[...]

I have tried to do a direct comparison between the two DSPs, the 6414
and the 6203.  This was possible, since we have prototype boards
equipped with each of the DSPs.

Two different versions of the DSP code were used:

-  The first one is based on the loop26 program. In this case, it is known
   that loop26 is not well optimized for the 64x.

   The already existing histogramming part of the DSP code (as shown
   at the December 2000 ROD meeting) was however modified in order to:
   a) optimize it for the 64x (remove the cross-path stalls); and
   b) handle histograms with 16-bit bins for low and medium gain, and
      32-bit bins for high gain (instead of the previous 32-bit bins
      for any gain).

   In this case, cycle count differences (between the two DSPs) for the
   histogramming part, are only due to the 6414 cache architecture.

-  The second code based on loop16, a "brand new" program I developed
   and optimized for the 6414 last August.  It is designed to perform
   exactly the same computation as loop26.  Simulation studies done
   last September suggested that loop16 (the new code) should be twice
   as fast, on a 600 MHz 6414, than loop26 on a 300 MHz 6203.  This code
   has been verified on the hardware (the 6414 Demo board) just a few
   weeks ago.

   In this case, the histogramming part of the code, had to be modified
   since loop16 writes the energy in integer format rather than
   floating-point.  Also this histogramming code ideally (i.e. no cache
   inefficiency) needs 6 cycles for each of the 128 channels, as opposed
   to 6 cycles per channel-above-threshold for the histogramming code
   associated with loop26.

The DSP computation time has been obtained, in real time, by using one
of the DSP counters (the 6203 counter counts at 1/4 of the CPU frequency;
on the 6414 it counts at 1/8 of the CPU frequency).  This is a nice
alternative to using an oscilloscope, since the time per event is
now stored in the DSP output data along with the results of the
computation, and can be read out via VME.

The histogramming code allows to fill histograms only for channels
above a user-specified energy threshold. For all measurements,
this threshold was set to 1.9 ADC counts.
(The main optimal filtering code, loop26 or loop16, still computes
 energy, time and chi2 for all channels).

The main result is shown in the attached postscript file.
Event by event, I plot the computation time (in microseconds)
versus the number of channels above the energy threshold
(for which histograms are filled).

There are 4 main families of points:

- blue    is for loop26 on the 300 MHz 6203 silicon

- black   is for loop26 on the 600 MHz 6414 silicon

- red     is for loop16 on the 600 MHz 6414 silicon

- magenta is for loop16 - device simulator results (Code Composer)
          Here I plot only the first 64 events.

The three "horizontal" groups are for the optimal filtering
computation (loop26 or loop16) alone;

The four other groups, showing an increase in time as a function
of the number of channels histogrammed, show the total processing
time (optimal filtering plus histogramming).

The 6203 behaviour is a "straight line" since there is no cache;
for the 6414, the processing time fluctuates and is in fact
data-dependent.

Note also that the 6414 simulator does not exactly reproduce
the silicon.  However, if the simulator did not take into
account the cache at all, the magenta plot would be flat.

So, even if the 6414 is not twice as fast as the 6203 (especially
when histograms come into play) it is still slightly better, and
has another advantage - a bigger memory.


S.Simion - June 2002