README.md

Romp

Basic examples using OpenMP with R, for C, C++, F77, and Fortran 2003.

Examples

There are 3 examples, using each of C, C++ (Rcpp), F77, and F2003.

Hello World

A basic OpenMP hello world. Note that the order of thread printing is not guaranteed.

Sum

Sums up a numeric vector.

Sweep

Sweeps a numeric vector from a numeric matrix of the same length as the number of rows of that matrix. Equivalent to calling sweep(x, STATS=vec, MARGIN=1, FUN="-") in R.

Benchmarks

I wouldn't take the numbers here too seriously, especially for the sum example, where they appear to be statistically identical. The languages separate on the sweep example, though possibly for implementation reasons.

The benchmarks are located at Romp/inst/benchmarks/ of the source tree. All tests were performed using:

Sum

         test replications elapsed relative
5  f90_sum(x)          100   1.178    1.000
4  f77_sum(x)          100   1.179    1.001
2    c_sum(x)          100   1.186    1.007
3 rcpp_sum(x)          100   1.196    1.015
1      sum(x)          100   2.605    2.211

Sweep

                test replications elapsed relative
4  f77_sweep(x, vec)          100   5.242    1.000
5  f90_sweep(x, vec)          100   5.315    1.014
2    c_sweep(x, vec)          100   5.354    1.021
3 rcpp_sweep(x, vec)          100  14.966    2.855
1    r_sweep(x, vec)          100  52.330    9.983

Primes Below an Integer

                 test replications elapsed relative
1    c_primesbelow(n)          100   9.054    1.000
3  f90_primesbelow(n)          100   9.629    1.064
2  f77_primesbelow(n)          100   9.656    1.066
4 rcpp_primesbelow(n)          100  10.391    1.148

Integration with R

If you are unfamiliar with integrating C, C++, or Fortran into R, then the following will hopefully be of use to you.

C

We use the .Call() interface. There is a (nearly) deprecated .C() interface, which you should not use, as it has serious performance loss compared to the .Call() interface.

If you are interested in some simplifications of R's C interface but don't want to jump to C++, you might consider taking a look at the C project RNACI, which is also available as a header-only library.

C++

Here we use Rcpp. Using RcppAttributes (noted by the // [[Rcpp::export]] calls), we can write something that looks very much like C++, and use Rcpp's (R function) compileAttributes() to generate C-level and R-level R wrappers. The script Romp/resrc does this.

Fortran

For both F77 and F90+, integration with R is non-trivial. As with C, there is a (nearly) deprecated interface .Fortran() which you should not use due to its large performance overhead. Instead, you should:

  1. Write your Fortran code.
  2. Write a C wrapper of the Fortran code using R's .Call() interface (or Rcpp if you prefer).
  3. Call the C code from R using .Call(().

As noted, you can use Rcpp in lieu of R's basic C interface for wrapping Fortran code, though I seriously recommend against it. Bringing in C++ can complicate linking, among other things, and for wrapping C/Fortran code, in my opinion, Rcpp brings little to the table (C++ is another story!). If you are interested in some simplifications of R's C interface but don't want to jump to use Rcpp, you might consider taking a look at RNACI.

It is difficult to reliably use F90+ functions and subroutines which live in modules without the use of the F2003 iso_c_binding module (which has been supported by every compiler for ages). This package gives some nice examples of how to use module code.

A final note about F90+ in particular, is that some of the advice in Writing R Extensions is not always entirely accurate.



wrathematics/Romp documentation built on May 4, 2019, 9:48 a.m.