KDE/kdeedu/kstars/kstars

[Khudyakov Alexey]

SVN commit 1045844 by khudyakov:

Remove radian caching from dms. It's not really sensible to
cache results of multiplication. It's better to save some space

CCMAIL: kstars-devel@...

 M  +0 -14     dms.cpp  
 M  +7 -11     dms.h  


--- trunk/KDE/kdeedu/kstars/kstars/dms.cpp #1045843:1045844
@@ -27,14 +27,12 @@
 void dms::setD( const double &x ) {
     D = x;
     scDirty = true;
-    rDirty = true;
 }
 
 void dms::setD(const int &d, const int &m, const int &s, const int &ms) {
     D = (double)abs(d) + ((double)m + ((double)s + (double)ms/1000.)/60.)/60.;
     if (d<0) {D = -1.0*D;}
     scDirty = true;
-    rDirty = true;
 }
 
 void dms::setH( const double &x ) {
@@ -45,14 +43,11 @@
     D = 15.0*((double)abs(h) + ((double)m + ((double)s + (double)ms/1000.)/60.)/60.);
     if (h<0) {D = -1.0*D;}
     scDirty = true;
-    rDirty = true;
 }
 
 void dms::setRadians( const double &Rad ) {
     setD( Rad/DegToRad );
-    Radians = Rad;
     scDirty = true;
-    rDirty = true;
 }
 
 bool dms::setFromString( const QString &str, bool isDeg ) {
@@ -254,15 +249,6 @@
     cosa = Cos;
 }
 
-const double& dms::radians( void ) const {
-    if ( rDirty ) {
-        Radians = D*DegToRad;
-        rDirty = false;
-    }
-
-    return Radians;
-}
-
 const dms dms::reduce( void ) const {
     double a = D;
     while (a<0.0) {a += 360.0;}
--- trunk/KDE/kdeedu/kstars/kstars/dms.h #1045843:1045844
@@ -279,7 +279,7 @@
      *@return the Sine of the angle.
      *@sa cos()
      */
-    const double& sin( void ) const;
+    const double& sin() const;
 
     /**@short Compute the Angle's Cosine.
      *
@@ -289,15 +289,12 @@
      *@return the Cosine of the angle.
      *@sa sin()
      */
-    const double& cos( void ) const;
+    const double& cos() const;
 
     /**@short Express the angle in radians.
-     *The computed Radians value is stored internally.  On subsequent calls,
-     *the stored value is returned directly (unless the angle's value has
-     *changed).
-     *@return the angle in radians (double)
-     */
-    const double& radians( void ) const;
+     * @return the angle in radians (double)
+     */
+    double radians() const { return D*DegToRad; }
 
     /**@short Set angle according to the argument, in radians.
      *
@@ -310,7 +307,7 @@
     /**return the equivalent angle between 0 and 360 degrees.
      *@warning does not change the value of the parent angle itself.
      */
-    const dms reduce( void ) const;
+    const dms reduce() const;
 
     /**@return a nicely-formatted string representation of the angle
      *in degrees, arcminutes, and arcseconds.
@@ -351,9 +348,8 @@
 private:
     double D;
 
-    mutable double Radians;
     mutable double Sin, Cos;
-    mutable bool scDirty, rDirty;
+    mutable bool scDirty;
 };
 
 #endif
_______________________________________________
Kstars-devel mailing list
Kstars-devel@...
https://mail.kde.org/mailman/listinfo/kstars-devel

[Khudyakov Alexey]

В сообщении от 07 ноября 2009 00:13:21 Alexey Khudyakov написал:
> SVN commit 1045844 by khudyakov:
>
> Remove radian caching from dms. It's not really sensible to
> cache results of multiplication. It's better to save some space
>
Follow up on DMS caching

Here nice plot showing caching efficiency as function of time. Efficiency is
defined as ratio of all cache hits to all cache misses so far. I made two
measurements. In 'idle' I simply left KStars running for few minutes. In
'active' I dragged sky around, zoomed in/out.

