Sk4f: add floor()

src/opts/SkNx_neon.h

Index: src/opts/SkNx_neon.h
diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index 2cb8eb348d448e1aa9af4685c9739fb9123f46e7..cdc4615849556ff185afd59f15c82e8d5a5320b5 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -10,6 +10,17 @@
 
 #define SKNX_IS_FAST
 
+// ARMv8 has vrndmq_f32 to floor 4 floats.  Here we emulate it:
+//   - round by adding (1<<23) with our sign, then subtracting it;
+//   - if that rounded value is bigger than our input, subtract 1.
+static inline float32x4_t armv7_vrndmq_f32(float32x4_t v) {
+    auto sign = vandq_u32((uint32x4_t)v, vdupq_n_u32(1<<31));
+    auto bias = (float32x4_t)(vorrq_u32((uint32x4_t)vdupq_n_f32(1<<23), sign));
+    auto rounded = vsubq_f32(vaddq_f32(v, bias), bias);
+    auto too_big = vcgtq_f32(rounded, v);
+    return vsubq_f32(rounded, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdupq_n_f32(1)));
+}
+
 // Well, this is absurd.  The shifts require compile-time constant arguments.
 
 #define SHIFT8(op, v, bits) switch(bits) { \
@@ -161,6 +172,14 @@ public:
     static SkNx Max(const SkNx& l, const SkNx& r) { return vmaxq_f32(l.fVec, r.fVec); }
 
     SkNx abs() const { return vabsq_f32(fVec); }
+    SkNx floor() const {
+    #if defined(SK_CPU_ARM64)
+        return vrndmq_f32(fVec);
+    #else
+        return armv7_vrndmq_f32(fVec);
+    #endif
+    }
+
 
     SkNx rsqrt0() const { return vrsqrteq_f32(fVec); }
     SkNx rsqrt1() const {