On Mon, 30 Mar 2009, Ross Ridge wrote: > >2) half-precision floats; > > Do you mean just conversion only support, like Sandra Loosemore's > proposed ARM patch, or full arithmetic support like any other scalar or > vector type?
Arithmetic via converting to float, doing arithmetic on that and converting back to half-precision gives you the same semantics as IEEE arithmetic directly on half-precision, except for the special case of converting double to half-precision where you can get double rounding if you go via float. (This is because the precision of float is strictly more than twice that of half-precision. Of course the above only applies to single arithmetic operations; more complicated combinations would need truncation to half-precision at each step rather than doing everything in float.) So if you have hardware float operations but not hardware half-precision you probably want to use those hardware float operations for arithmetic rather than having any separate half-precision arithmetic implementation. There are at least three possibilities for semantics: the ARM semantics where promotion to float happens at the language level; full IEEE arithmetic with every operation having a result in the range and precision of half-precision; or C99 excess precision where the semantic type is half-precision but the values are expressed to the range and precision of float. -- Joseph S. Myers jos...@codesourcery.com
