Enable AMF Decoder, HWContext AMF, vpp_amf filter by Hotactioncop · Pull Request #6475 · HandBrake/HandBrake

Hotactioncop · 2024-12-10T00:04:28Z

Adds hwcontext_amf, enabling a shared AMF context for encoders,
decoders, and AMF-based filters, without copy to the host memory.

Benefits:

Optimizations for direct video memory access from CPU
Significant performance boost in full AMF pipelines with filters
Integration of GPU filters like VPP, Super Resolution, and
Compression Artefact Removal(in future plans)
VCN power management control for decoders.
Ability to specify which VCN instance to use for decoding
(like for encoder)
AMD will soon introduce full AMF API for multimedia accelerator MA35D
- With AMF API, integration will be much easier:
  GPU and the accelerator will have the same API
- including encoder, decoder, scaler, color converter

Tested on:

Windows 10+ (via MinGW)
Ubuntu Linux

galad87 · 2024-12-10T06:57:00Z

I have a feeling it would be better to wait until the patches are integrated into FFmpeg, and after #6446 is merged.

stickz · 2024-12-21T23:20:32Z

Thanks so much for this! I'm running the AMF decoder with SVT-AV1 on my AMD Ryzen 9 9950x. It works great! I just had to update my bios and disable PBO to fix the BSOD problem.

Hotactioncop · 2025-02-21T17:32:08Z

Asking for patch review.
What changed since last PR:
Updated to latest HandBrake master.
Updated contriib/ffmpeg patches to latest HWcontext, which is implemented into FFmpeg master.
Implemented GUI checkbox to enable AMF decoder on Linux.
Fixed crashes on Linux for copy_plane when default bytewidth larger than the source frame linesize.

galad87

Added a bunch of random comments, didn't test this yet. But it would be better to replicate what NVenc/NVdec are doing (even if it was an hit and run and no one is keeping it working), instead of QSV.

galad87 · 2025-02-28T16:19:38Z

libhb/common.c

    job->qsv.decode                = !!(title->video_decode_support &
                                        HB_DECODE_SUPPORT_QSV);
 #endif
+    memset(&job->amf, 0, sizeof(job->amf));


hb_job_t is already zeroed on creation, this shouldn't be needed.

Thanks, this will be applied in the upcoming fix commit.

galad87 · 2025-02-28T16:22:50Z

libhb/decavcodec.c

            hb_dict_set(settings, "format", hb_value_string(av_get_pix_fmt_name(pv->job->input_pix_fmt)));
            hb_avfilter_append_dict(filters, "scale_cuda", settings);
        }
+        else if ((pv->frame->hw_frames_ctx && pv->job && pv->job->hw_pix_fmt == AV_PIX_FMT_AMF_SURFACE) &&


A rotate filter is needed in this function too, if not all those video recorded on phones with a rotation metadata will fail to encode properly

Currently, AMF Hardware filters do not support rotation.
I will add an extra condition if (pv->title->rotation == HB_ROTATION_0) in the upcoming fix commit.

galad87 · 2025-02-28T16:23:56Z

libhb/decavcodec.c

                // HWAccel falled back to the software decoder
                av_frame_ref(pv->frame, pv->hw_frame);
                av_frame_unref(pv->hw_frame);
-                if (ret < 0)


Accidentally removed, will be returned in the next fix commit.

galad87 · 2025-02-28T16:24:28Z

libhb/decavcodec.c

-    } while (ret >= 0);
+        if(send_packet)
+        {
+            av_usleep(100);// wait 0.1 ms before resubmitting


Please don't add random usleep around.

A sleep is required if the submit input is full, to prevent unnecessary CPU usage.

galad87 · 2025-02-28T16:25:37Z

libhb/decavcodec.c

+        if(name != NULL)
+        {
+            pv->codec = avcodec_find_decoder_by_name(name);
+            if(pv->job) // no need to copy output for preview


I feel this is not a good way, what if we start to use hw filters for previews in the future?

A reasonable remark. Will be removed in the next fix commit.

galad87 · 2025-02-28T16:28:24Z

libhb/decavcodec.c

        }
    }
 #endif
+#if HB_PROJECT_FEATURE_VCE


The other hardware encoders check if the hwaccel falled back to software decoder, wouldn't that work for amfdec too?

galad87 · 2025-02-28T16:28:46Z

libhb/handbrake/common.h

+    // AMF-specific settings
+    struct
+    {
+        int num_sw_filters;


Do we need to store these numbers?

We need to store this variable because it is used to check whether software filters are present.
If this variable is not 0, it means that software filters are present, and the pipeline can't be fully hardware-accelerated. However, since HandBrake is designed for maximum compatibility, we use int instead of bool.

galad87 · 2025-02-28T16:34:37Z

libhb/vce_common.c

+{
+    return job->hw_pix_fmt == AV_PIX_FMT_AMF_SURFACE;
+}
+hb_buffer_t *  hb_vce_copy_avframe_to_video_buffer(hb_job_t *job, AVFrame *frame, AVRational time_base)


Is this function needed? Looks the same as hb_avframe_to_video_buffer

The hw frame to sw frame conversion happens before, so there shouldn't be any need to check for sw_filters or whatever.

galad87 · 2025-02-28T16:37:36Z

libhb/decavcodec.c

+    if(w->hw_device_ctx)
    {
-        pv->codec = avcodec_find_decoder_by_name(pv->qsv.codec_name);
+        const char* name = hb_vce_decode_get_codec_name(w->codec_param);


Would't be better to use the hwaccel framework like nvdec and videotoolboxdec? Don't try to copy the QSV implementation, because it was added before hwaccel, and it's doing custom things that hopefully will go away soon.

The HWContext and AMF decoder is not hardware accelerated, it is an offload decoder.
Implementation of HWaccel AMF in hw_configs for default codecs was blocked by ffmpeg maintainers:
https://patchwork.ffmpeg.org/project/ffmpeg/patch/20240214015515.1027-2-ovchinnikov.dmitrii@gmail.com/

The issue with standalone hw decoders is that they are missing features, and are never in sync with the built-in parsers and hwaccel decoders. For example HDR dynamic metadata are ignored, and others metadata too. It's a pain because you never know what works and what doesn't.

That ultimately leads to a bad user experience when features that should be transparent and just work, don't.

Geoknysis · 2025-02-28T20:44:46Z

There may be a bug that requires fixing before the feature is fully functional. Per the below, the output is sub-optimal and the work speed is slower. Below each image is the respective activity log for the encode. I also noticed that with the amfdecode option selected, when selecting the constant framerate video option, HandBrake crashes.

I used the official h265 1080p preset for each. I compiled HandBrake with https://github.com/Hotactioncop/HandBrake/tree/Enable_AMF_decode source.

Activity.log (official_h265_vcn_1080p_amfdecode).txt

Activity.log (official_h265_vcn_1080p_no_amfdecode).txt

Hotactioncop · 2025-03-21T15:37:56Z

There may be a bug that requires fixing before the feature is fully functional. Per the below, the output is sub-optimal and the work speed is slower. Below each image is the respective activity log for the encode. I also noticed that with the amfdecode option selected, when selecting the constant framerate video option, HandBrake crashes.

I used the official h265 1080p preset for each. I compiled HandBrake with https://github.com/Hotactioncop/HandBrake/tree/Enable_AMF_decode source.

Thank you for testing this PR!
I noticed that you are seeing different frames compared to my results.
To help me investigate and reproduce the issue, could you please provide the following details?

Input video file (or a short sample that reproduces the issue).
Your software setup, including:

Linnux version
GPU model
Driver version
Any specific build configuration or flags used
Your HandBrake settings, especially the encoding parameters you used.

With my setup and a prepared video file matching the characteristics in your log file, I haven't encountered any errors.
Having your exact input and environment details will help me debug this properly.
Thanks!

Mine181 · 2025-08-25T19:38:04Z

any update on this merge request?

Hotactioncop · 2025-08-27T12:21:40Z

Thanks for following up!
The patch didn’t fit anymore after recent HandBrake updates, so It is rewritten to match the new code.
It is currently under testing, and I will update the PR shortly.

sr55 added the Enhancement label Dec 29, 2024

sr55 added this to the 1.10.0 milestone Dec 29, 2024

sr55 mentioned this pull request Feb 19, 2025

Intel QuickSync decoding not used with AMD VCE encoding #6653

Closed

Hotactioncop force-pushed the Enable_AMF_decode branch from e8687d0 to fc8b54d Compare February 21, 2025 17:15

Enable AMF Decoder, HWContext AMF, vpp_amf filter

310f497

Hotactioncop force-pushed the Enable_AMF_decode branch from fc8b54d to 310f497 Compare February 28, 2025 12:19

galad87 reviewed Feb 28, 2025

View reviewed changes

galad87 removed this from the 1.10.0 milestone Jun 11, 2025

galad87 mentioned this pull request Nov 11, 2025

Command line is lacking call to AMD VCE HandBrake/HandBrake-docs#247

Closed

Conversation

Hotactioncop commented Dec 10, 2024

Uh oh!

galad87 commented Dec 10, 2024

Uh oh!

stickz commented Dec 21, 2024

Uh oh!

Hotactioncop commented Feb 21, 2025

Uh oh!

galad87 left a comment

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Hotactioncop Mar 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Geoknysis commented Feb 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Hotactioncop commented Mar 21, 2025

Uh oh!

Mine181 commented Aug 25, 2025

Uh oh!

Hotactioncop commented Aug 27, 2025

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

6 participants

Hotactioncop Mar 21, 2025 •

edited

Loading

Geoknysis commented Feb 28, 2025 •

edited

Loading