[Libva] [PATCH v3 intel-driver 3/9] decoder: h264: simplify and optimize reference frame store updates.
Gwenole Beauchesne
gb.devel at gmail.com
Thu Jun 5 22:22:20 PDT 2014
2014-06-06 3:45 GMT+02:00 Zhao, Yakui <yakui.zhao at intel.com>:
> On Thu, 2014-06-05 at 10:20 -0600, Gwenole Beauchesne wrote:
>> 2014-06-05 17:29 GMT+02:00 Xiang, Haihao <haihao.xiang at intel.com>:
>> >
>> >
>> >> -----Original Message-----
>> >> From: Gwenole Beauchesne [mailto:gb.devel at gmail.com]
>> >> Sent: Thursday, June 05, 2014 9:29 PM
>> >> To: Xiang, Haihao
>> >> Cc: Zhao, Yakui; libva at lists.freedesktop.org
>> >> Subject: Re: [Libva] [PATCH v3 intel-driver 3/9] decoder: h264: simplify and
>> >> optimize reference frame store updates.
>> >>
>> >> Hi,
>> >>
>> >> 2014-06-05 14:44 GMT+02:00 Xiang, Haihao <haihao.xiang at intel.com>:
>> >> >
>> >> >
>> >> >> -----Original Message-----
>> >> >> From: Libva [mailto:libva-bounces at lists.freedesktop.org] On Behalf Of
>> >> >> Gwenole Beauchesne
>> >> >> Sent: Thursday, June 05, 2014 5:20 PM
>> >> >> To: Zhao, Yakui
>> >> >> Cc: libva at lists.freedesktop.org
>> >> >> Subject: Re: [Libva] [PATCH v3 intel-driver 3/9] decoder: h264:
>> >> >> simplify and optimize reference frame store updates.
>> >> >>
>> >> >> 2014-06-04 7:30 GMT+02:00 Zhao, Yakui <yakui.zhao at intel.com>:
>> >> >> > On Tue, 2014-06-03 at 22:59 -0600, Gwenole Beauchesne wrote:
>> >> >> >> 2014-06-04 4:11 GMT+02:00 Zhao, Yakui <yakui.zhao at intel.com>:
>> >> >> >> > On Tue, 2014-06-03 at 10:43 -0600, Gwenole Beauchesne wrote:
>> >> >> >> >> Simplify and optimize the update process of the reference frame
>> >> store.
>> >> >> >> >> Use less iterations to look up existing objects. Use a cache to
>> >> >> >> >> store the free'd slots.
>> >> >> >> >>
>> >> >> >> >> Prerequisite: the reference_objects[] array was previously
>> >> >> >> >> arranged in a way that the element at index i is exactly the
>> >> >> >> >> object_surface that corresponds to the VA surface identified by
>> >> >> >> >> the VAPictureH264.picture_id located at index i in the
>> >> >> >> >> ReferenceFrames[]
>> >> >> array.
>> >> >> >> >
>> >> >> >> > As a whole, this patch looks good to me. Based on the hardware
>> >> >> >> > requirement, it will be better that one reference picture later
>> >> >> >> > used is not assigned for the other purpose even when it is not
>> >> >> >> > used in current reference frame list. In such case this patch
>> >> >> >> > does some optimization about DPB, which can defer the usage of
>> >> >> >> > one picture later used that is not in current reference frame.especially
>> >> when:
>> >> >> >> > one picture is not used during decoding frame X while it is used
>> >> >> >> > during
>> >> >> decoding frame X + 1.
>> >> >> >>
>> >> >> >> The case you mention was supposed to be fixed with the other patch
>> >> >> >> series where the re-factorisation is a pre-requisite. :) A very
>> >> >> >> strict GenFrameStoreContext needs to be maintained with the set of
>> >> >> >> used_frames[] as it is today, and an additional set of
>> >> >> >> free_frames[] *candidates*. Though, this case is not covered by
>> >> >> >> the conformance suite, and I haven't found any real world case
>> >> >> >> exhibiting that, neither did I try to generate one yet. So we can defer this
>> >> one.
>> >> >> >>
>> >> >> >> > Although it still has no improvment for the following scenario,
>> >> >> >> > it still is an improvement.
>> >> >> >> > >One picture is not used during decoding frame X and X + 1
>> >> >> >> > while it is used during decoding frame X + 2.
>> >> >> >> >
>> >> >> >> > Of course the perfect solution is that the upper layer can
>> >> >> >> > assure that one picture later used is still added into the DPB
>> >> >> >> > of current frame although it is not used in current reference list.
