Unify some SkNx code

 - one base case and one N=1 case instead of two each (or three with doubles)
 - use SkNx_cast instead of FromBytes/toBytes
 - 4-at-a-time Sk4f::ToBytes becomes a special standalone Sk4f_ToBytes

If I did everything right, this'll be perf- and pixel- neutral.

https://gold.skia.org/search2?issue=1526523003&unt=true&query=source_type%3Dgm&master=false

BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review URL: https://codereview.chromium.org/1526523003

11 files changed, 188 insertions, 249 deletions

diff --git a/bench/Sk4fBench.cpp b/bench/Sk4fBench.cpp
index b2f2b4a58e..1daa3f2973 100644
--- a/bench/Sk4fBench.cpp
+++ b/bench/Sk4fBench.cpp
@@ -33,8 +33,8 @@ struct Sk4fBytesRoundtripBench : public Benchmark {
         for (int i = 0; i < loops; i++) {
             uint32_t color = lcg_rand(&seed),
                      back;
-            auto f = Sk4f::FromBytes((const uint8_t*)&color);
-            f.toBytes((uint8_t*)&back);
+            auto f = SkNx_cast<float>(Sk4b::Load((const uint8_t*)&color));
+            SkNx_cast<uint8_t>(f).store((uint8_t*)&back);
             junk ^= back;
         }
         blackhole ^= junk;
@@ -62,7 +62,7 @@ struct Sk4fGradientBench : public Benchmark {
                  c = b + dcdx,
                  d = c + dcdx;
             for (size_t i = 0; i < SK_ARRAY_COUNT(fDevice); i += 4) {
-                Sk4f::ToBytes((uint8_t*)(fDevice+i), a, b, c, d);
+                Sk4f_ToBytes((uint8_t*)(fDevice+i), a, b, c, d);
                 a = a + dcdx4;
                 b = b + dcdx4;
                 c = c + dcdx4;
diff --git a/src/core/SkNx.h b/src/core/SkNx.h
index 36a645cac1..a28898b258 100644
--- a/src/core/SkNx.h
+++ b/src/core/SkNx.h
@@ -9,8 +9,7 @@
 #define SkNx_DEFINED
 
 
-#define SKNX_NO_SIMDx  // Remove the x to disable SIMD for all SkNx types.
-
+//#define SKNX_NO_SIMD
 
 #include "SkScalar.h"
 #include "SkTypes.h"
@@ -25,7 +24,8 @@
 namespace {
 
 // The default implementations just fall back on a pair of size N/2.
-
+// These support the union of operations we might do to ints and floats, but
+// platform specializations might support fewer (e.g. no float <<, no int /).
 template <int N, typename T>
 class SkNx {
 public:
@@ -55,78 +55,11 @@ public:
     SkNx operator + (const SkNx& o) const { return SkNx(fLo + o.fLo, fHi + o.fHi); }
     SkNx operator - (const SkNx& o) const { return SkNx(fLo - o.fLo, fHi - o.fHi); }
     SkNx operator * (const SkNx& o) const { return SkNx(fLo * o.fLo, fHi * o.fHi); }
+    SkNx operator / (const SkNx& o) const { return SkNx(fLo / o.fLo, fHi / o.fHi); }
 
     SkNx operator << (int bits) const { return SkNx(fLo << bits, fHi << bits); }
     SkNx operator >> (int bits) const { return SkNx(fLo >> bits, fHi >> bits); }
 
