vo_gpu_next: use pl deinterlace shader for bwdif instead of sw filter by llyyr · Pull Request #17394 · mpv-player/mpv

llyyr · 2026-02-10T06:53:40Z

No description provided.

dancingmirrors · 2026-02-10T18:29:51Z

Yes, for some reason bwdif_vulkan stopped working for me after around FFmpeg 7.1, just outputs a green screen instead. No idea what happened but libplacebo's deinterlacing works.

kasper93 · 2026-02-26T02:30:30Z

This should be implemented such that it's possible to output full frame rate (60i -> 60p), instead of every other frame.

Headcrabed · 2026-04-15T19:06:41Z

Any updates?

llyyr · 2026-04-15T21:27:52Z

No, and I'm quite busy irl so I won't have time for this for a few months still. If anyone wants to pick this up they're welcome

llyyr · 2026-06-08T17:20:34Z

Rebased on top of #18099

This is ready for testing/review now

kasper93 · 2026-06-08T17:31:17Z

This is ready for testing/review now

This should be implemented such that it's possible to output full frame rate (60i -> 60p), instead of every other frame.

llyyr · 2026-06-08T17:41:47Z

This should be implemented such that it's possible to output full frame rate (60i -> 60p), instead of every other frame.

unless I'm misunderstanding it, documentation of pl_source_frame.first_field already says libplacebo is doing the doubling internally. Can you tell me what I need to do here or how I can check if it's working right?

kasper93 · 2026-06-08T17:55:19Z

This should be implemented such that it's possible to output full frame rate (60i -> 60p), instead of every other frame.

unless I'm misunderstanding it, documentation of pl_source_frame.first_field already says libplacebo is doing the doubling internally. Can you tell me what I need to do here or how I can check if it's working right?

It does, but you have to drive rendering twice as fast. Two times per each frame to output full frame rate, else you are dropping every other field (output frame).

Get this sample and test it. You should be able to see all consecutive numbers. You can pause and frame step, both forward and backward frame step should work correctly. https://github.com/bbgdzxng1/interlace-detection-test-patterns/raw/9d163e87b222e2711ab5fae805d80b2e655e04c1/bt601-525_480_interlaced_bff.ts

EDIT: You can find more samples at https://github.com/bbgdzxng1/interlace-detection-test-patterns/

llyyr · 2026-06-08T18:16:54Z

From those samples, even current vaapi/vulkan deint filters skip half the fields on gpu-next. On gpu, 60i becomes 60p correctly. I'll need to check why

llyyr · 2026-06-15T12:36:42Z

This outputs full frame rate now. Though I'm not sure if there's a simpler way to do this than adding a dummy video filter

kasper93 · 2026-06-27T20:07:37Z

    pars->params.frame_mixer = opts->interpolation ? map_scaler(p, SCALER_TSCALE) : NULL;

+    pars->params.deinterlace_params = fopts->deinterlace != 0 ? &pars->deinterlace_params : NULL;
+    pars->deinterlace_params.algo = PL_DEINTERLACE_BWDIF;


This should be an option, not hardcoded.

https://mpv.io/manual/master/#options-deinterlace

Deinterlace option says bwdif is inserted typically, and making this an option would be out of scope for this PR as it would need to change for hardware deinterlacers too

The limitation of autofilter, should not be transposed to gpu-next.

Added option to select it

kasper93 · 2026-06-27T23:30:42Z

frame stepping backwards doesn't work. Also quality of image after deinterlacing is not good, you can see it especially in =bob mode, which should just show alternating fields. It looks like chroma is not aligned with luma correctly. Look at the bottom of the numbers, heavily de-saturated. This needs to be fixed before we can enable it by default.

Hardware bob:

libplacebo bob / bilinear:

llyyr · 2026-06-28T19:12:54Z

Also quality of image after deinterlacing is not good, you can see it especially in =bob mode, which should just show alternating fields. It looks like chroma is not aligned with luma correctly. Look at the bottom of the numbers, heavily de-saturated. This needs to be fixed before we can enable it by default.

Can be reproduced in plplay too, would need to fix libplacebo implementation first. From what I can tell, only yadif similar to how the software implementation does. bob and bwdif have artifacts that aren't present with the software filter.

frame stepping backwards doesn't work.

Will look into this after the libplacebo side is fixed

kasper93 · 2026-06-28T19:22:03Z

Can be reproduced in plplay too, would need to fix libplacebo implementation first. From what I can tell, only yadif similar to how the software implementation does. bob and bwdif have artifacts that aren't present with the software filter.

To avoid having to manually compare this, I think we should add a test in libplacebo, that compares output to software reference, we can make this be enable only if building with ffmpeg. Naturally due to all pipeline being in float, it won't be bitexact, but we can compute PSNR. The result should be very close, give the same configured algorithm and sampling method.

kasper93 · 2026-07-26T17:47:27Z

            .pts = mpi->pts,
-            .duration = can_interpolate ? frame->approx_duration : 0,
+            .duration = can_interpolate ? frame->approx_duration :
+                        use_fields ? mpi->pkt_duration : 0,


In some cases we might not have pkt_duration, not sure if that's critical, but we could use frame->approx_duration which is our estimation.

