try to speed-up maprect + round2i + contains

We call roundOut in a few places. If we can get SkNx::Ceil we could efficiently implement that as well.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2133413002
CQ_INCLUDE_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review-Url: https://codereview.chromium.org/2133413002

3 files changed, 1137 insertions, 0 deletions

diff --git a/include/private/SkNx.h b/include/private/SkNx.h
new file mode 100644
index 0000000000..8d1b29dfaa
--- /dev/null
+++ b/include/private/SkNx.h
@@ -0,0 +1,310 @@
+/*
+ * 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
+
+//#define SKNX_NO_SIMD
+
+#include "SkScalar.h"
+#include "SkTypes.h"
+#include <limits>
+#include <math.h>
+#include <type_traits>
+
+#define SI static inline
+
+// The default SkNx<N,T> just proxies down to a pair of SkNx<N/2, T>.
+template <int N, typename T>
+struct SkNx {
+    typedef SkNx<N/2, T> Half;
+
+    Half fLo, fHi;
+
+    SkNx() = default;
+    SkNx(const Half& lo, const Half& hi) : fLo(lo), fHi(hi) {}
+
+    SkNx(T v) : fLo(v), fHi(v) {}
+
+    SkNx(T a, T b)           : fLo(a)  , fHi(b)   { static_assert(N==2, ""); }
+    SkNx(T a, T b, T c, T d) : fLo(a,b), fHi(c,d) { static_assert(N==4, ""); }
+    SkNx(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) {
+        static_assert(N==8, "");
+    }
+    SkNx(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) {
+        static_assert(N==16, "");
+    }
+
+    T operator[](int k) const {
+        SkASSERT(0 <= k && k < N);
+        return k < N/2 ? fLo[k] : fHi[k-N/2];
+    }
+
+    static SkNx Load(const void* vptr) {
+        auto ptr = (const char*)vptr;
+        return { Half::Load(ptr), Half::Load(ptr + N/2*sizeof(T)) };
+    }
+    void store(void* vptr) const {
+        auto ptr = (char*)vptr;
+        fLo.store(ptr);
+        fHi.store(ptr + N/2*sizeof(T));
+    }
+
+    bool anyTrue() const { return fLo.anyTrue() || fHi.anyTrue(); }
+    bool allTrue() const { return fLo.allTrue() && fHi.allTrue(); }
+
+    SkNx    abs() const { return { fLo.   abs(), fHi.   abs() }; }
+    SkNx   sqrt() const { return { fLo.  sqrt(), fHi.  sqrt() }; }
+    SkNx  rsqrt() const { return { fLo. rsqrt(), fHi. rsqrt() }; }
+    SkNx  floor() const { return { fLo. floor(), fHi. floor() }; }
+    SkNx invert() const { return { fLo.invert(), fHi.invert() }; }
+
+    SkNx operator!() const { return { !fLo, !fHi }; }
+    SkNx operator-() const { return { -fLo, -fHi }; }
+    SkNx operator~() const { return { ~fLo, ~fHi }; }
+
+    SkNx operator<<(int bits) const { return { fLo << bits, fHi << bits }; }
+    SkNx operator>>(int bits) const { return { fLo >> bits, fHi >> bits }; }
+
+    SkNx operator+(const SkNx& y) const { return { fLo + y.fLo, fHi + y.fHi }; }
+    SkNx operator-(const SkNx& y) const { return { fLo - y.fLo, fHi - y.fHi }; }
+    SkNx operator*(const SkNx& y) const { return { fLo * y.fLo, fHi * y.fHi }; }
+    SkNx operator/(const SkNx& y) const { return { fLo / y.fLo, fHi / y.fHi }; }
+
+    SkNx operator&(const SkNx& y) const { return { fLo & y.fLo, fHi & y.fHi }; }
+    SkNx operator|(const SkNx& y) const { return { fLo | y.fLo, fHi | y.fHi }; }
+    SkNx operator^(const SkNx& y) const { return { fLo ^ y.fLo, fHi ^ y.fHi }; }
+
+    SkNx operator==(const SkNx& y) const { return { fLo == y.fLo, fHi == y.fHi }; }
+    SkNx operator!=(const SkNx& y) const { return { fLo != y.fLo, fHi != y.fHi }; }
+    SkNx operator<=(const SkNx& y) const { return { fLo <= y.fLo, fHi <= y.fHi }; }
+    SkNx operator>=(const SkNx& y) const { return { fLo >= y.fLo, fHi >= y.fHi }; }
+    SkNx operator< (const SkNx& y) const { return { fLo <  y.fLo, fHi <  y.fHi }; }
+    SkNx operator> (const SkNx& y) const { return { fLo >  y.fLo, fHi >  y.fHi }; }
+
+    SkNx saturatedAdd(const SkNx& y) const {
+        return { fLo.saturatedAdd(y.fLo), fHi.saturatedAdd(y.fHi) };
+    }
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return { fLo.thenElse(t.fLo, e.fLo), fHi.thenElse(t.fHi, e.fHi) };
+    }
+
+    static SkNx Min(const SkNx& x, const SkNx& y) {
+        return { Half::Min(x.fLo, y.fLo), Half::Min(x.fHi, y.fHi) };
+    }
+    static SkNx Max(const SkNx& x, const SkNx& y) {
+        return { Half::Max(x.fLo, y.fLo), Half::Max(x.fHi, y.fHi) };
+    }
+};
+
+// The N -> N/2 recursion bottoms out at N == 1, a scalar value.
