Sk4x_sse.h

src/core/Sk4x.h

Index: src/core/Sk4x.h
diff --git a/src/core/Sk4x.h b/src/core/Sk4x.h
index 5b15560ede46cdcab4792974658cf134c15f27b1..7c246dffc01b7265c44772f2328f0ad385adba1c 100644
--- a/src/core/Sk4x.h
+++ b/src/core/Sk4x.h
@@ -4,13 +4,22 @@
 #include "SkTypes.h"
 
 #define SK4X_PREAMBLE 1
-    #include "Sk4x_portable.h"
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+        #include "Sk4x_sse.h"
+    #else
+        #include "Sk4x_portable.h"
+    #endif
 #undef SK4X_PREAMBLE
 
 template <typename T> class Sk4x;
 typedef Sk4x<float>   Sk4f;
 typedef Sk4x<int32_t> Sk4i;
 
+// Some Sk4x methods are implemented only for Sk4f or Sk4i.
+// They might be unavailable, really slow, or just a bad idea.
+// Talk to mtklein if you find yourself unable to link and
+// really need one of those methods.
+
 template <typename T> class Sk4x {
 public:
     Sk4x();  // Uninitialized; use Sk4x(0,0,0,0) for zero.
@@ -34,6 +43,7 @@ public:
     Sk4x   bitNot()            const;
     Sk4x   bitAnd(const Sk4x&) const;
     Sk4x    bitOr(const Sk4x&) const;
+    // TODO: Sk4x bitAndNot(const Sk4x&) const; is efficient in SSE.
     Sk4x      add(const Sk4x&) const;
     Sk4x subtract(const Sk4x&) const;
     Sk4x multiply(const Sk4x&) const;
@@ -56,15 +66,27 @@ public:
     static Sk4x XYAB(const Sk4x& xyzw, const Sk4x& abcd);
     static Sk4x ZWCD(const Sk4x& xyzw, const Sk4x& abcd);
 
+    // TODO: these are particularly efficient in SSE.  Useful?  Also efficient in NEON?
+    // static Sk4x XAYB(const Sk4x& xyzw, const Sk4x& abcd);
+    // static Sk4x ZCWD(const Sk4x& xyzw, const Sk4x& abcd);
+
 private:
     // It's handy to have Sk4f and Sk4i be mutual friends.
     template <typename S> friend class Sk4x;
 
 #define SK4X_PRIVATE 1
-    #include "Sk4x_portable.h"
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+        #include "Sk4x_sse.h"
+    #else
+        #include "Sk4x_portable.h"
+    #endif
 #undef SK4X_PRIVATE
 };
 
-#include "Sk4x_portable.h"
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    #include "Sk4x_sse.h"
+#else
+    #include "Sk4x_portable.h"
+#endif
 
 #endif//Sk4x_DEFINED