-    static SkNx Min(const SkNx& a, const SkNx& b) {
-        return SkNx(SkNx<N/2, T>::Min(a.fLo, b.fLo), SkNx<N/2, T>::Min(a.fHi, b.fHi));
-    }
-    SkNx operator < (const SkNx& o) const { return SkNx(fLo < o.fLo, fHi < o.fHi); }
-
-    template <int k> T kth() const {
-        SkASSERT(0 <= k && k < N);
-        return k < N/2 ? fLo.template kth<k>() : fHi.template kth<k-N/2>();
-    }
-
-    bool allTrue() const { return fLo.allTrue() && fHi.allTrue(); }
-    bool anyTrue() const { return fLo.anyTrue() || fHi.anyTrue(); }
-    SkNx thenElse(const SkNx& t, const SkNx& e) const {
-        return SkNx(fLo.thenElse(t.fLo, e.fLo), fHi.thenElse(t.fHi, e.fHi));
-    }
-
-protected:
-    REQUIRE(0 == (N & (N-1)));
-
-    SkNx<N/2, T> fLo, fHi;
-};
-
-template <int N>
-class SkNx<N,float> {
-public:
-    SkNx() {}
-    SkNx(float val) : fLo(val),  fHi(val) {}
-    static SkNx Load(const float vals[N]) {
-        return SkNx(SkNx<N/2, float>::Load(vals), SkNx<N/2, float>::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 SkNx FromBytes(const uint8_t bytes[N]) {
-        return SkNx(SkNx<N/2, float>::FromBytes(bytes), SkNx<N/2, float>::FromBytes(bytes+N/2));
-    }
-
-    SkNx(float a, float b)                   : fLo(a),   fHi(b)   { REQUIRE(N==2); }
-    SkNx(float a, float b, float c, float d) : fLo(a,b), fHi(c,d) { REQUIRE(N==4); }
-    SkNx(float a, float b, float c, float d, float e, float f, float g, float h)
-        : fLo(a,b,c,d)
-        , fHi(e,f,g,h) { REQUIRE(N==8); }
-
-    void store(float vals[N]) const {
-        fLo.store(vals);
-        fHi.store(vals+N/2);
-    }
-    // Please see note on FromBytes().
-    // Clamps to [0.0,255.0] floats and truncates to [0,255] bytes.
-    void toBytes(uint8_t bytes[N]) const {
-        fLo.toBytes(bytes);
-        fHi.toBytes(bytes+N/2);
-    }
-
-    // Some implementations can do this faster.
-    static void ToBytes(uint8_t bytes[4*N],
-                        const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
-        a.toBytes(bytes+0*N);
-        b.toBytes(bytes+1*N);
-        c.toBytes(bytes+2*N);
-        d.toBytes(bytes+3*N);
-    }
-
-    SkNx operator + (const SkNx& o) const { return SkNx(fLo + o.fLo, fHi + o.fHi); }
-    SkNx operator - (const SkNx& o) const { return SkNx(fLo - o.fLo, fHi - o.fHi); }
-    SkNx operator * (const SkNx& o) const { return SkNx(fLo * o.fLo, fHi * o.fHi); }
-    SkNx operator / (const SkNx& o) const { return SkNx(fLo / o.fLo, fHi / o.fHi); }
-
     SkNx operator == (const SkNx& o) const { return SkNx(fLo == o.fLo, fHi == o.fHi); }
     SkNx operator != (const SkNx& o) const { return SkNx(fLo != o.fLo, fHi != o.fHi); }
     SkNx operator  < (const SkNx& o) const { return SkNx(fLo  < o.fLo, fHi  < o.fHi); }
@@ -134,15 +67,14 @@ public:
     SkNx operator <= (const SkNx& o) const { return SkNx(fLo <= o.fLo, fHi <= o.fHi); }
     SkNx operator >= (const SkNx& o) const { return SkNx(fLo >= o.fLo, fHi >= o.fHi); }
 
-    static SkNx Min(const SkNx& l, const SkNx& r) {
-        return SkNx(SkNx<N/2, float>::Min(l.fLo, r.fLo), SkNx<N/2, float>::Min(l.fHi, r.fHi));
+    static SkNx Min(const SkNx& a, const SkNx& b) {
+        return SkNx(SkNx<N/2, T>::Min(a.fLo, b.fLo), SkNx<N/2, T>::Min(a.fHi, b.fHi));
     }
-    static SkNx Max(const SkNx& l, const SkNx& r) {
-        return SkNx(SkNx<N/2, float>::Max(l.fLo, r.fLo), SkNx<N/2, float>::Max(l.fHi, r.fHi));
+    static SkNx Max(const SkNx& a, const SkNx& b) {
+        return SkNx(SkNx<N/2, T>::Max(a.fLo, b.fLo), SkNx<N/2, T>::Max(a.fHi, b.fHi));
     }
 
-    SkNx  sqrt() const { return SkNx(fLo. sqrt(), fHi. sqrt()); }
-
+    SkNx sqrt() const { return SkNx(fLo.sqrt(), fHi.sqrt()); }
     // Generally, increasing precision, increasing cost.
     SkNx rsqrt0() const { return SkNx(fLo.rsqrt0(), fHi.rsqrt0()); }
     SkNx rsqrt1() const { return SkNx(fLo.rsqrt1(), fHi.rsqrt1()); }
@@ -151,7 +83,7 @@ public:
     SkNx       invert() const { return SkNx(fLo.      invert(), fHi.      invert()); }
     SkNx approxInvert() const { return SkNx(fLo.approxInvert(), fHi.approxInvert()); }
 
