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

src/opts/SkNx_neon.h

Index: src/opts/SkNx_neon.h
diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index 6b216827a8f007449258d18edf4d85811ac1bc96..f1deabc5febacb4d2f688c9dae61df581f2cfc2e 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -81,20 +81,21 @@ public:
     static SkNf Min(const SkNf& l, const SkNf& r) { return vmin_f32(l.fVec, r.fVec); }
     static SkNf Max(const SkNf& l, const SkNf& r) { return vmax_f32(l.fVec, r.fVec); }
 
-    SkNf rsqrt() const {
-        float32x2_t est0 = vrsqrte_f32(fVec),
-                    est1 = vmul_f32(vrsqrts_f32(fVec, vmul_f32(est0, est0)), est0);
-        return est1;
+    SkNf rsqrt0() const { return vrsqrte_f32(fVec); }
+    SkNf rsqrt1() const {
+        float32x2_t est0 = this->rsqrt0().fVec;
+        return vmul_f32(vrsqrts_f32(fVec, vmul_f32(est0, est0)), est0);
+    }
+    SkNf rsqrt2() const {
+        float32x2_t est1 = this->rsqrt1().fVec;
+        return vmul_f32(vrsqrts_f32(fVec, vmul_f32(est1, est1)), est1);
     }
 
     SkNf sqrt() const {
     #if defined(SK_CPU_ARM64)
         return vsqrt_f32(fVec);
     #else
-        float32x2_t est1 = this->rsqrt().fVec,
-        // An extra step of Newton's method to refine the estimate of 1/sqrt(this).
-                    est2 = vmul_f32(vrsqrts_f32(fVec, vmul_f32(est1, est1)), est1);
-        return vmul_f32(fVec, est2);
+        return *this * this->rsqrt2();
     #endif
     }
 
@@ -151,10 +152,15 @@ public:
     static SkNf Max(const SkNf& l, const SkNf& r) { return vmaxq_f64(l.fVec, r.fVec); }
 
     SkNf  sqrt() const { return vsqrtq_f64(fVec);  }
-    SkNf rsqrt() const {
-        float64x2_t est0 = vrsqrteq_f64(fVec),
-                    est1 = vmulq_f64(vrsqrtsq_f64(fVec, vmulq_f64(est0, est0)), est0);
-        return est1;
+
+    SkNf rsqrt0() const { return vrsqrteq_f64(fVec); }
+    SkNf rsqrt1() const {
+        float64x2_t est0 = this->rsqrt0().fVec;
+        return vmulq_f64(vrsqrtsq_f64(fVec, vmulq_f64(est0, est0)), est0);
+    }
+    SkNf rsqrt2() const {
+        float64x2_t est1 = this->rsqrt1().fVec;
+        return vmulq_f64(vrsqrtsq_f64(fVec, vmulq_f64(est1, est1)), est1);
     }
 
     SkNf approxInvert() const {
@@ -269,20 +275,21 @@ public:
     static SkNf Min(const SkNf& l, const SkNf& r) { return vminq_f32(l.fVec, r.fVec); }
     static SkNf Max(const SkNf& l, const SkNf& r) { return vmaxq_f32(l.fVec, r.fVec); }
 
-    SkNf rsqrt() const {
-        float32x4_t est0 = vrsqrteq_f32(fVec),
-                    est1 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0);
-        return est1;
+    SkNf rsqrt0() const { return vrsqrteq_f32(fVec); }
+    SkNf rsqrt1() const {
+        float32x4_t est0 = this->rsqrt0().fVec;
+        return vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0);
+    }
+    SkNf rsqrt2() const {
+        float32x4_t est1 = this->rsqrt1().fVec;
+        return vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1);
     }
 
     SkNf sqrt() const {
     #if defined(SK_CPU_ARM64)
         return vsqrtq_f32(fVec);
     #else
-        float32x4_t est1 = this->rsqrt().fVec,
-        // An extra step of Newton's method to refine the estimate of 1/sqrt(this).
-                    est2 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1);
-        return vmulq_f32(fVec, est2);
+        return *this * this->rsqrt2();
     #endif
     }