ARMCC: Rounding floats to integers

Information in this knowledgebase article applies to:

• Keil MDK
• Arm Compiler 5

QUESTION

I'm using a Cortex-M4 based target device and now want to convert floating-point variables to rounded integer numbers, for example:

float fvalue = 99.999f;
int32_t i32value;

int main(void)
{
  i32value = fvalue;  // result 99 instead of 100
.....

I'm using Arm Compiler 5, that translates the code above as follows:

VLDR     s0,[r0,#0]
VCVT.S32.F32 s0,s0
LDR      r0,|L3.304|
VSTR     s0,[r0,#0]

The value of i32value is truncated and rounded to 99, instead of 100. 

Is it possible to make the Arm Compiler 5 use the VCVTR instruction instead of VCVT to get the rounded value 100, instead of 99?

RESOLUTION

Note, that according to the C99 standards, all conversions from float to integer are truncated.

As written here "VCVT, VCVTR between floating-point and integer", the floating-point to integer operation normally uses the Round towards Zero rounding mode, but can optionally use the rounding mode specified by the FPSCR.

To have VCVTR instead of VCVT, you can use the function _ffix_r().

Moreover, you can use these two C99 functions fegetround() and fesetround() to get and set a different round mode.

For example:

#include <math.h>
#include <rt_fp.h>
#include <fenv.h>

...
   int mode = fegetround();
   i32value = _ffix_r(fvalue);  // result 100 instead of 99
   fesetround(mode);
...

This can enable you to use one round mode for only one conversion. You can also use fesetround() to change the round mode for the entire project.

Note that in order to use fegetround() and fesetround(), the Microlib option in µVision should be disabled.

Last Reviewed: Thursday, November 5, 2020