-    template <int k> float kth() const {
+    template <int k> T kth() const {
         SkASSERT(0 <= k && k < N);
         return k < N/2 ? fLo.template kth<k>() : fHi.template kth<k-N/2>();
     }
@@ -164,14 +96,11 @@ public:
 
 protected:
     REQUIRE(0 == (N & (N-1)));
-    SkNx(const SkNx<N/2, float>& lo, const SkNx<N/2, float>& hi) : fLo(lo), fHi(hi) {}
 
-    SkNx<N/2, float> fLo, fHi;
+    SkNx<N/2, T> fLo, fHi;
 };
 
-
 // Bottom out the default implementations with scalars when nothing's been specialized.
-
 template <typename T>
 class SkNx<1,T> {
 public:
@@ -190,42 +119,11 @@ public:
     SkNx operator + (const SkNx& o) const { return SkNx(fVal + o.fVal); }
     SkNx operator - (const SkNx& o) const { return SkNx(fVal - o.fVal); }
     SkNx operator * (const SkNx& o) const { return SkNx(fVal * o.fVal); }
+    SkNx operator / (const SkNx& o) const { return SkNx(fVal / o.fVal); }
 
     SkNx operator << (int bits) const { return SkNx(fVal << bits); }
     SkNx operator >> (int bits) const { return SkNx(fVal >> bits); }
 
-    static SkNx Min(const SkNx& a, const SkNx& b) { return SkNx(SkTMin(a.fVal, b.fVal)); }
-    SkNx operator <(const SkNx& o) const { return SkNx(fVal < o.fVal); }
-
-    template <int k> T kth() const {
-        SkASSERT(0 == k);
-        return fVal;
-    }
-
-    bool allTrue() const { return fVal; }
-    bool anyTrue() const { return fVal; }
-    SkNx thenElse(const SkNx& t, const SkNx& e) const { return fVal ? t : e; }
-
-protected:
-    T fVal;
-};
-
-template <>
-class SkNx<1,float> {
-public:
-    SkNx() {}
-    SkNx(float val) : fVal(val) {}
-    static SkNx Load(const float vals[1]) { return SkNx(vals[0]); }
-    static SkNx FromBytes(const uint8_t bytes[1]) { return SkNx((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)); }
-
-    SkNx operator + (const SkNx& o) const { return SkNx(fVal + o.fVal); }
-    SkNx operator - (const SkNx& o) const { return SkNx(fVal - o.fVal); }
-    SkNx operator * (const SkNx& o) const { return SkNx(fVal * o.fVal); }
-    SkNx operator / (const SkNx& o) const { return SkNx(fVal / o.fVal); }
-
     SkNx operator == (const SkNx& o) const { return SkNx(fVal == o.fVal); }
     SkNx operator != (const SkNx& o) const { return SkNx(fVal != o.fVal); }
     SkNx operator  < (const SkNx& o) const { return SkNx(fVal  < o.fVal); }
@@ -233,33 +131,31 @@ public:
     SkNx operator <= (const SkNx& o) const { return SkNx(fVal <= o.fVal); }
     SkNx operator >= (const SkNx& o) const { return SkNx(fVal >= o.fVal); }
 
-    static SkNx Min(const SkNx& l, const SkNx& r) { return SkNx(SkTMin(l.fVal, r.fVal)); }
-    static SkNx Max(const SkNx& l, const SkNx& r) { return SkNx(SkTMax(l.fVal, r.fVal)); }
+    static SkNx Min(const SkNx& a, const SkNx& b) { return SkNx(SkTMin(a.fVal, b.fVal)); }
+    static SkNx Max(const SkNx& a, const SkNx& b) { return SkNx(SkTMax(a.fVal, b.fVal)); }
 
-    SkNx  sqrt() const { return SkNx(sqrtf(fVal));        }
-    SkNx rsqrt0() const { return SkNx(1.0f / sqrtf(fVal)); }
+    SkNx  sqrt () const { return SkNx(Sqrt(fVal)); }
+    SkNx rsqrt0() const { return this->sqrt().invert();  }
     SkNx rsqrt1() const { return this->rsqrt0(); }
     SkNx rsqrt2() const { return this->rsqrt1(); }
 
-    SkNx       invert() const { return SkNx(1.0f / fVal); }
+    SkNx       invert() const { return SkNx(1) / SkNx(fVal); }
     SkNx approxInvert() const { return this->invert();    }
 
-    template <int k> float kth() const {
-        SkASSERT(k == 0);
+    template <int k> T kth() const {
+        SkASSERT(0 == k);
         return fVal;
     }
 
