[Discuss-gnuradio] complex dotprod speedup

[Eric Blossom]

Thanks to some serious SSE and 3DNow! hacking by Stephane Fillod, we
now have a much faster version of the complex/complex/complex dot
product function.  This function is at the bottom of the
gr.freq_xlating_fir_filter_ccc function.  This function performs
channel selection and digital downconversion on complex data.  This is
really handy when using the USRP, since we're dealing with complex
data on the host.

Below is a selection of benchmark results on different machines.  One
thing that I find interesting is the wide variation between the
generic and SIMD times as a function of machine microarchitecture.
The generic implementations are C++ with partial loop unrolling.

Eric


taps     = number of filter taps
input    = number of input samples
cpu      = combined user+sys cpu time
taps/sec = derived performance measure.  Higher is better.


=== This machine is a Pentium M (1.4 GHz) ===

address@hidden tests]$ ./benchmark_dotprod_ccc
   generic: taps:  256  input: 4e+07  cpu: 121.476  taps/sec:   8.43e+07
       SSE: taps:  256  input: 4e+07  cpu:  39.010  taps/sec:  2.625e+08

=== This machine is an Athlon MP 1800+ (1.5 GHz) ===

address@hidden tests]$ ./benchmark_dotprod_ccc
   generic: taps:  256  input: 4e+07  cpu: 118.090  taps/sec:  8.671e+07
 3DNow!Ext: taps:  256  input: 4e+07  cpu:  29.705  taps/sec:  3.447e+08
    3DNow!: taps:  256  input: 4e+07  cpu:  33.213  taps/sec:  3.083e+08
       SSE: taps:  256  input: 4e+07  cpu:  37.242  taps/sec:   2.75e+08

=== This machine is a Pentium 4 (1.7 GHz) ===

address@hidden tests]$ ./benchmark_dotprod_ccc
   generic: taps:  256  input: 4e+07  cpu: 156.850  taps/sec:  6.529e+07
       SSE: taps:  256  input: 4e+07  cpu:  23.241  taps/sec:  4.406e+08