Performance improvements for sending small images (tested with 64x64) (…

…#174)

- directly compress them in the call thread
- keep JPEG out buffer allocated to save one memory allocation

deflect/ImageJpegCompressor.cpp

@@ -109,27 +109,25 @@ QByteArray ImageJpegCompressor::computeJpeg(const ImageWrapper& sourceImage,
  const int tjPitch = sourceImage.width * sourceImage.getBytesPerPixel();
  const int tjHeight = imageRegion.height();
  const int tjPixelFormat = _getTurboJpegFormat(sourceImage.pixelFormat);
- unsigned char* tjJpegBuf = nullptr;
- unsigned long tjJpegSize = 0;
+
  const int tjJpegSubsamp = _getTurboJpegSubsamp(sourceImage.subsampling);
+ unsigned long tjJpegSize = tjBufSize(tjWidth, tjHeight, tjJpegSubsamp);
+
+ _tjJpegBuf.resize(tjJpegSize);
+
  const int tjJpegQual = sourceImage.compressionQuality;
- const int tjFlags = 0; // or: TJFLAG_BOTTOMUP
+ const int tjFlags = TJFLAG_NOREALLOC; // or: TJFLAG_BOTTOMUP
 
+ auto ptr = _tjJpegBuf.data();
  int err = tjCompress2(_tjHandle, tjSrcBuffer, tjWidth, tjPitch, tjHeight,
- tjPixelFormat, &tjJpegBuf, &tjJpegSize, tjJpegSubsamp,
+ tjPixelFormat, &ptr, &tjJpegSize, tjJpegSubsamp,
  tjJpegQual, tjFlags);
  if (err != 0)
  {
  std::cerr << "libjpeg-turbo image conversion failure" << std::endl;
  return QByteArray();
  }
 
- // move the JPEG buffer to a byte array
- const QByteArray jpegData((char*)tjJpegBuf, tjJpegSize);
-
- // free the libjpeg-turbo allocated memory
- tjFree(tjJpegBuf);
-
- return jpegData;
+ return QByteArray((const char*)ptr, tjJpegSize);
 }
 }

deflect/ImageJpegCompressor.h

@@ -71,6 +71,7 @@ class ImageJpegCompressor
 
 private:
  tjhandle _tjHandle;
+ std::vector<unsigned char> _tjJpegBuf;
 };
 }
 

deflect/ImageSegmenter.cpp

@@ -69,6 +69,21 @@ bool ImageSegmenter::generate(const ImageWrapper& image, const Handler& handler)
  return _generateRaw(image, handler);
 }
 
+Segment ImageSegmenter::compressSingleSegment(const ImageWrapper& image)
+{
+#ifdef DEFLECT_USE_LIBJPEGTURBO
+ auto segments = _generateSegments(image);
+ if (segments.size() > 1)
+ throw std::runtime_error(
+ "compressSingleSegment only works for small images");
+ ImageSegmenter::_computeJpeg(segments[0], false);
+ return segments[0];
+#else
+ throw std::runtime_error(
+ "LibJpegTurbo not available, needed for compressSingleSegment");
+#endif
+}
+
 void ImageSegmenter::setNominalSegmentDimensions(const uint width,
  const uint height)
 {
@@ -85,7 +100,7 @@ bool ImageSegmenter::_generateJpeg(const ImageWrapper& image,
 
  // start creating JPEGs for each segment, in parallel
  QtConcurrent::map(segments, std::bind(&ImageSegmenter::_computeJpeg, this,
- std::placeholders::_1));
+ std::placeholders::_1, true));
 
  // Sending compressed jpeg segments while they arrive in the queue.
  // Note: Qt insists that sending (by calling handler()) should happen
@@ -108,7 +123,7 @@ bool ImageSegmenter::_generateJpeg(const ImageWrapper& image,
 #endif
 }
 
-void ImageSegmenter::_computeJpeg(Segment& segment)
+void ImageSegmenter::_computeJpeg(Segment& segment, const bool sendSegment)
 {
 #ifdef DEFLECT_USE_LIBJPEGTURBO
  QRect imageRegion(segment.parameters.x - segment.sourceImage->x,
@@ -124,7 +139,8 @@ void ImageSegmenter::_computeJpeg(Segment& segment)
  segment.imageData =
  compressor.localData().computeJpeg(*segment.sourceImage, imageRegion);
  segment.parameters.dataType = DataType::jpeg;
- _sendQueue.enqueue(segment);
+ if (sendSegment)
+ _sendQueue.enqueue(segment);
 #endif
 }
 

deflect/ImageSegmenter.h

@@ -91,6 +91,16 @@ class ImageSegmenter
  */
  DEFLECT_API void setNominalSegmentDimensions(uint width, uint height);
 
+ /**
+ * For a small input image (tested with 64x64, possible for <=512 as well),
+ * directly compress it to a single segment which will be enqueued for
+ * sending.
+ *
+ * @param image The image to be compressed
+ * @return the compressed segment
+ */
+ DEFLECT_API Segment compressSingleSegment(const ImageWrapper& image);
+
 private:
  struct SegmentationInfo
  {
@@ -105,7 +115,7 @@ class ImageSegmenter
  };
 
  bool _generateJpeg(const ImageWrapper& image, const Handler& handler);
- void _computeJpeg(Segment& task);
+ void _computeJpeg(Segment& task, bool sendSegment);
  bool _generateRaw(const ImageWrapper& image, const Handler& handler) const;
 
  Segments _generateSegments(const ImageWrapper& image) const;

deflect/StreamSendWorker.cpp

@@ -49,6 +49,7 @@
 namespace
 {
 const unsigned int SEGMENT_SIZE = 512;
+const unsigned int SMALL_IMAGE_SIZE = 64;
 }
 
 namespace deflect
@@ -128,7 +129,17 @@ Stream::Future StreamSendWorker::enqueueImage(const ImageWrapper& image,
  return promise.get_future();
  }
 
- auto tasks = std::vector<Task>{[this, image] { return _sendImage(image); }};
+ std::vector<Task> tasks;
+
+ if (image.width <= SMALL_IMAGE_SIZE && image.height <= SMALL_IMAGE_SIZE &&
+ image.compressionPolicy == COMPRESSION_ON)
+ {
+ auto segment = _imageSegmenter.compressSingleSegment(image);
+ tasks.emplace_back([this, segment] { return _sendSegment(segment); });
+ }
+ else
+ tasks.emplace_back([this, image] { return _sendImage(image); });
+
  if (finish)
  tasks.emplace_back([this] { return _sendFinish(); });
 

doc/Changelog.md

@@ -5,6 +5,9 @@ Changelog {#Changelog}
 
 ### 0.13.1 (git master)
 
+* [174](https://github.com/BlueBrain/Deflect/pull/174):
+ Performance improvements for sending small images (tested with 64x64):
+ directly compress them in the call thread
 * [173](https://github.com/BlueBrain/Deflect/pull/173):
  Fix server: disabling system proxy which are default starting Qt 5.8