-    bool allTrue() const { return this->pun() != 0; }
-    bool anyTrue() const { return this->pun() != 0; }
-    SkNx thenElse(const SkNx& t, const SkNx& e) const { return this->pun() ? t : e; }
+    bool allTrue() const { return fVal != 0; }
+    bool anyTrue() const { return fVal != 0; }
+    SkNx thenElse(const SkNx& t, const SkNx& e) const { return fVal != 0 ? t : e; }
 
 protected:
-    uint32_t pun() const {
-        union { float f; uint32_t i; } pun = { fVal };
-        return pun.i;
-    }
+    static double Sqrt(double val) { return ::sqrt (val); }
+    static float  Sqrt(float  val) { return ::sqrtf(val); }
 
-    float fVal;
+    T fVal;
 };
 
 // This default implementation can be specialized by ../opts/SkNx_foo.h
@@ -296,7 +192,7 @@ SkNx<N,D> SkNx_cast_fallback(const SkNx<N,S>& src, SkIntSequence<Ix...>) {
 
 // This is a generic cast between two SkNx with the same number of elements N.  E.g.
 //   Sk4b bs = ...;                    // Load 4 bytes.
-//   Sk4f fs = SkNx_cast<float>(bs);   // (This will replace SkNf::FromBytes() one day.)
+//   Sk4f fs = SkNx_cast<float>(bs);   // Cast each byte to a float.
 //   Sk4i is = SkNx_cast<int>(fs);     // Cast each float to int.
 // This can be specialized in ../opts/SkNx_foo.h if there's a better platform-specific cast.
 template <typename D, typename S, int N>
@@ -306,20 +202,6 @@ SkNx<N,D> SkNx_cast(const SkNx<N,S>& src) {
 
 }  // namespace
 
-
-// Include platform specific specializations if available.
-#ifndef SKNX_NO_SIMD
-    #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 SkNx<2, float> Sk2f;
 typedef SkNx<2, float> Sk2s;
 typedef SkNx<4, float> Sk4f;
@@ -327,10 +209,34 @@ typedef SkNx<4, float> Sk4s;
 typedef SkNx<8, float> Sk8f;
 typedef SkNx<8, float> Sk8s;
 
-typedef SkNx<8,  uint16_t> Sk8h;
+typedef SkNx< 4, uint16_t> Sk4h;
+typedef SkNx< 8, uint16_t> Sk8h;
 typedef SkNx<16, uint16_t> Sk16h;
+
+typedef SkNx< 4, uint8_t>  Sk4b;
+typedef SkNx< 8, uint8_t>  Sk8b;
 typedef SkNx<16, uint8_t>  Sk16b;
 
 typedef SkNx<4, int> Sk4i;
 
+// Include platform specific specializations if available.
+#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX
+    #include "../opts/SkNx_avx.h"
+#elif !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    #include "../opts/SkNx_sse.h"
+#elif !defined(SKNX_NO_SIMD) && defined(SK_ARM_HAS_NEON)
+    #include "../opts/SkNx_neon.h"
+#else
+    static inline
+    void Sk4f_ToBytes(uint8_t p[16], const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {
+        SkNx_cast<uint8_t>(a).store(p+ 0);
+        SkNx_cast<uint8_t>(b).store(p+ 4);
+        SkNx_cast<uint8_t>(c).store(p+ 8);
+        SkNx_cast<uint8_t>(d).store(p+12);
+    }
+#endif
+
+#undef REQUIRE
+
+
 #endif//SkNx_DEFINED
diff --git a/src/effects/SkColorMatrixFilter.cpp b/src/effects/SkColorMatrixFilter.cpp
index 3f0f12f2d8..a7b2b7a2f7 100644
--- a/src/effects/SkColorMatrixFilter.cpp
+++ b/src/effects/SkColorMatrixFilter.cpp
@@ -258,7 +258,7 @@ static Sk4f clamp_0_1(const Sk4f& x) {
 
 static SkPMColor round(const Sk4f& x) {
     SkPMColor c;
-    (x * Sk4f(255) + Sk4f(0.5f)).toBytes((uint8_t*)&c);
+    SkNx_cast<uint8_t>(x * Sk4f(255) + Sk4f(0.5f)).store((uint8_t*)&c);
     return c;
 }
 
@@ -296,7 +296,7 @@ void SkColorMatrixFilter::filterSpan(const SkPMColor src[], int count, SkPMColor
                 continue;
             }
 
