LGTM ( i tested only on Windows )

you can see build logs with nasm is available and nasm is unavailable

result output of test code is like below

SIMD Support: YES

Frames captured (VideoCapture) :   100    Average FPS:     170.8    Average time per frame:      5.85 ms
Frames captured (imread)       :   100    Average FPS:     109.0    Average time per frame:      9.17 ms

SIMD Support: NO

Frames captured (VideoCapture) :   100    Average FPS:     171.1    Average time per frame:      5.84 ms
Frames captured (imread)       :   100    Average FPS:      82.9    Average time per frame:     12.06 ms

FYI: perf tests can report status and performance statistics with to xml with flag --gtest_output=xml:report_file_name.xml. The xml reports could be compared with summary.py script: https://github.com/opencv/opencv/blob/4.x/modules/ts/misc/summary.py

I have changed some parts of the CMakeLists take into account your comments. But i am thinking that is not what you want exactly. Can you give me more info about that and if possible, could you look at last workflow errors? In the macOS ARM build, 3 tests are failed. I did not understand why this is the case because these pass in my computer. @asmorkalov

I guess the simd related codes do not appear in the relevant place in Solution Explorer. something needs to be done to solve it

Geometric mean (ms)

@ocpalo, we really need to finalize this PR by the end of GSoC. Please, be more active

So far, we have verified in the following OS and architecture.

• macOS x64(Intel CPU)
• macOS ARM64
• Ubuntu ARM64
• Ubuntu x64
• Windows x64

Performance comparision between SIMD enabled vs not. Ubuntu x64

Geometric mean (ms)

@asmorkalov Hi,

Vadim informed me in the last meeting about CPU type and optimization baseline. I have a few comments about this. Let me know what you think so I can do the changes according to your comments. In the 3rdparty/libjpeg-turbo/CMakeLists.txt you have commented before that i need to reuse OpenCV variables as x86_64,x86 etc. Then alalek wrote a comment that about we should do minimal changes if code is borrowed from upstream(which is true). Should I still use the openCV variables or do not use to keep align with the upstream.

My second question is, as you said before some CPU's does not support SIMD instructions. Since we provide this SIMD extension as optional(enabled by default), the user can disable this extension. Should I worry about this? I think it will make CMakeLists complicated.

Hello @ocpalo Thanks for the great job! I reviewed CMakeLists in libjpeg-turbo repos and tested previous version of the patch. Findings:

• Libjpeg-turbo cmake checks NEON/altivec/DSPr2 support with C compiler checks in CMake. So we are in save, if the code is build, for example, for amrv6. NEON part is not included and CMake reports warning about this. I tested it with old Raspberry Pi and it passes tests correctly. For the builds where NEON support is enabled on compiler side with -mfpu=neon NEON is enabled with libjpeg turbo too. I tested it with my Jetson TK1 board (armv7 with NEON). So everything works well with C compiler indirectly. I do not have Power and MIPS boards to check, but the CMake approach is similar and I do not see any reasons for suspicious.
• Libjpeg-turbo always has AVX2 assembly included into build for Intel systems, but the tests do not fail on old system without AVX2 support. Most probably it's dispatched in runtime.

So What I propose to do:

• Presume original CMake as much as possible. Do not wary about duplicate system checks and other specifics.
• Make WITH_SIMD cache variable with value (NOT CV_DISABLE_OPTIMIZATION) to have precise control for testing.
• Allow libjpeg-turbo for other Unixes like BSD, Solaris, SunOS, etc. It's already supported by original CMake and I do not see any reason to drop them, if all tools hve been found.

@asmorkalov Thanks for the feedback!

• I have created a another flag to enable SIMD which is WITH_LIBJPEG_TURBO_SIMD and it is value is NOT CV_DISABLE_OPTIMIZATION. WITH_SIMD flag is depends on that flag. I think this satisfy the second item in your propose list.
• I did not drop other Unixes support on purpose. If you are talking about this line, it is related with TurboJPEG api and in OpenCV it is not enabled anyway. Other than that, I updated the compiler flags to enable maximum optimization in SunOS|Pro and clang.

I am trying to solve how to catch if old systems does not support AVX2 assembly support at compile time.

CMake emits warning when libjpeg-turbo SIMD Extension is enabled but could not activated due to no support. Hence this seems like workflow fails. Should I change it WARNING to the STATUS?

Also, I was planning to add the following lines to the CMake to see if SIMD extension is supported on compile time. But could not be sure if this is right or not.

if(X86_64)
  if(NOT "${CPU_BASELINE_FINAL}" MATCHES "SSE2|AVX2" OR NOT "${CPU_DISPATCH}" MATCHES "SSE2|AVX2")
    set(WITH_SIMD 0)
  endif()
endif()
if(X86)
  if(NOT "${CPU_BASELINE_FINAL}" MATCHES "MMX|AVX2|SSE2" OR NOT "${CPU_DISPATCH}" MATCHES "MMX|AVX2|SSE2")
    set(WITH_SIMD 0)
  endif()
endif()

If someone confirm, i will push it.

Only "CPU_BASELINE_FINAL" should be used. Also take a look on other 3rdparty file hows to do this right.

CPU_DISPATCH is for dispatching in "runtime" - probably can't be used here due to missing checks (and missing compiler flags).

I believe NOT CV_DISABLE_OPTIMIZATION is enough as suggested by @asmorkalov .

If NOT CV_DISABLE_OPTIMIZATION is enough, then I think pr is ready.

ENABLE_LIBJPEG_TURBO_SIMD "Include SIMD extensions for libjpeg-turbo, if available for this platform" (NOT CV_DISABLE_OPTIMIZATION)

SIMD extension is controlled via ENABLE_LIBJPEG_TURBO_SIMD flag. And this flags value is NOT CV_DISABLE_OPTIMIZATION. If there are no magic behind this value, we are good.

cc @alalek @asmorkalov

Thanks, I appreciate for the support and the help!