vf_d3d11vpp: add support for scaling

[kasper93]

No description provided.

[kasper93]

I've refactored options a bit to avoid hardcoded scaling size. @billxc: please take a look if that works for you.

[github-actions]

Download the artifacts for this pull request:

Windows

• mpv-i686-w64-mingw32
• mpv-x86_64-w64-mingw32
• mpv-x86_64-windows-msvc

macOS

• mpv-macos-12-intel
• mpv-macos-13-intel
• mpv-macos-14-arm

[billxc]

looks good for me， thanks for the work！@kasper93

[IC0hO]

For my 4K screen, when using the super resolution feature, to achieve an output resolution of 4K, the scale value should ideally be set to 2.0 for a 1080p video and 3.0 for a 720p video. Could you add a feature to automatically set the scale values?

[Andarwinux]

For my 4K screen, when using the super resolution feature, to achieve an output resolution of 4K, the scale value should ideally be set to 2.0 for a 1080p video and 3.0 for a 720p video. Could you add a feature to automatically set the scale values?

This can be done with lua script, although it would be nice to have native support.

function autovsr()
    display_width = mp.get_property_native("display-width")
    video_width = mp.get_property_native("width")
    display_height = mp.get_property_native("display-height")
    video_height = mp.get_property_native("height")
    mp.set_property_native("vf","")
    pixfmt = mp.get_property_native("video-params/hw-pixelformat") or mp.get_property_native("video-params/pixelformat")
    if video_width ~= nil and display_width ~= nil then 
        scale = math.max(display_width,display_height) / math.max(video_width,video_height)
        scale = scale - scale % 0.1
        if scale > 1 then
            if pixfmt == "nv12" then
                mp.set_property_native("vf","d3d11vpp:scaling-mode=nvidia:scale="..scale)
            elseif pixfmt == "yuv420p" then
                mp.set_property_native("vf","format=nv12,format=d3d11,d3d11vpp:scaling-mode=nvidia:scale="..scale)
            end
        end
    end
end
mp.observe_property("video-params/pixelformat", "native", autovsr)

Update: Improved the script, now automatically adapts to different hwdec or swdec.

[wadixx]

I couldn't make it work. Isn't below commands in mpv.config enough?

vf-add=d3d11vpp
vf=d3d11vpp=scale=1.0
vf=d3d11vpp=scaling-mode=nvidia

[kasper93]

I couldn't make it work. Isn't below commands in mpv.config enough?

vf-add=d3d11vpp vf=d3d11vpp=scale=1.0 vf=d3d11vpp=scaling-mode=nvidia

vf=d3d11vpp=scale=2:scaling-mode=nvidia it has to be one filter.

[kasper93]

This can be done with lua script, although it would be nice to have native support.

Same as --vf=scale don't use window dimensions. vf is before output, we don't even have window yet. And it it correct way to feedback the required size after the window is created in lua script. Otherwise patches welcome, but I will not hack this around in mpv code.

[hooke007]

I meet with an issue with d3d11vpp filter.
If a cpu filter appended. How should I correctly make it work.

vf-append = d3d11vpp=scale=2.0
vf-append = vflip # How to make it work?

[wadixx]

I couldn't make it work. Isn't below commands in mpv.config enough?
vf-add=d3d11vpp vf=d3d11vpp=scale=1.0 vf=d3d11vpp=scaling-mode=nvidia

vf=d3d11vpp=scale=2:scaling-mode=nvidia it has to be one filter.

I added lines below still no luck. Already tried with different resolutions. Mpv registers d3d11vpp as filter but it doesn't trigger super resolution status in Nvidia Control panel. In case of conflict i tried it with fresh install with no other configuration other than below.

vf-add=d3d11vpp
vf=d3d11vpp=scale=2:scaling-mode=nvidia

[kasper93]

I meet with an issue with d3d11vpp filter.
If a cpu filter appended. How should I correctly make it work.

vf-append = d3d11vpp=scale=2.0
vf-append = vflip # How to make it work?

Frame download doesn't work currently, because they are not allocated by ffmpeg and we don't have proper context to use for that. Will see what can be done about it.

