massive performance improvements. zrle measured at 2.5x faster (no measurements for tight, but should be the same or better).

git-svn-id: svn://svn.code.sf.net/p/tigervnc/code/trunk@4742 3789f03b-4d11-0410-bbf8-ca57d06f2519

6 files changed, 60 insertions, 90 deletions

diff --git a/java/com/tigervnc/rdr/InStream.java b/java/com/tigervnc/rdr/InStream.java
index 803f55f6..60967991 100644
--- a/java/com/tigervnc/rdr/InStream.java
+++ b/java/com/tigervnc/rdr/InStream.java
@@ -130,20 +130,30 @@ abstract public class InStream {
                                      int b1 = b[ptr++]; int b2 = b[ptr++];
                                      return b0 << 16 | b1 << 8 | b2; }
 
-  public final int readPixel(int bytesPerPixel, boolean e) {
-    int[] pix = new int[4];
-    for (int i=0; i < bytesPerPixel; i++)
-      pix[i] = readU8();
-    if (e) {
-      return pix[0] << 16 | pix[1] << 8 | pix[2] | (0xff << 24);
+  public final int readPixel(int bytesPerPixel, boolean bigEndian) {
+    byte[] pix = new byte[4];
+    readBytes(pix, 0, bytesPerPixel);
+
+    if (bigEndian) {
+      return 0xff000000 | (pix[0] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[2] & 0xff);
     } else {
-      return pix[2] << 16 | pix[1] << 8 | pix[0] | (0xff << 24);
+      return 0xff000000 | (pix[2] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[0] & 0xff);
     }
   }
 
-  public final void readPixels(int[] buf, int length, int bytesPerPixel, boolean e) {
-    for (int i = 0; i < length; i++)
-      buf[i] = readPixel(bytesPerPixel, e);
+  public final void readPixels(int[] buf, int length, int bytesPerPixel, boolean bigEndian) {
+    int npixels = length*bytesPerPixel;
+    byte[] pixels = new byte[npixels];
+    readBytes(pixels, 0, npixels);
+    for (int i = 0; i < length; i++) {
+      byte[] pix = new byte[4];
+      System.arraycopy(pixels, i*bytesPerPixel, pix, 0, bytesPerPixel);
+      if (bigEndian) {
+        buf[i] = 0xff000000 | (pix[0] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[2] & 0xff);
+      } else {
+        buf[i] = 0xff000000 | (pix[2] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[0] & 0xff);
+      }
+    }
   }
 
   public final int readCompactLength() {
diff --git a/java/com/tigervnc/rdr/JavaInStream.java b/java/com/tigervnc/rdr/JavaInStream.java
index 5c350369..d2eda520 100644
--- a/java/com/tigervnc/rdr/JavaInStream.java
+++ b/java/com/tigervnc/rdr/JavaInStream.java
@@ -32,6 +32,7 @@ public class JavaInStream extends InStream {
     bufSize = bufSize_;
     b = new byte[bufSize];
     ptr = end = offset = 0;
+    timing = false;
     timeWaitedIn100us = 5;
     timedKbits = 0;
   }
@@ -108,12 +109,17 @@ public class JavaInStream extends InStream {
   }
 
   private int read(byte[] buf, int bufPtr, int len, boolean wait) {
-    try {
       long before = 0;
       if (timing)
         before = System.nanoTime();
 
-      int n = jis.read(buf, bufPtr, len);
+      int n = -1;
+      try {
+        n = jis.read(buf, bufPtr, len);
+      } catch (java.io.IOException e) {
+        throw new IOException(e);
+      }
+
       if (n < 0) throw new EndOfStream();
 
       if (timing) {
@@ -123,8 +129,11 @@ public class JavaInStream extends InStream {
 
         // limit rate to between 10kbit/s and 40Mbit/s
 
-        if (newTimeWaited > newKbits*1000) newTimeWaited = newKbits*1000;
-        if (newTimeWaited < newKbits/4)    newTimeWaited = newKbits/4;
+        if (newTimeWaited > newKbits*1000) {
+          newTimeWaited = newKbits*1000;
+        } else if (newTimeWaited < newKbits/4) {
+          newTimeWaited = newKbits/4;
+        }
 
         timeWaitedIn100us += newTimeWaited;
         timedKbits += newKbits;
@@ -132,9 +141,6 @@ public class JavaInStream extends InStream {
 
       return n;
 
