Add Sk4f::ToBytes(uint8_t[16], Sk4f, Sk4f, Sk4f, Sk4f)

This is a big speedup for float -> byte.  E.g. gradient_linear_clamp_3color:
 x86-64 147µs -> 103µs  (Broadwell MBP)
 arm64 2.03ms -> 648µs  (Galaxy S6)
 armv7 1.12ms -> 489µs  (Galaxy S6, same device!)

BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;client.skia.android:Test-Android-GCC-Nexus9-CPU-Denver-Arm64-Debug-Trybot

Review URL: https://codereview.chromium.org/1483953002

5 files changed, 37 insertions, 9 deletions

diff --git a/bench/Sk4fBench.cpp b/bench/Sk4fBench.cpp
index 02ac14591d..b2f2b4a58e 100644
--- a/bench/Sk4fBench.cpp
+++ b/bench/Sk4fBench.cpp
@@ -62,10 +62,7 @@ struct Sk4fGradientBench : public Benchmark {
                  c = b + dcdx,
                  d = c + dcdx;
             for (size_t i = 0; i < SK_ARRAY_COUNT(fDevice); i += 4) {
-                a.toBytes((uint8_t*)(fDevice+i+0));
-                b.toBytes((uint8_t*)(fDevice+i+1));
-                c.toBytes((uint8_t*)(fDevice+i+2));
-                d.toBytes((uint8_t*)(fDevice+i+3));
+                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 b9b67704d4..36a645cac1 100644
--- a/src/core/SkNx.h
+++ b/src/core/SkNx.h
@@ -113,6 +113,15 @@ public:
         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); }
diff --git a/src/effects/gradients/SkLinearGradient.cpp b/src/effects/gradients/SkLinearGradient.cpp
index f47b6ab30e..b224474536 100644
--- a/src/effects/gradients/SkLinearGradient.cpp
+++ b/src/effects/gradients/SkLinearGradient.cpp
@@ -748,10 +748,15 @@ template <bool apply_alpha> void ramp(SkPMColor dstC[], int n, const Sk4f& c, co
     Sk4f cd2 = cd0 + dc2;
     Sk4f cd3 = cd1 + dc2;
     while (n >= 4) {
-        *dstC++ = trunc_from_255<apply_alpha>(cd0);
-        *dstC++ = trunc_from_255<apply_alpha>(cd1);
-        *dstC++ = trunc_from_255<apply_alpha>(cd2);
-        *dstC++ = trunc_from_255<apply_alpha>(cd3);
+        if (!apply_alpha) {
+            Sk4f::ToBytes((uint8_t*)dstC, cd0, cd1, cd2, cd3);
+            dstC += 4;
+        } else {
+            *dstC++ = trunc_from_255<apply_alpha>(cd0);
+            *dstC++ = trunc_from_255<apply_alpha>(cd1);
+            *dstC++ = trunc_from_255<apply_alpha>(cd2);
+            *dstC++ = trunc_from_255<apply_alpha>(cd3);
+        }
         cd0 = cd0 + dc4;
         cd1 = cd1 + dc4;
         cd2 = cd2 + dc4;
@@ -861,7 +866,7 @@ void SkLinearGradient::LinearGradientContext::shade4_dx_clamp(SkPMColor dstC[],
         ramp<apply_alpha>(dstC, n, c, dc, dither0, dither1);
         dstC += n;
         SkASSERT(dstC <= endDstC);
-        
+
         if (n & 1) {
             SkTSwap(dither0, dither1);
         }
diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index a03f0be674..6fe6137e5f 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -165,6 +165,14 @@ public:
         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);
diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index f0ccd3f7b5..a4f8656536 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -132,6 +132,15 @@ public:
         *(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); }
     SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }