Closed michaelweylandt closed 6 years ago
Pushed a new version which i) uses the fastLmPure
functions and ii) switches to microbenchmark
Here's example output:
> source("lmBenchmark.R")
lm benchmark for n = 100000 and p = 40: nrep = 20
Method Relative Elapsed (milliseconds)
4 LDLt 1.000000 16.59547
5 LLt 1.001358 16.61801
10 SymmEig 2.177705 36.14004
8 QR 7.230222 119.98892
7 PivQR 7.772614 128.99018
6 lm.fit 13.522861 224.41823
1 arma 17.717057 294.02287
2 GESDD 31.514008 522.98975
9 SVD 56.295458 934.24953
3 GSL 112.837376 1872.58917
Thank you! Almost there -- the microbenchmark
use requires a bit more care. Maybe tarball it at your end and run R CMD check --as-cran
on it for good measure.
Now you must add r-cran-microbenchmark
here to appease Travis.
Sorry about that - I didn't get a build-failed email from Travis.
Done now -- Travis passes with a WARNING
about missing vignette source and a NOTE
about large directories. Both appear to be unrelated to this change.
Yes, they are constants. In the setup I prefer at Travis, vignettes are skipped, and you get idea about directory sizes.
The RcppEigen benchmark script calls a C++ function which is no longer provided by RcppArmadillo (since mid-August 2017). This updates the benchmark script to call the RcppAttributes-generated entry point
_RcppArmadillo_fastLm_impl