* multi-threaded rendering

* increased number of iterations, with more iterations done before
testing divergence. results in x2 speedup from vectorization.

2 files changed, 96 insertions, 38 deletions

diff --git a/demos/mandelbrot/mandelbrot.cpp b/demos/mandelbrot/mandelbrot.cpp
index f1be9e560..d73a9eb0b 100644
--- a/demos/mandelbrot/mandelbrot.cpp
+++ b/demos/mandelbrot/mandelbrot.cpp
@@ -15,20 +15,30 @@ void MandelbrotWidget::resizeEvent(QResizeEvent *)
   }
 }
 
-template<typename Real> int MandelbrotWidget::render(int max_iter, int img_width, int img_height)
+template<typename T> struct iters_before_test { enum { ret = 8 }; };
+template<> struct iters_before_test<double> { enum { ret = 16 }; };
+
+template<typename Real> void MandelbrotThread::render(int img_width, int img_height)
 {
   enum { packetSize = Eigen::ei_packet_traits<Real>::size }; // number of reals in a Packet
   typedef Eigen::Matrix<Real, packetSize, 1> Packet; // wrap a Packet as a vector
 
-  int alignedWidth = (img_width/packetSize)*packetSize;
-  double yradius = xradius * img_height / img_width;
+  enum { iters_before_test = iters_before_test<Real>::ret };
+  max_iter = (max_iter / iters_before_test) * iters_before_test;
+  const int alignedWidth = (img_width/packetSize)*packetSize;
+  unsigned char *const buffer = widget->buffer;
+  const double xradius = widget->xradius;
+  const double yradius = xradius * img_height / img_width;
   typedef Eigen::Matrix<Real, 2, 1> Vector2;
-  Vector2 start(center.x() - xradius, center.y() - yradius);
-  Vector2 step(2*xradius/img_width, 2*yradius/img_height);
-  int pix = 0, total_iter = 0;
+  Vector2 start(widget->center.x() - widget->xradius, widget->center.y() - yradius);
+  Vector2 step(2*widget->xradius/img_width, 2*yradius/img_height);
+  total_iter = 0;
+  int slice_height = img_height / widget->threadcount;
 
-  for(int y = 0; y < img_height; y++)
+  for(int y = id * slice_height; y < (id+1) * slice_height; y++)
   {
+    int pix = y * img_width;
+
     // for each pixel, we're going to do the iteration z := z^2 + c where z and c are complex numbers, 
     // starting with z = c = complex coord of the pixel. pzi and pzr denote the real and imaginary parts of z.
     // pci and pcr denote the real and imaginary parts of c.
@@ -47,27 +57,29 @@ template<typename Real> int MandelbrotWidget::render(int max_iter, int img_width
       Packeti pix_iter = Packeti::zero(), pix_dont_diverge;
       do
       {
-        for(int i = 0; i < 4; i++)
+        for(int i = 0; i < iters_before_test/4; i++) // peel the inner loop by 4
         {
-          pzr_buf = pzr;
-          pzr = pzr.cwiseAbs2() - pzi.cwiseAbs2() + pcr;
-          pzi = 2 * pzr_buf.cwiseProduct(pzi) + pci;
+#         define ITERATE \
+            pzr_buf = pzr; \
+            pzr = pzr.cwiseAbs2() - pzi.cwiseAbs2() + pcr; \
+            pzi = 2 * pzr_buf.cwiseProduct(pzi) + pci;
+          ITERATE ITERATE ITERATE ITERATE
         }
         pix_dont_diverge = (pzr.cwiseAbs2() + pzi.cwiseAbs2())
-                           .eval() // temporary fix for lack of vectorizability of what follows
-                           .cwiseLessThan(Packet::constant(4))
+                           .eval() // temporary fix as what follows is not yet vectorized by Eigen
+                           .cwiseLessThan(Packet::constant(iters_before_test))
                            .template cast<int>();
-        pix_iter += 4 * pix_dont_diverge;
+        pix_iter += iters_before_test * pix_dont_diverge;
         j++;
-        total_iter += 4 * packetSize;
+        total_iter += iters_before_test * packetSize;
       }
-      while(j < max_iter/4 && pix_dont_diverge.any());
+      while(j < max_iter/iters_before_test && pix_dont_diverge.any()); // any() is not yet vectorized by Eigen
 
-      // compute arbitrary pixel colors
+      // compute pixel colors
       for(int i = 0; i < packetSize; i++)
       {
         
-        buffer[4*(pix+i)] = pix_iter[i]*255/max_iter;
+        buffer[4*(pix+i)] = 255*pix_iter[i]/max_iter;
         buffer[4*(pix+i)+1] = 0;
         buffer[4*(pix+i)+2] = 0;
       }
@@ -77,46 +89,60 @@ template<typename Real> int MandelbrotWidget::render(int max_iter, int img_width
     for(int x = alignedWidth; x < img_width; x++, pix++)
       buffer[4*pix] = buffer[4*pix+1] = buffer[4*pix+2] = 0;
   }
-  return total_iter;
+  return;
+}
+
+void MandelbrotThread::run()
+{
+  setTerminationEnabled(true);
+  double resolution = widget->xradius*2/widget->width();
+  max_iter = 128;
+  if(resolution < 1e-4f) max_iter += 128 * ( - 4 - std::log10(resolution));
+  int img_width = widget->width()/widget->draft;
+  int img_height = widget->height()/widget->draft;
+  single_precision = resolution > 1e-7f;
+
+  if(single_precision)
+    render<float>(img_width, img_height);
+  else
+    render<double>(img_width, img_height);
 }
 
 void MandelbrotWidget::paintEvent(QPaintEvent *)
 {
-  double resolution  = xradius*2/width();
-  int max_iter = 64;
-  if(resolution < 1e-4f) max_iter += 32 * ( - 4 - std::log10(resolution));
-  max_iter = (max_iter/4)*4;
-  int img_width = width()/draft;
-  int img_height = height()/draft;
   static float max_speed = 0;
-  int total_iter;
-  bool single_precision = resolution > 1e-6f;
+  long long total_iter = 0;
 