>> >> >> >>
>> >> >> >> No, this is not going to happen. This would be an implementation
>> >> >> >> error of the standard.
>> >> >> >
>> >> >> > In theory it is no problem for normal H264. But it is true for MVC
>> >> >> > case as some reference frames for view 1 is not used when decoding view
>> >> 0.
>> >> >>
>> >> >> When I said this was not going to happen, this concerned the "perfect
>> >> >> solution" you proposed, which is not one. That would be a workaround
>> >> >> to driver bugs, and an error of implementation of the standard. Of
>> >> >> course, the DPB holds all the necessary view components, but the
>> >> >> VAPictureH264.ReferenceFrames[] is not exactly the DPB, and it cannot be.
>> >> >
>> >> > The number of elements in VAPictureH264.ReferenceFrames[] is 16, so it
>> >> > is enough to hold All reference surfaces for H.264. and the patch works as
>> >> before, I am OK for your patch.
>> >> >
>> >> >>
>> >> >> The case you mention is what I said a paragraph above, and the only
>> >> "perfect"
>> >> >> way to solve it is to correctly track the references in the driver
>> >> >> side. This was what the GenFrameStoreContext was for. However, as I
>> >> >> mentioned, there is no supporting case in the conformance test suite,
>> >> >> and I haven't come by a real-world sample exhibiting this.
>> >> >
>> >> > The case Yakui mentioned is not for H.264 as a surface should be in
>> >> > DPB if it is referenced by later Surface.
>> >> >
>> >> > Xiaowei told us each view uses a separated DPB buffer for MVC
>> >>
>> >> There is no such thing in MVC. There is a unique DPB with a size scalable by a
>> >> log2 number of views.
>> >>
>> >> > A case might occurs as below:
>> >> >
>> >> > View0: f0, f1, f2, f3, ...
>> >> > View1: s0, s1, s2, s3, ...
>> >> >
>> >> > f0 is referenced by f1, but f0 isn't referenced by other surfaces in view0, f1
>> >> is reference by f2.
>> >> > F0 is also referenced by s1 , both f0 and f1 are referenced by S2 in
>> >> > view1
>> >> >
>> >> > When decoding f1, f0 is added into View0's DPB, so it is added into
>> >> > frame store too, the frame store index for f0 is *0*.
>> >> >
>> >> > When decoding s1, f0 is added into view1's DPB, it has been in the frame
>> >> store, the frame store index for f0 is still *0*.
>> >> >
>> >> > When decoding f2, f0 is removed from view0's DPB and f1 is added into
>> >> > view0's DPB, so f0 is removed from frame store and f1 is added into
>> >> > frame store too, now the frame store index For f1 is 0,
>> >> >
>> >> > When decoding s2, f0 is still in View1's DPB so it will be added into
>> >> > the frame store again, and now the frame store index for f0 is *1*
>> >> >
>> >> > The frame store index for f0 is changed from 0 to 1 and the decoding result
>> >> will be wrong.
>> >>
>> >> That's exactly the thing we are discussing since the beginning. :)
>> >
>> > The above case doesn't exist if MVC uses a unique DBP buffer as f0 is still in the DBP after
>> > Decoding f2, hence the driver won't remove f0 from the frame store.
>>
>> Not necessarily. Anyway, your described case does not exist at all
>> because you can only have inter-view prediction with frames from the
>> same access unit. So, you cannot have f0+f1 referenced by s2, only f2
>> and s0+s1 could be referenced by s2. But anyway, that's the idea. A
>> valid example could be f1+f2 for s2, then f1+f2 for f3, then f3+s1+s2
>> for s3.
>>
>> Because of the additional constraint I gave that ReferenceFrames[]
>> cannot always be the DPB because MVC is not limited to 2 views to
>> begin with, the case I describe is what will give issues. This is
>> because you cannot simply always append inter-view reference
>> components in the ReferenceFrames[] array because this is not scalable
>> for MultiView High with a higher number of views. This could work for
>> Stereo High because the DPB is always capped to 16 frames at most.
>
>> Besides, you'd better see the ReferenceFrames[] array as RefPicList0[]
>> U RefPicList1[] rather than the DPB. The DPB, as per the H.264 spec
>> definition, could have more than 16 entries for Multiview High
>> profile. And I don't really want to workaround/do different things for
>> Stereo High profile vs. Multiview High profile. So, the
>> ReferenceFrames[] is to be filled in with the reference frames
>> effectively used (L0+L1) for decoding the current picture.
