Hallo Rob!

Replying to a message of Rob Basler to Jonathan De Boyne Pollar:

 RB> Actually, I am doing a considerable amount of "floating point" math
 RB> using fixed point integer math to do 3-D projections.  The reason I
 RB> don't want to use floating point is that integer math is faster than
 RB> floating point EVEN WITH A COPROCESSOR especially for addition and
 RB> subtraction, 

Yes, additions are faster in integer math. However, at least on a pentium,
multiplication and (correct) division are faster using floating point math,
since integer mul/div are actually executed by the FPU with all the
necessary conversions before and after and no pipelining. Whereas with FPU
instructions, you can start a pipelined add/sub/mul operation every cycle
(successive muls every other cycle).

And pentium just follows the common trend. In case of a Power[PC] or other
RISC, you loose at least half of the processor performance and half of the
available registers, if you replace floating point ops by integer ops. All
thease beasts are designed to execute floating point in parallel with
integer, and multiplication is usually done much faster using floating
point.

  Gruss, Andreas

--- FleetStreet 0.97wb