use Sk4f for matrix math

Need to land SK_SUPPORT_LEGACY_SCALAR_MAPPOINTS in chrome to suppress Affine
version which causes slight differences (which will need to be rebaselined)

BUG=skia:

Review URL: https://codereview.chromium.org/1045493002

4 files changed, 74 insertions, 195 deletions

diff --git a/bench/MatrixBench.cpp b/bench/MatrixBench.cpp
index 68195e70e4..e3e5c703e8 100644
--- a/bench/MatrixBench.cpp
+++ b/bench/MatrixBench.cpp
@@ -278,10 +278,9 @@ protected:
         N = 32
     };
     SkPoint fSrc[N], fDst[N];
-    const bool fNewWay;
 public:
-    MapPointsMatrixBench(const char name[], const SkMatrix& m, bool newWay)
-        : MatrixBench(name), fM(m), fNewWay(newWay)
+    MapPointsMatrixBench(const char name[], const SkMatrix& m)
+        : MatrixBench(name), fM(m)
     {
         SkRandom rand;
         for (int i = 0; i < N; ++i) {
@@ -290,21 +289,13 @@ public:
     }
 
     void performTest() override {
-        if (fNewWay) {
-            for (int i = 0; i < 1000000; ++i) {
-                fM.mapPts(fDst, fSrc, N);
-            }
-        } else {
-            for (int i = 0; i < 1000000; ++i) {
-                fM.mapPoints(fDst, fSrc, N);
-            }
+        for (int i = 0; i < 1000000; ++i) {
+            fM.mapPoints(fDst, fSrc, N);
         }
     }
 };
-DEF_BENCH( return new MapPointsMatrixBench("mappts_trans0", make_trans(), false); )
-DEF_BENCH( return new MapPointsMatrixBench("mappts_trans1", make_trans(), true); )
-DEF_BENCH( return new MapPointsMatrixBench("mappts_scale0", make_scale(), false); )
-DEF_BENCH( return new MapPointsMatrixBench("mappts_scale1", make_scale(), true); )
-DEF_BENCH( return new MapPointsMatrixBench("mappts_afine0", make_afine(), false); )
-DEF_BENCH( return new MapPointsMatrixBench("mappts_afine1", make_afine(), true); )
+DEF_BENCH( return new MapPointsMatrixBench("mappoints_identity", SkMatrix::I()); )
+DEF_BENCH( return new MapPointsMatrixBench("mappoints_trans", make_trans()); )
+DEF_BENCH( return new MapPointsMatrixBench("mappoints_scale", make_scale()); )
+DEF_BENCH( return new MapPointsMatrixBench("mappoints_affine", make_afine()); )
 
diff --git a/include/core/SkMatrix.h b/include/core/SkMatrix.h
index 0408634ead..1c79d22121 100644
--- a/include/core/SkMatrix.h
+++ b/include/core/SkMatrix.h
@@ -413,7 +413,12 @@ public:
         @param count The number of points in src to read, and then transform
                      into dst.
     */
-    void mapPoints(SkPoint dst[], const SkPoint src[], int count) const;
+    void mapPoints(SkPoint dst[], const SkPoint src[], int count) const {
+        SkASSERT((dst && src && count > 0) || 0 == count);
+        // no partial overlap
+        SkASSERT(src == dst || &dst[count] <= &src[0] || &src[count] <= &dst[0]);
+        this->getMapPtsProc()(*this, dst, src, count);
+    }
 
     /** Apply this matrix to the array of points, overwriting it with the
         transformed values.
@@ -426,10 +431,6 @@ public:
         this->mapPoints(pts, pts, count);
     }
 
-    void mapPts(SkPoint dst[], const SkPoint src[], int count) const {
-        gMapVPtsProcs[this->getType() & 0xF](*this, dst, src, count);
-    }
-
     /** Like mapPoints but with custom byte stride between the points. Stride
      *  should be a multiple of sizeof(SkScalar).
      */
@@ -801,12 +802,9 @@ private:
                              int count);
     static void Persp_pts(const SkMatrix&, SkPoint dst[], const SkPoint[], int);
 
