Barry Smith <[email protected]> writes: > Note also that the Intel floating point registers are 80 bits wide > while double precision numbers in memory are 64 bit. Thus moving a > partially computed sum to memory and then back to register wipes > out 16 bits of the number.
This comment refers to the x87 unit which fell out of favor when SSE provided better performance and vectorization. It is only used now if you compile for x86 (32-bit) without SSE support.
pgpEBlj0ktlnV.pgp
Description: PGP signature
