R/speed_analysis_R_vs_C.R

In andreas50/utsOperators: Moving Averages and Other Rolling Operators for Unevenly Spaced Time Series

#############################################
# Speed analysis of R vs. C implementations #
#############################################

# Hardware: i7-2600, 32GB RAM
# Software: Windows 7 Pro 64bit, R 3.5.1, gcc-4.9.3
# Date:  8/2018

### sma(..., interpolation="last")
# -) for a moderate-length time series, the C implementation is ~110 times faster
if (0) {
  ts1 <- ex_uts3()
  width <- ddays(100)
  
  # R vs. C: 2.11s vs. 1.60s
  system.time(for (j in 1:2000) sma_last_R(ts1, width))
  system.time(for (j in 1:20000) sma(ts1, width, interpolation="last"))
  
  # Profile C implementation
  # -) ~15% of time spent in C implementation
  # -) argument checking takes up most of the time
  Rprof(interval=0.01)
  for (j in 1:5e4) sma(ts1, width, interpolation="last")
  Rprof(NULL)
  summaryRprof()
}


### sma(..., interpolation="linear")
# -) for a moderate-length time series, the C implementation is ~180 times faster
if (0) {
  ts1 <- ex_uts3()
  width <- ddays(100)
  
  # R vs. C: 3.73s vs. 1.72s
  system.time(for (j in 1:2000) sma_linear_R(ts1, width))
  system.time(for (j in 1:20000) sma(ts1, width, interpolation="linear"))
  
  # With and without inlined helper functions: 1.87s vs. 1.92s
  # -) just using "static" seems to give the same speed increase, because the compiler automatically inlines helper functions, even if not asked to
  
  # Profile C implementation
  # -) ~20% of time spent in C implementation
  # -) argument checking takes up most of the time
  Rprof(interval=0.01)
  for (j in 1:5e4) sma(ts1, width, interpolation="linear")
  Rprof(NULL)
  summaryRprof()
}