| Re: [AD] Color convertors |
[ Thread Index |
Date Index
| More lists.liballeg.org/allegro-developers Archives
]
Eric Botcazou wrote:
Of course. The windowed driver already makes use of a trick to ensure that
the window (position) is always aligned on a 4-byte boundary, that is
2-pixel alignment on 15/16-bit desktops and 4-pixel alignment on 24-bit
desktop.
Going with 8 byte alignment will benefit the MMX code, since memory
read/writes won't need to be split in 2 by the memory controler. I don't
know how this affects non-686s, but on the 686 the cost is twice that of
reading on aligned boundaries (since the memory controlers does 2 reads then
combines the result). Athlons seem to be immune to unligned memory accesses
though.
The non-MMX code will sufer a slight penalty since I'll be needing an
extra register.
Playing with the legacy 80386 registers is a very fun game ;-)
Actually, it turns out I didn't need one!
In the 24->32 non-mmx code, the penalty is 4.5 cycles/line on Pentiums, so I
don't think it'll affect speed much.
--
- Robert J Ohannessian
"Microsoft code is probably O(n^20)" (my CS prof)
http://pages.infinit.net/voidstar/