daz : Denormals are zeros

The tool to change the CPU flag about denormals number.

DAZ (Denormals-Are-Zero) treats denormal inputs as zero
FTZ (Flush-To-Zero) writes zero for denormal outputs

Application

On x86-64 CPUs with 64-bit programs, the CPUs SSE unit performs the floating point operations. When it comes to calculate with denormal (aka. subnormal) numbers, there are performance penalties.

If your specific use-case doesn't require highest accuracy with small numbers, these can be treated as - or rounded to - zero. This is achieved by setting the CPU-flags. When doing so, the calculations won't be slowed down by factors!

See Wikipedia for additional infos on denormal numbers.

In python, especially NumPy functions show a measurable benefit.

Usage

basic functional use:

import daz
daz.set_ftz()
daz.set_daz()
daz.unset_ftz()
daz.unset_daz()
daz.get_ftz()
daz.get_daz()

alternative 1:

from daz import DAZ
# prev_daz: bool = DAZ.set_daz(daz: bool | None = True)
# prev_ftz: bool = DAZ.set_ftz(ftz: bool | None = True)
prev_daz = DAZ.set_daz(True)
prev_ftz = DAZ.set_ftz()

alternative 2:

from daz import DAZ

# DAZ(daz: bool = True, ftz: bool = True)
with DAZ():
    # daz and ftz set True
    pass

with DAZ(False, True):
    # daz unset, but ftz set True
    pass

hayguen/daz

daz : Denormals are zeros

Application

Usage