GPGPU Programming for Games and Science

218 GPGPU Programming for Games and Science

For 32-bit floating-point numbers, some of the deviations from the rules are

about the handling of quiet or signaling NaNs. Two variations that potentially

have more impact when comparing the results from the GPU and from an FPU

associated with the CPU are the following.

• The default rounding rule for IEEE 754-2008 is round-to-nearest ties-

to-even. This rule is discussed in Section 2.5.2.1. The idea is that a

floating-point operation produces the floating-point number nearest to

the theoretical result. When the theoretical result has a fractional part

that is one-half, you round down if the integer part is