core: arm64: v_round() works with round to nearest, ties to even.

modules/core/include/opencv2/core/hal/intrin_neon.hpp

@@ -1997,12 +1997,12 @@ inline v_int32x4 v_trunc(const v_float32x4& a)
 inline v_int32x4 v_round(const v_float64x2& a)
 {
     static const int32x2_t zero = vdup_n_s32(0);
-    return v_int32x4(vcombine_s32(vmovn_s64(vcvtaq_s64_f64(a.val)), zero));
+    return v_int32x4(vcombine_s32(vmovn_s64(vcvtnq_s64_f64(a.val)), zero));
 }
 
 inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
 {
-    return v_int32x4(vcombine_s32(vmovn_s64(vcvtaq_s64_f64(a.val)), vmovn_s64(vcvtaq_s64_f64(b.val))));
+    return v_int32x4(vcombine_s32(vmovn_s64(vcvtnq_s64_f64(a.val)), vmovn_s64(vcvtnq_s64_f64(b.val))));
 }
 
 inline v_int32x4 v_floor(const v_float64x2& a)

modules/core/test/test_intrin_utils.hpp

@@ -1475,12 +1475,15 @@ template<typename R> struct TheTest
     TheTest & test_float_math()
     {
         typedef typename V_RegTraits<R>::round_reg Ri;
-        Data<R> data1, data2, data3;
+        Data<R> data1, data1_border, data2, data3;
+        // See https://github.com/opencv/opencv/issues/24213
+        data1_border *= 0.5;
         data1 *= 1.1;
         data2 += 10;
-        R a1 = data1, a2 = data2, a3 = data3;
+        R a1 = data1, a1_border = data1_border, a2 = data2, a3 = data3;
 
         Data<Ri> resB = v_round(a1),
+                 resB_border = v_round(a1_border),
                  resC = v_trunc(a1),
                  resD = v_floor(a1),
                  resE = v_ceil(a1);
@@ -1493,6 +1496,7 @@ template<typename R> struct TheTest
         {
             SCOPED_TRACE(cv::format("i=%d", i));
             EXPECT_EQ(cvRound(data1[i]), resB[i]);
+            EXPECT_EQ(cvRound(data1_border[i]), resB_border[i]);
             EXPECT_EQ((typename VTraits<Ri>::lane_type)data1[i], resC[i]);
             EXPECT_EQ(cvFloor(data1[i]), resD[i]);
             EXPECT_EQ(cvCeil(data1[i]), resE[i]);