aboutsummaryrefslogtreecommitdiffstats
path: root/examples/using_double_on_avr/double_conversion.h
diff options
context:
space:
mode:
Diffstat (limited to 'examples/using_double_on_avr/double_conversion.h')
-rw-r--r--examples/using_double_on_avr/double_conversion.h26
1 files changed, 26 insertions, 0 deletions
diff --git a/examples/using_double_on_avr/double_conversion.h b/examples/using_double_on_avr/double_conversion.h
new file mode 100644
index 00000000..62b6a8ae
--- /dev/null
+++ b/examples/using_double_on_avr/double_conversion.h
@@ -0,0 +1,26 @@
+/* AVR-GCC does not have real double datatype. Instead its double
+ * is equal to float, i.e. 32 bit value. If you need to communicate
+ * with other systems that use double in their .proto files, you
+ * need to do some conversion.
+ *
+ * These functions use bitwise operations to mangle floats into doubles
+ * and then store them in uint64_t datatype.
+ */
+
+#ifndef DOUBLE_CONVERSION
+#define DOUBLE_CONVERSION
+
+#include <stdint.h>
+
+/* Convert native 4-byte float into a 8-byte double. */
+extern uint64_t float_to_double(float value);
+
+/* Convert 8-byte double into native 4-byte float.
+ * Values are rounded to nearest, 0.5 away from zero.
+ * Overflowing values are converted to Inf or -Inf.
+ */
+extern float double_to_float(uint64_t value);
+
+
+#endif
+