speedup haircubics

src/core/SkGeometry.cpp

 /*
  * Copyright 2006 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkGeometry.h"
 #include "SkMatrix.h"
 #include "SkNx.h"
 
+#if 0
 static Sk2s from_point(const SkPoint& point) {
     return Sk2s::Load(&point.fX);
 }
 
 static SkPoint to_point(const Sk2s& x) {
     SkPoint point;
     x.store(&point.fX);
     return point;
 }
+#endif
 
 static SkVector to_vector(const Sk2s& x) {
     SkVector vector;
     x.store(&vector.fX);
     return vector;
 }
 
 /** If defined, this makes eval_quad and eval_cubic do more setup (sometimes
     involving integer multiplies by 2 or 3, but fewer calls to SkScalarMul.
     May also introduce overflow of fixed when we compute our setup.
@@ skipping 414 matching lines @@
 
     dst[0] = src[0];
     dst[1] = to_point(ab);
     dst[2] = to_point(abc);
     dst[3] = to_point(abcd);
     dst[4] = to_point(bcd);
     dst[5] = to_point(cd);
     dst[6] = src[3];
 }
 
+void SkCubicToPoly(const SkPoint pts[4], SkPoint coeff[4]) {
+    Sk2s p0 = from_point(pts[0]);
+    Sk2s p1 = from_point(pts[1]);
+    Sk2s p2 = from_point(pts[2]);
+    Sk2s p3 = from_point(pts[3]);
+
+    const Sk2s three(3);
+    Sk2s p1minusp2 = p1 - p2;
+
+    Sk2s D = p0;
+    Sk2s A = p3 + three * p1minusp2 - D;
+    Sk2s B = three * (D - p1minusp2 - p1);
+    Sk2s C = three * (p1 - D);
+
+    coeff[0] = to_point(A);
+    coeff[1] = to_point(B);
+    coeff[2] = to_point(C);
+    coeff[3] = to_point(D);
+}
+
+SkPoint SkEvalCubicPolyAt(const SkPoint coeff[4], SkScalar t) {
+    Sk2s pt = sk2s_cubic_eval(from_point(coeff[0]), from_point(coeff[1]),
+                              from_point(coeff[2]), from_point(coeff[3]), Sk2s(t));
+    return to_point(pt);
+}
+
 /*  http://code.google.com/p/skia/issues/detail?id=32
 
     This test code would fail when we didn't check the return result of
     valid_unit_divide in SkChopCubicAt(... tValues[], int roots). The reason is
     that after the first chop, the parameters to valid_unit_divide are equal
     (thanks to finite float precision and rounding in the subtracts). Thus
     even though the 2nd tValue looks < 1.0, after we renormalize it, we end
     up with 1.0, hence the need to check and just return the last cubic as
     a degenerate clump of 4 points in the sampe place.
 
@@ skipping 1069 matching lines @@
         matrix.preScale(SK_Scalar1, -SK_Scalar1);
     }
     if (userMatrix) {
         matrix.postConcat(*userMatrix);
     }
     for (int i = 0; i < conicCount; ++i) {
         matrix.mapPoints(dst[i].fPts, 3);
     }
     return conicCount;
 }