fix cast arm packing test

src/layer/arm/cast_arm.cpp

@@ -49,7 +49,11 @@ int Cast_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
 #if __ARM_NEON
     if (elempack % 4 == 0)
     {
+#if (__ARM_FP & 2)
         if (!cpu_support_arm_vfpv4() && (type_from == 2 || type_to == 2))
+#else
+        if (type_from == 2 || type_to == 2)
+#endif // (__ARM_FP & 2)
         {
             // no fp16 conversion instruction, fallback
             return Cast::forward(bottom_blob, top_blob, opt);

src/layer/arm/packing_arm.cpp

@@ -34,6 +34,7 @@ int Packing_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
         return Packing::forward(bottom_blob, top_blob, opt);
     }
 
+    size_t elemsize = bottom_blob.elemsize;
     int elempack = bottom_blob.elempack;
 
     if (elempack == out_elempack)
@@ -42,6 +43,11 @@ int Packing_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
         return 0;
     }
 
+    if (elemsize / elempack != 4u)
+    {
+        return Packing::forward(bottom_blob, top_blob, opt);
+    }
+
     bool pack1to4 = elempack == 1 && out_elempack == 4;
     bool pack4to1 = elempack == 4 && out_elempack == 1;
 
@@ -54,7 +60,6 @@ int Packing_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
     int h = bottom_blob.h;
     int channels = bottom_blob.c;
     int dims = bottom_blob.dims;
-    size_t elemsize = bottom_blob.elemsize;
 
     if (!use_padding)
     {

src/mat.cpp

@@ -113,58 +113,6 @@ void Mat::substract_mean_normalize(const float* mean_vals, const float* norm_val
     delete op;
 }
 
-// convert half precision floating point to float
-static float half2float(unsigned short value)
-{
-    // 1 : 5 : 10
-    unsigned short sign = (value & 0x8000) >> 15;
-    unsigned short exponent = (value & 0x7c00) >> 10;
-    unsigned short significand = value & 0x03FF;
-
-//     fprintf(stderr, "%d %d %d\n", sign, exponent, significand);
-
-    // 1 : 8 : 23
-    union
-    {
-        unsigned int u;
-        float f;
-    } tmp;
-    if (exponent == 0)
-    {
-        if (significand == 0)
-        {
-            // zero
-            tmp.u = (sign << 31);
-        }
-        else
-        {
-            // denormal
-            exponent = 0;
-            // find non-zero bit
-            while ((significand & 0x200) == 0)
-            {
-                significand <<= 1;
-                exponent++;
-            }
-            significand <<= 1;
-            significand &= 0x3FF;
-            tmp.u = (sign << 31) | ((-exponent + (-15 + 127)) << 23) | (significand << 13);
-        }
-    }
-    else if (exponent == 0x1F)
-    {
-        // infinity or NaN
-        tmp.u = (sign << 31) | (0xFF << 23) | (significand << 13);
-    }
-    else
-    {
-        // normalized
-        tmp.u = (sign << 31) | ((exponent + (-15 + 127)) << 23) | (significand << 13);
-    }
-
-    return tmp.f;
-}
-
 Mat Mat::from_float16(const unsigned short* data, int size)
 {
     Mat m(size);
@@ -224,7 +172,7 @@ Mat Mat::from_float16(const unsigned short* data, int size)
 #endif // __ARM_NEON
     for (; remain>0; remain--)
     {
-        *ptr = half2float(*data);
+        *ptr = float16_to_float32(*data);
 
         data++;
         ptr++;