instruction sets (like, AVX512DQ) which has 1:1 mapping to compiler flags / processor features.

Groups, like AVX512 are not properly supported for now. Perhaps AVX512 can be replaced with AVX512-KNL (Knights Landing), AVX512-SKX (Skylake with BW, DQ, VL), AVX512-CNL (CannonLake with additional IFMA, VBMI) groups. But I'm not sure that it is a good idea.
BTW, Intel compiler defined this flag "COMMON-AVX512", "MIC-AVX512", "CORE-AVX512" (DQ is here)

So probably we should start from atomic features. Groups can be added later.

(I'm quite aware of the Intel roadmap/instructions since that's my dayjob ;-) )

512DQ is a reasonable baseline in terms of what is shipping/being used by people, where the CNL changes can be add-ons similar to how FMA3 is for AVX2.

Great! I believe you can propose better solution.

I just mean that current patch doesn't work as expected:

$ cmake -DCPU_BASELINE=AVX512 <opencv_src_dir>
$ make
...
.../opencv/modules/core/src/system.cpp: In member function ‘void cv::HWFeatures::initialize()’:
.../build/opencv/cv_cpu_config.h:51:7: error: ‘CV_CPU_AVX512’ was not declared in this scope
     , CV_CPU_AVX512 \
       ^
.../opencv/modules/core/src/system.cpp:531:37: note: in expansion of macro ‘CV_CPU_BASELINE_FEATURES’
         int baseline_features[] = { CV_CPU_BASELINE_FEATURES };
                                     ^
.../build/opencv/cv_cpu_config.h:51:7: note: suggested alternative: ‘CV_CPU_AVX2’
     , CV_CPU_AVX512 \
       ^
...opencv/modules/core/src/system.cpp:531:37: note: in expansion of macro ‘CV_CPU_BASELINE_FEATURES’
         int baseline_features[] = { CV_CPU_BASELINE_FEATURES };
                                     ^

This is reproducer for all platforms (including non-AVX512).

Build problem reproducer for AVX512 systems is quite straightforward:

$ CXXFLAGS="-march=native" cmake <opencv_src_dir>
$ make
... error message as above ...

(can be emulated via SDE tool: sde -skx -env 'CXXFLAGS' ' -march=native' -- cmake ../../dev)

We need to fix these builds before merging.

BTW, OpenCV knows these AVX512 CPU capabilities (there is no "AVX512", but there is "AVX_512DQ").

In this case this line should be adopted too (to avoid compiler generation of non-supported instructions, see #6990).
But in this case you can't use 512DQ instruction in code because of missing compiler flags.

My suggestion is to rename current "AVX512" => "AVX_512DQ" and fix compiler flags. I believe it is enough to support the current patch.

Could you take a look on these changes: alalek@pr10416_r ?

• looks like AVX-512DQ intrinsics are not used yet (switched to AVX-512F). If AVX-512DQ is necessary let me know about this.
• I have no access to AVX512 capable machine at this moment so I tested this via SDE tool only (for "-knl" and "-sdx" targets)

test your patch on top and it works; updated this PR.

Thank you for checking!
I will take a look on the current build failures.

Looking at the details .. it's not quite there...
the performance really comes with avx512vl, not just avx512f

This usually means that compiler optimizes other code (without direct intrinsic calls) by yourself, and "-mavx512f" option is not enough.

Does build in this way works well?

cmake -DCPU_BASELINE=NATIVE <opencv_src_dir>
or
CXXFLAGS="-march=native" cmake <opencv_src_dir>

Thank you for explanation!

I will take a look on this.

BTW, What CMake options do you use? (What compiler?)

@fenrus75 Please take a look on #10463 (AVX512_SKX).