[CrendKing]

When using this, I always get two warning messages

[w][d3d11vpp] No fitting video processor found, picking #0.
[w][autoconvert] Unexpected AVFrame/imgfmt hardware context mismatch.

I could eliminate the first warning by specifying the mode like mode=blend (only mocomp does not work). I guess it doesn't matter as long as I only have one video card?

The second warning seems to come from

mpv/filters/f_autoconvert.c

Line 190 in e3a9ce2

mp_warn(log, "Unexpected AVFrame/imgfmt hardware context mismatch.\n");

. Does that mean the frame image becomes software after the filter?

[Johnnycyan]

I can't for the life of me get this to actually work. I have this in my mpv.conf:

vf-add=d3d11vpp
vf=d3d11vpp=scale=2:scaling-mode=nvidia

Also tried without the first line. I don't get any logging even saying it's trying to enable it. Though it shows that it's set in the info section, it doesn't actually do anything and Nvidia reports it is inactive.

[Andarwinux]

RTX Video Super Resolution will not be active when a game is using NVIDIA Image Scaling (NIS), Dynamic Super Resolution (DSR) or Deep Learning Dynamic Super Resolution (DLDSR)

[Johnnycyan]

Well none of those are the case for me, and other applications have no issues enabling it

[kasper93]

vf-add=d3d11vpp

Why?

[Johnnycyan]

vf-add=d3d11vpp

Why?

Because I saw that elsewhere. But as I said I tried without that line and it still did nothing

[kasper93]

Have you read documentation? Do you have hardware decoding enabled?

gpu-api=d3d11
hwdec=d3d11va

[Johnnycyan]

Have you read documentation? Do you have hardware decoding enabled?

gpu-api=d3d11
hwdec=d3d11va

Ah I was missing hwdec=d3d11va thank you. Another question, does this not support RTX VSR HDR? Or is there something else I'm missing for that? Cause it just does the scaling right now.

[kasper93]

does this not support RTX VSR HDR?

No. I don't have hardware to test this, so you have to wait for someone else to implement or at least poc how to configure this.

[Johnnycyan]

Okay thank you!

[hooke007]

Do you have hardware decoding enabled

Well, I found actually it could work with swdec if the source is yuv420p10.

[Andarwinux]

Just add format=nv12, any pixfmt will work with d3d11vpp

[hooke007]

Just add format=nv12, any pixfmt will work with d3d11vpp

That's not the issue here. I think autoconvert is broken in some place. yuv420p8 should also work without users manual setting.

[kasper93]

Frame download doesn't work currently, because they are not allocated by ffmpeg and we don't have proper context to use for that. Will see what can be done about it.

Fixed by #14746

Just add format=nv12, any pixfmt will work with d3d11vpp

That's not the issue here. I think autoconvert is broken in some place. yuv420p8 should also work without users manual setting.

Not really broken, autoconvert will not inject software on cpu format conversion before hwupload. So it currently fails with

hwupload: Unable to find a compatible upload format for yuv420p

and while it could automatically do the conversion, I think it is better to let users explicitly state it, to avoid possibly slow on CPU conversion.

Currently if you want to have this filter work with software input/output you can do

vf-append=format=nv12
vf-append=d3d11vpp=scale=2
vf-append=format=nv12
vf-append=vflip

Well, I found actually it could work with swdec if the source is yuv420p10.

Yes, because there is direct upload format compatibility in this case. Also I need to update this filter to output also rgb10a2 when needed.

[kasper93]

That's not the issue here. I think autoconvert is broken in some place. yuv420p8 should also work without users manual setting.

autoconvert is also now fixed in #14746

[Damole-wer]

Is there a way to disable upscaling part for Nvidia VSR? I'm only interested in artifact reduction, since upscaling is only slightly better than lanczos.
It seems to support this (at least at 1080p), according to this blog post. Works in chrome. I've tried --vf=d3d11vpp:scale=1:scaling-mode=nvidia, but the feature doesn't get enabled and shows "Status: Inactive" in NVCP, checked at 480p, 720p and 1080p.
It works with a scale=1.002, but distorts the image quite a bit and obviously not an ideal way.
Thank you in advance!

[kasper93]

@Damole-wer: Try #15071, should help in this case.