CODEWITHSUNDEEP
floating-pointieee-754underflow
Laplace
Laplace

Reputation: 19

Flush to Zero when a computation results in a denormal number in linux

A computation in my C code is producing a gradual underflow, and when it happens the program is terminating with SIGFPE. How can I flush the result to zero when a gradual underflow (Denormal) results from a computation, and not terminate the execution? (I am working on a redhat linux machine). Thanks.

Upvotes: 1

Views: 2486

Answers (1)

Brett Hale
Brett Hale

Reputation: 22318

You haven't specified the architecture - I'm going to take a guess that it's a relatively recent x86[-64], in which case you can manipulate the SSE control register using _mm_getcsr, _mm_setcsr, specified in the <xmmintrin.h> (or <immintrin.h>) header.

The 'flush-to-zero' bit is set with 0x8000, and 'denormals-are-zero' (for inputs / src) is set with 0x0040.

_mm_setcsr(_mm_getcsr() | 0x8040); or with <pmmintrin.h> (SSE3) :

_mm_setcsr(_mm_getcsr() | (_MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON));

This might make it easier to determine the source of the underflow, but it shouldn't be considered a solution, since the FP environment is no longer IEEE-754 compliant.

Upvotes: 4

Related Questions