-    } catch (java.io.IOException e) {
-      throw new IOException(e);
-    }
   }
   private int read(byte[] buf, int bufPtr, int len) { return read(buf, bufPtr, len, true); }
 
diff --git a/java/com/tigervnc/rfb/RREDecoder.java b/java/com/tigervnc/rfb/RREDecoder.java
index e0ff5cfe..f15ac109 100644
--- a/java/com/tigervnc/rfb/RREDecoder.java
+++ b/java/com/tigervnc/rfb/RREDecoder.java
@@ -38,7 +38,7 @@ public class RREDecoder extends Decoder {
       int y = is.readU16();
       int w = is.readU16();
       int h = is.readU16();
-      handler.fillRect(new Rect(r.tl.x+x, r.tl.y+y, r.tl.x+x+w, r.tl.y+y+h), pix);
+      handler.fillRect(new Rect(x, y, w, h), pix);
     }
   }
 
diff --git a/java/com/tigervnc/rfb/Rect.java b/java/com/tigervnc/rfb/Rect.java
index fab4f5dd..4aa69c14 100644
--- a/java/com/tigervnc/rfb/Rect.java
+++ b/java/com/tigervnc/rfb/Rect.java
@@ -76,7 +76,12 @@ public class Rect {
   public final boolean overlaps(Rect r) {
     return tl.x < r.br.x && tl.y < r.br.y && br.x > r.tl.x && br.y > r.tl.y;
   }
-  public final int area() {return is_empty() ? 0 : (br.x-tl.x)*(br.y-tl.y);}
+  public final int area() {
+    int area = (br.x-tl.x)*(br.y-tl.y);
+    if (area > 0)
+      return area;
+    return 0;
+  }
   public final Point dimensions() {return new Point(width(), height());}
   public final int width() {return br.x-tl.x;}
   public final int height() {return br.y-tl.y;}
diff --git a/java/com/tigervnc/rfb/TightDecoder.java b/java/com/tigervnc/rfb/TightDecoder.java
index 24828fbb..2b97110d 100644
--- a/java/com/tigervnc/rfb/TightDecoder.java
+++ b/java/com/tigervnc/rfb/TightDecoder.java
@@ -78,15 +78,7 @@ public class TightDecoder extends Decoder {
     if (comp_ctl == rfbTightFill) {
       int pix;
       if (cutZeros) {
-        byte[] elem = new byte[3];
-        is.readBytes(elem, 0, 3);
-        if (bigEndian) {
-          pix =
-            (elem[2] & 0xFF) << 16 | (elem[1] & 0xFF) << 8 | (elem[0] & 0xFF) | 0xFF << 24;
-        } else {
-          pix =
-            (elem[0] & 0xFF) << 16 | (elem[1] & 0xFF) << 8 | (elem[2] & 0xFF) | 0xFF << 24;
-        }
+        pix = is.readPixel(3, false);
       } else {
         pix = (bpp == 8) ? is.readOpaque8() : is.readOpaque24B();
       }
@@ -135,17 +127,7 @@ public class TightDecoder extends Decoder {
       case rfbTightFilterPalette:
         palSize = is.readU8() + 1;
         if (cutZeros) {
-          byte[] elem = new byte[3];
-          for (int i = 0; i < palSize; i++) {
-            is.readBytes(elem, 0, 3);
-            if (bigEndian) {
-              palette[i] =
-                (elem[2] & 0xFF) << 16 | (elem[1] & 0xFF) << 8 | (elem[0] & 0xFF) | 0xFF << 24;
-            } else {
-              palette[i] =
-                (elem[0] & 0xFF) << 16 | (elem[1] & 0xFF) << 8 | (elem[2] & 0xFF) | 0xFF << 24;
-            }
-          }
+          is.readPixels(palette, palSize, 3, false);
         } else {
           for (int i = 0; i < palSize; i++) {
             palette[i] = (bpp == 8) ? is.readOpaque8() : is.readOpaque24B();
@@ -196,17 +178,7 @@ public class TightDecoder extends Decoder {
         }
       } else {
         if (cutZeros) {
-          byte[] elem = new byte[3];
-          for (int i = 0; i < r.area(); i++) {
-            input.readBytes(elem, 0, 3);
-            if (bigEndian) {
-              buf[i] =
-                (elem[2] & 0xFF) << 16 | (elem[1] & 0xFF) << 8 | (elem[0] & 0xFF) | 0xFF << 24;
-            } else {
-              buf[i] =
-                (elem[0] & 0xFF) << 16 | (elem[1] & 0xFF) << 8 | (elem[2] & 0xFF) | 0xFF << 24;
-            }
-          }
+          input.readPixels(buf, r.area(), 3, false);
         } else {
           for (int ptr=0; ptr < dataSize; ptr++)
             buf[ptr] = input.readU8();
@@ -264,14 +236,12 @@ public class TightDecoder extends Decoder {
     int[] thisRow = new int[TIGHT_MAX_WIDTH * 3];
     int[] pix = new int[3];
     int[] est = new int[3];
+    PixelFormat myFormat = handler.cp.pf();
 
