lrwiman <[EMAIL PROTECTED]> wrote:
>
> It has been mentioned several times recently that factoring is all integer
> work, and LL testing is nearly all floating point.
>
> It is my understanding that on intel CPU's, these are done on separate parts of
> the CPU. Would it increase net performance to do factoring and LL assignments
> at the same time?
>
If we can execute the two codes together, the functional units
will be better utilized. But we will need the combined memory bandwidth
for both codes, and enough cache for both working sets.
On Pentiums, a bigger concern is the number of registers --
most floating point intensive algorithms need several integer registers
for loop counters and address computations, so the integer code will
lack these. The two algorithms must be designed so the
control (i.e., branch) logic is identical -- we don't
want the integer code to be calling a subroutine
which executes the Euclidean GCD algorithm and repeatedly tests whether
the latest remainder is zero, while the floating point code
loops over the FFT output to square each element, for example.
When the control logic is identical for two algorithms, and
the subscripting (i.e., memory access) patterns are similar,
it may be feasible to merge integer and floating point algorithms
for improved functional unit utilization. For example, Ernst Mayer
is writing code which will do an integer FFT and a floating point
FFT of the same length concurrently.
________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
