let the performance test pass on ARM

  * use round-to-neaerest in div of carotene

3rdparty/carotene/src/div.cpp

@@ -51,6 +51,13 @@ namespace {
 
 #ifdef CAROTENE_NEON
 
+inline float32x4_t vroundq(const float32x4_t& v)
+{
+    const int32x4_t signMask = vdupq_n_s32(1 << 31), half = vreinterpretq_s32_f32(vdupq_n_f32(0.5f));
+    float32x4_t v_addition = vreinterpretq_f32_s32(vorrq_s32(half, vandq_s32(signMask, vreinterpretq_s32_f32(v))));
+    return vaddq_f32(v, v_addition);
+}
+
 template <typename T>
 inline T divSaturateQ(const T &v1, const T &v2, const float scale)
 {
@@ -62,10 +69,10 @@ inline T divSaturateQ(const T &v1, const T &v2, const float scale)
 }
 template <>
 inline int32x4_t divSaturateQ<int32x4_t>(const int32x4_t &v1, const int32x4_t &v2, const float scale)
-{ return vcvtq_s32_f32(vmulq_f32(vmulq_n_f32(vcvtq_f32_s32(v1), scale), internal::vrecpq_f32(vcvtq_f32_s32(v2)))); }
+{ return vcvtq_s32_f32(vroundq(vmulq_f32(vmulq_n_f32(vcvtq_f32_s32(v1), scale), internal::vrecpq_f32(vcvtq_f32_s32(v2))))); }
 template <>
 inline uint32x4_t divSaturateQ<uint32x4_t>(const uint32x4_t &v1, const uint32x4_t &v2, const float scale)
-{ return vcvtq_u32_f32(vmulq_f32(vmulq_n_f32(vcvtq_f32_u32(v1), scale), internal::vrecpq_f32(vcvtq_f32_u32(v2)))); }
+{ return vcvtq_u32_f32(vroundq(vmulq_f32(vmulq_n_f32(vcvtq_f32_u32(v1), scale), internal::vrecpq_f32(vcvtq_f32_u32(v2))))); }
 
 template <typename T>
 inline T divSaturate(const T &v1, const T &v2, const float scale)