fix(bench): del ctypes/bytes objects immediately in callback to prevent OOM on 1GB device

tests/benchmark_nv12_decode.py

@@ -183,20 +183,23 @@ def _on_sample(sink) -> Gst.FlowReturn:
         except Exception:
             pass
 
-    # Measure from_buffer_copy() cost (the CPU copy the app does)
+    # Measure extract_dup (GStreamer buf → Python bytes) + from_buffer_copy
+    # (Python bytes → ctypes array for SDL upload).  del objects immediately
+    # after timing so CPython's ref-counting frees the 3+ MB allocations at
+    # once rather than letting them accumulate across frames (OOM on 1 GB device).
     t0 = time.monotonic()
     raw = buf.extract_dup(0, buf.get_size())
-    copy_us = (time.monotonic() - t0) * 1e6
+    extract_us = (time.monotonic() - t0) * 1e6
 
     if fmt_str == "NV12":
-        y_size  = int(info.stride[0]) * int(info.height)
-        # Simulate the single from_buffer_copy the app now does:
-        # ONE copy of the full Y+UV buffer, then ctypes.byref for the UV offset.
+        y_size = int(info.stride[0]) * int(info.height)
         t1 = time.monotonic()
         arr = (ctypes.c_ubyte * len(raw)).from_buffer_copy(raw)
-        y_ptr  = ctypes.cast(arr, ctypes.POINTER(ctypes.c_ubyte))
-        uv_ptr = ctypes.cast(ctypes.byref(arr, y_size), ctypes.POINTER(ctypes.c_ubyte))
-        copy_us = (time.monotonic() - t1) * 1e6
+        copy_us = extract_us + (time.monotonic() - t1) * 1e6
+        del arr   # free 3 MB ctypes array immediately
+    else:
+        copy_us = extract_us
+    del raw       # free 3 MB bytes object immediately
 
     with stats.lock:
         stats.total_frames += 1