+template <typename T>
+struct SkNx<1,T> {
+    T fVal;
+
+    SkNx() = default;
+    SkNx(T v) : fVal(v) {}
+
+    T operator[](int k) const {
+        SkASSERT(k == 0);
+        return fVal;
+    }
+
+    static SkNx Load(const void* ptr) {
+        SkNx v;
+        memcpy(&v, ptr, sizeof(T));
+        return v;
+    }
+    void store(void* ptr) const { memcpy(ptr, &fVal, sizeof(T)); }
+
+    bool anyTrue() const { return fVal != 0; }
+    bool allTrue() const { return fVal != 0; }
+
+    SkNx    abs() const { return Abs(fVal); }
+    SkNx   sqrt() const { return Sqrt(fVal); }
+    SkNx  rsqrt() const { return T(1) / this->sqrt(); }
+    SkNx  floor() const { return Floor(fVal); }
+    SkNx invert() const { return T(1) / *this; }
+
+    SkNx operator!() const { return !fVal; }
+    SkNx operator-() const { return -fVal; }
+    SkNx operator~() const { return FromBits(~ToBits(fVal)); }
+
+    SkNx operator<<(int bits) const { return fVal << bits; }
+    SkNx operator>>(int bits) const { return fVal >> bits; }
+
+    SkNx operator+(const SkNx& y) const { return fVal + y.fVal; }
+    SkNx operator-(const SkNx& y) const { return fVal - y.fVal; }
+    SkNx operator*(const SkNx& y) const { return fVal * y.fVal; }
+    SkNx operator/(const SkNx& y) const { return fVal / y.fVal; }
+
+    SkNx operator&(const SkNx& y) const { return FromBits(ToBits(fVal) & ToBits(y.fVal)); }
+    SkNx operator|(const SkNx& y) const { return FromBits(ToBits(fVal) | ToBits(y.fVal)); }
+    SkNx operator^(const SkNx& y) const { return FromBits(ToBits(fVal) ^ ToBits(y.fVal)); }
+
+    SkNx operator==(const SkNx& y) const { return FromBits(fVal == y.fVal ? ~0 : 0); }
+    SkNx operator!=(const SkNx& y) const { return FromBits(fVal != y.fVal ? ~0 : 0); }
+    SkNx operator<=(const SkNx& y) const { return FromBits(fVal <= y.fVal ? ~0 : 0); }
+    SkNx operator>=(const SkNx& y) const { return FromBits(fVal >= y.fVal ? ~0 : 0); }
+    SkNx operator< (const SkNx& y) const { return FromBits(fVal <  y.fVal ? ~0 : 0); }
+    SkNx operator> (const SkNx& y) const { return FromBits(fVal >  y.fVal ? ~0 : 0); }
+
+    static SkNx Min(const SkNx& x, const SkNx& y) { return x.fVal < y.fVal ? x : y; }
+    static SkNx Max(const SkNx& x, const SkNx& y) { return x.fVal > y.fVal ? x : y; }
+
+    SkNx saturatedAdd(const SkNx& y) const {
+        static_assert(std::is_unsigned<T>::value, "");
+        T sum = fVal + y.fVal;
+        return sum < fVal ? std::numeric_limits<T>::max() : sum;
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const { return fVal != 0 ? t : e; }
+
+private:
+    // Helper functions to choose the right float/double methods.  (In <cmath> madness lies...)
+    static float   Abs(float val) { return  ::fabsf(val); }
+    static float  Sqrt(float val) { return  ::sqrtf(val); }
+    static float Floor(float val) { return ::floorf(val); }
+
+    static double   Abs(double val) { return  ::fabs(val); }
+    static double  Sqrt(double val) { return  ::sqrt(val); }
+    static double Floor(double val) { return ::floor(val); }
+
+    // Helper functions for working with floats/doubles as bit patterns.
+    template <typename U> static U ToBits(U v) { return v; }
+    static int32_t ToBits(float  v) { int32_t bits; memcpy(&bits, &v, sizeof(v)); return bits; }
+    static int64_t ToBits(double v) { int64_t bits; memcpy(&bits, &v, sizeof(v)); return bits; }
+
+    template <typename Bits> static T FromBits(Bits bits) {
+        static_assert(std::is_pod<T   >::value &&
+                      std::is_pod<Bits>::value &&
+                      sizeof(T) <= sizeof(Bits), "");
+        T val;
+        memcpy(&val, &bits, sizeof(T));
+        return val;
+    }
+};
+
+// Allow scalars on the left or right of binary operators, and things like +=, &=, etc.
