Sk2x for NEON

Also decreases the precision of Sk4f::rsqrt() for speed, keeping Sk4f::sqrt() the same:
instead of doing two estimation steps in rsqrt(), do one there and one more in sqrt().

Tests pass on my Nexus 7.  float64x2_t is still a TODO for when I get a hold of a Nexus 9.

BUG=skia:

Review URL: https://codereview.chromium.org/1018423003

1 files changed, 9 insertions, 5 deletions

diff --git a/src/opts/Sk4x_neon.h b/src/opts/Sk4x_neon.h
index 2851fb31a4..3f35fe785b 100644
--- a/src/opts/Sk4x_neon.h
+++ b/src/opts/Sk4x_neon.h
@@ -82,13 +82,17 @@ M(Sk4f) divide  (const Sk4f& o) const {
 }
 
 M(Sk4f) rsqrt() const {
-    float32x4_t est0 = vrsqrteq_f32(fVec);
-    float32x4_t est1 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0);
-    float32x4_t est2 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1);
-    return est2;
+    float32x4_t est0 = vrsqrteq_f32(fVec),
+                est1 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0);
+    return est1;
 }
 
-M(Sk4f)  sqrt() const { return this->multiply(this->rsqrt()); }
+M(Sk4f)  sqrt() const {
+    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);
+}
 
 M(Sk4i) equal           (const Sk4f& o) const { return vreinterpretq_s32_u32(vceqq_f32(fVec, o.fVec)); }
 M(Sk4i) notEqual        (const Sk4f& o) const { return vreinterpretq_s32_u32(vmvnq_u32(vceqq_f32(fVec, o.fVec))); }