Dramatically improve "Fast" linear-RGB conversions.

This is following the discussion in #18.
Documentation and a notebook for deriving the constants will follow.

colors.go

@@ -333,11 +333,23 @@ func (col Color) LinearRgb() (r, g, b float64) {
     return
 }
 
+// A much faster and still quite precise linearization using a 6th-order Taylor approximation.
+// See the accompanying Jupyter notebook for derivation of the constants.
+func linearize_fast(v float64) float64 {
+    v1 := v - 0.5
+    v2 := v1*v1
+    v3 := v2*v1
+    v4 := v2*v2
+    //v5 := v3*v2
+    return -0.248750514614486 + 0.925583310193438*v + 1.16740237321695*v2 + 0.280457026598666*v3 - 0.0757991963780179*v4 //+ 0.0437040411548932*v5
+}
+
 // FastLinearRgb is much faster than and almost as accurate as LinearRgb.
+// BUT it is important to NOTE that they only produce good results for valid colors r,g,b in [0,1].
 func (col Color) FastLinearRgb() (r, g, b float64) {
-    r = math.Pow(col.R, 2.2)
-    g = math.Pow(col.G, 2.2)
-    b = math.Pow(col.B, 2.2)
+    r = linearize_fast(col.R)
+    g = linearize_fast(col.G)
+    b = linearize_fast(col.B)
     return
 }
 
@@ -353,9 +365,35 @@ func LinearRgb(r, g, b float64) Color {
     return Color{delinearize(r), delinearize(g), delinearize(b)}
 }
 
+func delinearize_fast(v float64) float64 {
+    if v > 0.2 {
+        v1 := v - 0.6
+        v2 := v1*v1
+        v3 := v2*v1
+        v4 := v2*v2
+        v5 := v3*v2
+        return 0.442430344268235 + 0.592178981271708*v - 0.287864782562636*v2 + 0.253214392068985*v3 - 0.272557158129811*v4 + 0.325554383321718*v5
+    } else if v > 0.03 {
+        v1 := v - 0.115
+        v2 := v1*v1
+        v3 := v2*v1
+        v4 := v2*v2
+        v5 := v3*v2
+        return 0.194915592891669 + 1.55227076330229*v - 3.93691860257828*v2 + 18.0679839248761*v3 - 101.468750302746*v4 + 632.341487393927*v5
+    } else {
+        v1 := v - 0.015
+        v2 := v1*v1
+        v3 := v2*v1
+        v4 := v2*v2
+        v5 := v3*v2
+        return 0.0519565234928877 + 5.09316778537561*v - 99.0338180489702*v2 + 3484.52322764895*v3 - 150028.083412663*v4 + 7168008.42971613*v5
+    }
+}
+
 // FastLinearRgb is much faster than and almost as accurate as LinearRgb.
+// BUT it is important to NOTE that they only produce good results for valid inputs r,g,b in [0,1].
 func FastLinearRgb(r, g, b float64) Color {
-    return Color{math.Pow(r, 1.0/2.2), math.Pow(g, 1.0/2.2), math.Pow(b, 1.0/2.2)}
+    return Color{delinearize_fast(r), delinearize_fast(g), delinearize_fast(b)}
 }
 
 // XyzToLinearRgb converts from CIE XYZ-space to Linear RGB space.

colors_test.go

@@ -2,11 +2,14 @@ package colorful
 
 import (
     "math"
+    "math/rand"
     "strings"
     "testing"
     "image/color"
 )
 
+var bench_result float64  // Dummy for benchmarks to avoid optimization
+
 // Checks whether the relative error is below eps
 func almosteq_eps(v1, v2, eps float64) bool {
     if math.Abs(v1) > delta {
@@ -197,7 +200,70 @@ func TestHexConversion(t *testing.T) {
 
 /// Linear ///
 //////////////
-// TODO (implicitly tested by XYZ)
+
+// LinearRgb itself is implicitly tested by XYZ conversions below (they use it).
+// So what we do here is just test that the FastLinearRgb approximation is "good enough"
+func TestFastLinearRgb(t *testing.T) {
+    const eps = 6.0/255.0  // We want that "within 6 RGB values total" is "good enough".
+
+    for r := 0.0 ; r < 256.0 ; r++ {
+        for g := 0.0 ; g < 256.0 ; g++ {
+            for b := 0.0 ; b < 256.0 ; b++ {
+                c := Color{r/255.0, g/255.0, b/255.0}
+                r_want, g_want, b_want := c.LinearRgb()
+                r_appr, g_appr, b_appr := c.FastLinearRgb()
+                dr, dg, db := math.Abs(r_want-r_appr), math.Abs(g_want-g_appr), math.Abs(b_want-b_appr)
+                if dr+dg+db > eps {
+                    t.Errorf("FastLinearRgb not precise enough for %v: differences are (%v, %v, %v), allowed total difference is %v", c, dr, dg, db, eps)
+                    return
+                }
+
+                c_want := LinearRgb(r/255.0, g/255.0, b/255.0)
+                c_appr := FastLinearRgb(r/255.0, g/255.0, b/255.0)
+                dr, dg, db = math.Abs(c_want.R-c_appr.R), math.Abs(c_want.G-c_appr.G), math.Abs(c_want.B-c_appr.B)
+                if dr+dg+db > eps {
+                    t.Errorf("FastLinearRgb not precise enough for (%v, %v, %v): differences are (%v, %v, %v), allowed total difference is %v", r, g, b, dr, dg, db, eps)
+                    return
+                }
+            }
+        }
+    }
+}
+
+// Also include some benchmarks to make sure the `Fast` versions are actually significantly faster!
+// (Sounds silly, but the original ones weren't!)
+
+func BenchmarkColorToLinear(bench *testing.B) {
+    var r, g, b float64
+    for n := 0; n < bench.N; n++ {
+        r, g, b = Color{rand.Float64(), rand.Float64(), rand.Float64()}.LinearRgb()
+    }
+    bench_result = r + g + b
+}
+
+func BenchmarkFastColorToLinear(bench *testing.B) {
+    var r, g, b float64
+    for n := 0; n < bench.N; n++ {
+        r, g, b = Color{rand.Float64(), rand.Float64(), rand.Float64()}.FastLinearRgb()
+    }
+    bench_result = r + g + b
+}
+
+func BenchmarkLinearToColor(bench *testing.B) {
+    var c Color
+    for n := 0; n < bench.N; n++ {
+        c = LinearRgb(rand.Float64(), rand.Float64(), rand.Float64())
+    }
+    bench_result = c.R + c.G + c.B
+}
+
+func BenchmarkFastLinearToColor(bench *testing.B) {
+    var c Color
+    for n := 0; n < bench.N; n++ {
+        c = FastLinearRgb(rand.Float64(), rand.Float64(), rand.Float64())
+    }
+    bench_result = c.R + c.G + c.B
+}
 
 /// XYZ ///
 ///////////