tests: fix benchmark pipelines (decodebin + rank steering, LD_PRELOAD note)

tests/benchmark_decode.py

@@ -224,17 +224,26 @@ def main() -> None:
         size = os.path.getsize(video_path)
         print(f"  Using existing: {video_path} ({size // 1024}KB)")
 
+    def _set_mpp_rank(rank: int) -> None:
+        """Temporarily set mppvideodec factory rank to steer decodebin selection."""
+        factory = reg.find_feature("mppvideodec", Gst.ElementFactory.__gtype__)
+        if factory:
+            factory.set_rank(rank)
+
     # --- benchmark runs -----------------------------------------------------
     _section("Benchmarks (sync=false, as fast as possible)")
 
     results = []
 
+    # Pipeline note: qtdemux has dynamic src pads so we use decodebin.
+    # decodebin rank manipulation steers it toward SW or HW decoder.
+    GST_RANK_NONE = 0
+    MPP_RANK_HIGH = 257  # above avdec_h264 (256)
+
     # 1. SW fakesink — pure decode throughput
     if av_ok:
-        pipe = (
-            f"filesrc location={video_path} ! qtdemux ! h264parse "
-            f"! avdec_h264 ! fakesink sync=false"
-        )
+        _set_mpp_rank(GST_RANK_NONE)  # ensure avdec_h264 wins
+        pipe = f"filesrc location={video_path} ! decodebin ! fakesink sync=false"
         results.append(_run_benchmark(
             "SW fakesink (avdec_h264 → discard)",
             video_path, pipe, is_appsink=False, Gst=Gst, GLib=GLib,
@@ -242,10 +251,10 @@ def main() -> None:
 
     # 2. SW appsink — full BGRA path (as used by app without HW decode)
     if av_ok:
+        _set_mpp_rank(GST_RANK_NONE)
         pipe = (
-            f"filesrc location={video_path} ! qtdemux ! h264parse "
-            f"! avdec_h264 ! videoconvert "
-            f"! video/x-raw,format=BGRA "
+            f"filesrc location={video_path} ! decodebin "
+            f"! videoconvert ! video/x-raw,format=BGRA "
             f"! appsink name=bench_sink"
         )
         results.append(_run_benchmark(
@@ -255,10 +264,8 @@ def main() -> None:
 
     # 3. HW fakesink — pure MPP decode throughput
     if mpp_ok:
-        pipe = (
-            f"filesrc location={video_path} ! qtdemux ! h264parse "
-            f"! mppvideodec ! fakesink sync=false"
-        )
+        _set_mpp_rank(MPP_RANK_HIGH)  # prefer mppvideodec
+        pipe = f"filesrc location={video_path} ! decodebin ! fakesink sync=false"
         results.append(_run_benchmark(
             "HW fakesink (mppvideodec → discard)",
             video_path, pipe, is_appsink=False, Gst=Gst, GLib=GLib,
@@ -266,9 +273,9 @@ def main() -> None:
 
     # 4. HW appsink NV12 — full app path with HW + zero-copy SDL upload
     if mpp_ok:
+        _set_mpp_rank(MPP_RANK_HIGH)
         pipe = (
-            f"filesrc location={video_path} ! qtdemux ! h264parse "
-            f"! mppvideodec "
+            f"filesrc location={video_path} ! decodebin "
             f"! appsink name=bench_sink caps=video/x-raw,format=NV12"
         )
         results.append(_run_benchmark(
@@ -276,6 +283,8 @@ def main() -> None:
             video_path, pipe, is_appsink=True, Gst=Gst, GLib=GLib,
         ))
 
+    _set_mpp_rank(64)  # restore default marginal rank
+
     # --- summary table -------------------------------------------------------
     _section("Results")
     print(f"  {'Path':<45} {'FPS':>6}  {'CPU%':>5}")