-    static void Trans_vpts(const SkMatrix&, SkPoint dst[], const SkPoint[], int);
-    static void Scale_vpts(const SkMatrix&, SkPoint dst[], const SkPoint[], int);
     static void Affine_vpts(const SkMatrix&, SkPoint dst[], const SkPoint[], int);
 
     static const MapPtsProc gMapPtsProcs[];
-    static const MapPtsProc gMapVPtsProcs[];
 
     friend class SkPerspIter;
 };
diff --git a/src/core/SkMatrix.cpp b/src/core/SkMatrix.cpp
index a43e00f3af..0f56b724f5 100644
--- a/src/core/SkMatrix.cpp
+++ b/src/core/SkMatrix.cpp
@@ -8,6 +8,7 @@
 #include "SkMatrix.h"
 #include "SkFloatBits.h"
 #include "SkString.h"
+#include "Sk4x.h"
 
 #include <stddef.h>
 
@@ -867,61 +868,72 @@ bool SkMatrix::invertNonIdentity(SkMatrix* inv) const {
 
 ///////////////////////////////////////////////////////////////////////////////
 
-void SkMatrix::Identity_pts(const SkMatrix& m, SkPoint dst[],
-                            const SkPoint src[], int count) {
+void SkMatrix::Identity_pts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
     SkASSERT(m.getType() == 0);
 
-    if (dst != src && count > 0)
+    if (dst != src && count > 0) {
         memcpy(dst, src, count * sizeof(SkPoint));
-}
-
-void SkMatrix::Trans_pts(const SkMatrix& m, SkPoint dst[],
-                         const SkPoint src[], int count) {
-    SkASSERT(m.getType() == kTranslate_Mask);
-
-    if (count > 0) {
-        SkScalar tx = m.fMat[kMTransX];
-        SkScalar ty = m.fMat[kMTransY];
-        do {
-            dst->fY = src->fY + ty;
-            dst->fX = src->fX + tx;
-            src += 1;
-            dst += 1;
-        } while (--count);
     }
 }
 
-void SkMatrix::Scale_pts(const SkMatrix& m, SkPoint dst[],
-                         const SkPoint src[], int count) {
-    SkASSERT(m.getType() == kScale_Mask);
-
+void SkMatrix::Trans_pts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
+    SkASSERT(m.getType() <= kTranslate_Mask);
+    
     if (count > 0) {
-        SkScalar mx = m.fMat[kMScaleX];
-        SkScalar my = m.fMat[kMScaleY];
-        do {
-            dst->fY = src->fY * my;
-            dst->fX = src->fX * mx;
+        SkScalar tx = m.getTranslateX();
+        SkScalar ty = m.getTranslateY();
+        if (count & 1) {
+            dst->fX = src->fX + tx;
+            dst->fY = src->fY + ty;
             src += 1;
             dst += 1;
-        } while (--count);
+        }
+        Sk4f trans4(tx, ty, tx, ty);
+        count >>= 1;
+        if (count & 1) {
+            (Sk4f::Load(&src->fX) + trans4).store(&dst->fX);
+            src += 2;
+            dst += 2;
+        }
+        count >>= 1;
+        for (int i = 0; i < count; ++i) {
+            (Sk4f::Load(&src[0].fX) + trans4).store(&dst[0].fX);
+            (Sk4f::Load(&src[2].fX) + trans4).store(&dst[2].fX);
+            src += 4;
+            dst += 4;
+        }
     }
 }
 
-void SkMatrix::ScaleTrans_pts(const SkMatrix& m, SkPoint dst[],
-                              const SkPoint src[], int count) {
-    SkASSERT(m.getType() == (kScale_Mask | kTranslate_Mask));
-
+void SkMatrix::Scale_pts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
+    SkASSERT(m.getType() <= (kScale_Mask | kTranslate_Mask));
+    
     if (count > 0) {
-        SkScalar mx = m.fMat[kMScaleX];
-        SkScalar my = m.fMat[kMScaleY];
-        SkScalar tx = m.fMat[kMTransX];
-        SkScalar ty = m.fMat[kMTransY];
-        do {
-            dst->fY = src->fY * my + ty;
-            dst->fX = src->fX * mx + tx;
+        SkScalar tx = m.getTranslateX();
+        SkScalar ty = m.getTranslateY();
+        SkScalar sx = m.getScaleX();
+        SkScalar sy = m.getScaleY();
+        if (count & 1) {
+            dst->fX = src->fX * sx + tx;
+            dst->fY = src->fY * sy + ty;
             src += 1;
             dst += 1;
-        } while (--count);
+        }
+        Sk4f trans4(tx, ty, tx, ty);
+        Sk4f scale4(sx, sy, sx, sy);
+        count >>= 1;
+        if (count & 1) {
+            (Sk4f::Load(&src->fX) * scale4 + trans4).store(&dst->fX);
+            src += 2;
+            dst += 2;
+        }
+        count >>= 1;
+        for (int i = 0; i < count; ++i) {
+            (Sk4f::Load(&src[0].fX) * scale4 + trans4).store(&dst[0].fX);
+            (Sk4f::Load(&src[2].fX) * scale4 + trans4).store(&dst[2].fX);
+            src += 4;
+            dst += 4;
+        }
     }
 }
 
