comment

Author: NicolasHug

src/torchcodec/_core/BetaCudaDeviceInterface.cpp

@@ -1331,14 +1331,6 @@ void BetaCudaDeviceInterface::convert_av_frame_to_frame_output(
   //
   // In contrast, a PacketDecoder will always upload CPU frames before retuning
   // them because its contract is to respect its device parameter.
-  //
-  // TODO_API_BREAKDOWN UF P1: Should test mismatch between device param of
-  // PacketDecoder and ColorConversion - maybe we're fine not handling this?
-  // Should check CUDA-CPU, CPU-CUDA, and CUDA-CUDA (with different devices)
-  // cases.
-  // TODO_API_BREAKDOWN UF P1: OK but we want the ColorConverter to be
-  // standalone: can we feed it frames on CPU and then on GPU? Will it be OK
-  // with that? Does that influence the TODO just above?
   bool needs_upload = mode() == Mode::Both && decoding_on_cpu_;
 
   // `uploaded` owns the GPU buffer for as long as it's in scope, which covers

src/torchcodec/decoders/_blocks/_color_converter.py

@@ -54,6 +54,15 @@ def __init__(
             output_dtype=convert_output_dtype_to_str(output_dtype),
         )
 
+    # TODO_API_BREAKDOWN DESIGN P2: The frame device must match the converter
+    # device. We have two alternative options:
+    # - not take a device parameter in the constructor and make the converter
+    #   device-agnostic. It requires caching the interfaces on the Converter.
+    # - take a device parameter and always honor it: that means downloading or
+    #   uploading CPU frames when needed.
+    # I feel like maybe we want to allow a CPU frame to be CCed on the GPU and
+    # do the upload ourselves (the download makes no sense, it's super slow).
+    # Anyway, that can be done later.
     def convert(self, raw_frame: RawFrame) -> Frame:
         data = _blocks_convert_frame(self._handle, raw_frame._handle, raw_frame._device)
         if raw_frame.storage is not None: