                 COMMON BENCHMARKS FOR THE DSP56000/1

DSP56000 benchmark results for a number of common DSP building block
programs are shown.  These benchmarks are contained in Appendix B of the
DSP56000 Digital Signal Processor User's Manual (DSP56000UM/AD) or in the
Free Software section of Dr.BuB.  Note that none of these benchmarks use
memory-inefficient "straight-line" coding and none of them preclude the
use of interrupts.

Benchmark                                         Performance      
---------------------------------------------------------------------
N-tap real FIR filter with data shift         97.5 ns per tap

N-tap real, adaptive FIR filter               292.5 ns per tap
     with FIR and LMS update

N real, cascaded IIR biquad filter            390 ns per filter section
     sections (four coefficients)   

N-tap complex FIR filter with data shift      390 ns per tap
 
32-point complex FFT (radix 2,                68.1 us 
     looped code)

256-point complex FFT (radix 2,               0.706 ms
     looped code)

512-point complex FFT (radix 2,               2.04 ms
     looped code)

1024-point complex FFT                        3.39 ms
     (file: fftr2e.asm)

Matrix Multiplication [1x3][3x3]              1.7 us

Matrix Multiplication [10x10][10x10]          172 us

Two-dimensional convolution                   975 ns per output
     (3 x 3 coefficient mask)

Division                                      2.7 us 

Finding of maximum absolute value             195 ns per point
     and index in an array

Leroux-Gueguen LPC analysis:
    (file: lgsol1.asm)    8th order            46 us
                         10th order            61 us
                         16th order           117 us
==========================================================================

The next table shows how effectively the DSP56000 multiplier is utilized
in the kernel, or inner DO loop, of multiply intensive DSP algorithms.


                                           Multiply       DSP56000
Benchmark - Looped Kernel                 Operations       Cycles
-------------------------------------------------------------------
N term real FIR filter                        N               N
N term real Adaptive FIR filter              2N              3N
N term complex FIR filter                    4N              4N
N term cascaded biquad IIR filter            4N              4N
N complex multiplies                         4N              4N
Nth order power series (real)                 N              2N
N radix 2 FFT butterflies                    4N              6N
N 3 X 3 two dimensional convolutions         9N             10N


