speed-up SkMatrix::preScale by 3x, by special-casing it instead of just calling

concat. Inspired by the profile of the fishtank site



git-svn-id: http://skia.googlecode.com/svn/trunk@1462 2bbb7eff-a529-9590-31e7-b0007b416f81

bench/MatrixBench.cpp

@@ -0,0 +1,90 @@
+#include "SkBenchmark.h"
+#include "SkMatrix.h"
+#include "SkString.h"
+
+class MatrixBench : public SkBenchmark {
+    SkString    fName;
+    enum { N = 100000 };
+public:
+    MatrixBench(void* param, const char name[]) : INHERITED(param) {
+        fName.printf("matrix_%s", name);
+    }
+
+    virtual void performTest() = 0;
+
+protected:
+    virtual const char* onGetName() {
+        return fName.c_str();
+    }
+
+    virtual void onDraw(SkCanvas* canvas) {
+        for (int i = 0; i < N; i++) {
+            this->performTest();
+        }
+    }
+
+private:
+    typedef SkBenchmark INHERITED;
+};
+
+// we want to stop the compiler from eliminating code that it thinks is a no-op
+// so we have a non-static global we increment, hoping that will convince the
+// compiler to execute everything
+int gMatrixBench_NonStaticGlobal;
+
+#define always_do(pred)                     \
+    do {                                    \
+        if (pred) {                         \
+            ++gMatrixBench_NonStaticGlobal; \
+        }                                   \
+    } while (0)
+
+class EqualsMatrixBench : public MatrixBench {
+public:
+    EqualsMatrixBench(void* param) : INHERITED(param, "equals") {}
+protected:
+    virtual void performTest() {
+        SkMatrix m0, m1, m2;
+        
+        m0.reset();
+        m1.reset();
+        m2.reset();
+        always_do(m0 == m1);
+        always_do(m1 == m2);
+        always_do(m2 == m0);
+        always_do(m0.getType());
+        always_do(m1.getType());
+        always_do(m2.getType());
+    }
+private:
+    typedef MatrixBench INHERITED;
+};
+
+class ScaleMatrixBench : public MatrixBench {
+public:
+    ScaleMatrixBench(void* param) : INHERITED(param, "scale") {
+        
+        fM0.reset();
+        fM1.setScale(fSX, fSY);
+        fM2.setTranslate(fSX, fSY);
+        fSX = fSY = SkFloatToScalar(1.5f);
+    }
+protected:
+    virtual void performTest() {
+        SkMatrix m;
+        m = fM0; m.preScale(fSX, fSY);
+        m = fM1; m.preScale(fSX, fSY);
+        m = fM2; m.preScale(fSX, fSY);
+    }
+private:
+    SkMatrix fM0, fM1, fM2;
+    SkScalar fSX, fSY;
+    typedef MatrixBench INHERITED;
+};
+
+static SkBenchmark* M0(void* p) { return new EqualsMatrixBench(p); }
+static SkBenchmark* M1(void* p) { return new ScaleMatrixBench(p); }
+
+static BenchRegistry gReg0(M0);
+static BenchRegistry gReg1(M1);
+

include/core/SkMatrix.h

@@ -438,6 +438,7 @@ public:
     */
     bool fixedStepInX(SkScalar y, SkFixed* stepX, SkFixed* stepY) const;
 
+#ifdef SK_SCALAR_IS_FIXED
     friend bool operator==(const SkMatrix& a, const SkMatrix& b) {
         return memcmp(a.fMat, b.fMat, sizeof(a.fMat)) == 0;
     }
@@ -445,6 +446,12 @@ public:
     friend bool operator!=(const SkMatrix& a, const SkMatrix& b) {
         return memcmp(a.fMat, b.fMat, sizeof(a.fMat)) != 0;
     }
+#else
+    friend bool operator==(const SkMatrix& a, const SkMatrix& b);    
+    friend bool operator!=(const SkMatrix& a, const SkMatrix& b) {
+        return !(a == b);
+    }
+#endif
 
     enum {
         // flatten/unflatten will never return a value larger than this
@@ -489,6 +496,11 @@ private:
 
         kUnknown_Mask = 0x80,
 
+        kORableMasks =  kTranslate_Mask |
+                        kScale_Mask |
+                        kAffine_Mask |
+                        kPerspective_Mask,
+
         kAllMasks = kTranslate_Mask |
                     kScale_Mask |
                     kAffine_Mask |
@@ -507,6 +519,11 @@ private:
         fTypeMask = SkToU8(mask);
     }
 
+    void orTypeMask(int mask) {
+        SkASSERT((mask & kORableMasks) == mask);
+        fTypeMask = SkToU8(fTypeMask | mask);
+    }
+
     void clearTypeMask(int mask) {
         // only allow a valid mask
         SkASSERT((mask & kAllMasks) == mask);

src/core/SkMatrix.cpp

@@ -119,6 +119,21 @@ uint8_t SkMatrix::computeTypeMask() const {
 
 ///////////////////////////////////////////////////////////////////////////////
 
+#ifdef SK_SCALAR_IS_FLOAT
+
+bool operator==(const SkMatrix& a, const SkMatrix& b) {
+    const SkScalar* SK_RESTRICT ma = a.fMat;
+    const SkScalar* SK_RESTRICT mb = b.fMat;
+
+    return  ma[0] == mb[0] && ma[1] == mb[1] && ma[2] == mb[2] &&
+            ma[3] == mb[3] && ma[4] == mb[4] && ma[5] == mb[5] &&
+            ma[6] == mb[6] && ma[7] == mb[7] && ma[8] == mb[8];
+}
+
+#endif
+
+///////////////////////////////////////////////////////////////////////////////
+
 void SkMatrix::setTranslate(SkScalar dx, SkScalar dy) {
     if (SkScalarToCompareType(dx) || SkScalarToCompareType(dy)) {
         fMat[kMTransX] = dx;
@@ -202,9 +217,27 @@ bool SkMatrix::preScale(SkScalar sx, SkScalar sy, SkScalar px, SkScalar py) {
 }
 
 bool SkMatrix::preScale(SkScalar sx, SkScalar sy) {
+#ifdef SK_SCALAR_IS_FIXED
     SkMatrix    m;
     m.setScale(sx, sy);
     return this->preConcat(m);
+#else
+    // the assumption is that these multiplies are very cheap, and that
+    // a full concat and/or just computing the matrix type is more expensive.
+    // Also, the fixed-point case checks for overflow, but the float doesn't,
+    // so we can get away with these blind multiplies.
+
+    fMat[kMScaleX] = SkScalarMul(fMat[kMScaleX], sx);
+    fMat[kMSkewY] = SkScalarMul(fMat[kMSkewY],   sx);
+    fMat[kMPersp0] = SkScalarMul(fMat[kMPersp0], sx);
+
+    fMat[kMSkewX] = SkScalarMul(fMat[kMSkewX],   sy);
+    fMat[kMScaleY] = SkScalarMul(fMat[kMScaleY], sy);
+    fMat[kMPersp1] = SkScalarMul(fMat[kMPersp1], sy);
+
+    this->orTypeMask(kScale_Mask);
+    return true;
+#endif
 }
 
 bool SkMatrix::postScale(SkScalar sx, SkScalar sy, SkScalar px, SkScalar py) {