non-linear regression: fix fitting values very close to c

We have to ignore the values very close to c, since we are regression
the difference with c, not the absolute value, and that difference will
not actually get any smaller.

src/core/perfmodel/regression.c

@@ -20,6 +20,10 @@
 #define MAXREGITER	1000
 #define EPS 1.0e-10
 
+/* For measurements close to C, we do not want to try to fit, since we are
+   fitting the distance to C, which won't actually really get smaller */
+#define C_RADIUS 2
+
 static double compute_b(double c, unsigned n, unsigned *x, double *y)
 {
 	double b;
@@ -29,10 +33,14 @@ static double compute_b(double c, unsigned n, unsigned *x, double *y)
 	double sumx = 0.0;
 	double sumx2 = 0.0;
 	double sumy = 0.0;
+	unsigned nn = 0;
 
 	unsigned i;
 	for (i = 0; i < n; i++)
 	{
+		if ((y[i] - c) / c < C_RADIUS)
+			continue;
+
 		double xi = log(x[i]);
 		double yi = log(y[i]-c);
 
@@ -40,9 +48,11 @@ static double compute_b(double c, unsigned n, unsigned *x, double *y)
 		sumx += xi;
 		sumx2 += xi*xi;
 		sumy += yi;
+
+		nn += 1;
 	}
 
-	b = (n * sumxy - sumx * sumy) / (n*sumx2 - sumx*sumx);
+	b = (nn * sumxy - sumx * sumy) / (nn*sumx2 - sumx*sumx);
 
 	return b;
 }
@@ -54,18 +64,24 @@ static double compute_a(double c, double b, unsigned n, unsigned *x, double *y)
 	/* X = log (x) , Y = log (y - c) */
 	double sumx = 0.0;
 	double sumy = 0.0;
+	unsigned nn = 0;
 
 	unsigned i;
 	for (i = 0; i < n; i++)
 	{
+		if ((y[i] - c) / c < C_RADIUS)
+			continue;
+
 		double xi = log(x[i]);
 		double yi = log(y[i]-c);
 
 		sumx += xi;
 		sumy += yi;
+
+		nn += 1;
 	}
 
-	a = (sumy - b*sumx) / n;
+	a = (sumy - b*sumx) / nn;
 
 	return a;
 }
@@ -85,10 +101,13 @@ static double test_r(double c, unsigned n, unsigned *x, double *y)
 	double sumx2 = 0.0;
 	double sumy = 0.0;
 	double sumy2 = 0.0;
+	unsigned nn = 0;
 
 	unsigned i;
 	for (i = 0; i < n; i++)
 	{
+		if ((y[i] - c) / c < C_RADIUS)
+			continue;
 		double xi = log(x[i]);
 		double yi = log(y[i]-c);
 
@@ -99,6 +118,8 @@ static double test_r(double c, unsigned n, unsigned *x, double *y)
 		sumx2 += xi*xi;
 		sumy += yi;
 		sumy2 += yi*yi;
+
+		nn += 1;
 	}
 
 	//printf("sumxy %e\n", sumxy);
@@ -107,7 +128,7 @@ static double test_r(double c, unsigned n, unsigned *x, double *y)
 	//printf("sumy %e\n", sumy);
 	//printf("sumy2 %e\n", sumy2);
 
-	r = (n * sumxy - sumx * sumy) / sqrt( (n* sumx2 - sumx*sumx) * (n*sumy2 - sumy*sumy) );
+	r = (nn * sumxy - sumx * sumy) / sqrt( (nn* sumx2 - sumx*sumx) * (nn*sumy2 - sumy*sumy) );
 
 	return r;
 }

tests/perfmodels/non_linear_regression_based.c

@@ -21,7 +21,7 @@
  * Benchmark memset with a non-linear regression
  */
 
-#define START_LOG 8
+#define START_LOG 0
 #ifdef STARPU_QUICK_CHECK
 #define END_LOG 20
 #else