Patches on top of Radu's latest.

src/opts/SkPMFloat_neon.h

Index: src/opts/SkPMFloat_neon.h
diff --git a/src/opts/SkPMFloat_neon.h b/src/opts/SkPMFloat_neon.h
index 8bee5b551aebb708c09b7557ae74ea6a8618f083..ecb151f1fdb70dad1b18190281150ba29c4549d4 100644
--- a/src/opts/SkPMFloat_neon.h
+++ b/src/opts/SkPMFloat_neon.h
@@ -7,6 +7,8 @@
 
 namespace { // See SkPMFloat.h
 
+static_assert(SK_A32_SHIFT == 24, "This file assumes little-endian.");
+
 inline SkPMFloat::SkPMFloat(SkPMColor c) {
     SkPMColorAssert(c);
     uint8x8_t   fix8    = (uint8x8_t)vdup_n_u32(c);
@@ -28,8 +30,21 @@ inline SkPMColor SkPMFloat::round() const {
 }
 
 inline Sk4f SkPMFloat::alphas() const {
-    static_assert(SK_A32_SHIFT == 24, "Assuming little-endian.");
     return vdupq_lane_f32(vget_high_f32(fVec), 1);  // Duplicate high lane of high half i.e. lane 3.
 }
 
+inline SkPMFloat SkPMFloat::FromBGRx(SkColor c) {
+    uint8x8_t fix8 = (uint8x8_t)vdup_n_u32(c);
+#if defined(SK_PMCOLOR_IS_RGBA)
+    fix8 = vtbl1_u8(fix8, vcreate_u8(0x0300010203000102ULL));  // 03 00 01 02, 2x, i.e. swap R&B.
+#endif
+    uint16x8_t fix8_16 = vmovl_u8(fix8);
+    uint32x4_t fix8_32 = vmovl_u16(vget_low_u16(fix8_16));
+    fix8_32 = vsetq_lane_u32(0xFF, fix8_32, 3);  // Force alpha to 1.
+
+    SkPMFloat pmf = Sk4f(vmulq_f32(vcvtq_f32_u32(fix8_32), vdupq_n_f32(1.0f/255)));
+    SkASSERT(pmf.isValid());
+    return pmf;
+}
+
 }  // namespace