>
> If the ReferenceFrames[] only holds the RefPicList0 + RefPicList1
> instead of DPB, then it will trigger the issue Haihao and I mentioned
> several times in previous emails. Some reference picture is not used in
> RefPicList0/1 during decoding frame X while it will be used later in
> decoding frame X+1. So the upper middles will add it into the
> ReferenceFrames[] even although it is not used when decoding the current
> frame. If not, the driver still can trigger the corresponding issue
> although it can reduce the possibility of triggering issue after some
> workaround.
Good, so you start to understand what we were discussing and the
constraints around it. So, I am glad you finally see the neccessity to
some code refactoring for tracking the refs frames correctly. :)
So, see the patch11 of the new series, or go back to v2 with the
GenFrameStoreContext addition.
Regards,
Gwenole.
>> >> And there is no way to handle that from the unique GenFrameStore array as it
>> >> is today, even if you delay the marking as "free", or whatever.
>> >> You need a more elaborated data struct. And this is what was to be provided
>> >> with the GenFrameStoreContext and additional re-factoring so that I don't
>> >> have to copy code around again.
>> >>
>> >> I am working on an alternative that wouldn't need too many changes at this
>> >> time.
>> >>
>> >> Regards,
>> >> Gwenole.
>> >>
>> >> >
>> >> >>
>> >> >> >> >> Theory of operations:
>> >> >> >> >>
>> >> >> >> >> 1. Obsolete entries are removed first, i.e. entries in the internal DPB
>> >> >> >> >> that no longer have a match in the supplied ReferenceFrames[]
>> >> array.
>> >> >> >> >> That obsolete entry index is stored in a local cache: free_slots[].
>> >> >> >> >>
>> >> >> >> >> 2. This cache is completed with entries considered as "invalid" or "not
>> >> >> >> >> present", sequentially while traversing the frame store for
>> >> obsolete
>> >> >> >> >> entries. At the end of this removal process, the free_slots[] array
>> >> >> >> >> represents all possible indices in there that could be re-used for
>> >> >> >> >> new reference frames to track.
>> >> >> >> >>
>> >> >> >> >> 3. The list of ReferenceFrames[] objects is traversed for new entries
>> >> >> >> >> that are not already in the frame store. If an entry needs to be
>> >> >> >> >> added, it is placed at the index obtained from the next free_slots[]
>> >> >> >> >> element. There is no need to traverse the frame store array again,
>> >> >> >> >> the next available slot can be known from that free_slots[] cache.
>> >> >> >> >>
>> >> >> >> >> Signed-off-by: Gwenole Beauchesne
>> >> >> >> >> <gwenole.beauchesne at intel.com>
>> >> >> >> >> ---
>> >> >> >> >> src/i965_decoder_utils.c | 107
>> >> >> >> >> ++++++++++++++++++++--------------------------
>> >> >> >> >> 1 file changed, 46 insertions(+), 61 deletions(-)
>> >> >> >> >>
>> >> >> >> >> diff --git a/src/i965_decoder_utils.c
>> >> >> >> >> b/src/i965_decoder_utils.c index 2570be9..c686235 100644
>> >> >> >> >> --- a/src/i965_decoder_utils.c
>> >> >> >> >> +++ b/src/i965_decoder_utils.c
>> >> >> >> >> @@ -418,88 +418,73 @@ gen6_send_avc_ref_idx_state( }
>> >> >> >> >>
>> >> >> >> >> void
>> >> >> >> >> -intel_update_avc_frame_store_index(VADriverContextP ctx,
>> >> >> >> >> - struct decode_state
>> >> >> *decode_state,
>> >> >> >> >> -
>> >> >> VAPictureParameterBufferH264 *pic_param,
>> >> >> >> >> - GenFrameStore
>> >> >> frame_store[MAX_GEN_REFERENCE_FRAMES])
>> >> >> >> >> +intel_update_avc_frame_store_index(
>> >> >> >> >> + VADriverContextP ctx,
>> >> >> >> >> + struct decode_state *decode_state,
>> >> >> >> >> + VAPictureParameterBufferH264 *pic_param,
>> >> >> >> >> + GenFrameStore
>> >> >> frame_store[MAX_GEN_REFERENCE_FRAMES]
>> >> >> >> >> +)
>> >> >> >> >> {
>> >> >> >> >> - int i, j;
>> >> >> >> >> -
>> >> >> >> >> - assert(MAX_GEN_REFERENCE_FRAMES ==
>> >> >> ARRAY_ELEMS(pic_param->ReferenceFrames));
>> >> >> >> >> -
>> >> >> >> >> - for (i = 0; i < MAX_GEN_REFERENCE_FRAMES; i++) {
>> >> >> >> >> - int found = 0;
>> >> >> >> >> -
>> >> >> >> >> - if (frame_store[i].surface_id == VA_INVALID_ID ||
>> >> >> >> >> - frame_store[i].obj_surface == NULL)
>> >> >> >> >> + int free_slots[MAX_GEN_REFERENCE_FRAMES];
>> >> >> >> >> + int i, j, n, num_free_slots;
>> >> >> >> >> + bool found;
>> >> >> >> >> +
>> >> >> >> >
>> >> >> >> > It will be better to fill the free_slots through the following
>> >> >> >> > two loops instead of one loop, which is to try to defer the
>> >> >> >> > usage of the reference picture later used that is not used in
>> >> >> >> > current reference frame
>> >> >> list.