-            Sk4f srcf = Sk4f::FromBytes((const uint8_t*)&src_c) * Sk4f(1.0f/255);
+            Sk4f srcf = SkNx_cast<float>(Sk4b::Load((const uint8_t*)&src_c)) * Sk4f(1.0f/255);
 
             if (0xFF != SkGetPackedA32(src_c)) {
                 srcf = unpremul(srcf);
diff --git a/src/effects/gradients/SkLinearGradient.cpp b/src/effects/gradients/SkLinearGradient.cpp
index 53264e4d8c..8760029795 100644
--- a/src/effects/gradients/SkLinearGradient.cpp
+++ b/src/effects/gradients/SkLinearGradient.cpp
@@ -150,7 +150,7 @@ SkLinearGradient::LinearGradientContext::LinearGradientContext(
         const Sk4f scale(1, 1, 1, paintAlpha);
         for (int i = 0; i < count; ++i) {
             uint32_t c = SkSwizzle_Color_to_PMColor(shader.fOrigColors[i]);
-            rec[i].fColor = Sk4f::FromBytes((const uint8_t*)&c) * scale;
+            rec[i].fColor = SkNx_cast<float>(Sk4b::Load((const uint8_t*)&c)) * scale;
             if (i > 0) {
                 SkASSERT(rec[i - 1].fPos <= rec[i].fPos);
             }
@@ -162,7 +162,7 @@ SkLinearGradient::LinearGradientContext::LinearGradientContext(
         for (int i = 0; i < count; ++i) {
             SkPMColor pmc = SkPreMultiplyColor(shader.fOrigColors[i]);
             pmc = SkAlphaMulQ(pmc, alphaScale);
-            rec[i].fColor = Sk4f::FromBytes((const uint8_t*)&pmc);
+            rec[i].fColor = SkNx_cast<float>(Sk4b::Load((const uint8_t*)&pmc));
             if (i > 0) {
                 SkASSERT(rec[i - 1].fPos <= rec[i].fPos);
             }
@@ -699,7 +699,7 @@ find_backward(const SkLinearGradient::LinearGradientContext::Rec rec[], float ti
 
 template <bool apply_alpha> SkPMColor trunc_from_255(const Sk4f& x) {
     SkPMColor c;
-    x.toBytes((uint8_t*)&c);
+    SkNx_cast<uint8_t>(x).store((uint8_t*)&c);
     if (apply_alpha) {
         c = SkPreMultiplyARGB(SkGetPackedA32(c), SkGetPackedR32(c),
                               SkGetPackedG32(c), SkGetPackedB32(c));
@@ -751,7 +751,7 @@ template <bool apply_alpha> void ramp(SkPMColor dstC[], int n, const Sk4f& c, co
     Sk4f cd3 = cd1 + dc2;
     while (n >= 4) {
         if (!apply_alpha) {
-            Sk4f::ToBytes((uint8_t*)dstC, cd0, cd1, cd2, cd3);
+            Sk4f_ToBytes((uint8_t*)dstC, cd0, cd1, cd2, cd3);
             dstC += 4;
         } else {
             *dstC++ = trunc_from_255<apply_alpha>(cd0);
diff --git a/src/effects/gradients/SkRadialGradient.cpp b/src/effects/gradients/SkRadialGradient.cpp
index 52d06392e1..d734aa056a 100644
--- a/src/effects/gradients/SkRadialGradient.cpp
+++ b/src/effects/gradients/SkRadialGradient.cpp
@@ -307,7 +307,7 @@ void shadeSpan_radial_clamp2(SkScalar sfx, SkScalar sdx, SkScalar sfy, SkScalar
             dR = dR + ddR;
 
             uint8_t fi[4];
-            dist.toBytes(fi);
+            SkNx_cast<uint8_t>(dist).store(fi);
 
             for (int i = 0; i < 4; i++) {
                 *dstC++ = cache[toggle + fi[i]];
@@ -319,7 +319,7 @@ void shadeSpan_radial_clamp2(SkScalar sfx, SkScalar sdx, SkScalar sfy, SkScalar
             Sk4f dist = Sk4f::Min(fast_sqrt(R), max);
 
             uint8_t fi[4];
-            dist.toBytes(fi);
+            SkNx_cast<uint8_t>(dist).store(fi);
             for (int i = 0; i < count; i++) {
                 *dstC++ = cache[toggle + fi[i]];
                 toggle = next_dither_toggle(toggle);
diff --git a/src/opts/SkColorCubeFilter_opts.h b/src/opts/SkColorCubeFilter_opts.h
index 4c394051bb..ab8d1d4c9b 100644
--- a/src/opts/SkColorCubeFilter_opts.h
+++ b/src/opts/SkColorCubeFilter_opts.h
@@ -59,10 +59,10 @@ void color_cube_filter_span(const SkPMColor src[],
             const SkColor lutColor10 = colorCube[ix + i10];
             const SkColor lutColor11 = colorCube[ix + i11];
 
