[PATCH v3] drm/msm/dpu: improve DSC allocation

[Dmitry Baryshkov]

On Tue, 12 Dec 2023 at 02:03, Kuogee Hsieh <quic_khsieh at quicinc.com> wrote:
>
>
> On 12/11/2023 1:30 PM, Dmitry Baryshkov wrote:
> > On Mon, 11 Dec 2023 at 20:38, Kuogee Hsieh <quic_khsieh at quicinc.com> wrote:
> >> A DCE (Display Compression Engine) contains two DSC hard slice
> >> encoders. Each DCE start with even DSC encoder index followed by
> > "starts". But it will not be correct. The DCE doesn't start with the
> > DSC encoder. DCE consists of two DSC encoders, one has an odd index
> > and another one has an even index.
> >
> >> an odd DSC encoder index. Each encoder can work independently.
> >> But Only two DSC encoders from same DCE can be paired to work
> > only
> >
> >> together to support merge mode. In addition, the DSC with even
> > There are different merge modes. Here you are talking about the DSC merge mode.
> >
> >> index have to mapping to even pingpong index and DSC with odd
> > PINGPONG (end everywhere else).
> >
> > have to be mapped, should be used, etc.
> >
> >> index have to mapping to odd pingpong index at its data path.
> >> This patch improve DSC allocation mechanism with consideration
> > improves
> >
> >> of above factors.
> > of these factors.
> >
> >> Changes in V3:
> >> -- add dpu_rm_pingpong_dsc_check()
> >> -- for pair allocation use i += 2 at for loop
> >>
> >> Changes in V2:
> >>      -- split _dpu_rm_reserve_dsc() into _dpu_rm_reserve_dsc_single() and
> >>         _dpu_rm_reserve_dsc_pair()
> >>
> >> Fixes: f2803ee91a41 ("drm/msm/disp/dpu1: Add DSC support in RM")
> > This tag is incorrect. The patch should be split into two pieces. One
> > which fixes DSC allocation for DSC 1.1 encoders, where there were no
> > DCE blocks, another one which adds proper handling for DCE.
> > Unless the paired allocation requirement also applies to pre-DCE DSC
> > encoders. But in that case the commit message doesn't make any sense.
> >
> > I checked 4.x Qualcomm kernels. None of them contained any of these
> > restrictions for DSC blocks. Only the displaypack targeting 4.19
> > kernel got these changes. But it predates DCE pairs support.
>
> as I said earlier the rule of odd/even pp-index map to odd/even
> dsc-index is there since dsc v1.1.
>
> I think current code (including down stream code) works by luck to not
> encounter a configuration with two independence paths, one with single
> dsc and the other one use two dsc to support dsc merge mode.
>
> this patch is the fix to enforce this rule for both dsc v1.1 and v1.2
> and I will rework commit message yo have better description.

Good. Does this apply to paired allocation too? I think so, as the
techpack first got the paired allocation and only afterwards it has
got the DSC/PP idx check.

Regarding the patch itself. May I suggest an alternative approach,
which should work better, I think. At least it will not require
'deleting' the PP indices. First you preprocess the pp_to_enc_id array
and list all PP indices selected for this encoder. Then you work with
this array, matching PP and DSC blocks.




-- 
With best wishes
Dmitry