Re: ExactDotProduct
Dan Zuras Intervals wrote:
And the space involved is not memory. It is registers.
Many registers. Many many registers.
Even today when computers are cheap enough to throw away
with your clothes, registers are expensive in the design
of such machines.
Perhaps there will come a time when this is not so but
I suspect registers will always be expensive in any given
design relative to the cost & application of that design.
The reason for this is that even if a designer chooses to
design in many many registers there are always other things
that can be done with those registers that have a far
greater impact on performance than to dedicate them to the
use of accurate dot products.
Still, I believe Ulrich's approach might find its way into
a general purpose design some day if it were implemented
using a large set of registers that the computer could ALSO
use for other, much more common, purposes.
I believe this might be true, too, e.g., we see today a trend in larger and
larger Vector Processing Units (VPUs), where each register can accomodate
multiple data formats such as packed integers and floating-point values,
etc. All kinds of various useful operations can then be performed SIMD-style
on the operands in these registers.
Larrabee will (supposedly) have a set of 32 registers, each 512 bits wide...
a total of 16384 bits. Theoretically, that's already enough bits for three
binary64 complete formats, or 25 binary32 complete formats.
Nate