By: Dresdenboy (M.Waldhauer.delete@this.gmx.de), July 23, 2004 3:27 am
Room: Moderated Discussions
_Arthur (_Arthur@globetrotter.net) on 7/22/04 wrote:
---------------------------
>One could also microcode new instructions for something useful, like to speed up
>(slightly) a SETI search or DiVX encoding...
That's exactly, why I'm interested in that stuff (regarding Prime95). Such custom instructions could help speeding up parts of the FFT butterflies or help to circumvent MUL/IMUL inefficiencies (with fixed input and result registers) or a missing MUL to get only the upper half of a product. But to make that stuff useful, it's better to have a small routine in microcode than just one instruction because this could waste decode bandwidth. But maybe (according to a patent) the Pack pipeline stage is able to fill instruction slots which are not used by instructions coming from microcode ROM/RAM.
---------------------------
>One could also microcode new instructions for something useful, like to speed up
>(slightly) a SETI search or DiVX encoding...
That's exactly, why I'm interested in that stuff (regarding Prime95). Such custom instructions could help speeding up parts of the FFT butterflies or help to circumvent MUL/IMUL inefficiencies (with fixed input and result registers) or a missing MUL to get only the upper half of a product. But to make that stuff useful, it's better to have a small routine in microcode than just one instruction because this could waste decode bandwidth. But maybe (according to a patent) the Pack pipeline stage is able to fill instruction slots which are not used by instructions coming from microcode ROM/RAM.