Fix float16 denorm conversion, add rounding (#3848)

* Fix float16 denorm conversion, add rounding

* review comments

src/ngraph/type/float16.cpp

@@ -49,6 +49,7 @@ float16::float16(float value)
         uint32_t iv;
     };
     fv = value;
+    uint32_t hidden_one = 0x00800000;
     uint32_t sign = iv & 0x80000000;
     uint32_t biased_exp = (iv & 0x7F800000) >> 23;
     uint32_t raw_frac = (iv & 0x007FFFFF);
@@ -67,19 +68,20 @@ float16::float16(float value)
     }
     else if (exp < -14)
     {
-        // denorm or 0
+        // denorm
         biased_exp = 0;
-        raw_frac |= 0x00800000;
-        raw_frac = raw_frac >> (exp + 16);
+        raw_frac |= hidden_one;
+        uint32_t shift = (-15 - exp) + (23 - frac_size) + 1;
+        raw_frac = (raw_frac + (hidden_one >> (shift + 1))) >> shift;
     }
     else if (exp > 15)
     {
         biased_exp = 0x1F;
         raw_frac = 0;
     }
-    else
+    else if ((biased_exp != 0 || raw_frac != 0))
     {
-        raw_frac = raw_frac >> (23 - frac_size);
+        raw_frac = (raw_frac + 0x1000) >> (23 - frac_size);
         biased_exp = exp + exp_bias;
     }
     m_value = (sign >> 16) | (biased_exp << frac_size) | raw_frac;

test/float16.cpp

@@ -87,3 +87,14 @@ TEST(float16, assigns)
         EXPECT_EQ(f32arr[i], f16arr[i]);
     }
 }
+
+TEST(float16, values)
+{
+    std::vector<double> f32vec{2.73786e-05, 3.87722e-05, -0.0223043};
+    std::vector<uint16_t> intvals = {459, 650, 42422};
+    for (size_t i = 0; i < f32vec.size(); ++i)
+    {
+        float16 fp16val = f32vec.at(i);
+        EXPECT_EQ(intvals.at(i), fp16val.to_bits());
+    }
+}