@@ -1000,89 +1012,6 @@ void SkMatrix::Persp_pts(const SkMatrix& m, SkPoint dst[],
     }
 }
 
-const SkMatrix::MapPtsProc SkMatrix::gMapPtsProcs[] = {
-    SkMatrix::Identity_pts, SkMatrix::Trans_pts,
-    SkMatrix::Scale_pts,    SkMatrix::ScaleTrans_pts,
-    SkMatrix::Rot_pts,      SkMatrix::RotTrans_pts,
-    SkMatrix::Rot_pts,      SkMatrix::RotTrans_pts,
-    // repeat the persp proc 8 times
-    SkMatrix::Persp_pts,    SkMatrix::Persp_pts,
-    SkMatrix::Persp_pts,    SkMatrix::Persp_pts,
-    SkMatrix::Persp_pts,    SkMatrix::Persp_pts,
-    SkMatrix::Persp_pts,    SkMatrix::Persp_pts
-};
-
-void SkMatrix::mapPoints(SkPoint dst[], const SkPoint src[], int count) const {
-    SkASSERT((dst && src && count > 0) || 0 == count);
-    // no partial overlap
-    SkASSERT(src == dst || &dst[count] <= &src[0] || &src[count] <= &dst[0]);
-
-    this->getMapPtsProc()(*this, dst, src, count);
-}
-
-#include "Sk4x.h"
-
-void SkMatrix::Trans_vpts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
-    SkASSERT(m.getType() <= kTranslate_Mask);
-
-    if (count > 0) {
-        SkScalar tx = m.getTranslateX();
-        SkScalar ty = m.getTranslateY();
-        if (count & 1) {
-            dst->fX = src->fX + tx;
-            dst->fY = src->fY + ty;
-            src += 1;
-            dst += 1;
-        }
-        Sk4f trans4(tx, ty, tx, ty);
-        count >>= 1;
-        if (count & 1) {
-            (Sk4f::Load(&src->fX) + trans4).store(&dst->fX);
-            src += 2;
-            dst += 2;
-        }
-        count >>= 1;
-        for (int i = 0; i < count; ++i) {
-            (Sk4f::Load(&src[0].fX) + trans4).store(&dst[0].fX);
-            (Sk4f::Load(&src[2].fX) + trans4).store(&dst[2].fX);
-            src += 4;
-            dst += 4;
-        }
-    }
-}
-
-void SkMatrix::Scale_vpts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
-    SkASSERT(m.getType() <= (kScale_Mask | kTranslate_Mask));
-
-    if (count > 0) {
-        SkScalar tx = m.getTranslateX();
-        SkScalar ty = m.getTranslateY();
-        SkScalar sx = m.getScaleX();
-        SkScalar sy = m.getScaleY();
-        if (count & 1) {
-            dst->fX = src->fX * sx + tx;
-            dst->fY = src->fY * sy + ty;
-            src += 1;
-            dst += 1;
-        }
-        Sk4f trans4(tx, ty, tx, ty);
-        Sk4f scale4(sx, sy, sx, sy);
-        count >>= 1;
-        if (count & 1) {
-            (Sk4f::Load(&src->fX) * scale4 + trans4).store(&dst->fX);
-            src += 2;
-            dst += 2;
-        }
-        count >>= 1;
-        for (int i = 0; i < count; ++i) {
-            (Sk4f::Load(&src[0].fX) * scale4 + trans4).store(&dst[0].fX);
-            (Sk4f::Load(&src[2].fX) * scale4 + trans4).store(&dst[2].fX);
-            src += 4;
-            dst += 4;
-        }
-    }
-}
-
 void SkMatrix::Affine_vpts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
     SkASSERT(m.getType() != kPerspective_Mask);
 
