DB>Ok. By load do you mean load 64 bits from memory into the FPU? Then by 
store
DB>do you mean storing those 64 bits from the FPU into memory?
Yes.  Using the FILD and FIST instructions.
They'll work even on an 8086 XT with an 8087.
(I should point out that the 64 bit integers in the FPU are really only
partially implemented.  All you can do is load and store them.  All the
integer FPU operations work on 16 or 32 bit integers.)
DB>If so, by taking 64bits at a time, you could transfer memory -> FPU -> 
video
DB>memory? Would that be faster even with the intermediate FPU step?
Frankly, I have no idea.  I guess it would depend on the efficiency of
the FPU, and that varies greatly from processor to processor, and
manufacturer to manufacturer, and how good of a video card you have, and
how fast of a CPU.
I've never timed something like that, but... I'd guess that using normal
32 bit rep sto would be at least as fast, and probably faster.
BUT, and this is a big BUT... video cards are slow.  It's probably that
the normal rep store would be able to run faster than what the video
card can accept the data.  That was certainly true of most ISA cards,
and even many VLB cards.
Somewhere or other, I've got a video card benchmark program that shows
how many writes per second you can do to it, and then compares that with
how many writes per second you can do to regular memory.  Regular memory
is considerably faster.
So,... I don't think it would help.  The CPU's rep stor is probably
already capable of running faster than the video card can accept the
data, so even if using the FPU was faster, it wouldn't do any good.
--- QScan/PCB v1.19b / 01-0162
---------------