float xfermodes (burn, dodge, softlight) in Sk8f, possibly using AVX.

src/core/SkNx.h

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkNx_DEFINED
 #define SkNx_DEFINED
 
@@ skipping 13 matching lines @@
 // will probably pick one randomly for us, which is rarely correct).
 namespace {
 
 // The default implementations just fall back on a pair of size N/2.
 
 template <int N, typename T>
 class SkNi {
 public:
     SkNi() {}
     SkNi(const SkNi<N/2, T>& lo, const SkNi<N/2, T>& hi) : fLo(lo), fHi(hi) {}
-    explicit SkNi(T val) : fLo(val), fHi(val) {}
+    SkNi(T val) : fLo(val), fHi(val) {}
     static SkNi Load(const T vals[N]) {
         return SkNi(SkNi<N/2,T>::Load(vals), SkNi<N/2,T>::Load(vals+N/2));
     }
 
     SkNi(T a, T b)                                : fLo(a),       fHi(b)       { REQUIRE(N==2); }
     SkNi(T a, T b, T c, T d)                      : fLo(a,b),     fHi(c,d)     { REQUIRE(N==4); }
     SkNi(T a, T b, T c, T d,  T e, T f, T g, T h) : fLo(a,b,c,d), fHi(e,f,g,h) { REQUIRE(N==8); }
     SkNi(T a, T b, T c, T d,  T e, T f, T g, T h,
          T i, T j, T k, T l,  T m, T n, T o, T p)
         : fLo(a,b,c,d, e,f,g,h), fHi(i,j,k,l, m,n,o,p) { REQUIRE(N==16); }
@@ skipping 33 matching lines @@
 protected:
     REQUIRE(0 == (N & (N-1)));
 
     SkNi<N/2, T> fLo, fHi;
 };
 
 template <int N>
 class SkNf {
 public:
     SkNf() {}
-    explicit SkNf(float val) : fLo(val),  fHi(val) {}
+    SkNf(float val) : fLo(val),  fHi(val) {}
     static SkNf Load(const float vals[N]) {
         return SkNf(SkNf<N/2>::Load(vals), SkNf<N/2>::Load(vals+N/2));
     }
     // FromBytes() and toBytes() specializations may assume their argument is N-byte aligned.
     // E.g. Sk4f::FromBytes() may assume it's reading from a 4-byte-aligned pointer.
     // Converts [0,255] bytes to [0.0, 255.0] floats.
     static SkNf FromBytes(const uint8_t bytes[N]) {
         return SkNf(SkNf<N/2>::FromBytes(bytes), SkNf<N/2>::FromBytes(bytes+N/2));
     }
 
@@ skipping 61 matching lines @@
     SkNf<N/2> fLo, fHi;
 };
 
 
 // Bottom out the default implementations with scalars when nothing's been specialized.
 
 template <typename T>
 class SkNi<1,T> {
 public:
     SkNi() {}
-    explicit SkNi(T val) : fVal(val) {}
+    SkNi(T val) : fVal(val) {}
     static SkNi Load(const T vals[1]) { return SkNi(vals[0]); }
 
     void store(T vals[1]) const { vals[0] = fVal; }
 
     SkNi saturatedAdd(const SkNi& o) const {
         SkASSERT((T)(~0) > 0); // TODO: support signed T
         T sum = fVal + o.fVal;
         return SkNi(sum < fVal ? (T)(~0) : sum);
     }
 
@@ skipping 17 matching lines @@
     SkNi thenElse(const SkNi& t, const SkNi& e) const { return fVal ? t : e; }
 
 protected:
     T fVal;
 };
 
 template <>
 class SkNf<1> {
 public:
     SkNf() {}
-    explicit SkNf(float val) : fVal(val) {}
+    SkNf(float val) : fVal(val) {}
     static SkNf Load(const float vals[1]) { return SkNf(vals[0]); }
     static SkNf FromBytes(const uint8_t bytes[1]) { return SkNf((float)bytes[0]); }
 
     void store(float vals[1]) const { vals[0] = fVal; }
     void toBytes(uint8_t bytes[1]) const { bytes[0] = (uint8_t)(SkTMin(fVal, 255.0f)); }
 
     SkNf operator + (const SkNf& o) const { return SkNf(fVal + o.fVal); }
     SkNf operator - (const SkNf& o) const { return SkNf(fVal - o.fVal); }
     SkNf operator * (const SkNf& o) const { return SkNf(fVal * o.fVal); }
     SkNf operator / (const SkNf& o) const { return SkNf(fVal / o.fVal); }
@@ skipping 48 matching lines @@
 
 // A reminder alias that shuffles can be used to duplicate a single index across a vector.
 template <int Ix, typename SkNx>
 inline SkNx SkNx_dup(const SkNx& src) { return SkNx_shuffle<Ix>(src); }
 
 }  // namespace
 
 
 // Include platform specific specializations if available.
 #ifndef SKNX_NO_SIMD
-    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX
+        #include "../opts/SkNx_avx.h"
+    #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
         #include "../opts/SkNx_sse.h"
     #elif defined(SK_ARM_HAS_NEON)
         #include "../opts/SkNx_neon.h"
     #endif
 #endif
 
 #undef REQUIRE
 
 typedef SkNf<2> Sk2f;
 typedef SkNf<2> Sk2s;
-
 typedef SkNf<4> Sk4f;
 typedef SkNf<4> Sk4s;
+typedef SkNf<8> Sk8f;
+typedef SkNf<8> Sk8s;
 
 typedef SkNi<8,  uint16_t> Sk8h;
 typedef SkNi<16, uint16_t> Sk16h;
 typedef SkNi<16, uint8_t>  Sk16b;
 
 #endif//SkNx_DEFINED