Add sk_float_rsqrt with SSE + NEON fast paths.

bench/MathBench.cpp

 #include "SkBenchmark.h"
 #include "SkColorPriv.h"
 #include "SkMatrix.h"
 #include "SkRandom.h"
 #include "SkString.h"
 #include "SkPaint.h"
 
 static float sk_fsel(float pred, float result_ge, float result_lt) {
     return pred >= 0 ? result_ge : result_lt;
 }
@@ skipping 74 matching lines @@
                               const float* SK_RESTRICT src,
                               int count) {
         for (int i = 0; i < count; ++i) {
             dst[i] = src[i] + 1;
         }
     }
 private:
     typedef MathBench INHERITED;
 };
 
+class SkRSqrtMathBench : public MathBench {
+public:
+    SkRSqrtMathBench() : INHERITED("sk_float_rsqrt") {}
+protected:
+    virtual void performTest(float* SK_RESTRICT dst,
+                              const float* SK_RESTRICT src,
+                              int count) {
+        for (int i = 0; i < count; ++i) {
+            dst[i] = sk_float_rsqrt(src[i]);
+        }
+    }
+private:
+    typedef MathBench INHERITED;
+};
+
+
 class SlowISqrtMathBench : public MathBench {
 public:
     SlowISqrtMathBench() : INHERITED("slowIsqrt") {}
 protected:
     virtual void performTest(float* SK_RESTRICT dst,
                               const float* SK_RESTRICT src,
                               int count) {
         for (int i = 0; i < count; ++i) {
             dst[i] = 1.0f / sk_float_sqrt(src[i]);
         }
@@ skipping 438 matching lines @@
 DEF_BENCH(return new DivModBench<uint64_t>("uint64_t"))
 
 DEF_BENCH(return new DivModBench<int8_t>("int8_t"))
 DEF_BENCH(return new DivModBench<int16_t>("int16_t"))
 DEF_BENCH(return new DivModBench<int32_t>("int32_t"))
 DEF_BENCH(return new DivModBench<int64_t>("int64_t"))
 
 ///////////////////////////////////////////////////////////////////////////////
 
 DEF_BENCH( return new NoOpMathBench(); )
+DEF_BENCH( return new SkRSqrtMathBench(); )
 DEF_BENCH( return new SlowISqrtMathBench(); )
 DEF_BENCH( return new FastISqrtMathBench(); )
 DEF_BENCH( return new QMul64Bench(); )
 DEF_BENCH( return new QMul32Bench(); )
 
 DEF_BENCH( return new IsFiniteBench(-1); )
 DEF_BENCH( return new IsFiniteBench(0); )
 DEF_BENCH( return new IsFiniteBench(1); )
 DEF_BENCH( return new IsFiniteBench(2); )
 DEF_BENCH( return new IsFiniteBench(3); )
 DEF_BENCH( return new IsFiniteBench(4); )
 DEF_BENCH( return new IsFiniteBench(5); )
 
 DEF_BENCH( return new FloorBench(false); )
 DEF_BENCH( return new FloorBench(true); )
 
 DEF_BENCH( return new CLZBench(false); )
 DEF_BENCH( return new CLZBench(true); )
 
 DEF_BENCH( return new NormalizeBench(); )
 
 DEF_BENCH( return new FixedMathBench(); )