>> >> >> >> > a. check the invalid surface_id or NULL surface and then add it
>> >> >> >> > into free_slots.
>> >> >> >> > b. check whether one picture is used by the current reference
>> >> >> >> > frame list. If not, add it into the free_slots.
>> >> >> >>
>> >> >> >> The previous algorithm was filling in the holes. This one does the
>> >> >> >> same. I don't see where your proposal could bring an improvement?
>> >> >> >
>> >> >> > In your patch you use the free_slots to find one free slot to hold
>> >> >> > a new reference picture in DPB. Right? It will firstly try to find
>> >> >> > one free_slot with invalid picture_id or surface, which is used to
>> >> >> > hold a new reference picture in DPB.
>> >> >>
>> >> >> No, it will take whatever comes first between:
>> >> >> - a slot with an invalid picture_id or surface ;
>> >> >> - or a slot with a retired reference picture candidate, i.e.
>> >> >> something that cannot be found in the ReferenceFrames[] array.
>> >> >> And this works exactly the same way as before. No functional change yet.
>> >> >>
>> >> >> > It can remove the possibility that one picture later used is
>> >> >> > immediately used to hold a new reference picture in DPB. So based
>> >> >> > on the above point IMO it is better to find the free_slots through
>> >> >> > two steps to hold the new reference picture in DPB.
>> >> >>
>> >> >> This would only hide further this situation, but OK.
>> >> >>
>> >> >> Regards,
>> >> >> Gwenole.
>> >> >>
>> >> >> >> >> + /* Remove obsolete entries from the internal DPB */
>> >> >> >> >> + for (i = 0, n = 0; i < MAX_GEN_REFERENCE_FRAMES; i++) {
>> >> >> >> >> + GenFrameStore * const fs = &frame_store[i];
>> >> >> >> >> + if (fs->surface_id == VA_INVALID_ID || !fs->obj_surface) {
>> >> >> >> >> + free_slots[n++] = i;
>> >> >> >> >> continue;
>> >> >> >> >> + }
>> >> >> >> >>
>> >> >> >> >> - assert(frame_store[i].frame_store_id != -1);
>> >> >> >> >> -
>> >> >> >> >> - for (j = 0; j < MAX_GEN_REFERENCE_FRAMES; j++) {
>> >> >> >> >> - VAPictureH264 *ref_pic =
>> >> >> &pic_param->ReferenceFrames[j];
>> >> >> >> >> - if (ref_pic->flags & VA_PICTURE_H264_INVALID)
>> >> >> >> >> + // Find whether the current entry is still a valid
>> >> >> >> >> + reference
>> >> >> frame
>> >> >> >> >> + found = false;
>> >> >> >> >> + for (j = 0; j <
>> >> >> >> >> + ARRAY_ELEMS(decode_state->reference_objects);
>> >> >> j++) {
>> >> >> >> >> + struct object_surface * const obj_surface =
>> >> >> >> >> + decode_state->reference_objects[j];
>> >> >> >> >> + if (!obj_surface)
>> >> >> >> >> continue;
>> >> >> >> >> -
>> >> >> >> >> - if (frame_store[i].surface_id == ref_pic->picture_id) {
>> >> >> >> >> - found = 1;
>> >> >> >> >> + if ((found = obj_surface == fs->obj_surface))
>> >> >> >> >> break;
>> >> >> >> >> - }
>> >> >> >> >> }
>> >> >> >> >>
>> >> >> >> >> - /* remove it from the internal DPB */
>> >> >> >> >> + // ... or remove it
>> >> >> >> >> if (!found) {
>> >> >> >> >> - frame_store[i].surface_id = VA_INVALID_ID;
>> >> >> >> >> - frame_store[i].frame_store_id = -1;
>> >> >> >> >> - frame_store[i].obj_surface = NULL;
>> >> >> >> >> + fs->surface_id = VA_INVALID_ID;
>> >> >> >> >> + fs->obj_surface = NULL;
>> >> >> >> >> + fs->frame_store_id = -1;
