Sk4f: floor() via int32_t roundtrip.

src/opts/SkNx_neon.h

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkNx_neon_DEFINED
 #define SkNx_neon_DEFINED
 
 #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.
+//   - roundtrip through integers via truncation
+//   - subtract 1 if that's too big (possible for negative values).
+// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
 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)));
+    auto roundtrip = vcvtq_f32_s32(vcvtq_s32_f32(v));
+    auto too_big = vcgtq_f32(roundtrip, v);
+    return vsubq_f32(roundtrip, (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) { \
     case  1: return op(v,  1);  case  2: return op(v,  2);  case  3: return op(v,  3); \
     case  4: return op(v,  4);  case  5: return op(v,  5);  case  6: return op(v,  6); \
     case  7: return op(v,  7); \
     } return fVec
 
@@ skipping 356 matching lines @@
 
 template<> inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
     return vget_low_u16(vmovl_u8(src.fVec));
 }
 
 template<> inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
     return vmovn_u16(vcombine_u16(src.fVec, src.fVec));
 }
 
 #endif//SkNx_neon_DEFINED