Split rsqrt into rsqrt{0,1,2}, with increasing cost and precision on ARM

src/core/SkNx.h

Index: src/core/SkNx.h
diff --git a/src/core/SkNx.h b/src/core/SkNx.h
index 65b5b97a83e536cc928b897f17653b8c5622143f..4cfc6e3d9ae05aff619c590f9ef57ff5fe9d2e10 100644
--- a/src/core/SkNx.h
+++ b/src/core/SkNx.h
@@ -119,7 +119,11 @@ public:
     }
 
     SkNf  sqrt() const { return SkNf(fLo. sqrt(), fHi. sqrt()); }
-    SkNf rsqrt() const { return SkNf(fLo.rsqrt(), fHi.rsqrt()); }
+
+    // Generally, increasing precision, increasing cost.
+    SkNf rsqrt0() const { return SkNf(fLo.rsqrt0(), fHi.rsqrt0()); }
+    SkNf rsqrt1() const { return SkNf(fLo.rsqrt1(), fHi.rsqrt1()); }
+    SkNf rsqrt2() const { return SkNf(fLo.rsqrt2(), fHi.rsqrt2()); }
 
     SkNf       invert() const { return SkNf(fLo.      invert(), fHi.      invert()); }
     SkNf approxInvert() const { return SkNf(fLo.approxInvert(), fHi.approxInvert()); }
@@ -207,7 +211,9 @@ public:
     static SkNf Max(const SkNf& l, const SkNf& r) { return SkNf(SkTMax(l.fVal, r.fVal)); }
 
     SkNf  sqrt() const { return SkNf(Sqrt(fVal));        }
-    SkNf rsqrt() const { return SkNf((T)1 / Sqrt(fVal)); }
+    SkNf rsqrt0() const { return SkNf((T)1 / Sqrt(fVal)); }
+    SkNf rsqrt1() const { return this->rsqrt0(); }
+    SkNf rsqrt2() const { return this->rsqrt1(); }
 
     SkNf       invert() const { return SkNf((T)1 / fVal); }
     SkNf approxInvert() const { return this->invert();    }