On 01/08/2014 07:43 AM, Julian Gehring wrote:
Hi,
R-2.13 introduced the byte code compiler package 'compiler' [1], which can be
used to precompile the R code of a package at installation time (using the
ByteCompile field in the Description file or '--byte-compile' with R CMD
INSTALL). I have been using this lately with my own packages, and had some
notable improvements (and also did not find a case where it decreased the
performance). I was wondering if anyone has extensive experience with this.
Specifically, are there good reasons not to use this by default for packages?
my prejudice (i.e., no actual data) is that improvements from byte compilation
point to code that could be vectorized (for even greater improvement) instead --
for instance, replacing a for loop over a vector with a vectorized alternative
library(compiler)
library(microbenchmark)
f = function(n) { i = 0; for (j in seq_len(n)) i = i + 1; i }
fc = cmpfun(f)
v = function(n) { i = integer(n); i[] = 1; sum(i) }
identical(f(1e5), v(1e5))
## [1] TRUE
microbenchmark(f(1e5), fc(1e5), v(1e5), times=10)
## Unit: milliseconds
## expr min lq median uq max neval
## f(1e+05) 32.246407 36.119512 37.028699 37.451135 37.894748 10
## fc(1e+05) 13.777743 14.610523 15.151645 15.303884 15.580075 10
## v(1e+05) 2.628862 2.644882 2.803085 2.862608 3.041894 10
The speed-up in the example (2x) seems to be more or less typical of 'good'
improvement (what is 'noticeable' in your case?) from compiling functions, and
for lots of cases there can easily be improvements of order(s) of magnitude by
bringing the code in line with R's design principles.
I think byte compilation is intended to be fully compatible with R, so there
shouldn't be any 'errors' introduced (and if there were, these would be fixed).
There is a cost to byte compilation, but this is paid at the time of
installation rather than loading and is not really noticeable except for large
packages. If your code is well-written and there are still improvements from
byte compilation, then it seems reasonable to use byte compilation as a
package-level setting.
Martin
Best wishes
Julian
[1] http://dirk.eddelbuettel.com/blog/2011/04/12/
_______________________________________________
Bioc-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/bioc-devel
--
Computational Biology / Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N.
PO Box 19024 Seattle, WA 98109
Location: Arnold Building M1 B861
Phone: (206) 667-2793
_______________________________________________
Bioc-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/bioc-devel
