Correct reference for fma for RTZ devices for half (#2269)

The reference was always being calculated using round to nearest even
for fma when testing half precision.

Fixes https://github.com/KhronosGroup/OpenCL-CTS/issues/2224

test_conformance/math_brute_force/reference_math.cpp

@@ -2734,6 +2734,34 @@ static double round_to_nearest_even_double(cl_ulong hi, cl_ulong lo,
     return u.d;
 }
 
+static double round_toward_zero_double(cl_ulong hi, cl_ulong lo, int exponent)
+{
+    union {
+        cl_ulong u;
+        cl_double d;
+    } u;
+
+    // edges
+    if (exponent > 1023) return CL_DBL_MAX;
+    if (exponent <= -1074) return 0.0;
+
+    // Figure out which bits go where
+    int shift = 11;
+    if (exponent < -1022)
+    {
+        shift -= 1022 + exponent; // subnormal: shift is not 52
+        exponent = -1023; //              set exponent to 0
+    }
+    else
+        hi &= 0x7fffffffffffffffULL; // normal: leading bit is implicit. Remove
+                                     // it.
+
+    // Assemble the double (round toward zero)
+    u.u = (hi >> shift) | ((cl_ulong)(exponent + 1023) << 52);
+
+    return u.d;
+}
+
 // Shift right.  Bits lost on the right will be OR'd together and OR'd with the
 // LSB
 static inline void shift_right_sticky_128(cl_ulong *hi, cl_ulong *lo, int shift)
@@ -2966,7 +2994,14 @@ long double reference_fmal(long double x, long double y, long double z)
     }
 
     // round
-    ua.d = round_to_nearest_even_double(hi, lo, exponent);
+    if (gIsInRTZMode)
+    {
+        ua.d = round_toward_zero_double(hi, lo, exponent);
+    }
+    else
+    {
+        ua.d = round_to_nearest_even_double(hi, lo, exponent);
+    }
 
     // Set the sign
     ua.u |= sign;