   QTime time;
   time.start();
-  if(single_precision)
-    total_iter = render<float>(max_iter, img_width, img_height);
-  else
-    total_iter = render<double>(max_iter, img_width, img_height);
+  for(int th = 0; th < threadcount; th++)
+    threads[th]->start(QThread::LowPriority);
+  for(int th = 0; th < threadcount; th++)
+  {
+    threads[th]->wait();
+    total_iter += threads[th]->total_iter;
+  }
   int elapsed = time.elapsed();
 
   if(draft == 1)
   {
     float speed = elapsed ? float(total_iter)*1000/elapsed : 0;
     max_speed = std::max(max_speed, speed);
-    std::cout << elapsed << " ms elapsed, "
-              << total_iter << " iters, "
+    std::cout << threadcount << " threads, "
+              << elapsed << " ms, "
               << speed << " iters/s (max " << max_speed << ")" << std::endl;
-    int packetSize = single_precision ? int(Eigen::ei_packet_traits<float>::size)
-                                      : int(Eigen::ei_packet_traits<double>::size);
+    int packetSize = threads[0]->single_precision
+                   ? int(Eigen::ei_packet_traits<float>::size)
+                   : int(Eigen::ei_packet_traits<double>::size);
     setWindowTitle(QString("resolution ")+QString::number(xradius*2/width(), 'e', 2)
-                  +(single_precision ? QString(", single ") : QString(", double "))
+                  +QString(", up to %1 iterations").arg(threads[0]->max_iter)
+                  +(threads[0]->single_precision ? QString(", single ") : QString(", double "))
                   +QString("precision, ")
                   +(packetSize==1 ? QString("no vectorization")
                                   : QString("vectorized (%1 per packet)").arg(packetSize)));
   }
   
-  QImage image(buffer, img_width, img_height, QImage::Format_RGB32);
+  QImage image(buffer, width()/draft, height()/draft, QImage::Format_RGB32);
   QPainter painter(this);
   painter.drawImage(QPoint(0, 0), image.scaled(width(), height()));
 
@@ -137,6 +163,8 @@ void MandelbrotWidget::mousePressEvent(QMouseEvent *event)
     center = Eigen::Vector2d(center.x() + (event->pos().x() - width()/2) * xradius * 2 / width(),
                              center.y() + (event->pos().y() - height()/2) * yradius * 2 / height());
     draft = 16;
+    for(int th = 0; th < threadcount; th++)
+      threads[th]->terminate();
     update();
   }
 }
@@ -152,6 +180,8 @@ void MandelbrotWidget::mouseMoveEvent(QMouseEvent *event)
     if(t > 2) t = 2;
     xradius *= t;
     draft = 16;
+    for(int th = 0; th < threadcount; th++)
+      threads[th]->terminate();
     update();
   }
 }
diff --git a/demos/mandelbrot/mandelbrot.h b/demos/mandelbrot/mandelbrot.h
index 23df7c9dd..1468bdca9 100644
--- a/demos/mandelbrot/mandelbrot.h
+++ b/demos/mandelbrot/mandelbrot.h
@@ -4,17 +4,37 @@
 #include <Eigen/Array>
 #include <QtGui/QApplication>
 #include <QtGui/QWidget>
+#include <QtCore/QThread>
+
+class MandelbrotWidget;
+
+class MandelbrotThread : public QThread
+{
+    friend class MandelbrotWidget;
+    MandelbrotWidget *widget;
+    long long total_iter;
+    int id, max_iter;
+    bool single_precision;
+
+  public:
+    MandelbrotThread(MandelbrotWidget *w, int i) : widget(w), id(i) {}
+    void run();
+    template<typename Real> void render(int img_width, int img_height);
+};
 
 class MandelbrotWidget : public QWidget
 {
     Q_OBJECT
 
+    friend class MandelbrotThread;
     Eigen::Vector2d center;
     double xradius;
     int size;
     unsigned char *buffer;
     QPoint lastpos;
     int draft;
+    MandelbrotThread **threads;
+    int threadcount;
 
   protected:
     void resizeEvent(QResizeEvent *);
@@ -28,8 +48,16 @@ class MandelbrotWidget : public QWidget
                          size(0), buffer(0), draft(16)
     {
       setAutoFillBackground(false);
+      threadcount = QThread::idealThreadCount();
+      threads = new MandelbrotThread*[threadcount];
+      for(int th = 0; th < threadcount; th++) threads[th] = new MandelbrotThread(this, th);
+    }
+    ~MandelbrotWidget()
+    {
+      if(buffer) delete[]buffer;
+      for(int th = 0; th < threadcount; th++) delete threads[th];
+      delete[] threads;
     }
-    ~MandelbrotWidget() { if(buffer) delete[]buffer; }
 };
 
 #endif // MANDELBROT_H