New upstream version 7.0.2

[nyanmisaka]

Vulkan patches are not yet ready.

Changes

• New upstream version 7.0.2
• Rename package to jellyin-ffmpeg7
• Update build scripts and dependencies

Issues

• Fixes Upgrade to FFmpeg 7.0 #391

[gnattu]

I'm going to make a version merges two of them. But this is good enough for now.

[gnattu]

With 7.0.1+backported patches ffmpeg now correctly set the buffer attachments for videotoolbox and now the tonemap pipeline needs to be slightly modified because the old one works with some assumptions with wrong colors, and the HDR attachments is not properly removed doing tone-mapping to 10bit now.

[nyanmisaka]

and the HDR attachments is not properly removed doing tone-mapping to 10bit now.

Is CVBufferRemoveAttachment() needed here?

[gnattu]

and the HDR attachments is not properly removed doing tone-mapping to 10bit now.

Is CVBufferRemoveAttachment() needed here?

We need to override the kCVImageBufferTransferFunctionKey I think, but I will look at this later. Once this one is merged I will submit a PR fix it.

[gnattu]

Well interesting, scale_vt is still producing wrong color without manually hwupload. I guess the upstream missed up something once again.

[nyanmisaka]

Well interesting, scale_vt is still producing wrong color without manually hwupload. I guess the upstream missed up something once again.

These two fixes have been included in the backport patches.
FFmpeg/FFmpeg@1fa7554
FFmpeg/FFmpeg@cd9ceae

Is this possible related to the newly added SDK version check?
FFmpeg/FFmpeg@85706f5
FFmpeg/FFmpeg@2fc37c4
FFmpeg/FFmpeg@ca7fcf5

[gnattu]

Well interesting, scale_vt is still producing wrong color without manually hwupload. I guess the upstream missed up something once again.

These two fixes have been included in the backport patches. FFmpeg/FFmpeg@1fa7554 FFmpeg/FFmpeg@cd9ceae

Is this possible related to the newly added SDK version check? FFmpeg/FFmpeg@85706f5 FFmpeg/FFmpeg@2fc37c4 FFmpeg/FFmpeg@ca7fcf5

The CGColorSpace generated from its provided buffer is wrong and creates an SDR colorspace.

[nyanmisaka]

The CGColorSpace generated from its provided buffer is wrong and creates an SDR colorspace.

Is this also reproducible in the ffmpeg master branch? I saw your discussion on the mailing list and thought it was fixed.

[gnattu]

The CGColorSpace generated from its provided buffer is wrong and creates an SDR colorspace.

Is this also reproducible in the ffmpeg master branch? I saw your discussion on the mailing list and thought it was fixed.

I identified the problem but I found that it is really funny. The whole sdk version checking drama is to get the attachments from the pixel buffer, but the attachments get from that buffer will always creates a non-HDR buffer. The very first implementation allocated another dedicated dictionary which only specifies the color primaries, the transfer function and the color matrix, which will properly create an HDR colorspace and make scale_vt happy. I need to submit this to the upstream I think.

[gnattu]

The good news is that, a framebuffer with properly set attachments brings slightly performance improvements over hwupload, instead of regression in earlier version. So maybe we can get rid of hwupload in jellyfin 10.10.

[gnattu]

Patch submitted to upstream: https://ffmpeg.org//pipermail/ffmpeg-devel/2024-August/331960.html

[nyanmisaka]

Patch submitted to upstream: https://ffmpeg.org//pipermail/ffmpeg-devel/2024-August/331960.html

defined but not used?

static CFDictionaryRef vt_cv_buffer_copy_attachments(CVBufferRef buffer,
                                                     CVAttachmentMode attachment_mode)

redundant braces?

#if (TARGET_OS_OSX && __MAC_OS_X_VERSION_MAX_ALLOWED >= 100800) || \
    (TARGET_OS_IOS && __IPHONE_OS_VERSION_MAX_ALLOWED >= 100000)
    if (__builtin_available(macOS 10.8, iOS 10, *)) {
        colorspace = CVImageBufferCreateColorSpaceFromAttachments(attachments);
    }
#endif

[gnattu]

Patch submitted to upstream: https://ffmpeg.org//pipermail/ffmpeg-devel/2024-August/331960.html

defined but not used?

static CFDictionaryRef vt_cv_buffer_copy_attachments(CVBufferRef buffer,
                                                     CVAttachmentMode attachment_mode)

This can be used because the real cause is simpler than I originally thought.

redundant braces?

#if (TARGET_OS_OSX && __MAC_OS_X_VERSION_MAX_ALLOWED >= 100800) || \
    (TARGET_OS_IOS && __IPHONE_OS_VERSION_MAX_ALLOWED >= 100000)
    if (__builtin_available(macOS 10.8, iOS 10, *)) {
        colorspace = CVImageBufferCreateColorSpaceFromAttachments(attachments);
    }
#endif

CVImageBufferCreateColorSpaceFromAttachments is only available in macOS 10.8 so this is still necessary.

[gnattu]

See: https://github.com/FFmpeg/FFmpeg/blob/1b2a925e94c772c59a88c03c1654bddf6aff0ca2/libavutil/hwcontext_videotoolbox.c#L600

The color primaries value is wrongly assigned to color transfer key and that's why the attachments copied from the pixel buffer will always create an SDR colorspace. After correcting this the original code works properly.

[nyanmisaka]

See: https://github.com/FFmpeg/FFmpeg/blob/1b2a925e94c772c59a88c03c1654bddf6aff0ca2/libavutil/hwcontext_videotoolbox.c#L600

The color primaries value is wrongly assigned to color transfer key and that's why the attachments copied from the pixel buffer will always create an SDR colorspace. After correcting this the original code works properly.

colorxxx is hard to distinguish at a glance... It's necessary to use underscores or camelCase.

[gnattu]

Patch submitted to upstream: https://ffmpeg.org//pipermail/ffmpeg-devel/2024-August/331960.html

Replaced by: https://ffmpeg.org//pipermail/ffmpeg-devel/2024-August/331962.html

[gnattu]

This version currently does not contain the nop vt_device_derive because it is not needed for upstream. But our current server code uses hwupload=derive_device=videotoolbox a lot. Should we add that as well to support current server?

[nyanmisaka]

This version currently does not contain the nop vt_device_derive because it is not needed for upstream. But our current server code uses hwupload=derive_device=videotoolbox a lot. Should we add that as well to support current server?

I remembered it. Will re-add it as a separate patch since upstream will never accept it.

[nyanmisaka]

Merging to jellyfin-7.0 branch for now.