Fix cuda

Author: scotts

src/torchcodec/_core/CpuDeviceInterface.cpp

@@ -94,6 +94,9 @@ void CpuDeviceInterface::initialize(
   if (areTransformsSwScaleCompatible &&
       (userRequestedSwScale || isWidthSwScaleCompatible)) {
     colorConversionLibrary_ = ColorConversionLibrary::SWSCALE;
+
+    // SCOTT NEXT TODO: set swsFlags_
+
   } else {
     colorConversionLibrary_ = ColorConversionLibrary::FILTERGRAPH;
 
@@ -112,6 +115,8 @@ void CpuDeviceInterface::initialize(
       filters_ = filters.str();
     }
   }
+
+  initialized_ = true;
 }
 
 // Note [preAllocatedOutputTensor with swscale and filtergraph]:
@@ -127,6 +132,7 @@ void CpuDeviceInterface::convertAVFrameToFrameOutput(
     UniqueAVFrame& avFrame,
     FrameOutput& frameOutput,
     std::optional<torch::Tensor> preAllocatedOutputTensor) {
+  TORCH_CHECK(initialized_, "CpuDeviceInterface was not initialized.");
   if (preAllocatedOutputTensor.has_value()) {
     auto shape = preAllocatedOutputTensor.value().sizes();
     TORCH_CHECK(
@@ -167,7 +173,7 @@ void CpuDeviceInterface::convertAVFrameToFrameOutput(
       prevSwsFrameContext_ = swsFrameContext;
     }
     int resultHeight =
-        convertAVFrameToTensorUsingSwsScale(avFrame, outputTensor);
+        convertAVFrameToTensorUsingSwScale(avFrame, outputTensor);
     // If this check failed, it would mean that the frame wasn't reshaped to
     // the expected height.
     // TODO: Can we do the same check for width?
@@ -227,7 +233,7 @@ void CpuDeviceInterface::convertAVFrameToFrameOutput(
   }
 }
 
-int CpuDeviceInterface::convertAVFrameToTensorUsingSwsScale(
+int CpuDeviceInterface::convertAVFrameToTensorUsingSwScale(
     const UniqueAVFrame& avFrame,
     torch::Tensor& outputTensor) {
   uint8_t* pointers[4] = {

src/torchcodec/_core/CpuDeviceInterface.h

@@ -37,7 +37,7 @@ class CpuDeviceInterface : public DeviceInterface {
           std::nullopt) override;
 
  private:
-  int convertAVFrameToTensorUsingSwsScale(
+  int convertAVFrameToTensorUsingSwScale(
       const UniqueAVFrame& avFrame,
       torch::Tensor& outputTensor);
 
@@ -71,6 +71,10 @@ class CpuDeviceInterface : public DeviceInterface {
   AVRational timeBase_;
   FrameDims outputDims_;
 
+  // If we use swscale for resizing, the flags control the resizing algorithm.
+  // We exclusively get the value from the ResizeTransform.
+  int swsFlags_ = 0;
+
   // The copy filter just copies the input to the output. Computationally, it
   // should be a no-op. If we get no user-provided transforms, we will use the
   // copy filter.
@@ -85,6 +89,8 @@ class CpuDeviceInterface : public DeviceInterface {
   // be created before decoding a new frame.
   SwsFrameContext prevSwsFrameContext_;
   FiltersContext prevFiltersContext_;
+
+  bool initialized_ = false;
 };
 
 } // namespace facebook::torchcodec

src/torchcodec/_core/CudaDeviceInterface.cpp

@@ -216,10 +216,17 @@ void CudaDeviceInterface::convertAVFrameToFrameOutput(
     // Typically that happens if the video's encoder isn't supported by NVDEC.
     // Below, we choose to convert the frame's color-space using the CPU
     // codepath, and send it back to the GPU at the very end.
+    //
     // TODO: A possibly better solution would be to send the frame to the GPU
     // first, and do the color conversion there.
+    //
+    // TODO: If we're going to keep this around, we should probably cache it?
     auto cpuDevice = torch::Device(torch::kCPU);
     auto cpuInterface = createDeviceInterface(cpuDevice);
+    TORCH_CHECK(
+        cpuInterface != nullptr, "Failed to create CPU device interface");
+    cpuDeviceInterface->initialize(
+        nullptr, VideoStreamOptions(), {}, timeBase_, outputDims_);
 
     FrameOutput cpuFrameOutput;
     cpuInterface->convertAVFrameToFrameOutput(