SKPMFloat: we can beat the naive loops when clamping

src/opts/SkPMFloat_SSE2.h

Index: src/opts/SkPMFloat_SSE2.h
diff --git a/src/opts/SkPMFloat_SSE2.h b/src/opts/SkPMFloat_SSE2.h
index e14f94fe135682cffdf8d1a3f28cab6bf6aeb4a6..7298b4da3c6b67ddd9611b73ef3ef0ab91bc3a92 100644
--- a/src/opts/SkPMFloat_SSE2.h
+++ b/src/opts/SkPMFloat_SSE2.h
@@ -19,7 +19,7 @@ inline SkPMFloat::SkPMFloat(SkPMColor c) {
 
 inline SkPMColor SkPMFloat::get() const {
     SkASSERT(this->isValid());
-    return this->clamped();  // At the moment, we don't know anything faster.
+    return this->clamped();  // Haven't beaten this yet.
 }
 
 inline SkPMColor SkPMFloat::clamped() const {
@@ -30,3 +30,30 @@ inline SkPMColor SkPMFloat::clamped() const {
     SkPMColorAssert(c);
     return c;
 }
+
+inline void SkPMFloat::From4PMColors(SkPMFloat floats[4], const SkPMColor colors[4]) {
+    // Haven't beaten this yet.
+    for (int i = 0; i < 4; i++) { floats[i] = FromPMColor(colors[i]); }
+}
+
+inline void SkPMFloat::To4PMColors(SkPMColor colors[4], const SkPMFloat floats[4]) {
+    SkASSERT(floats[0].isValid() && floats[1].isValid()
+          && floats[2].isValid() && floats[3].isValid());
+    // Haven't beaten this yet.
+    ClampTo4PMColors(colors, floats);
+}
+
+inline void SkPMFloat::ClampTo4PMColors(SkPMColor colors[4], const SkPMFloat floats[4]) {
+    // Same as _SSSE3.h's.  We use 3 _mm_packus_epi16() where the naive loop uses 8.
+    __m128i c0 = _mm_cvtps_epi32(_mm_load_ps(floats[0].fColor)),  // _mm_cvtps_epi32 rounds for us!
+            c1 = _mm_cvtps_epi32(_mm_load_ps(floats[1].fColor)),
+            c2 = _mm_cvtps_epi32(_mm_load_ps(floats[2].fColor)),
+            c3 = _mm_cvtps_epi32(_mm_load_ps(floats[3].fColor));
+    __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1),
+                                     _mm_packus_epi16(c2, c3));
+    _mm_storeu_si128((__m128i*)colors, c3210);
+    SkPMColorAssert(colors[0]);
+    SkPMColorAssert(colors[1]);
+    SkPMColorAssert(colors[2]);
+    SkPMColorAssert(colors[3]);
+}