brush up divSaturate of carotene

  * use rounding to nearest same as divSaturateQ

Author: tomoaki0705

3rdparty/carotene/src/div.cpp

@@ -74,17 +74,24 @@ template <>
 inline uint32x4_t divSaturateQ<uint32x4_t>(const uint32x4_t &v1, const uint32x4_t &v2, const float scale)
 { return vcvtq_u32_f32(vroundq(vmulq_f32(vmulq_n_f32(vcvtq_f32_u32(v1), scale), internal::vrecpq_f32(vcvtq_f32_u32(v2))))); }
 
+inline float32x2_t vround(const float32x2_t& v)
+{
+    const int32x2_t signMask = vdup_n_s32(1 << 31), half = vreinterpret_s32_f32(vdup_n_f32(0.5f));
+    float32x2_t v_addition = vreinterpret_f32_s32(vorr_s32(half, vand_s32(signMask, vreinterpret_s32_f32(v))));
+    return vadd_f32(v, v_addition);
+}
+
 template <typename T>
 inline T divSaturate(const T &v1, const T &v2, const float scale)
 {
     return internal::vqmovn(divSaturateQ(internal::vmovl(v1), internal::vmovl(v2), scale));
 }
 template <>
 inline int32x2_t divSaturate<int32x2_t>(const int32x2_t &v1, const int32x2_t &v2, const float scale)
-{ return vcvt_s32_f32(vmul_f32(vmul_n_f32(vcvt_f32_s32(v1), scale), internal::vrecp_f32(vcvt_f32_s32(v2)))); }
+{ return vcvt_s32_f32(vround(vmul_f32(vmul_n_f32(vcvt_f32_s32(v1), scale), internal::vrecp_f32(vcvt_f32_s32(v2))))); }
 template <>
 inline uint32x2_t divSaturate<uint32x2_t>(const uint32x2_t &v1, const uint32x2_t &v2, const float scale)
-{ return vcvt_u32_f32(vmul_f32(vmul_n_f32(vcvt_f32_u32(v1), scale), internal::vrecp_f32(vcvt_f32_u32(v2)))); }
+{ return vcvt_u32_f32(vround(vmul_f32(vmul_n_f32(vcvt_f32_u32(v1), scale), internal::vrecp_f32(vcvt_f32_u32(v2))))); }
 
 
 template <typename T>