[AVC VDEnc] Target usage changes with free media-driver

[AntonGrishin]

Hi,

original issue is Intel-Media-SDK/MediaSDK#1735

short description
W/ free/non-free driver we have different AVC low power encoding quality

./sample_encode h264   -w 176 -h 144 -f 30 -cqp -qpi 30 -qpp 30 -qpb 30 -qsv-ff \ 
-i AUD_MW_E.yuv -o AUD_MW_E.264

Analysis from original issue:

Issue reproduced with latest media-driver master.
By driver dumps:
With free driver sets TU = 7 by default, instead of TU = 4 w/ non-free.
I found that with -DENABLE_NONFREE_KERNELS=OFF, also enables flag _FULL_OPEN_SOURCE.

This code is cause of issue:

#ifdef _FULL_OPEN_SOURCE
    if (!GFX_IS_PRODUCT(m_encodeCtx->pMediaCtx->platform, IGFX_ICELAKE_LP))
    {
        if (m_encodeCtx->targetUsage >= 1 && m_encodeCtx->targetUsage <= 2)
        {
            m_encodeCtx->targetUsage = 4;
        }
        else if (m_encodeCtx->targetUsage >= 3 &&m_encodeCtx->targetUsage <= 5)
        {
            m_encodeCtx->targetUsage = 7;
        }
    }
#endif

I checked with m_encodeCtx->targetUsage = 4; in second else if block. Files w/ free/non-free drivers are bit match.

Link to code:

media-driver/media_driver/linux/common/codec/ddi/media_ddi_encode_avc.cpp

Line 566 in a331ceb

#ifdef _FULL_OPEN_SOURCE

As I know, VDEnc not depends on kernels (we used CQP mode)
So why media-driver changes this field?
Also, based on condition in if block - why TU doesn`t change only for ICL?

Regards,
Anton

[XinfengZhang]

it is because ICL LP full open source driver need kernel for HME when it is target usage 1. but now , there are no such limitation----HME media kernel already was open sourced. will change it soon.

[dvrogozh]

HME kernel was open sourced only for ICL. However, I still think that we should not remap target usages even for those platforms where HME is not open sourced (BDW/SKL): is that possible to just skip the kernel for TU1 if it is not available? I guess this should make quality(TU1)=quality(TU4)?

[AntonGrishin]

Any updates?

[daleksan]

When you plan to merge #1013 ?