     // Allocate netbuf and read in data
     int[] netbuf = new int[dataSize];
-    for (int i = 0; i < dataSize; i++)
-      netbuf[i] = is.readU8();
-    //is.readBytes(netbuf, 0, dataSize);
+    is.readBytes(netbuf, 0, dataSize);
 
-    PixelFormat myFormat = handler.cp.pf();
     int rectHeight = r.height();
     int rectWidth = r.width();
 
@@ -282,11 +252,9 @@ public class TightDecoder extends Decoder {
         thisRow[c] = pix[c];
       }
       if (myFormat.bigEndian) {
-        buf[y*rectWidth] =
-          (pix[0] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[2] & 0xFF) | 0xFF << 24;
+        buf[y*rectWidth] =  0xff000000 | (pix[2] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[0] & 0xff);
       } else {
-        buf[y*rectWidth] =
-          (pix[2] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[0] & 0xFF) | 0xFF << 24;
+        buf[y*rectWidth] =  0xff000000 | (pix[0] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[2] & 0xff);
       }
 
       /* Remaining pixels of a row */
@@ -302,11 +270,9 @@ public class TightDecoder extends Decoder {
           thisRow[x*3+c] = pix[c];
         }
         if (myFormat.bigEndian) {
-          buf[y*rectWidth] =
-            (pix[2] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[0] & 0xFF) | 0xFF << 24;
+          buf[y*rectWidth+x] =  0xff000000 | (pix[2] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[0] & 0xff);
         } else {
-          buf[y*rectWidth] =
-            (pix[0] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[2] & 0xFF) | 0xFF << 24;
+          buf[y*rectWidth+x] =  0xff000000 | (pix[0] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[2] & 0xff);
         }
       }
 
@@ -321,25 +287,21 @@ public class TightDecoder extends Decoder {
     int[] thisRow = new int[TIGHT_MAX_WIDTH];
     int[] pix = new int[3];
     int[] est = new int[3];
+    PixelFormat myFormat = handler.cp.pf();
 
     // Allocate netbuf and read in data
     int[] netbuf = new int[dataSize];
-    for (int i = 0; i < dataSize; i++)
-      netbuf[i] = is.readU8();
-    //is.readBytes(netbuf, 0, dataSize);
+    is.readBytes(netbuf, 0, dataSize);
 