-            Sk4f  sum = Sk4f::FromBytes((const uint8_t*)&lutColor00) * g0b0;
-            sum = sum + Sk4f::FromBytes((const uint8_t*)&lutColor01) * g0b1;
-            sum = sum + Sk4f::FromBytes((const uint8_t*)&lutColor10) * g1b0;
-            sum = sum + Sk4f::FromBytes((const uint8_t*)&lutColor11) * g1b1;
+            Sk4f  sum = SkNx_cast<float>(Sk4b::Load((const uint8_t*)&lutColor00)) * g0b0;
+            sum = sum + SkNx_cast<float>(Sk4b::Load((const uint8_t*)&lutColor01)) * g0b1;
+            sum = sum + SkNx_cast<float>(Sk4b::Load((const uint8_t*)&lutColor10)) * g1b0;
+            sum = sum + SkNx_cast<float>(Sk4b::Load((const uint8_t*)&lutColor11)) * g1b1;
             color = color + sum * Sk4f((float)colorToFactors[x][r]);
         }
         if (a != 255) {
@@ -74,7 +74,7 @@ void color_cube_filter_span(const SkPMColor src[],
         color = SkNx_shuffle<2,1,0,3>(color);
     #endif
         uint8_t* dstBytes = (uint8_t*)(dst+i);
-        color.toBytes(dstBytes);
+        SkNx_cast<uint8_t>(color).store(dstBytes);
         dstBytes[SK_A32_SHIFT/8] = a;
     }
 }
diff --git a/src/opts/SkNx_avx.h b/src/opts/SkNx_avx.h
index 6236769652..f635181a92 100644
--- a/src/opts/SkNx_avx.h
+++ b/src/opts/SkNx_avx.h
@@ -26,27 +26,10 @@ public:
     SkNx(float val) : fVec(_mm256_set1_ps(val)) {}
     static SkNx Load(const float vals[8]) { return _mm256_loadu_ps(vals); }
 
-    static SkNx FromBytes(const uint8_t bytes[8]) {
-        __m128i fix8  = _mm_loadl_epi64((const __m128i*)bytes),
-                fix16 = _mm_unpacklo_epi8 (fix8 , _mm_setzero_si128()),
-                 lo32 = _mm_unpacklo_epi16(fix16, _mm_setzero_si128()),
-                 hi32 = _mm_unpackhi_epi16(fix16, _mm_setzero_si128());
-        __m256i fix32 = _mm256_insertf128_si256(_mm256_castsi128_si256(lo32), hi32, 1);
-        return _mm256_cvtepi32_ps(fix32);
-    }
-
     SkNx(float a, float b, float c, float d,
          float e, float f, float g, float h) : fVec(_mm256_setr_ps(a,b,c,d,e,f,g,h)) {}
 
     void store(float vals[8]) const { _mm256_storeu_ps(vals, fVec); }
-    void toBytes(uint8_t bytes[8]) const {
-        __m256i fix32 = _mm256_cvttps_epi32(fVec);
-        __m128i  lo32 = _mm256_extractf128_si256(fix32, 0),
-                 hi32 = _mm256_extractf128_si256(fix32, 1),
-                fix16 = _mm_packus_epi32(lo32, hi32),
-                fix8  = _mm_packus_epi16(fix16, fix16);
-        _mm_storel_epi64((__m128i*)bytes, fix8);
-    }
 
     SkNx operator + (const SkNx& o) const { return _mm256_add_ps(fVec, o.fVec); }
     SkNx operator - (const SkNx& o) const { return _mm256_sub_ps(fVec, o.fVec); }
@@ -87,6 +70,25 @@ public:
     __m256 fVec;
 };
 
+template<> inline Sk8b SkNx_cast<uint8_t, float, 8>(const Sk8f& src) {
+    __m256i _32 = _mm256_cvttps_epi32(src.fVec);
+    __m128i  lo = _mm256_extractf128_si256(_32, 0),
+             hi = _mm256_extractf128_si256(_32, 1),
+            _16 = _mm_packus_epi32(lo, hi);
+    return _mm_packus_epi16(_16, _16);
+}
+
+template<> inline Sk8f SkNx_cast<float, uint8_t, 8>(const Sk8b& src) {
+    /* TODO lo = _mm_cvtepu8_epi32(src.fVec),
+     *      hi = _mm_cvtepu8_epi32(_mm_srli_si128(src.fVec, 4))
+     */
+    __m128i _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()),
+             lo = _mm_unpacklo_epi16(_16, _mm_setzero_si128()),
+             hi = _mm_unpackhi_epi16(_16, _mm_setzero_si128());
+    __m256i _32 = _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1);
+    return _mm256_cvtepi32_ps(_32);
+}
+
 }  // namespace
 
 #endif//SkNx_avx_DEFINED
diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index 43409bee44..0955cb2b5b 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -150,31 +150,9 @@ public:
     SkNx() {}
     SkNx(float val)           : fVec(vdupq_n_f32(val)) {}
     static SkNx Load(const float vals[4]) { return vld1q_f32(vals); }
-    static SkNx FromBytes(const uint8_t vals[4]) {
-        uint8x8_t   fix8    = (uint8x8_t)vld1_dup_u32((const uint32_t*)vals);
-        uint16x8_t  fix8_16 = vmovl_u8(fix8);
-        uint32x4_t  fix8_32 = vmovl_u16(vget_low_u16(fix8_16));
-        return SkNx(vcvtq_f32_u32(fix8_32));
-    }
-
     SkNx(float a, float b, float c, float d) { fVec = (float32x4_t) { a, b, c, d }; }
 
     void store(float vals[4]) const { vst1q_f32(vals, fVec); }
-    void toBytes(uint8_t bytes[4]) const {
-        uint32x4_t  fix8_32 = vcvtq_u32_f32(fVec);
-        uint16x4_t  fix8_16 = vqmovn_u32(fix8_32);
-        uint8x8_t   fix8    = vqmovn_u16(vcombine_u16(fix8_16, vdup_n_u16(0)));
-        vst1_lane_u32((uint32_t*)bytes, (uint32x2_t)fix8, 0);
-    }
-
-    static void ToBytes(uint8_t bytes[16],
-                        const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
-        vst1q_u8(bytes, 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]);
-    }
-
     SkNx approxInvert() const {
         float32x4_t est0 = vrecpeq_f32(fVec),
                     est1 = vmulq_f32(vrecpsq_f32(est0, fVec), est0);
@@ -288,6 +266,24 @@ public:
 };
 
 template <>
+class SkNx<4, uint8_t> {
+public:
+    SkNx(const uint8x8_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    static SkNx Load(const uint8_t vals[4]) {
+        return (uint8x8_t)vld1_dup_u32((const uint32_t*)vals);
+    }
+    void store(uint8_t vals[4]) const {
+        return vst1_lane_u32((uint32_t*)vals, (uint32x2_t)fVec, 0);
+    }
+
+    // TODO as needed
+
+    uint8x8_t fVec;
+};
+
+template <>
 class SkNx<16, uint8_t> {
 public:
     SkNx(const uint8x16_t& vec) : fVec(vec) {}
@@ -329,11 +325,30 @@ public:
 #undef SHIFT16
 #undef SHIFT8
 
-template<>
-inline SkNx<4, int> SkNx_cast<int, float, 4>(const SkNx<4, float>& src) {
+template<> inline Sk4i SkNx_cast<int, float, 4>(const Sk4f& src) {
     return vcvtq_s32_f32(src.fVec);
 }
 
+template<> inline Sk4b SkNx_cast<uint8_t, float, 4>(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, 4>(const Sk4b& src) {
+    uint16x8_t _16 = vmovl_u8 (src.fVec) ;
+    uint32x4_t _32 = vmovl_u16(vget_low_u16(_16));
+    return vcvtq_f32_u32(_32);
+}
+
+static inline void Sk4f_ToBytes(uint8_t bytes[16],
+                                const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {
+    vst1q_u8(bytes, 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]);
+}
+
 }  // namespace
 
 #endif//SkNx_neon_DEFINED
diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index 9092f44480..f3952a9768 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -111,37 +111,9 @@ public:
     SkNx(float val)           : fVec( _mm_set1_ps(val) ) {}
     static SkNx Load(const float vals[4]) { return _mm_loadu_ps(vals); }
 
-    static SkNx FromBytes(const uint8_t bytes[4]) {
-        __m128i fix8 = _mm_cvtsi32_si128(*(const int*)bytes);
-    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
-        const char _ = ~0;  // Zero these bytes.
-        __m128i fix8_32 = _mm_shuffle_epi8(fix8, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_));
-    #else
-        __m128i fix8_16 = _mm_unpacklo_epi8 (fix8,    _mm_setzero_si128()),
-                fix8_32 = _mm_unpacklo_epi16(fix8_16, _mm_setzero_si128());
-    #endif
-        return SkNx(_mm_cvtepi32_ps(fix8_32));
-        // TODO: use _mm_cvtepu8_epi32 w/SSE4.1?
-    }
-
     SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
 
     void store(float vals[4]) const { _mm_storeu_ps(vals, fVec); }
-    void toBytes(uint8_t bytes[4]) const {
-        __m128i fix8_32 = _mm_cvttps_epi32(fVec),
-                fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),
-                fix8    = _mm_packus_epi16(fix8_16, fix8_16);
-        *(int*)bytes = _mm_cvtsi128_si32(fix8);
-    }
-
-    static void ToBytes(uint8_t bytes[16],
-                        const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
-        _mm_storeu_si128((__m128i*)bytes,
-                         _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))));
-    }
 
