Fix the two piecewise-linear regression calculation

#66

The current implementation of the regression calculation has these flaws:

When processing (x[0], y[0]), L1 must be any line through (x[0], y[0]) which meets
L2 at a point (x’, y’) where x[0] < x' < x[1]. L1 has no error.

When processing (x[n - 2], y[n - 2]), L2 must be any line through (x[n - 1], y[n - 1])
which meets L1 at a point (x’, y’) where x[n - 2] < x' < x[n - 1]. L2 has no error.

The lambda calculation is incorrect. It includes a term called H which is equal
to C - I. Looking at the algorithm of Kundu/Ubhaya, this should be just C.

lambda should to be used with calculating L1 and (1 - lambda) should to be used
with calculating L2. Currently (1 - lambda) is used in calculating L1 and L2.

The current calculation has this condition if (t1 != t2) continue; This condition
is almost always true even if t1 and t2 are essentiallyEqual.

Author: shallawa

MotionMark/resources/runner/motionmark.js

@@ -162,11 +162,11 @@
         var calculation = regressionResult.regression;
         result[Strings.json.complexity] = {};
         result[Strings.json.complexity][Strings.json.regressions.segment1] = [
-            [regressionResult.minComplexity, calculation.s1 + calculation.t1 * regressionResult.minComplexity],
-            [calculation.complexity, calculation.s1 + calculation.t1 * calculation.complexity]
+            [regressionResult.minComplexity, desiredFrameLength],
+            [calculation.complexity, desiredFrameLength]
         ];
         result[Strings.json.complexity][Strings.json.regressions.segment2] = [
-            [calculation.complexity, calculation.s2 + calculation.t2 * calculation.complexity],
+            [calculation.complexity, desiredFrameLength],
             [regressionResult.maxComplexity, calculation.s2 + calculation.t2 * regressionResult.maxComplexity]
         ];
         result[Strings.json.complexity][Strings.json.complexity] = calculation.complexity;

MotionMark/resources/statistics.js

@@ -195,6 +195,8 @@ Regression = Utilities.createClass(
             e: Number.MAX_VALUE
         };
 
+        this.complexity = 0;
+
         if (options.preferredProfile == Strings.json.profiles.flat) {
             this._calculateRegression(samples, {
                 s1: options.desiredFrameLength,
@@ -213,7 +215,6 @@ Regression = Utilities.createClass(
         this.startIndex = Math.min(startIndex, endIndex);
         this.endIndex = Math.max(startIndex, endIndex);
 
-        this.complexity = this._complexity();
         this.s1 = this.segment1.s;
         this.t1 = this.segment1.t;
         this.s2 = this.segment2.s;
@@ -242,10 +243,30 @@ Regression = Utilities.createClass(
         return this.segment1.e + this.segment2.e;
     },
 
-    _setOptimal: function(segment1, segment2, options) {
+    _areEssentiallyEqual: function(n1, n2) {
+        const epsilon = 0.001;
+        return Math.abs(n1 - n2) < epsilon;
+    },
+
+    _setOptimal: function(segment1, segment2, xp, x, options) {
         if (segment1.e + segment2.e > this.segment1.e + this.segment2.e)
             return false;
 
+        let complexity = this._complexity();
+        if (!this._areEssentiallyEqual(this.segment1.t, this.segment2.t)) {
+            // If segment1 and segment2 are not parallel, then they have to meet
+            // at complexity such that xp < complexity < x.
+            if (!(complexity >= xp && complexity <= x))
+                return false;
+        } else {
+            // If segment1 and segment2 are parallel, then they have to form one
+            // single line.
+            if (!this._areEssentiallyEqual(this.segment1.s, this.segment2.s))
+                return false;
+        }
+
+        this.complexity = complexity;
+
         this.segment1.s = options.s1 !== undefined ? options.s1 : segment1.s;
         this.segment1.t = options.t1 !== undefined ? options.t1 : segment1.t;
         this.segment1.e = segment1.e;
@@ -330,9 +351,10 @@ Regression = Utilities.createClass(
 
             let segment1;
             let segment2;
+            let xp = (j == 0) ? 0 : sortedSamples[j - 1][complexityIndex];
 
             if (j == 0) {
-                // Let segment1 be any line through (x0, y0) which meets segment2 at
+                // Let segment1 be any line through (x[0], y[0]) which meets segment2 at
                 // a point (x’, y’) where x[0] < x' < x[1]. segment1 has no error.
                 let xMid = (x + sortedSamples[j + 1][complexityIndex]) / 2;
                 let yMid = s2 + t2 * xMid;
@@ -373,16 +395,18 @@ Regression = Utilities.createClass(
                 };
             }
 
-            if (this._setOptimal(segment1, segment2, options))
+            if (this._setOptimal(segment1, segment2, xp, x, options))
                 continue
 
+            // These values remove the influence of this sample
             let G = A + B * x - C * y;
-            let H = D + E * x - F * y;
+            let J = D + E * x - F * y;
 
             let I = c1 - 2 * b1 * x + a1 * xx;
             let K = c2 - 2 * b2 * x + a2 * xx;
 
-            let lambda = (G * F + G * K - H * C) / (I * H + G * K);
+            // Calculate lambda, which divides the weight of this sample between the two lines
+            let lambda = (G * F + G * K - J * C) / (I * J + G * K);
             if (!(lambda > 0 && lambda < 1))
                 continue;
 
@@ -402,7 +426,7 @@ Regression = Utilities.createClass(
                 e: (k2 + a2 * s2 * s2 + c2 * t2 * t2 - 2 * d2 * s2 - 2 * h2 * t2 + 2 * b2 * s2 * t2) + lambda1 * Math.pow(y - (s2 + t2 * x), 2)
             };
 
-            this._setOptimal(segment1, segment2, options);
+            this._setOptimal(segment1, segment2, xp, x, options);
         }
     }
 });