conversions: Fix verification of half subnormal cases with FTZ (#1914)

test_conformance/conversions/basic_test_conversions.cpp

@@ -1005,14 +1005,6 @@ double SubtractTime(uint64_t endTime, uint64_t startTime)
 }
 #endif
 
-static void setAllowZ(uint8_t *allow, uint32_t *x, cl_uint count)
-{
-    cl_uint i;
-    for (i = 0; i < count; ++i)
-        allow[i] |= (uint8_t)((x[i] & 0x7f800000U) == 0);
-}
-
-
 void MapResultValuesComplete(const std::unique_ptr<CalcRefValsBase> &ptr);
 
 void CL_CALLBACK CalcReferenceValuesComplete(cl_event e, cl_int status,
@@ -1337,15 +1329,13 @@ cl_int PrepareReference(cl_uint job_id, cl_uint thread_id, void *p)
         // Decide if we allow a zero result in addition to the correctly rounded
         // one
         memset(a, 0, count);
-        if (gForceFTZ)
+        if (gForceFTZ && (inType == kfloat || outType == kfloat))
         {
-            if (inType == kfloat || outType == kfloat)
+            info->set_allow_zero_array((uint8_t *)a, d, s, count);
-                setAllowZ((uint8_t *)a, (uint32_t *)s, count);
         }
-        if (gForceHalfFTZ)
+        if (gForceHalfFTZ && (inType == khalf || outType == khalf))
         {
-            if (inType == khalf || outType == khalf)
+            info->set_allow_zero_array((uint8_t *)a, d, s, count);
-                setAllowZ((uint8_t *)a, (uint32_t *)s, count);
         }
     }
     else

test_conformance/conversions/conversions_data_info.h

@@ -30,6 +30,7 @@ extern roundingMode qcom_rm;
 
 #include <CL/cl_half.h>
 
+#include "harness/conversions.h"
 #include "harness/mt19937.h"
 #include "harness/rounding_mode.h"
 #include "harness/typeWrappers.h"
@@ -82,6 +83,7 @@ struct DataInitBase : public DataInitInfo
     virtual void conv_array(void *out, void *in, size_t n) {}
     virtual void conv_array_sat(void *out, void *in, size_t n) {}
     virtual void init(const cl_uint &, const cl_uint &) {}
+    virtual void set_allow_zero_array(uint8_t *allow, void *out, void *in, size_t n) {}
 };
 
 template <typename InType, typename OutType, bool InFP, bool OutFP>
@@ -99,6 +101,9 @@ struct DataInfoSpec : public DataInitBase
     void conv(OutType *out, InType *in);
     void conv_sat(OutType *out, InType *in);
 
+    // Decide if we allow a zero result in addition to the correctly rounded one
+    void set_allow_zero(uint8_t *allow, OutType *out, InType *in);
+
     // min/max ranges for output type of data
     std::pair<OutType, OutType> ranges;
 
@@ -130,6 +135,10 @@ struct DataInfoSpec : public DataInitBase
     }
 
     void init(const cl_uint &, const cl_uint &) override;
+    void set_allow_zero_array(uint8_t *allow, void *out, void *in, size_t n) override {
+        for (size_t i = 0; i < n; i++)
+            set_allow_zero(&allow[i], &((OutType *)out)[i], &((InType *)in)[i]);
+    }
     InType clamp(const InType &);
     inline float fclamp(float lo, float v, float hi)
     {
@@ -717,6 +726,33 @@ void DataInfoSpec<InType, OutType, InFP, OutFP>::conv_sat(OutType *out,
     }
 }
 
+template <typename InType, typename OutType, bool InFP, bool OutFP>
+void DataInfoSpec<InType, OutType, InFP, OutFP>::set_allow_zero(uint8_t *allow,
+                                                                OutType *out,
+                                                                InType *in)
+{
+  // from double
+  if (std::is_same<InType, cl_double>::value)
+    *allow |= IsDoubleSubnormal(*in);
+  // from float
+  if (std::is_same<InType, cl_float>::value)
+    *allow |= IsFloatSubnormal(*in);
+  // from half
+  if (is_in_half())
+    *allow |= IsHalfSubnormal(*in);
+
+  // handle the cases that the converted result is subnormal
+  // from double
+  if (std::is_same<OutType, cl_double>::value)
+    *allow |= IsDoubleSubnormal(*out);
+  // from float
+  if (std::is_same<OutType, cl_float>::value)
+    *allow |= IsFloatSubnormal(*out);
+  // from half
+  if (is_out_half())
+    *allow |= IsHalfSubnormal(*out);
+}
+
 template <typename InType, typename OutType, bool InFP, bool OutFP>
 void DataInfoSpec<InType, OutType, InFP, OutFP>::init(const cl_uint &job_id,
                                                       const cl_uint &thread_id)