-    PixelFormat myFormat = handler.cp.pf();
     int rectHeight = r.height();
     int rectWidth = r.width();
 
     for (y = 0; y < rectHeight; y++) {
       /* First pixel in a row */
       if (myFormat.bigEndian) {
-        buf[y*rectWidth] =
-          (pix[2] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[0] & 0xFF) | 0xFF << 24;
+        buf[y*rectWidth] =  0xff000000 | (pix[2] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[0] & 0xff);
       } else {
-        buf[y*rectWidth] =
-          (pix[0] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[2] & 0xFF) | 0xFF << 24;
+        buf[y*rectWidth] =  0xff000000 | (pix[0] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[2] & 0xff);
       }
       for (c = 0; c < 3; c++)
         pix[c] += prevRow[c];
@@ -363,13 +325,10 @@ public class TightDecoder extends Decoder {
         System.arraycopy(thisRow, x*3, pix, 0, pix.length);
 
         if (myFormat.bigEndian) {
-          buf[y*rectWidth+x] =
-            (pix[2] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[0] & 0xFF) | 0xFF << 24;
+          buf[y*rectWidth+x] =  0xff000000 | (pix[2] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[0] & 0xff);
         } else {
-          buf[y*rectWidth+x] =
-            (pix[0] & 0xFF) << 16 | (pix[1] & 0xFF) << 8 | (pix[2] & 0xFF) | 0xFF << 24;
+          buf[y*rectWidth+x] =  0xff000000 | (pix[0] & 0xff)<<16 | (pix[1] & 0xff)<<8 | (pix[2] & 0xff);
         }
-
       }
 
       System.arraycopy(thisRow, 0, prevRow, 0, prevRow.length);
diff --git a/java/com/tigervnc/rfb/ZRLEDecoder.java b/java/com/tigervnc/rfb/ZRLEDecoder.java
index 4f740f9d..f9a44c0f 100644
--- a/java/com/tigervnc/rfb/ZRLEDecoder.java
+++ b/java/com/tigervnc/rfb/ZRLEDecoder.java
@@ -30,7 +30,8 @@ public class ZRLEDecoder extends Decoder {
   public void readRect(Rect r, CMsgHandler handler) {
     InStream is = reader.getInStream();
     int[] buf = reader.getImageBuf(64 * 64 * 4);
-    int bytesPerPixel = handler.cp.pf().bpp / 8;
+    int bpp = handler.cp.pf().bpp;
+    int bytesPerPixel = (bpp > 24 ? 3 : bpp / 8);
     boolean bigEndian = handler.cp.pf().bigEndian;
 
     int length = is.readU32();
@@ -50,13 +51,7 @@ public class ZRLEDecoder extends Decoder {
         int palSize = mode & 127;
         int[] palette = new int[128];
 
-        if (bytesPerPixel > 1) {
-          zis.readPixels(palette, palSize, 3, bigEndian);
-        } else {
-          for (int i = 0; i < palSize; i++) {
-            palette[i] = zis.readPixel(bytesPerPixel, bigEndian);
-          }
-        }
+        zis.readPixels(palette, palSize, bytesPerPixel, bigEndian);
         
         if (palSize == 1) {
           int pix = palette[0];
@@ -69,11 +64,7 @@ public class ZRLEDecoder extends Decoder {
 
             // raw
 
-            if (bytesPerPixel > 1) {
-              zis.readPixels(buf, t.area(), 3, bigEndian);
-            } else {
-              zis.readPixels(buf, t.area(), bytesPerPixel, bigEndian);
-            }
+            zis.readPixels(buf, t.area(), bytesPerPixel, bigEndian);
 
           } else {
 
@@ -109,8 +100,7 @@ public class ZRLEDecoder extends Decoder {
             int ptr = 0;
             int end = ptr + t.area();
             while (ptr < end) {
-              int pix = (bytesPerPixel > 1 ? zis.readPixel(3, bigEndian) : 
-                zis.readPixel(bytesPerPixel, bigEndian));
+              int pix = zis.readPixel(bytesPerPixel, bigEndian);
               int len = 1;
               int b;
               do {