+#define V template <int N, typename T> SI SkNx<N,T>
+    V operator+ (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) +  y; }
+    V operator- (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) -  y; }
+    V operator* (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) *  y; }
+    V operator/ (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) /  y; }
+    V operator& (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) &  y; }
+    V operator| (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) |  y; }
+    V operator^ (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) ^  y; }
+    V operator==(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) == y; }
+    V operator!=(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) != y; }
+    V operator<=(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) <= y; }
+    V operator>=(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) >= y; }
+    V operator< (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) <  y; }
+    V operator> (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) >  y; }
+
+    V operator+ (const SkNx<N,T>& x, T y) { return x +  SkNx<N,T>(y); }
+    V operator- (const SkNx<N,T>& x, T y) { return x -  SkNx<N,T>(y); }
+    V operator* (const SkNx<N,T>& x, T y) { return x *  SkNx<N,T>(y); }
+    V operator/ (const SkNx<N,T>& x, T y) { return x /  SkNx<N,T>(y); }
+    V operator& (const SkNx<N,T>& x, T y) { return x &  SkNx<N,T>(y); }
+    V operator| (const SkNx<N,T>& x, T y) { return x |  SkNx<N,T>(y); }
+    V operator^ (const SkNx<N,T>& x, T y) { return x ^  SkNx<N,T>(y); }
+    V operator==(const SkNx<N,T>& x, T y) { return x == SkNx<N,T>(y); }
+    V operator!=(const SkNx<N,T>& x, T y) { return x != SkNx<N,T>(y); }
+    V operator<=(const SkNx<N,T>& x, T y) { return x <= SkNx<N,T>(y); }
+    V operator>=(const SkNx<N,T>& x, T y) { return x >= SkNx<N,T>(y); }
+    V operator< (const SkNx<N,T>& x, T y) { return x <  SkNx<N,T>(y); }
+    V operator> (const SkNx<N,T>& x, T y) { return x >  SkNx<N,T>(y); }
+
+    V& operator<<=(SkNx<N,T>& x, int bits) { return (x = x << bits); }
+    V& operator>>=(SkNx<N,T>& x, int bits) { return (x = x >> bits); }
+
+    V& operator +=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x + y); }
+    V& operator -=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x - y); }
+    V& operator *=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x * y); }
+    V& operator /=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x / y); }
+    V& operator &=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x & y); }
+    V& operator |=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x | y); }
+    V& operator ^=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x ^ y); }
+
+    V& operator +=(SkNx<N,T>& x, T y) { return (x = x + SkNx<N,T>(y)); }
+    V& operator -=(SkNx<N,T>& x, T y) { return (x = x - SkNx<N,T>(y)); }
+    V& operator *=(SkNx<N,T>& x, T y) { return (x = x * SkNx<N,T>(y)); }
+    V& operator /=(SkNx<N,T>& x, T y) { return (x = x / SkNx<N,T>(y)); }
+    V& operator &=(SkNx<N,T>& x, T y) { return (x = x & SkNx<N,T>(y)); }
+    V& operator |=(SkNx<N,T>& x, T y) { return (x = x | SkNx<N,T>(y)); }
+    V& operator ^=(SkNx<N,T>& x, T y) { return (x = x ^ SkNx<N,T>(y)); }
+#undef V
+
+// SkNx<N,T> ~~> SkNx<N/2,T> + SkNx<N/2,T>
+template <int N, typename T>
+SI void SkNx_split(const SkNx<N,T>& v, SkNx<N/2,T>* lo, SkNx<N/2,T>* hi) {
+    *lo = v.fLo;
+    *hi = v.fHi;
+}
+
+// SkNx<N/2,T> + SkNx<N/2,T> ~~> SkNx<N,T>
+template <int N, typename T>
+SI SkNx<N*2,T> SkNx_join(const SkNx<N,T>& lo, const SkNx<N,T>& hi) {
+    return { lo, hi };
+}
+
+// A very generic shuffle.  Can reorder, duplicate, contract, expand...
+//    Sk4f v = { R,G,B,A };
+//    SkNx_shuffle<2,1,0,3>(v)         ~~> {B,G,R,A}
+//    SkNx_shuffle<2,1>(v)             ~~> {B,G}
+//    SkNx_shuffle<2,1,2,1,2,1,2,1>(v) ~~> {B,G,B,G,B,G,B,G}
+//    SkNx_shuffle<3,3,3,3>(v)         ~~> {A,A,A,A}
+template <int... Ix, int N, typename T>
+SI SkNx<sizeof...(Ix),T> SkNx_shuffle(const SkNx<N,T>& v) {
+    return { v[Ix]... };
+}
+
+// Cast from SkNx<N, Src> to SkNx<N, Dst>, as if you called static_cast<Dst>(Src).
+template <typename Dst, typename Src, int N>
+SI SkNx<N,Dst> SkNx_cast(const SkNx<N,Src>& v) {
+    return { SkNx_cast<Dst>(v.fLo), SkNx_cast<Dst>(v.fHi) };
+}
+template <typename Dst, typename Src>
+SI SkNx<1,Dst> SkNx_cast(const SkNx<1,Src>& v) {
+    return static_cast<Dst>(v.fVal);
+}
+
+typedef SkNx<2,     float> Sk2f;
+typedef SkNx<4,     float> Sk4f;
+typedef SkNx<8,     float> Sk8f;
+typedef SkNx<16,    float> Sk16f;
+
+typedef SkNx<2,  SkScalar> Sk2s;
+typedef SkNx<4,  SkScalar> Sk4s;
+typedef SkNx<8,  SkScalar> Sk8s;
+typedef SkNx<16, SkScalar> Sk16s;
+
+typedef SkNx<4,   uint8_t> Sk4b;
+typedef SkNx<8,   uint8_t> Sk8b;
+typedef SkNx<16,  uint8_t> Sk16b;
+
+typedef SkNx<4,  uint16_t> Sk4h;
+typedef SkNx<8,  uint16_t> Sk8h;
+typedef SkNx<16, uint16_t> Sk16h;
+
+typedef SkNx<4,       int> Sk4i;
+
+// Include platform specific specializations if available.
