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

5 files changed, 126 insertions, 78 deletions

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 {