@@ -1113,11 +1042,16 @@ void SkMatrix::Affine_vpts(const SkMatrix& m, SkPoint dst[], const SkPoint src[]
     }
 }
 
-const SkMatrix::MapPtsProc SkMatrix::gMapVPtsProcs[] = {
-    SkMatrix::Identity_pts, SkMatrix::Trans_vpts,
-    SkMatrix::Scale_vpts,   SkMatrix::Scale_vpts,
+const SkMatrix::MapPtsProc SkMatrix::gMapPtsProcs[] = {
+    SkMatrix::Identity_pts, SkMatrix::Trans_pts,
+    SkMatrix::Scale_pts,   SkMatrix::Scale_pts,
+#ifdef SK_SUPPORT_LEGACY_SCALAR_MAPPOINTS
+    SkMatrix::Rot_pts,      SkMatrix::RotTrans_pts,
+    SkMatrix::Rot_pts,      SkMatrix::RotTrans_pts,
+#else
     SkMatrix::Affine_vpts,  SkMatrix::Affine_vpts,
     SkMatrix::Affine_vpts,  SkMatrix::Affine_vpts,
+#endif
     // repeat the persp proc 8 times
     SkMatrix::Persp_pts,    SkMatrix::Persp_pts,
     SkMatrix::Persp_pts,    SkMatrix::Persp_pts,
diff --git a/tests/MatrixTest.cpp b/tests/MatrixTest.cpp
index 35306b3275..95d33acabd 100644
--- a/tests/MatrixTest.cpp
+++ b/tests/MatrixTest.cpp
@@ -799,48 +799,6 @@ static void test_decompScale(skiatest::Reporter* reporter) {
     REPORTER_ASSERT(reporter, !check_decompScale(m));
 }
 
-static void test_mappts(skiatest::Reporter* reporter, const SkMatrix& m, const char type[], int n) {
-    const int MAX = 100;
-    SkPoint src[MAX];
-    SkPoint dst0[MAX], dst1[MAX];
-    SkASSERT(n <= MAX);
-
-    SkRandom rand;
-    for (int i = 0; i < n; ++i) {
-        src[i].fX = rand.nextSScalar1() * 100;
-        src[i].fY = rand.nextSScalar1() * 100;
-    }
-
-    m.mapPoints(dst0, src, n);
-    m.mapPts(   dst1, src, n);
-    for (int i = 0; i < n; ++i) {
-        bool eq = SkScalarNearlyEqual(dst0[i].fX, dst1[i].fX) &&
-                  SkScalarNearlyEqual(dst0[i].fY, dst1[i].fY);
-        if (!eq) {
-            SkDebugf("%s [%d] points (%g %g) pts (%g %g)\n", type, i, dst0[i].fX, dst0[i].fY, dst1[i].fX, dst1[i].fY);
-            REPORTER_ASSERT(reporter, eq);
-        }
-    }
-}
-
-static void test_mappts(skiatest::Reporter* reporter) {
-    const int counts[] = { 0, 1, 2, 3, 4, 100 };
-    for (size_t i = 0; i < SK_ARRAY_COUNT(counts); ++i) {
-        const int n = counts[i];
-        SkMatrix m;
-        m.reset();
-        test_mappts(reporter, m, "ident", n);
-        m.setTranslate(2, 3);
-        test_mappts(reporter, m, "trans", n);
-        m.postScale(2, 0.5f);
-        test_mappts(reporter, m, "scale", n);
-        m.postRotate(35);
-        test_mappts(reporter, m, "affine", n);
-        m.setPerspX(0.1f);
-        test_mappts(reporter, m, "persp", n);
-    }
-}
-
 DEF_TEST(Matrix, reporter) {
     SkMatrix    mat, inverse, iden1, iden2;
 
@@ -961,8 +919,6 @@ DEF_TEST(Matrix, reporter) {
     test_set9(reporter);
 
     test_decompScale(reporter);
-
-    test_mappts(reporter);
 }
 
 DEF_TEST(Matrix_Concat, r) {