+#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    #include "SkNx_sse.h"
+#elif !defined(SKNX_NO_SIMD) && defined(SK_ARM_HAS_NEON)
+    #include "SkNx_neon.h"
+#endif
+
+SI void Sk4f_ToBytes(uint8_t p[16], const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {
+    SkNx_cast<uint8_t>(SkNx_join(SkNx_join(a,b), SkNx_join(c,d))).store(p);
+}
+
+#undef SI
+
+#endif//SkNx_DEFINED
diff --git a/include/private/SkNx_neon.h b/include/private/SkNx_neon.h
new file mode 100644
index 0000000000..bb81cded18
--- /dev/null
+++ b/include/private/SkNx_neon.h
@@ -0,0 +1,453 @@
+/*
+ * 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_neon_DEFINED
+#define SkNx_neon_DEFINED
+
+#include <arm_neon.h>
+
+#define SKNX_IS_FAST
+
+// ARMv8 has vrndmq_f32 to floor 4 floats.  Here we emulate it:
+//   - roundtrip through integers via truncation
+//   - subtract 1 if that's too big (possible for negative values).
+// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
+static inline float32x4_t armv7_vrndmq_f32(float32x4_t v) {
+    auto roundtrip = vcvtq_f32_s32(vcvtq_s32_f32(v));
+    auto too_big = vcgtq_f32(roundtrip, v);
+    return vsubq_f32(roundtrip, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdupq_n_f32(1)));
+}
+
+// Well, this is absurd.  The shifts require compile-time constant arguments.
+
+#define SHIFT8(op, v, bits) switch(bits) { \
+    case  1: return op(v,  1);  case  2: return op(v,  2);  case  3: return op(v,  3); \
+    case  4: return op(v,  4);  case  5: return op(v,  5);  case  6: return op(v,  6); \
+    case  7: return op(v,  7); \
+    } return fVec
+
+#define SHIFT16(op, v, bits) if (bits < 8) { SHIFT8(op, v, bits); } switch(bits) { \
+                                case  8: return op(v,  8);  case  9: return op(v,  9); \
+    case 10: return op(v, 10);  case 11: return op(v, 11);  case 12: return op(v, 12); \
+    case 13: return op(v, 13);  case 14: return op(v, 14);  case 15: return op(v, 15); \
+    } return fVec
+
+#define SHIFT32(op, v, bits) if (bits < 16) { SHIFT16(op, v, bits); } switch(bits) { \
+    case 16: return op(v, 16);  case 17: return op(v, 17);  case 18: return op(v, 18); \
+    case 19: return op(v, 19);  case 20: return op(v, 20);  case 21: return op(v, 21); \
+    case 22: return op(v, 22);  case 23: return op(v, 23);  case 24: return op(v, 24); \
+    case 25: return op(v, 25);  case 26: return op(v, 26);  case 27: return op(v, 27); \
+    case 28: return op(v, 28);  case 29: return op(v, 29);  case 30: return op(v, 30); \
+    case 31: return op(v, 31); } return fVec
+
+template <>
+class SkNx<2, float> {
+public:
+    SkNx(float32x2_t vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(float val)           : fVec(vdup_n_f32(val)) {}
+    static SkNx Load(const void* ptr) { return vld1_f32((const float*)ptr); }
+    SkNx(float a, float b) { fVec = (float32x2_t) { a, b }; }
+
+    void store(void* ptr) const { vst1_f32((float*)ptr, fVec); }
+
+    SkNx invert() const {
+        float32x2_t est0 = vrecpe_f32(fVec),
+                    est1 = vmul_f32(vrecps_f32(est0, fVec), est0);
+        return est1;
+    }
+
+    SkNx operator + (const SkNx& o) const { return vadd_f32(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsub_f32(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return vmul_f32(fVec, o.fVec); }
+    SkNx operator / (const SkNx& o) const {
+    #if defined(SK_CPU_ARM64)
+        return vdiv_f32(fVec, o.fVec);
+    #else
+        float32x2_t est0 = vrecpe_f32(o.fVec),
+                    est1 = vmul_f32(vrecps_f32(est0, o.fVec), est0),
+                    est2 = vmul_f32(vrecps_f32(est1, o.fVec), est1);
+        return vmul_f32(fVec, est2);
+    #endif
+    }
+
+    SkNx operator == (const SkNx& o) const { return vreinterpret_f32_u32(vceq_f32(fVec, o.fVec)); }
+    SkNx operator  < (const SkNx& o) const { return vreinterpret_f32_u32(vclt_f32(fVec, o.fVec)); }
+    SkNx operator  > (const SkNx& o) const { return vreinterpret_f32_u32(vcgt_f32(fVec, o.fVec)); }
+    SkNx operator <= (const SkNx& o) const { return vreinterpret_f32_u32(vcle_f32(fVec, o.fVec)); }
+    SkNx operator >= (const SkNx& o) const { return vreinterpret_f32_u32(vcge_f32(fVec, o.fVec)); }
+    SkNx operator != (const SkNx& o) const {
+        return vreinterpret_f32_u32(vmvn_u32(vceq_f32(fVec, o.fVec)));
+    }
+
+    static SkNx Min(const SkNx& l, const SkNx& r) { return vmin_f32(l.fVec, r.fVec); }
+    static SkNx Max(const SkNx& l, const SkNx& r) { return vmax_f32(l.fVec, r.fVec); }
+
+    SkNx rsqrt() const {