>> >> >> >> >> + free_slots[n++] = i;
>> >> >> >> >> }
>> >> >> >> >> }
>> >> >> >> >> + num_free_slots = n;
>> >> >> >> >>
>> >> >> >> >> - for (i = 0; i < MAX_GEN_REFERENCE_FRAMES; i++) {
>> >> >> >> >> - VAPictureH264 *ref_pic = &pic_param->ReferenceFrames[i];
>> >> >> >> >> - int found = 0;
>> >> >> >> >> -
>> >> >> >> >> - if (ref_pic->flags & VA_PICTURE_H264_INVALID ||
>> >> >> >> >> - ref_pic->picture_id == VA_INVALID_SURFACE ||
>> >> >> >> >> - decode_state->reference_objects[i] == NULL)
>> >> >> >> >> + /* Append the new reference frames */
>> >> >> >> >> + for (i = 0, n = 0; i <
>> >> >> >> >> + ARRAY_ELEMS(decode_state->reference_objects);
>> >> >> i++) {
>> >> >> >> >> + struct object_surface * const obj_surface =
>> >> >> >> >> + decode_state->reference_objects[i];
>> >> >> >> >> + if (!obj_surface)
>> >> >> >> >> continue;
>> >> >> >> >>
>> >> >> >> >> + // Find whether the current frame is not already in
>> >> >> >> >> + our frame
>> >> >> store
>> >> >> >> >> + found = false;
>> >> >> >> >> for (j = 0; j < MAX_GEN_REFERENCE_FRAMES; j++) {
>> >> >> >> >> - if (frame_store[j].surface_id == ref_pic->picture_id) {
>> >> >> >> >> - found = 1;
>> >> >> >> >> + if ((found = frame_store[j].obj_surface ==
>> >> >> >> >> + obj_surface))
>> >> >> >> >> break;
>> >> >> >> >> - }
>> >> >> >> >> }
>> >> >> >> >>
>> >> >> >> >> - /* add the new reference frame into the internal DPB */
>> >> >> >> >> + // ... or add it
>> >> >> >> >> if (!found) {
>> >> >> >> >> - int frame_idx;
>> >> >> >> >> - int slot_found;
>> >> >> >> >> - struct object_surface *obj_surface =
>> >> >> decode_state->reference_objects[i];
>> >> >> >> >> -
>> >> >> >> >> - slot_found = 0;
>> >> >> >> >> - frame_idx = -1;
>> >> >> >> >> - /* Find a free frame store index */
>> >> >> >> >> - for (j = 0; j < MAX_GEN_REFERENCE_FRAMES; j++) {
>> >> >> >> >> - if (frame_store[j].surface_id == VA_INVALID_ID ||
>> >> >> >> >> - frame_store[j].obj_surface == NULL) {
>> >> >> >> >> - frame_idx = j;
>> >> >> >> >> - slot_found = 1;
>> >> >> >> >> - break;
>> >> >> >> >> - }
>> >> >> >> >> + if (n < num_free_slots) {
>> >> >> >> >> + GenFrameStore * const fs =
>> >> >> &frame_store[free_slots[n++]];
>> >> >> >> >> + fs->surface_id = obj_surface->base.id;
>> >> >> >> >> + fs->obj_surface = obj_surface;
>> >> >> >> >> + fs->frame_store_id = fs - frame_store;
>> >> >> >> >> + }
>> >> >> >> >> + else {
>> >> >> >> >> + WARN_ONCE("No free slot found for DPB
>> >> >> >> >> + reference list!!!\n");
>> >> >> >> >> }
>> >> >> >> >> -
>> >> >> >> >> -
>> >> >> >> >> - if (slot_found) {
>> >> >> >> >> - frame_store[j].surface_id = ref_pic->picture_id;
>> >> >> >> >> - frame_store[j].frame_store_id = frame_idx;
>> >> >> >> >> - frame_store[j].obj_surface = obj_surface;
>> >> >> >> >> - } else {
>> >> >> >> >> - WARN_ONCE("Not free slot for DPB reference
>> >> >> list!!!\n");
>> >> >> >> >> - }
>> >> >> >> >> }
>> >> >> >> >> }
>> >> >> >> >> -
>> >> >> >> >> }
>> >> >> >> >>
>> >> >> >> >> void
>> >> >> >> >
>> >> >> >> >
>> >> >> >
>> >> >> >
>> >> >> _______________________________________________
>> >> >> Libva mailing list
>> >> >> Libva at lists.freedesktop.org
>> >> >> http://lists.freedesktop.org/mailman/listinfo/libva
>
>
More information about the Libva
mailing list