Ratio is quite low, it's in 0.5-0.6 range. Or in other words only about 1 call
in 3 is cached. This leads me to question efficiency of sin/cos caching.

In attachment plot and patch used for measurements


[cache_efficiency.patch]

---

diff --git a/kstars/kstars/dms.cpp b/kstars/kstars/dms.cpp
index 875a46b..66b4d21 100644
--- a/kstars/kstars/dms.cpp
+++ b/kstars/kstars/dms.cpp
@@ -24,6 +24,15 @@
 #include <kglobal.h>
 #include <klocale.h>
 
+#include <QTime>
+namespace {
+    int cacheHit  = 0;
+    int cacheMiss = 0;
+    int cnt = 0;
+
+    QTime t0;
+}
+
 void dms::setD( const double &x ) {
     D = x;
     scDirty = true;
@@ -212,8 +221,9 @@ const double& dms::sin( void ) const {
     if ( scDirty ) {
         double s,c;
         SinCos( s, c );
+    } else {
+        cacheHit++;
     }
-
     return Sin;
 }
 
@@ -221,8 +231,9 @@ const double& dms::cos( void ) const {
     if ( scDirty ) {
         double s,c;
         SinCos( s, c );
+    } else {
+        cacheHit++;
     }
-
     return Cos;
 }
 
@@ -233,6 +244,8 @@ void dms::SinCos( double &sina, double &cosa ) const {
      * GNU extension sincos() is defined.
      */
 
+    if( !t0.isValid() )
+        t0 = QTime::currentTime();
     if ( scDirty ) {
                 #ifdef __GLIBC__
                 #if ( __GLIBC__ >= 2 && __GLIBC_MINOR__ >=1 && !defined(__UCLIBC__))
@@ -249,6 +262,15 @@ void dms::SinCos( double &sina, double &cosa ) const {
         Cos = ::cos( radians() );
  #endif
         scDirty = false;
+        cacheMiss++;
+    } else {
+        cacheHit++;
+    }
+    if( cnt++ == 5000 ) {
+        cnt = 0;
+        kDebug() << QString("t=%1, hit=%2, miss=%3, ratio=%4").
+            arg( t0.msecsTo( QTime::currentTime() ) ).
+            arg(cacheHit).arg(cacheMiss).arg(((double)cacheHit)/cacheMiss);
     }
     sina = Sin;
     cosa = Cos;

---

_______________________________________________
Kstars-devel mailing list
Kstars-devel@...
https://mail.kde.org/mailman/listinfo/kstars-devel

[Khudyakov Alexey]

В сообщении от 07 ноября 2009 16:31:35 Khudyakov Alexey написал:

> В сообщении от 07 ноября 2009 00:13:21 Alexey Khudyakov написал:
> > SVN commit 1045844 by khudyakov:
> >
> > Remove radian caching from dms. It's not really sensible to
> > cache results of multiplication. It's better to save some space
>
> Follow up on DMS caching
>
> Here nice plot showing caching efficiency as function of time. Efficiency
>  is defined as ratio of all cache hits to all cache misses so far. I made
>  two measurements. In 'idle' I simply left KStars running for few minutes.
>  In 'active' I dragged sky around, zoomed in/out.
>
> Ratio is quite low, it's in 0.5-0.6 range. Or in other words only about 1
>  call in 3 is cached. This leads me to question efficiency of sin/cos
>  caching.
>

Indeed profiling shows that in case of dms caching is pessimisation. When I
replaced current caching version of sincos with inlined call to sincos it
become about 1.5 times faster. So I propose to remove caching.

Also it decrease size of dms to size of double and make it cheap to
pass/return it by value.
_______________________________________________
Kstars-devel mailing list
Kstars-devel@...
https://mail.kde.org/mailman/listinfo/kstars-devel