+        float32x2_t est0 = vrsqrte_f32(fVec);
+        return vmul_f32(vrsqrts_f32(fVec, vmul_f32(est0, est0)), est0);
+    }
+
+    SkNx sqrt() const {
+    #if defined(SK_CPU_ARM64)
+        return vsqrt_f32(fVec);
+    #else
+        float32x2_t est0 = vrsqrte_f32(fVec),
+                    est1 = vmul_f32(vrsqrts_f32(fVec, vmul_f32(est0, est0)), est0),
+                    est2 = vmul_f32(vrsqrts_f32(fVec, vmul_f32(est1, est1)), est1);
+        return vmul_f32(fVec, est2);
+    #endif
+    }
+
+    float operator[](int k) const {
+        SkASSERT(0 <= k && k < 2);
+        union { float32x2_t v; float fs[2]; } pun = {fVec};
+        return pun.fs[k&1];
+    }
+
+    bool allTrue() const {
+        auto v = vreinterpret_u32_f32(fVec);
+        return vget_lane_u32(v,0) && vget_lane_u32(v,1);
+    }
+    bool anyTrue() const {
+        auto v = vreinterpret_u32_f32(fVec);
+        return vget_lane_u32(v,0) || vget_lane_u32(v,1);
+    }
+
+    float32x2_t fVec;
+};
+
+template <>
+class SkNx<4, float> {
+public:
+    SkNx(float32x4_t vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(float val)           : fVec(vdupq_n_f32(val)) {}
+    static SkNx Load(const void* ptr) { return vld1q_f32((const float*)ptr); }
+    SkNx(float a, float b, float c, float d) { fVec = (float32x4_t) { a, b, c, d }; }
+
+    void store(void* ptr) const { vst1q_f32((float*)ptr, fVec); }
+    SkNx invert() const {
+        float32x4_t est0 = vrecpeq_f32(fVec),
+                    est1 = vmulq_f32(vrecpsq_f32(est0, fVec), est0);
+        return est1;
+    }
+
+    SkNx operator + (const SkNx& o) const { return vaddq_f32(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsubq_f32(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return vmulq_f32(fVec, o.fVec); }
+    SkNx operator / (const SkNx& o) const {
+    #if defined(SK_CPU_ARM64)
+        return vdivq_f32(fVec, o.fVec);
+    #else
+        float32x4_t est0 = vrecpeq_f32(o.fVec),
+                    est1 = vmulq_f32(vrecpsq_f32(est0, o.fVec), est0),
+                    est2 = vmulq_f32(vrecpsq_f32(est1, o.fVec), est1);
+        return vmulq_f32(fVec, est2);
+    #endif
+    }
+
+    SkNx operator==(const SkNx& o) const { return vreinterpretq_f32_u32(vceqq_f32(fVec, o.fVec)); }
+    SkNx operator <(const SkNx& o) const { return vreinterpretq_f32_u32(vcltq_f32(fVec, o.fVec)); }
+    SkNx operator >(const SkNx& o) const { return vreinterpretq_f32_u32(vcgtq_f32(fVec, o.fVec)); }
+    SkNx operator<=(const SkNx& o) const { return vreinterpretq_f32_u32(vcleq_f32(fVec, o.fVec)); }
+    SkNx operator>=(const SkNx& o) const { return vreinterpretq_f32_u32(vcgeq_f32(fVec, o.fVec)); }
+    SkNx operator!=(const SkNx& o) const {
+        return vreinterpretq_f32_u32(vmvnq_u32(vceqq_f32(fVec, o.fVec)));
+    }
+
+    static SkNx Min(const SkNx& l, const SkNx& r) { return vminq_f32(l.fVec, r.fVec); }
+    static SkNx Max(const SkNx& l, const SkNx& r) { return vmaxq_f32(l.fVec, r.fVec); }
+
+    SkNx abs() const { return vabsq_f32(fVec); }
+    SkNx floor() const {
+    #if defined(SK_CPU_ARM64)
+        return vrndmq_f32(fVec);
+    #else
+        return armv7_vrndmq_f32(fVec);
+    #endif
+    }
+
+
+    SkNx rsqrt() const {
+        float32x4_t est0 = vrsqrteq_f32(fVec);
+        return vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0);
+    }
+
+    SkNx sqrt() const {
+    #if defined(SK_CPU_ARM64)
+        return vsqrtq_f32(fVec);
+    #else
+        float32x4_t est0 = vrsqrteq_f32(fVec),
+                    est1 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0),
+                    est2 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1);
+        return vmulq_f32(fVec, est2);
+    #endif
+    }
+
+    float operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { float32x4_t v; float fs[4]; } pun = {fVec};
+        return pun.fs[k&3];
+    }
+
+    bool allTrue() const {
+        auto v = vreinterpretq_u32_f32(fVec);
+        return vgetq_lane_u32(v,0) && vgetq_lane_u32(v,1)
+            && vgetq_lane_u32(v,2) && vgetq_lane_u32(v,3);
+    }
+    bool anyTrue() const {
+        auto v = vreinterpretq_u32_f32(fVec);
+        return vgetq_lane_u32(v,0) || vgetq_lane_u32(v,1)
+            || vgetq_lane_u32(v,2) || vgetq_lane_u32(v,3);
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return vbslq_f32(vreinterpretq_u32_f32(fVec), t.fVec, e.fVec);
+    }
+
+    float32x4_t fVec;
+};
+
+// It's possible that for our current use cases, representing this as
+// half a uint16x8_t might be better than representing it as a uint16x4_t.
