search

ambisonictoolkit / Hilbert

SC3 Hilbert Transform Quark Library
GNU General Public License v3.0
5 stars 1 forks source link

Benchmark CPU results #3

Open joslloand opened 7 years ago

joslloand commented 7 years ago

https://github.com/mtmccrea/Hilbert/issues/10

Quick benchmarks:

  • late 2016 Macbook Pro
  • 50 instances of [real, imag] output

Weaver method notes // s.options.numWireBufs_(64*16); // required for 50 instances of HilbertW (pv_brickwall implementation)

  • *ar, DelayN (real) & PV_brickwall (imag, size: 2048): avg 9.2%, pk 17.5%, avg fluctuates quite a bit

PDN method notes // it appears the sin/cos parts are reversed or the imaginary part has its phase inverted // magnitude response not so great at high freqs... especially the FOS version

  • *ar, FOS: avg 14%, pk 19.5%
  • *ar1, SOS: avg 11%, pk 14.7%
  • ar2, SOS, refactored: avg 11%, pk 14.7% (same as ar1)

Hartley method notes // currently uses the DelayL version for the real part...

  • *ar, DelayL & Convolution2 (size: 2048) : avg 4.7%, pk 17.6%, disparity is LARGE
joslloand commented 7 years ago

Since these measurements were made, the implementations have changed.

New measurements are required for documentation.