use table of procs (and unrolling) to speed up mapPts

BUG=skia:

Review URL: https://codereview.chromium.org/1040783002

3 files changed, 64 insertions, 39 deletions

diff --git a/bench/MatrixBench.cpp b/bench/MatrixBench.cpp
index 6bd1e2a8a6..68195e70e4 100644
--- a/bench/MatrixBench.cpp
+++ b/bench/MatrixBench.cpp
@@ -267,8 +267,7 @@ DEF_BENCH( return new InvertMapRectMatrixBench(
 
 ///////////////////////////////////////////////////////////////////////////////
 
-static SkMatrix make_ident() { SkMatrix m; m.reset(); return m; }
-static SkMatrix make_trans() { SkMatrix m; m.setTranslate(2, 3); return m; }
+static SkMatrix make_trans() { return SkMatrix::MakeTrans(2, 3); }
 static SkMatrix make_scale() { SkMatrix m(make_trans()); m.postScale(1.5f, 0.5f); return m; }
 static SkMatrix make_afine() { SkMatrix m(make_trans()); m.postRotate(15); return m; }
 
@@ -302,8 +301,6 @@ public:
         }
     }
 };
-DEF_BENCH( return new MapPointsMatrixBench("mappts_ident0", make_ident(), false); )
-DEF_BENCH( return new MapPointsMatrixBench("mappts_ident1", make_ident(), true); )
 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); )
diff --git a/include/core/SkMatrix.h b/include/core/SkMatrix.h
index 89bd7935c9..0408634ead 100644
--- a/include/core/SkMatrix.h
+++ b/include/core/SkMatrix.h
@@ -426,7 +426,9 @@ public:
         this->mapPoints(pts, pts, count);
     }
 
-    void mapPts(SkPoint dst[], const SkPoint src[], int count) const;
+    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).
@@ -799,7 +801,12 @@ 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 81b8960855..a43e00f3af 100644
--- a/src/core/SkMatrix.cpp
+++ b/src/core/SkMatrix.cpp
@@ -1022,22 +1022,12 @@ void SkMatrix::mapPoints(SkPoint dst[], const SkPoint src[], int count) const {
 
 #include "Sk4x.h"
 
-void SkMatrix::mapPts(SkPoint dst[], const SkPoint src[], int count) const {
-    if (count <= 0) {
-        return;
-    }
+void SkMatrix::Trans_vpts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
+    SkASSERT(m.getType() <= kTranslate_Mask);
 
-    unsigned mask = this->getType() & 0xF;
-    
-    if (SkMatrix::kIdentity_Mask == mask) {
-        if (src != dst) {
-            memcpy(dst, src, count * sizeof(SkPoint));
-        }
-        return;
-    }
-    if (SkMatrix::kTranslate_Mask == mask) {
-        SkScalar tx = this->getTranslateX();
-        SkScalar ty = this->getTranslateY();
+    if (count > 0) {
+        SkScalar tx = m.getTranslateX();
+        SkScalar ty = m.getTranslateY();
         if (count & 1) {
             dst->fX = src->fX + tx;
             dst->fY = src->fY + ty;
@@ -1046,18 +1036,29 @@ void SkMatrix::mapPts(SkPoint dst[], const SkPoint src[], int count) const {
         }
         Sk4f trans4(tx, ty, tx, ty);
         count >>= 1;
-        for (int i = 0; i < count; ++i) {
+        if (count & 1) {
             (Sk4f::Load(&src->fX) + trans4).store(&dst->fX);
             src += 2;
             dst += 2;
         }
-        return;
+        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;
+        }
     }
-    if (mask <= SkMatrix::kScale_Mask + SkMatrix::kTranslate_Mask) {
-        SkScalar tx = this->getTranslateX();
-        SkScalar ty = this->getTranslateY();
-        SkScalar sx = this->getScaleX();
-        SkScalar sy = this->getScaleY();
+}
+
+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;
@@ -1067,20 +1068,31 @@ void SkMatrix::mapPts(SkPoint dst[], const SkPoint src[], int count) const {
         Sk4f trans4(tx, ty, tx, ty);
         Sk4f scale4(sx, sy, sx, sy);
         count >>= 1;
-        for (int i = 0; i < count; ++i) {
+        if (count & 1) {
             (Sk4f::Load(&src->fX) * scale4 + trans4).store(&dst->fX);
             src += 2;
             dst += 2;
         }
-        return;
+        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;
+        }
     }
-    if (mask < SkMatrix::kPerspective_Mask) {   // affine
-        SkScalar tx = this->getTranslateX();
-        SkScalar ty = this->getTranslateY();
-        SkScalar sx = this->getScaleX();
-        SkScalar sy = this->getScaleY();
-        SkScalar kx = this->getSkewX();
-        SkScalar ky = this->getSkewY();
+}
+
+void SkMatrix::Affine_vpts(const SkMatrix& m, SkPoint dst[], const SkPoint src[], int count) {
+    SkASSERT(m.getType() != kPerspective_Mask);
+
+    if (count > 0) {
+        SkScalar tx = m.getTranslateX();
+        SkScalar ty = m.getTranslateY();
+        SkScalar sx = m.getScaleX();
+        SkScalar sy = m.getScaleY();
+        SkScalar kx = m.getSkewX();
+        SkScalar ky = m.getSkewY();
         if (count & 1) {
             dst->set(src->fX * sx + src->fY * kx + tx,
                      src->fX * ky + src->fY * sy + ty);
@@ -1098,12 +1110,21 @@ void SkMatrix::mapPts(SkPoint dst[], const SkPoint src[], int count) const {
             src += 2;
             dst += 2;
         }
-        return;
     }
-    // fall through for perspective
-    this->mapPoints(dst, src, count);
 }
 
+const SkMatrix::MapPtsProc SkMatrix::gMapVPtsProcs[] = {
+    SkMatrix::Identity_pts, SkMatrix::Trans_vpts,
+    SkMatrix::Scale_vpts,   SkMatrix::Scale_vpts,
+    SkMatrix::Affine_vpts,  SkMatrix::Affine_vpts,
+    SkMatrix::Affine_vpts,  SkMatrix::Affine_vpts,
+    // 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::mapHomogeneousPoints(SkScalar dst[], const SkScalar src[], int count) const {