+// It'd make conversion to Sk4b one step simpler.
+template <>
+class SkNx<4, uint16_t> {
+public:
+    SkNx(const uint16x4_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint16_t val) : fVec(vdup_n_u16(val)) {}
+    static SkNx Load(const void* ptr) { return vld1_u16((const uint16_t*)ptr); }
+
+    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) {
+        fVec = (uint16x4_t) { a,b,c,d };
+    }
+
+    void store(void* ptr) const { vst1_u16((uint16_t*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return vadd_u16(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsub_u16(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return vmul_u16(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { SHIFT16(vshl_n_u16, fVec, bits); }
+    SkNx operator >> (int bits) const { SHIFT16(vshr_n_u16, fVec, bits); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) { return vmin_u16(a.fVec, b.fVec); }
+
+    uint16_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { uint16x4_t v; uint16_t us[4]; } pun = {fVec};
+        return pun.us[k&3];
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return vbsl_u16(fVec, t.fVec, e.fVec);
+    }
+
+    uint16x4_t fVec;
+};
+
+template <>
+class SkNx<8, uint16_t> {
+public:
+    SkNx(const uint16x8_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint16_t val) : fVec(vdupq_n_u16(val)) {}
+    static SkNx Load(const void* ptr) { return vld1q_u16((const uint16_t*)ptr); }
+
+    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
+         uint16_t e, uint16_t f, uint16_t g, uint16_t h) {
+        fVec = (uint16x8_t) { a,b,c,d, e,f,g,h };
+    }
+
+    void store(void* ptr) const { vst1q_u16((uint16_t*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return vaddq_u16(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsubq_u16(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return vmulq_u16(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { SHIFT16(vshlq_n_u16, fVec, bits); }
+    SkNx operator >> (int bits) const { SHIFT16(vshrq_n_u16, fVec, bits); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) { return vminq_u16(a.fVec, b.fVec); }
+
+    uint16_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 8);
+        union { uint16x8_t v; uint16_t us[8]; } pun = {fVec};
+        return pun.us[k&7];
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return vbslq_u16(fVec, t.fVec, e.fVec);
+    }
+
+    uint16x8_t fVec;
+};
+
+template <>
+class SkNx<4, uint8_t> {
+public:
+    SkNx(const uint8x8_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d) {
+        fVec = (uint8x8_t){a,b,c,d, 0,0,0,0};
+    }
+    static SkNx Load(const void* ptr) {
+        return (uint8x8_t)vld1_dup_u32((const uint32_t*)ptr);
+    }
+    void store(void* ptr) const {
+        return vst1_lane_u32((uint32_t*)ptr, (uint32x2_t)fVec, 0);
+    }
+    uint8_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { uint8x8_t v; uint8_t us[8]; } pun = {fVec};
+        return pun.us[k&3];
+    }
+
+    // TODO as needed
+
+    uint8x8_t fVec;
+};
+
+template <>
+class SkNx<16, uint8_t> {
+public:
+    SkNx(const uint8x16_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint8_t val) : fVec(vdupq_n_u8(val)) {}
+    static SkNx Load(const void* ptr) { return vld1q_u8((const uint8_t*)ptr); }
+
+    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
+         uint8_t e, uint8_t f, uint8_t g, uint8_t h,
+         uint8_t i, uint8_t j, uint8_t k, uint8_t l,
+         uint8_t m, uint8_t n, uint8_t o, uint8_t p) {
+        fVec = (uint8x16_t) { a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p };
+    }
+
+    void store(void* ptr) const { vst1q_u8((uint8_t*)ptr, fVec); }
+
+    SkNx saturatedAdd(const SkNx& o) const { return vqaddq_u8(fVec, o.fVec); }
+
+    SkNx operator + (const SkNx& o) const { return vaddq_u8(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsubq_u8(fVec, o.fVec); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) { return vminq_u8(a.fVec, b.fVec); }
+    SkNx operator < (const SkNx& o) const { return vcltq_u8(fVec, o.fVec); }
+
+    uint8_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 16);
+        union { uint8x16_t v; uint8_t us[16]; } pun = {fVec};
+        return pun.us[k&15];
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return vbslq_u8(fVec, t.fVec, e.fVec);
+    }
+
+    uint8x16_t fVec;
+};
+
+template <>
+class SkNx<4, int> {
+public:
+    SkNx(const int32x4_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(int v) {
+        fVec = vdupq_n_s32(v);
+    }
+    SkNx(int a, int b, int c, int d) {
+        fVec = (int32x4_t){a,b,c,d};
+    }
+    static SkNx Load(const void* ptr) {
+        return vld1q_s32((const int32_t*)ptr);
+    }
+    void store(void* ptr) const {
+        return vst1q_s32((int32_t*)ptr, fVec);
+    }
+    int operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { int32x4_t v; int is[4]; } pun = {fVec};
+        return pun.is[k&3];
+    }
+
+    SkNx operator + (const SkNx& o) const { return vaddq_s32(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsubq_s32(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return vmulq_s32(fVec, o.fVec); }
+
+    SkNx operator | (const SkNx& o) const { return vorrq_s32(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { SHIFT32(vshlq_n_s32, fVec, bits); }