Isn't this the opposite? We should always have pkt_duration, but approx_duration is only computed in display synced modes

kasper93 · 2026-07-26T19:27:06Z

When deinterlacing is enabled you need to request one more frame (at least) in vo_set_queue_params, because we need to have future frame in pl_queue

llyyr · 2026-07-27T02:52:40Z

Fixed the backstep problem and changed it to request extra frame when deinterlacing. The alignment issues persist however and still need fixing

kasper93 · 2026-07-27T07:43:47Z

Fixed the backstep problem and changed it to request extra frame when deinterlacing. The alignment issues persist however and still need fixing

Do you still see issues after applying those changes https://code.videolan.org/videolan/libplacebo/-/merge_requests/874 ?

llyyr · 2026-07-27T07:55:05Z

Do you still see issues after applying those changes https://code.videolan.org/videolan/libplacebo/-/merge_requests/874 ?

renderer: compensate for interlaced chroma siting when deinterlacing fixes it but the first frame still has alignment issue. you can reproduce with bt601-525_480_interlaced_bff.ts --deinterlace=auto --mute --hwdec=no --pause

llyyr · 2026-07-27T10:33:05Z

renderer: compensate for interlaced chroma siting when deinterlacing fixes it but the first frame still has alignment issue. you can reproduce with bt601-525_480_interlaced_bff.ts --deinterlace=auto --mute --hwdec=no --pause

I've fixed this by pushing the previous frame to queue. This PR along with the above libplacebo MR works pretty well now

kasper93 · 2026-07-27T11:15:51Z

    // Optional decoded enhancement-layer frame
    struct mp_image *enhancement_layer;
+    // Previous source frame retained for VO-side deinterlacing
+    struct mp_image *deint_prev;


I don't think this is good place to hold such information. vavpp and d3d11 may by itself request more past frames. And while libplacebo currently only needs 1 past frame, it is not extensible.

If anything we shouldn't hold any dummy frames, but instead demux a frame before again. Same as hr-seek works, where we already jump way before the seek target and iterate from keyframe.

Either way, mp_image is not a good vessel to carry over past frames.

Getting the past frame isn't hard, refqueue already has it. The difficult part is plumbing it to the VO.

Does adding num_past_frames to vo_set_queue_params and past_frames to vo_frame make sense? Alternatively I was thinking of making vf_fieldrate always produce 1 frame before the target with MP_IMGFIELD_TEMPORAL_HISTORY or such, but this would need to be special cased to not affect all the vsync/display related things

Does adding num_past_frames to vo_set_queue_params and past_frames to vo_frame make sense?

Yes, I think this make sense. It mirrors future frames nicely.

I've implemented this

Used in next commits

This is used to emit frame per field, so the VO is asked to draw both fields.

This is used to signal when the VO can do deinterlacing itself, instead of using a filter

Allow setting the requested number of past frames via vo_set_queue_params, and make them available in vo_frame.

kasper93 reviewed Feb 10, 2026

View reviewed changes

Comment thread filters/f_auto_filters.c Outdated

llyyr force-pushed the pl-deint branch 4 times, most recently from 0939147 to 6da6403 Compare February 10, 2026 20:15

llyyr force-pushed the pl-deint branch from 6da6403 to 2b7e535 Compare June 8, 2026 17:20

llyyr marked this pull request as ready for review June 8, 2026 17:20

llyyr force-pushed the pl-deint branch from 2b7e535 to 99f680a Compare June 15, 2026 12:35

llyyr force-pushed the pl-deint branch 2 times, most recently from f0a5310 to 09c39d9 Compare June 15, 2026 12:38

kasper93 requested changes Jun 27, 2026

View reviewed changes

llyyr force-pushed the pl-deint branch 2 times, most recently from 25ad50c to fbca8de Compare June 27, 2026 22:03

kasper93 reviewed Jul 26, 2026

View reviewed changes

llyyr force-pushed the pl-deint branch from fbca8de to 1b4d205 Compare July 27, 2026 02:52

llyyr force-pushed the pl-deint branch 2 times, most recently from ed5a365 to 95a9785 Compare July 27, 2026 10:32

kasper93 reviewed Jul 27, 2026

View reviewed changes

llyyr added 7 commits July 29, 2026 18:35

vo_gpu_next: fix stray comma

822c6a2

mp_image: add MP_IMGFIELD_TICK_{FIRST,SECOND} defines

2d8a9b5

Used in next commits

vf_fieldrate: add this filter

569d27f

This is used to emit frame per field, so the VO is asked to draw both fields.

vo: add VO_CAP_DEINTERLACE

e478347

This is used to signal when the VO can do deinterlacing itself, instead of using a filter

video: allow VOs to request past frames

97d4e50

Allow setting the requested number of past frames via vo_set_queue_params, and make them available in vo_frame.

vo_gpu_next: use shader deinterlacer instead of software bwdif

0da5fc6

filters/f_auto_filters: recreate deint filter for auto/yes cycle

d09129c

llyyr force-pushed the pl-deint branch from 95a9785 to d09129c Compare July 29, 2026 20:24

Uh oh!

Conversation

llyyr commented Feb 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

dancingmirrors commented Feb 10, 2026

Uh oh!

kasper93 commented Feb 26, 2026

Uh oh!

Headcrabed commented Apr 15, 2026

Uh oh!

llyyr commented Apr 15, 2026

Uh oh!

llyyr commented Jun 8, 2026

Uh oh!

kasper93 commented Jun 8, 2026

Uh oh!

llyyr commented Jun 8, 2026

Uh oh!

kasper93 commented Jun 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llyyr commented Jun 8, 2026

Uh oh!

llyyr commented Jun 15, 2026

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kasper93 commented Jun 27, 2026

Uh oh!

llyyr commented Jun 28, 2026

Uh oh!

kasper93 commented Jun 28, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kasper93 commented Jul 26, 2026

Uh oh!

llyyr commented Jul 27, 2026

Uh oh!

kasper93 commented Jul 27, 2026

Uh oh!

llyyr commented Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llyyr commented Jul 27, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

llyyr commented Feb 10, 2026 •

edited

Loading

kasper93 commented Jun 8, 2026 •

edited

Loading

llyyr commented Jul 27, 2026 •

edited

Loading