     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); }
@@ -253,6 +225,34 @@ public:
 };
 
 template <>
+class SkNx<4, uint8_t> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    static SkNx Load(const uint8_t vals[4]) { return _mm_cvtsi32_si128(*(const int*)vals); }
+    void store(uint8_t vals[4]) const { *(int*)vals = _mm_cvtsi128_si32(fVec); }
+
+    // TODO as needed
+
+    __m128i fVec;
+};
+
+template <>
+class SkNx<8, uint8_t> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    static SkNx Load(const uint8_t vals[8]) { return _mm_loadl_epi64((const __m128i*)vals); }
+    void store(uint8_t vals[8]) const { _mm_storel_epi64((__m128i*)vals, fVec); }
+
+    // TODO as needed
+
+    __m128i fVec;
+};
+
+template <>
 class SkNx<16, uint8_t> {
 public:
     SkNx(const __m128i& vec) : fVec(vec) {}
@@ -296,11 +296,42 @@ public:
 };
 
 
-template<>
-inline SkNx<4, int> SkNx_cast<int, float, 4>(const SkNx<4, float>& src) {
+template<> inline Sk4i SkNx_cast<int, float, 4>(const Sk4f& src) {
     return _mm_cvttps_epi32(src.fVec);
 }
 
+template<> inline Sk4b SkNx_cast<uint8_t, float, 4>(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<> inline Sk4f SkNx_cast<float, uint8_t, 4>(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);
+}
+
+static inline void Sk4f_ToBytes(uint8_t bytes[16],
+                                const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {
+    _mm_storeu_si128((__m128i*)bytes,
+                     _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))));
+}
+
+
 }  // namespace
 
 #endif//SkNx_sse_DEFINED
diff --git a/src/opts/SkXfermode_opts.h b/src/opts/SkXfermode_opts.h
index 69f2b420f5..93559ab8eb 100644
--- a/src/opts/SkXfermode_opts.h
+++ b/src/opts/SkXfermode_opts.h
@@ -265,11 +265,11 @@ public:
 
 private:
     static Sk4f Load(SkPMColor c) {
-        return Sk4f::FromBytes((uint8_t*)&c) * Sk4f(1.0f/255);
+        return SkNx_cast<float>(Sk4b::Load((uint8_t*)&c)) * Sk4f(1.0f/255);
     }
     static SkPMColor Round(const Sk4f& f) {
         SkPMColor c;
-        (f * Sk4f(255) + Sk4f(0.5f)).toBytes((uint8_t*)&c);
+        SkNx_cast<uint8_t>(f * Sk4f(255) + Sk4f(0.5f)).store((uint8_t*)&c);
         return c;
     }
     inline SkPMColor xfer32(SkPMColor dst, SkPMColor src) const {
diff --git a/tests/SkNxTest.cpp b/tests/SkNxTest.cpp
index b3e03d0a9b..754e9800b1 100644
--- a/tests/SkNxTest.cpp
+++ b/tests/SkNxTest.cpp
@@ -207,21 +207,6 @@ DEF_TEST(Sk4px_widening, r) {
     REPORTER_ASSERT(r, 0 == memcmp(&wideLoHi, &wideLoHiAlt, sizeof(wideLoHi)));
 }
 
-DEF_TEST(Sk4f_toBytes, r) {
-    uint8_t bytes[4];
-
-    // toBytes truncates, not rounds.
-    Sk4f(0.7f).toBytes(bytes);
-    REPORTER_ASSERT(r, bytes[0] == 0);
-
-    // Clamping edge cases.
-    Sk4f(-2.0f, -0.7f, 255.9f, 256.0f).toBytes(bytes);
-    REPORTER_ASSERT(r, bytes[0] == 0);
-    REPORTER_ASSERT(r, bytes[1] == 0);
-    REPORTER_ASSERT(r, bytes[2] == 255);
-    REPORTER_ASSERT(r, bytes[3] == 255);
-}
-
 DEF_TEST(SkNx_cast, r) {
     Sk4f fs(-1.7f, -1.4f, 0.5f, 1.9f);
     Sk4i is = SkNx_cast<int>(fs);