+    SkNx operator >> (int bits) const { SHIFT32(vshrq_n_s32, fVec, bits); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) { return vminq_s32(a.fVec, b.fVec); }
+    // TODO as needed
+
+    int32x4_t fVec;
+};
+
+#undef SHIFT32
+#undef SHIFT16
+#undef SHIFT8
+
+template<> inline Sk4i SkNx_cast<int, float>(const Sk4f& src) {
+    return vcvtq_s32_f32(src.fVec);
+
+}
+template<> inline Sk4f SkNx_cast<float, int>(const Sk4i& src) {
+    return vcvtq_f32_s32(src.fVec);
+}
+
+template<> inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
+    return vqmovn_u32(vcvtq_u32_f32(src.fVec));
+}
+
+template<> inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) {
+    return vcvtq_f32_u32(vmovl_u16(src.fVec));
+}
+
+template<> inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
+    uint32x4_t _32 = vcvtq_u32_f32(src.fVec);
+    uint16x4_t _16 = vqmovn_u32(_32);
+    return vqmovn_u16(vcombine_u16(_16, _16));
+}
+
+template<> inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
+    uint16x8_t _16 = vmovl_u8 (src.fVec) ;
+    uint32x4_t _32 = vmovl_u16(vget_low_u16(_16));
+    return vcvtq_f32_u32(_32);
+}
+
+template<> inline Sk16b SkNx_cast<uint8_t, float>(const Sk16f& src) {
+    Sk8f ab, cd;
+    SkNx_split(src, &ab, &cd);
+
+    Sk4f a,b,c,d;
+    SkNx_split(ab, &a, &b);
+    SkNx_split(cd, &c, &d);
+    return vuzpq_u8(vuzpq_u8((uint8x16_t)vcvtq_u32_f32(a.fVec),
+                             (uint8x16_t)vcvtq_u32_f32(b.fVec)).val[0],
+                    vuzpq_u8((uint8x16_t)vcvtq_u32_f32(c.fVec),
+                             (uint8x16_t)vcvtq_u32_f32(d.fVec)).val[0]).val[0];
+}
+
+template<> inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
+    return vget_low_u16(vmovl_u8(src.fVec));
+}
+
+template<> inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
+    return vmovn_u16(vcombine_u16(src.fVec, src.fVec));
+}
+
+#endif//SkNx_neon_DEFINED
diff --git a/include/private/SkNx_sse.h b/include/private/SkNx_sse.h
new file mode 100644
index 0000000000..65d9873c5c
--- /dev/null
+++ b/include/private/SkNx_sse.h
@@ -0,0 +1,374 @@
+/*
+ * 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_sse_DEFINED
+#define SkNx_sse_DEFINED
+
+#include <immintrin.h>
+
+// This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent.
+// If you do, make sure this is in a static inline function... anywhere else risks violating ODR.
+
+#define SKNX_IS_FAST
+
+template <>
+class SkNx<2, float> {
+public:
+    SkNx(const __m128& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(float val) : fVec(_mm_set1_ps(val)) {}
+    static SkNx Load(const void* ptr) {
+        return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr));
+    }
+    SkNx(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {}
+
+    void store(void* ptr) const { _mm_storel_pi((__m64*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
+    SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
+
+    SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
+    SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
+    SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
+    SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
+    SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
+    SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
+
+    static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
+    static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
+
+    SkNx   sqrt() const { return _mm_sqrt_ps (fVec);  }
+    SkNx  rsqrt() const { return _mm_rsqrt_ps(fVec); }
+    SkNx invert() const { return _mm_rcp_ps(fVec); }
+
+    float operator[](int k) const {
+        SkASSERT(0 <= k && k < 2);
+        union { __m128 v; float fs[4]; } pun = {fVec};
+        return pun.fs[k&1];
+    }
+
+    bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
+    bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
+
+    __m128 fVec;
+};
+
+template <>
+class SkNx<4, float> {
+public:
+    SkNx(const __m128& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(float val)           : fVec( _mm_set1_ps(val) ) {}
+    static SkNx Load(const void* ptr) { return _mm_loadu_ps((const float*)ptr); }
+
+    SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
+
+    void store(void* ptr) const { _mm_storeu_ps((float*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
+    SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
+
+    SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
+    SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
+    SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
+    SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
+    SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
+    SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
+
+    static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
+    static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
+
+    SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
+    SkNx floor() const {
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
+        return _mm_floor_ps(fVec);
+    #else
+        // Emulate _mm_floor_ps() with SSE2:
+        //   - roundtrip through integers via truncation
+        //   - subtract 1 if that's too big (possible for negative values).
+        // This restricts the domain of our inputs to a maximum somehwere around 2^31.
+        // Seems plenty big.
+        __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(fVec));
+        __m128 too_big = _mm_cmpgt_ps(roundtrip, fVec);
+        return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
+    #endif
+    }
+
+    SkNx   sqrt() const { return _mm_sqrt_ps (fVec);  }
+    SkNx  rsqrt() const { return _mm_rsqrt_ps(fVec); }
+    SkNx invert() const { return _mm_rcp_ps(fVec); }
+
+    float operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { __m128 v; float fs[4]; } pun = {fVec};
+        return pun.fs[k&3];
+    }
+
+    bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(fVec)); }
+    bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(fVec)); }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
+        return _mm_blendv_ps(e.fVec, t.fVec, fVec);
+    #else
+        return _mm_or_ps(_mm_and_ps   (fVec, t.fVec),
+                         _mm_andnot_ps(fVec, e.fVec));
+    #endif
+    }
+
+    __m128 fVec;
+};
+
+template <>
+class SkNx<4, int> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(int val) : fVec(_mm_set1_epi32(val)) {}
+    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
+    SkNx(int a, int b, int c, int d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
+
+    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const {
+        __m128i mul20 = _mm_mul_epu32(fVec, o.fVec),
+                mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o.fVec, 4));
+        return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
+                                  _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
+    }
+
+    SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
+    SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
+
+    int operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { __m128i v; int is[4]; } pun = {fVec};
+        return pun.is[k&3];
+    }
+
+    __m128i fVec;
+};
+
+template <>
+class SkNx<4, uint16_t> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
+    static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
+    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {}
+
+    void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
+    SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
+
+    uint16_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { __m128i v; uint16_t us[8]; } pun = {fVec};
+        return pun.us[k&3];
+    }
+
+    __m128i fVec;
+};
+
+template <>
+class SkNx<8, uint16_t> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
+    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
+    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
+         uint16_t e, uint16_t f, uint16_t g, uint16_t h) : fVec(_mm_setr_epi16(a,b,c,d,e,f,g,h)) {}
+
+    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
+    SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) {
+        // No unsigned _mm_min_epu16, so we'll shift into a space where we can use the
+        // signed version, _mm_min_epi16, then shift back.
+        const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 or MSVC will whine.
+        const __m128i top_8x = _mm_set1_epi16(top);
+        return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x),
+                                                  _mm_sub_epi8(b.fVec, top_8x)));
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
+                            _mm_andnot_si128(fVec, e.fVec));
+    }
+
+    uint16_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 8);
+        union { __m128i v; uint16_t us[8]; } pun = {fVec};
+        return pun.us[k&7];
+    }
+
+    __m128i fVec;
+};
+
+template <>
+class SkNx<4, uint8_t> {
+public:
+    SkNx() {}
+    SkNx(const __m128i& vec) : fVec(vec) {}
+    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d)
+        : fVec(_mm_setr_epi8(a,b,c,d, 0,0,0,0, 0,0,0,0, 0,0,0,0)) {}
+
+
+    static SkNx Load(const void* ptr) { return _mm_cvtsi32_si128(*(const int*)ptr); }
+    void store(void* ptr) const { *(int*)ptr = _mm_cvtsi128_si32(fVec); }
+
+    uint8_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { __m128i v; uint8_t us[16]; } pun = {fVec};
+        return pun.us[k&3];
+    }
+
+    // TODO as needed
+
+    __m128i fVec;
+};
+
+template <>
+class SkNx<16, uint8_t> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
+    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
+    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
+         uint8_t e, uint8_t f, uint8_t g, uint8_t h,
+         uint8_t i, uint8_t j, uint8_t k, uint8_t l,
+         uint8_t m, uint8_t n, uint8_t o, uint8_t p)
+        : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {}
+
+    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
+
+    SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
+    SkNx operator < (const SkNx& o) const {
+        // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
+        auto flip = _mm_set1_epi8(char(0x80));
+        return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec));
+    }
+
+    uint8_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 16);
+        union { __m128i v; uint8_t us[16]; } pun = {fVec};
+        return pun.us[k&15];
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
+                            _mm_andnot_si128(fVec, e.fVec));
+    }
+
+    __m128i fVec;
+};
+
+template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) {
+    return _mm_cvtepi32_ps(src.fVec);
+}
+
+template <> /*static*/ inline Sk4i SkNx_cast<int, float>(const Sk4f& src) {
+    return _mm_cvttps_epi32(src.fVec);
+}
+
+template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
+    auto _32 = _mm_cvttps_epi32(src.fVec);
+    // Ideally we'd use _mm_packus_epi32 here.  But that's SSE4.1+.
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+    // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
+    const int _ = ~0;
+    return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
+#else
+    // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32:
+    _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000));
+    return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000));
+#endif
+}
+
+template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
+    auto _32 = _mm_cvttps_epi32(src.fVec);
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+    const int _ = ~0;
+    return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_));
+#else
+    auto _16 = _mm_packus_epi16(_32, _32);
+    return     _mm_packus_epi16(_16, _16);
+#endif
+}
+
+template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+    const int _ = ~0;
+    auto _32 = _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_));
+#else
+    auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()),
+         _32 = _mm_unpacklo_epi16(_16,     _mm_setzero_si128());
+#endif
+    return _mm_cvtepi32_ps(_32);
+}
+
+template<> /*static*/ inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) {
+    auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
+    return _mm_cvtepi32_ps(_32);
+}
+
+template<> /*static*/ inline Sk16b SkNx_cast<uint8_t, float>(const Sk16f& src) {
+    Sk8f ab, cd;
+    SkNx_split(src, &ab, &cd);
+
+    Sk4f a,b,c,d;
+    SkNx_split(ab, &a, &b);
+    SkNx_split(cd, &c, &d);
+
+    return _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec),
+                                             _mm_cvttps_epi32(b.fVec)),
+                            _mm_packus_epi16(_mm_cvttps_epi32(c.fVec),
+                                             _mm_cvttps_epi32(d.fVec)));
+}
+
+template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
+    return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
+}
+
+template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
+    return _mm_packus_epi16(src.fVec, src.fVec);
+}
+
+#endif//SkNx_sse_DEFINED