Add sk_float_rsqrt with SSE + NEON fast paths.

include/core/SkFloatingPoint.h

Index: include/core/SkFloatingPoint.h
diff --git a/include/core/SkFloatingPoint.h b/include/core/SkFloatingPoint.h
index 44a3eef98dc8c0272a324494e4ea0153dd4f7972..acc5d5d9d85ddcdf0e9773ddb99cf87f035792a8 100644
--- a/include/core/SkFloatingPoint.h
+++ b/include/core/SkFloatingPoint.h
@@ -96,4 +96,30 @@ extern const uint32_t gIEEENegativeInfinity;
 #define SK_FloatNaN                 (*SkTCast<const float*>(&gIEEENotANumber))
 #define SK_FloatInfinity            (*SkTCast<const float*>(&gIEEEInfinity))
 #define SK_FloatNegativeInfinity    (*SkTCast<const float*>(&gIEEENegativeInfinity))
+
+#ifdef __SSE__
+#include <xmmintrin.h>
+#endif
+
+// Fast, approximate inverse square root.
+// Compare to name-brand "1.0f / sk_float_sqrt(x)".
+inline float sk_float_rsqrt(float x) {

[reed1, 2013/11/07 22:28:44]

static ...

+// We want all this inlined, so we'll inline SIMD and just take the hit when we don't know we've got
+// it at compile time.  This is going to be too fast to productively hide behind a function pointer.
+#ifdef __SSE__
+    float result;
+    _mm_store_ss(&result, _mm_rsqrt_ss(_mm_set_ss(x)));
+    return result;
+#else
+// TODO(mtklein): NEON
+    float xhalf = 0.5f*x;
+    int i = *SkTCast<int*>(&x);
+    i = 0x5f3759df - (i>>1);
+    x = *SkTCast<float*>(&i);
+    x = x*(1.5f-xhalf*x*x);
+//    x = x*(1.5f-xhalf*x*x); // this line takes err from 10^-3 to 10^-6, but at what cost!?!?
+    return x;
+#endif
+}
+
 #endif