[PATCH 0/1] MR8409: mfmediaengine: Only forward the most recent seek time.

[Brendan McGrath (＠redmcg)]

MF Media Engine on Windows will only send one seek request at a time. It will queue the `seekTime` value of the most recent `SetCurrentTime` call and forward only that request to the Media Session after it completes its current seek (i.e. when it sends out the `MF_MEDIA_ENGINE_EVENT_SEEKED` notify event).

Wine currently forwards every `SetCurrentTime` call to the Media Session, which can result in a large queue of seeks, one performed after the other, resulting in a delay before completing the final seek request.

This fixes slow seek times in some worlds in VRChat.

[PATCH 1/1] mfmediaengine: Only forward the most recent seek time.

[Brendan McGrath]

From: Brendan McGrath

---
 dlls/mfmediaengine/main.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/dlls/mfmediaengine/main.c b/dlls/mfmediaengine/main.c
index 27249f09fa6..51d89bde79b 100644
--- a/dlls/mfmediaengine/main.c
+++ b/dlls/mfmediaengine/main.c
@@ -153,6 +153,7 @@ struct media_engine
     double default_playback_rate;
     double volume;
     double duration;
+    double next_seek;
     MF_MEDIA_ENGINE_NETWORK network_state;
     MF_MEDIA_ENGINE_ERR error_code;
     HRESULT extended_code;
@@ -911,6 +912,8 @@ static HRESULT WINAPI media_engine_callback_GetParameters(IMFAsyncCallback *ifac
     return E_NOTIMPL;
 }

+static HRESULT media_engine_set_current_time(struct media_engine *engine, double seektime);
+
 static HRESULT WINAPI media_engine_session_events_Invoke(IMFAsyncCallback *iface, IMFAsyncResult *result)
 {
     struct media_engine *engine = impl_from_session_events_IMFAsyncCallback(iface);
@@ -989,6 +992,8 @@ static HRESULT WINAPI media_engine_session_events_Invoke(IMFAsyncCallback *iface
                 media_engine_set_flag(engine, FLAGS_ENGINE_SEEKING | FLAGS_ENGINE_IS_ENDED, FALSE);
                 IMFMediaEngineNotify_EventNotify(engine->callback, MF_MEDIA_ENGINE_EVENT_SEEKED, 0, 0);
                 IMFMediaEngineNotify_EventNotify(engine->callback, MF_MEDIA_ENGINE_EVENT_TIMEUPDATE, 0, 0);
+                if (isfinite(engine->next_seek))
+                    media_engine_set_current_time(engine, engine->next_seek);
             }
             LeaveCriticalSection(&engine->cs);
             IMFMediaEngineNotify_EventNotify(engine->callback, MF_MEDIA_ENGINE_EVENT_PLAYING, 0, 0);
@@ -1865,6 +1870,14 @@ static HRESULT media_engine_set_current_time(struct media_engine *engine, double
     if (FAILED(hr) || !(caps & MFSESSIONCAP_SEEK))
         return hr;

+    if (engine->flags & FLAGS_ENGINE_SEEKING)
+    {
+        engine->next_seek = seektime;
+        return S_OK;
+    }
+
+    engine->next_seek = NAN;
+
     position.vt = VT_I8;
     position.hVal.QuadPart = min(max(0, seektime), engine->duration) * 10000000;

@@ -3378,6 +3391,7 @@ static HRESULT init_media_engine(DWORD flags, IMFAttributes *attributes, struct
     engine->playback_rate = 1.0;
     engine->volume = 1.0;
     engine->duration = NAN;
+    engine->next_seek = NAN;
     engine->video_frame.pts = MINLONGLONG;
     InitializeCriticalSection(&engine->cs);

[Nikolay Sivov (＠nsivov)]

This makes sense. Have you checked if media session itself does similar kind of throttling on Windows?

[Brendan McGrath (＠redmcg)]

On Tue Jun 24 20:56:25 2025 +0000, Nikolay Sivov wrote:
> This makes sense. Have you checked if media session itself does similar
> kind of throttling on Windows?

I did yes. I made multiple calls to `IMFMediaSession::Start` in a row and found they all reached the `IMFMediaSource`. But if I called `IMFMediaEngine::SetCurrentTime` multiple times in a row, only the first and last made it to the `IMFMediaSource`.

[Nikolay Sivov (＠nsivov)]

Nikolay Sivov (@nsivov) commented about dlls/mfmediaengine/main.c:
>  engine->playback_rate = 1.0;
>  engine->volume = 1.0;
>  engine->duration = NAN;
> - engine->next_seek = NAN;
> engine->video_frame.pts = MINLONGLONG;

NAN was used for duration as a special marker because it doesn't come from user input. For the "next_seek" we could use another flag probably, instead of relying on this special value. Let me know which way looks better to you.

[Brendan McGrath (＠redmcg)]

On Tue Jun 24 21:21:02 2025 +0000, Nikolay Sivov wrote:
> NAN was used for duration as a special marker because it doesn't come
> from user input. For the "next_seek" we could use another flag probably,
> instead of relying on this special value. Let me know which way looks
> better to you.

I actually started with '-1', but then I thought maybe the user _could_ provide that. So I went with NAN for the same reason; it's not a value that you would expect from user input.

[PATCH 0/1] MR8409: mfmediaengine: Only forward the most recent seek time. - approved

[Nikolay Sivov (＠nsivov)]

This merge request was approved by Nikolay Sivov.