Add Sk4f_RoundToInt

src/core/SkNx.h

Index: src/core/SkNx.h
diff --git a/src/core/SkNx.h b/src/core/SkNx.h
index d0c7f41f1cd1b4b3b8a7e64b4b8fc9a4c6d43408..6e453896c79ee71dc1906df6913a5ef27d9240a4 100644
--- a/src/core/SkNx.h
+++ b/src/core/SkNx.h
@@ -299,6 +299,22 @@ typedef SkNx<4,       int> Sk4i;
     #include "../opts/SkNx_sse.h"
 #elif !defined(SKNX_NO_SIMD) && defined(SK_ARM_HAS_NEON)
     #include "../opts/SkNx_neon.h"
+#else
+
+SI Sk4i Sk4f_RoundToInt(const Sk4f& x) {
+    return { sk_float_round2int(x[0]),

[mtklein, 2016/07/11 14:05:03]

If we can't implement this in terms of things from <math.h>  (e.g. lrintf) can
you leave a note about why we need to use Skia's homebrew instead?

[msarett, 2016/07/12 15:42:43]

lrintf() seems fine to me.

+             sk_float_round2int(x[1]),
+             sk_float_round2int(x[2]),
+             sk_float_round2int(x[3]), };
+}
+
+SI uint32_t Sk4f_RoundToByte(const Sk4f& x) {

[mtklein, 2016/07/11 14:05:03]

Why don't we simplify this by adding just one new call,

   Sk4i Sk4f_round(Sk4f);

and then make sure SkNx_cast<uint8_t>(Sk4i) is implemented well to handle the
rest of the trip down to bytes?

   SkNx_cast<uint8_t>(Sk4f_round(...)).store(&dst)

[msarett, 2016/07/12 15:42:43]

SGTM

+    return sk_float_round2int(x[0]) <<  0
+         | sk_float_round2int(x[1]) <<  8
+         | sk_float_round2int(x[2]) << 16
+         | sk_float_round2int(x[3]) << 24;
+}
+
 #endif
 
 SI void Sk4f_ToBytes(uint8_t p[16], const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {