Re: [PATCH] media: hevc: fix pictures lists type

[Benjamin Gaignard]

Le 27/08/2021 à 14:40, Ezequiel Garcia a écrit :
> On Fri, 27 Aug 2021 at 09:36, John Cox <jc@xxxxxxxxxxxxx> wrote:
> > > Le 27/08/2021 à 12:10, John Cox a écrit :
> > > > > Le 26/08/2021 à 18:09, Nicolas Dufresne a écrit :
> > > > > > Le lundi 23 août 2021 à 12:35 +0100, John Cox a écrit :
> > > > > > > Hi
> > > > > > >
> > > > > > > > Le 23/08/2021 à 11:50, John Cox a écrit :
> > > > > > > > > > The lists embedded Picture Order Count values which are s32 so their type
> > > > > > > > > > most be s32 and not u8.
> > > > > > > > > I'm not convinced that you can't calculate all of those lists from the
> > > > > > > > > info already contained in the DPB array so this is probably redundant
> > > > > > > > > info though I grant that having the list pre-calced might make your life
> > > > > > > > > easier, and the userland side will have calculated the lists to
> > > > > > > > > calculate other required things so it isn't much extra work for it.
> > > > > > > > Yes the userland have already compute these lists and the number of items
> > > > > > > > in each of them.
> > > > > > > > Build them in the kernel would means to also compute the values of NumPocStCurrBefore,
> > > > > > > > NumPocStCurrAfter, NumPocLtCurr, NumPocStCurrAfter, NumPocStCurrBefore and NumPocLtCurr
> > > > > > > > and that requires information (NumNegativePics, NumPositivePics...) not provided to the kernel.
> > > > > > > > Since it have to be done in userland anyway, I'm reluctant to modify the API to redo in the kernel.
> > > > > > > Well, fair enough, I'm not going to argue
> > > > > > >
> > > > > > > > > Even if you do need the lists wouldn't it be a better idea to have them
> > > > > > > > > as indices into the DPB (you can't have a frame in any of those lists
> > > > > > > > > that isn't in the DPB) which already contains POCs then it will still
> > > > > > > > > fit into u8 and be smaller?
> > > > > > > > Hantro HW works with indexes but I think it is more simple to send PoC rather than indexes.
> > > > > > > I'd disagree but as I don't use the info I'm not concerned. Though I
> > > > > > > think I should point out that when Hantro converts the POCs to indicies
> > > > > > > it compares the now s32 POC in these lists with the u16 POC in the DPB
> > > > > > > so you might need to fix that too; by std (8.3.1) no POC diff can be
> > > > > > > outside s16 so you can mask & compare or use u16 POCs in the lists or
> > > > > > > s32 in the DPB.
> > > > > > Fun fact, my interpretation with the API when I drafted GStreamer support was
> > > > > > that it was DPB indexes:
> > > > > >
> > > > > > https://gitlab.freedesktop.org/ndufresne/gst-plugins-bad/-/blob/hevc_wip/sys/v4l2codecs/gstv4l2codech265dec.c#L850
> > > > > >
> > > > > > It felt quite natural to be, since this is also how we pass references for l0/l1
> > > > > > (unused by hantro I guess).
> > > > > >
> > > > > > Looking at old rkvdec code as a refresher:
> > > > > >
> > > > > >      for (j = 0; j < run->num_slices; j++) {
> > > > > >                    sl_params = &run->slices_params[j];
> > > > > >                    dpb = sl_params->dpb;
> > > > > >
> > > > > >                    hw_ps = &priv_tbl->rps[j];
> > > > > >                    memset(hw_ps, 0, sizeof(*hw_ps));
> > > > > >
> > > > > >                    for (i = 0; i <= sl_params->num_ref_idx_l0_active_minus1; i++) {
> > > > > >                            WRITE_RPS(!!(dpb[sl_params->ref_idx_l0[i]].rps == V4L2_HEVC_DPB_ENTRY_RPS_LT_CURR),
> > > > > >                                      REF_PIC_LONG_TERM_L0(i));
> > > > > >                            WRITE_RPS(sl_params->ref_idx_l0[i], REF_PIC_IDX_L0(i));
> > > > > >                    }
> > > > > >
> > > > > >                    for (i = 0; i <= sl_params->num_ref_idx_l1_active_minus1; i++) {
> > > > > >                            WRITE_RPS(!!(dpb[sl_params->ref_idx_l1[i]].rps == V4L2_HEVC_DPB_ENTRY_RPS_LT_CURR),
> > > > > >                                      REF_PIC_LONG_TERM_L1(i));
> > > > > >                            WRITE_RPS(sl_params->ref_idx_l1[i], REF_PIC_IDX_L1(i));
> > > > > >                    }
> > > > > >
> > > > > >
> > > > > > This is code is clearly unsafe, but now I remember that dpb_entry has a flag
> > > > > > "rps". So we know from the DPB in which of the list the reference lives, if any.
> > > > > > In the case of RKVDEC the HW only cares to know if this is long term or not.
> > > > > >
> > > > > > So without looking at the spec, is that dpb represention enough to reconstruct
> > > > > > these array ? If we pass these array, shall we keep the rps flag ? I think a
> > > > > > little step back and cleanup will be needed. I doubt there is a single answer,
> > > > > > perhaps list what others do (VA, DXVA, NVDEC, Khronos, etc) and we can
> > > > > > collectively decide were we want V4L2 to sit ?
> > > > > I have done some tests with Hantro driver and look at the spec, the order of the PoC
> > > > > in the reference lists matters. You can deducted the order for DPB rps flags.
> > > > > I would suggest to remove rps flags to avoid information duplication.
> > > > I want the DPB rps member for long term reference marking.  I don't care
> > > > about before / after, but LTR can't be deduced from PoC and if you are
> > > > going to keep the member you might as well keep before / after.
> > > Ok so keep like it is.
> > > In this case my patch is enough, right ?
> The problem with the patch is that it breaks existing userspace.
> Currently, there's no upstreamed userspace so this is not a huge
> deal.
>
> However, it's definitely not a good practice. Even if these are
> staging controls, I think a proper action item is to start discussing
> what's missing on the HEVC interface as a whole, so it can be
> moved to stable.

I do agree I think it could the time to talk about moving the API to stable.
My plan is to get this patch merge before sending a RFC to move the API.

Benjamin

> Otherwise, it almost feels like bikeshading.
>
> Thanks,
> Ezequiel