[Question]: Can dual encoders be used to improve performance of a single video encoding?

[pbaykalov]

Can dual encoders be used to improve performance of a single video encoding?
I know that there is an explicit selection of encoder instance and it works well in OBS.
I've seen there is EnableEncoderSmartAccessVideo but is it able to use dual encoders of the same GPU at all?
What kind of parallelization is done with EnableEncoderSmartAccessVideo?
It's interesting that there is no encoder selection for FFMPEG, transcoding speed is trully horrendous with decoding and encoding running on a single VCN instance on my 7900XTX.

[MikhailAMD]

There are several topics here:

1. In single transcode scenario on GPU with two VCNs decode session should go to one VCN and encode session to another; should happen automatically in FFmpeg.
2. SmartAccessVideo technology: enables redirection of encoder and/or decoder sessions between two VCN instances or to VCN from the second GPU or iGPU/APU, decision is based on current VCN load. Once enabled, redirection happens under the hood. Helps in multisession use cases, Windows only.
3. With two VCNs available, split-frame encoding could improve performance, but only in simple cases. It can be enabled by AMF_VIDEO_ENCODER_HEVC_MULTI_HW_INSTANCE_ENCODE for HEVC and similar for AV1. Note, that 7900XTX has AV1 in one VCN only and AV1 parameter would not work. Many conditions are checked to decide if it can be enabled or not, to name few: 4K resolution or more, no PA, no HIGH_MOTION_QUALITY_BOOST. etc, etc. In the essence, this parameter is just a hint.
4. Performance can depend on many factors: use of PA, use of CPU filters etc, etc. Fully accelerated FFmpeg transcoding via AMF should give decent numbers. It only makes sense to discuss after we see the command line, performance results and source material.
5. Ensure latest FFmpeg and latest drivers.

[pbaykalov]

In single transcode scenario on GPU with two VCNs decode session should go to one VCN and encode session to another; should happen automatically in FFmpeg.

I am definitely not seeing anything like it with AV1 input and H264 output.
Now that you said it I see it with H264 input and H264 output.
Also I again do not see anything like it when using MPV with AV1 video as input for OBS recording in H264 for testing - only one VCN is busy.

Does it have anything to do with only one VCN having AV1?

Exact command which shows no parallel VCN usage for me:
ffmpeg -hwaccel d3d11va -i 'AV1 HIGH QUALITY.mkv' -c:v h264_amf -vf "crop=3840:2160:0:0,scale=1920:1080" -usage transcoding -rc cbr -b 6M temp.mkv
Or should it?

[MikhailAMD]

Several issues here:

• I assume you are running the latest FFmpeg
• It looks like you are missing a parameter that would keep data in video memory: -hwaccel_output_format d3d11
• d3d11va decoder produces frames in D3D11 textures that cannot be used in D3D11 shaders and AMF needs it in vpp (BIND flags). So, unnecessary copy happens. To mitigate we introduced a parameter for d3dva HW context but it was not yet merged: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20659
• You can try AMF decoder that should provide optimal path: ffmpeg -hwaccel amf -hwaccel_output_format amf
• We will update AMF FFmpeg Wiki

[pbaykalov]

@MikhailAMD running this command with ffmpeg 8.0.1 produces following error:
ffmpeg -hwaccel amf -hwaccel_output_format amf -i 'AV1 HIGH QUALITY.mkv' -c:v h264_amf -vf "crop=3840:2160:0:0,scale=1920:1080" -usage transcoding -rc cbr -b 6M temp.mkv

Impossible to convert between the formats supported by the filter 'Parsed_crop_0' and the filter 'auto_scale_0' Link 'Parsed_crop_0.default' -> 'auto_scale_0.default':

As I understand it there is vpp_amf filter which should be used but it cannot crop the input at all and neither can the decoder.

[pbaykalov]

This command obviously works and throughput is a lot better but I still only see just one VCN being used. Is it expected behaviour?

ffmpeg -hwaccel amf -hwaccel_output_format amf -i 'AV1 HIGH QUALITY.mkv' -c:v h264_amf -vf "vpp_amf=2240:1080" -usage transcoding -rc cbr -b 6M temp.mkv

[MikhailAMD]

Your GPU should use two VCNs in this case. Note that after introduction of HW scheduler in Windows two VCNs look like one codec block in all monitoring tools: Task Manager, GPUVIEW, etc. https://devblogs.microsoft.com/directx/hardware-accelerated-gpu-scheduling/
If HW scheduler is on, can you disable it and reboot? Then record short GPUVIEW log, ZIP it and share, I will check.
What are your input and actual output resolutions in files (mediainfo)?
What is your driver version?