How to set ICC attribute "fp-model precise" for a single function, to prevent associative optimizations?

Question

I am implementing Kahan summation, in a project that supports compilation with gcc47, gcc48, clang33, icc13, and icc14.

As part of this algorithm, I would like to disable optimizations that take advantage of the associativity of addition of real numbers. (Floating point operations are not associative.)

I would like to disable those optimizations only in the relevant function. I have figured out how to do this under gcc, using the ''no-associative-math'' attribute. How can I do this in icc or clang? I have searched without luck.

class KahanSummation
{
  // GCC declaration
  void operator()(float observation) __attribute__((__optimize__("no-associative-math")))
  {
    // Kahan summation implementation
  }
};

Other GCC attributes that would imply no-associative-math are no-unsafe-math-optimizations or no-fast-math.

Looking at an Intel presentation (PDF, slide 8) or another or another (PDF, slide 11), I want to set "fp-model precise" in ICC, for only this function. The compilers I care about are ICC 13 and ICC 14.

class KahanSummation
{
  // ICC or clang declaration
  void operator()(float observation) __attribute__((????))
  {
    // Kahan summation implementation
  }
};

__attribute__ is a GCC extension. Implement those functions in a separate translation unit and compile that using the appropriate switch? — T.C.

T.C. T.C. · Accepted Answer · 2014-06-23T04:52:03

__attribute__ is a GCC extension. Clang also supports that syntax to maintain some GCC compatibility, but the only optimization-related attribute it seems to support is optnone, which turns off all optimizations. ICC ~~has a few optimization pragmas, but nothing that does what you want to do, from what I can tell.~~ appears to support #pragma float_control for VC++ compatibility, although I can't find anything on exactly how that pragma is supposed to be used in ICC's documentation, so you'll have to use VC++'s.

What you can do, though, is to define the function you want in a separate translation unit (i.e., cpp file):

// Your header
class KahanSummation
{
  // Declaration
  void operator()(float observation);
};

// Separate cpp file - implements only this function
void KahanSummation::operator()(float observation)
{
    // Kahan summation implementation
}

You can then compile the separate file using whatever compiler option you need to use, and link the resulting object file to the rest of your program.

How to set ICC attribute "fp-model precise" for a single function, to prevent associative optimizations?

1 Answers