The fractional resampler is frequently the most computationally expensive block
in a transmitter (and even in a receiver). It is also notoriously difficult to
parallelize. GR comes with a few options for non-integer resampling:
polyphase, mmse, rational. I benchmarked all of them on a variety of machine
for a rate change of 1.023. Turns out the old fractional_resampler is a bit
faster than the polyphase. However, it was still too slow (could not keep up
at ~30 Msps).
A one-line change to the fractional_resampler_[cc/ff] increases throughput by
almost 3x. The current fractional resampler uses the basic filter kernel. The
filter kernel is really pretty clever - uses SIMD via VOLK and goes to great
lengths to keep everything aligned on cache boundaries. All that overhead is
worth it when the filter has lots of taps. However, the resampler filter is
exactly 8 taps. The overhead of calling into the filter kernel is much, much
larger than the work done in the kernel.
Simply replace the line:
out[oo] = d_resamp->interpolate(&in[ii], d_mu);
with
auto TAPS = taps[(int) d_mu];
out[oo] = in[ii+0] * TAPS[7] + in [ii+1] * TAPS[6] + ... in [ii+7] *
TAPS[0].
taps come from 'interpolator_taps.h'
I wish I could just post the diff myself, but company policy does not allow me
to.
Instant speedup of 2.5X (complex), 3X (real). It may be even faster if we used
a SIMD intrinsic, but compilers are really pretty good at optimizing these days!
In the immortal words of Donald Knuth: Premature optimization is the root of
all evil.
Eugene Grayver, Ph.D.
Principal Engineer
Digital Communications Implementation Division
(310) 336-1274
_______________________________________________
Discuss-gnuradio mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
