Initial C++ implementation of transforms (#902)

Author: scotts

src/torchcodec/_core/BetaCudaDeviceInterface.cpp

@@ -156,13 +156,20 @@ BetaCudaDeviceInterface::~BetaCudaDeviceInterface() {
   }
 }
 
-void BetaCudaDeviceInterface::initializeInterface(AVStream* avStream) {
+void BetaCudaDeviceInterface::initialize(const AVStream* avStream) {
   torch::Tensor dummyTensorForCudaInitialization = torch::empty(
       {1}, torch::TensorOptions().dtype(torch::kUInt8).device(device_));
 
   TORCH_CHECK(avStream != nullptr, "AVStream cannot be null");
   timeBase_ = avStream->time_base;
 
+  auto cudaDevice = torch::Device(torch::kCUDA);
+  defaultCudaInterface_ =
+      std::unique_ptr<DeviceInterface>(createDeviceInterface(cudaDevice));
+  AVCodecContext dummyCodecContext = {};
+  defaultCudaInterface_->initialize(avStream);
+  defaultCudaInterface_->registerHardwareDeviceWithCodec(&dummyCodecContext);
+
   const AVCodecParameters* codecpar = avStream->codecpar;
   TORCH_CHECK(codecpar != nullptr, "CodecParameters cannot be null");
 
@@ -523,8 +530,6 @@ void BetaCudaDeviceInterface::flush() {
 }
 
 void BetaCudaDeviceInterface::convertAVFrameToFrameOutput(
-    const VideoStreamOptions& videoStreamOptions,
-    const AVRational& timeBase,
     UniqueAVFrame& avFrame,
     FrameOutput& frameOutput,
     std::optional<torch::Tensor> preAllocatedOutputTensor) {
@@ -535,20 +540,8 @@ void BetaCudaDeviceInterface::convertAVFrameToFrameOutput(
   // TODONVDEC P1: we use the 'default' cuda device interface for color
   // conversion. That's a temporary hack to make things work. we should abstract
   // the color conversion stuff separately.
-  if (!defaultCudaInterface_) {
-    auto cudaDevice = torch::Device(torch::kCUDA);
-    defaultCudaInterface_ =
-        std::unique_ptr<DeviceInterface>(createDeviceInterface(cudaDevice));
-    AVCodecContext dummyCodecContext = {};
-    defaultCudaInterface_->initializeContext(&dummyCodecContext);
-  }
-
   defaultCudaInterface_->convertAVFrameToFrameOutput(
-      videoStreamOptions,
-      timeBase,
-      avFrame,
-      frameOutput,
-      preAllocatedOutputTensor);
+      avFrame, frameOutput, preAllocatedOutputTensor);
 }
 
 BetaCudaDeviceInterface::FrameBuffer::Slot*

src/torchcodec/_core/BetaCudaDeviceInterface.h

@@ -37,11 +37,9 @@ class BetaCudaDeviceInterface : public DeviceInterface {
   explicit BetaCudaDeviceInterface(const torch::Device& device);
   virtual ~BetaCudaDeviceInterface();
 
-  void initializeInterface(AVStream* stream) override;
+  void initialize(const AVStream* avStream) override;
 
   void convertAVFrameToFrameOutput(
-      const VideoStreamOptions& videoStreamOptions,
-      const AVRational& timeBase,
       UniqueAVFrame& avFrame,
       FrameOutput& frameOutput,
       std::optional<torch::Tensor> preAllocatedOutputTensor =

src/torchcodec/_core/CMakeLists.txt

@@ -95,6 +95,7 @@ function(make_torchcodec_libraries
         SingleStreamDecoder.cpp
         Encoder.cpp
         ValidationUtils.cpp
+        Transform.cpp
     )
 
     if(ENABLE_CUDA)