comments

Author: NicolasHug

src/torchcodec/_core/BetaCudaDeviceInterface.cpp

@@ -1003,8 +1003,6 @@ void BetaCudaDeviceInterface::convert_av_frame_to_frame_output(
   // may round them up to even.
   FrameDims original_dims(av_frame->height, av_frame->width);
 
-  // On the CPU fallback we own the GPU frame we just created; otherwise the
-  // input frame is already what we need, and it's only borrowed.
   UniqueAVFrame transferred_frame;
   if (cpu_fallback_) {
     AVPixelFormat target_pix_fmt = (output_dtype_ == OutputDtype::FLOAT32)

src/torchcodec/_core/DeviceInterface.h

@@ -98,11 +98,6 @@ class DeviceInterface {
     return requested_dtype;
   }
 
-  // The caller retains ownership of the frame, and keeps using it after this
-  // call returns (the building-block ops hand out frames that live in Python).
-  // Implementations must not take ownership of it, and must not free it.
-  // Implementations needing a different frame, e.g. after a format conversion,
-  // must allocate their own and leave this one alone.
   virtual void convert_av_frame_to_frame_output(
       const UniqueAVFrame& av_frame,
       FrameOutput& frame_output,