Merge EE1-1.1 code changes (JVET-AD0156)

Support low complexity NN loop filter of JVET-AD0156.

Before looking at the MR, can you please solve the following issues:

• rebase on HEAD of the branch
• solve build error in the pipeline (see #9601 )

We have resolved both of the issues

Please remove all references/support to tensorflow in the C++ part. Only SADL is used for inference.

We have removed all the references/support to tensorflow in C++ part

added 4 commits

• 07943282...763e1cf2 - 2 commits from branch jvet-ahg-nnvc:VTM-11.0_nnvc
• c7d95239 - Support low complexity NNLoopFilter of JVET-AD0156
• 10e88bf5 - Resolve build errors in pipeline

Compare with previous version

added 1 commit

• 55538743 - Remove tensorflow based inference support

Compare with previous version

please update date

please update all updated files

changed this line in version 4 of the diff

Updated files NNFilterSetLC.h and NNFilterSetLC.cpp in commit 2425f029

Can you put encoder related code in EncoderLib/EncNNFilterSetLC.cpp (double filterPictureRD)

Moved encoder specific source to EncNNFilterSetLC.h and EncNNFilterSetLC.cpp in commit 81613501

put sadl related include in cpp

Moved sadl related include to cpp in commit 81613501

add simd option on your inference cpp file in CMakeList to benefit from SIMD speed up (see other filterset file for example)

Added simd options for low complexity source files in commonlib and encoder CMakeList in commit 194ec964

added 3 commits

• 2425f029 - Updated copyright dates
• 81613501 - Separate encoder src and class for low complexity NN loop filter
• 194ec964 - cmake avx2/avx512 inclusion for LC loopfilter src files

Compare with previous version

New code seems fine, you need to solve the build issue though (missing include?)

Resolved this issue in commit 5417af7e. Apologize for this, I had forgotten to commit EncGop.h which had the changes to include new header file.

added 1 commit

• 5417af7e - Fixed EncNNFilterSetLC.h file include error in encoder

Compare with previous version

Thanks for the fix! There is still modified SADL files in the commits, are they necessary?

As mentioned in my detailed mail earlier we have two sets of inter models ,

• model2_v2.sadl and model3_v2.sadl inter models are best we got for sadl without rounding (5.05% luma gain over nnvc-4.0)
• the original EE-1.1 int16 model2.sadl and model3.sadl are untouched (5.2% luma gain over nnvc-4.0 with SADL rounding).

Would you like us to remove the original int16 model2 and model3 with rounding enabled ? And possibly rename model2_v2.sadl as model2.sadl and model3_v2.sadl as model3.sadl ?

yes I think this is the good option, it will clarify the code. We agreed during the meeting to first have the model with the default quantization policy so we can push models which are optimized for that.

okay, understood. Will do this change and commit again.

I have just pushed the commit with this change.

added 1 commit

• 0970b09d - Replaced original int16 inter sadl models with v2 and deleted v2 models

Compare with previous version

assigned to @Franck

merged

mentioned in commit db813484