Added very experimental implementation for mixed precision. Comm is done with f32 and comp with f64.

src/core/kernels/kernels.h

@@ -8,11 +8,13 @@
 #define MPI_GPUDIRECT_DISABLED (0)
 #endif // AC_MPI_ENABLED
 
+typedef float AcRealPacked;
+
 typedef struct {
     int3 dims;
-    AcReal* data;
+    AcRealPacked* data;
 
-    AcReal* data_pinned;
+    AcRealPacked* data_pinned;
     bool pinned = false; // Set if data was received to pinned memory
 } PackedData;