Float calculation in VS2010 and VS2012/VS2013 C++

Question

Hi,
There are some difference in float calculation in VS2010 and VS2013/VS2012.   (vc++)
VS2010 by default uses FPU instruction for float calculation. while VS2013 by default , uses SSE2 instruction for the same.
Can you suggest which is the most accurate result. FPU or SSE2
Regards,
Alphonse

Answer 1

That tends to depend on the code you write. Without being a floating point expert I'll say that I find SSE2 preferable because it provides consistent results. Consistency isn't the same thing as accuracy but it's usually more valuable.
Consider the following example adapted from "What Every Computer Scientist Should Know About Floating-Point Arithmetic" (required reading if you really care about
floating point accuracy):
void test(float a, float b) {
float q = a / b;
if (q == a / b)
printf("Equal\n");
else
printf("Not Equal\n");
int main() {
test(3, 7);
This code prints "Not Equal" when FPU is used and "Equal" when SSE2 is used. Needless to say, the SSE2 result is much more desirable.
The reason for this is that FPU tends to perform calculations at a higher precision than requested. The VC++ runtime configures the FPU to use double precision and as a result all float calculations are actually done at double precision and the result
is converted to float when required by the code (like when storing into a float variable). In the above example the right side of == will be a double precision value while the left side is a double precision value that was previously rounded to single precision.
If you use only doubles it's less likely to run into such issues but as far as I can tell it's not completely impossible. The FPU rounds only the mantissa part to double precision but leaves the exponent alone. Presumable this can cause issues if very
large values are involved.
SSE2 doesn't suffer from this problem, it has distinct instructions for single and double precision and VS2013 makes use of them. If you use floats the calculation is done using single precision, if you use doubles the calculation is done using double precision; nothing
less, nothing more.