Merge branch 'main' of github.com:pytorch/torchcodec

Author: scotts

src/torchcodec/decoders/_core/VideoDecoder.cpp

@@ -1339,7 +1339,10 @@ int VideoDecoder::convertFrameToBufferUsingSwsScale(
 
   int expectedOutputHeight = outputTensor.sizes()[0];
   int expectedOutputWidth = outputTensor.sizes()[1];
-  if (activeStream.swsContext.get() == nullptr) {
+  auto curFrameSwsContextKey = SwsContextKey{
+      frame->width, frame->height, frameFormat, expectedOutputWidth, expectedOutputHeight};
+  if (activeStream.swsContext.get() == nullptr ||
+      activeStream.swsContextKey != curFrameSwsContextKey) {
     SwsContext* swsContext = sws_getContext(
         frame->width,
         frame->height,
@@ -1373,6 +1376,7 @@ int VideoDecoder::convertFrameToBufferUsingSwsScale(
         brightness,
         contrast,
         saturation);
+    activeStream.swsContextKey = curFrameSwsContextKey;
     activeStream.swsContext.reset(swsContext);
   }
   SwsContext* swsContext = activeStream.swsContext.get();

src/torchcodec/decoders/_core/VideoDecoder.h

@@ -317,6 +317,15 @@ class VideoDecoder {
     AVFilterContext* sourceContext = nullptr;
     AVFilterContext* sinkContext = nullptr;
   };
+  struct SwsContextKey {
+    int decodedWidth;
+    int decodedHeight;
+    AVPixelFormat decodedFormat;
+    int outputWidth;
+    int outputHeight;
+    bool operator==(const SwsContextKey&) const = default;
+    bool operator!=(const SwsContextKey&) const = default;
+  };
   // Stores information for each stream.
   struct StreamInfo {
     int streamIndex = -1;
@@ -337,6 +346,7 @@ class VideoDecoder {
     ColorConversionLibrary colorConversionLibrary = FILTERGRAPH;
     std::vector<FrameInfo> keyFrames;
     std::vector<FrameInfo> allFrames;
+    SwsContextKey swsContextKey;
     UniqueSwsContext swsContext;
   };
   VideoDecoder();