RD test failure with reduced LAPACK

[eddelbuettel]

@kthohr I get this on R-devel with an embedded LAPACK on Ubuntu 17.10:

--------------------------- 
Test file: /home/edd/git/rcpparmadillo/RcppArmadillo.Rcheck/RcppArmadillo/unitTests/runit.Rlapack.R 
.setUp (before test.Rlapack): ERROR !! 
Error executing .setUp before test.Rlapack : Error in dyn.load("/tmp/Rtmp0RSN4J/sourceCpp-x86_64-pc-linux-gnu-0.12.16/sourcecpp_60e72fa3cfdb/sourceCpp_8.so") : 
  unable to load shared object '/tmp/Rtmp0RSN4J/sourceCpp-x86_64-pc-linux-gnu-0.12.16/sourcecpp_60e72fa3cfdb/sourceCpp_8.so':
  /tmp/Rtmp0RSN4J/sourceCpp-x86_64-pc-linux-gnu-0.12.16/sourcecpp_60e72fa3cfdb/sourceCpp_8.so: undefined symbol: zgbsv_
--------------------------- 

I presume this passed on your macOS box?

[kthohr]

Hmm... zgbsv is only used with band matrices (and is indeed missing on Rlapack). I will investigate why this is being called by my tests.

And yes passed on my macOS computer (which uses system Lapack).

[kthohr]

I just ran the tests on a Windows machine with Rlapack and everything passed.

Executing test function test.Rlapack  ...  done successfully.

Just to be clear: that R-devel is linking with Rlapack? If so, the crippled_lapack switch must be missing.

[kthohr]

I think this is the culprit: https://github.com/RcppCore/RcppArmadillo/blob/master/configure.ac#L85

The crippled_lapack switch needs to be enabled for all Rlapack cases.

[eddelbuettel]

Doh. I think you are spot on.

[kthohr]

Actually this PR might fix @baptiste 's issue, too.

#207 (comment)

[eddelbuettel]

But only platforms that do not use the embedded Rlapack -- on those he would have to add the lapack routine called by Armadillo in his package to fill-in. Or am I missing something here?

[kthohr]

From @baptiste 's description, the Linux build on CRAN failed due to a generic solve() call when the inputs were complex-valued. This is also what my unit test was doing.

From my understanding of the Armadillo code, solve() will alternate between different solvers depending on the properties of the input matrices, which includes the band matrix case, but this special case isn't required to return a solution (it's just used to exploit a simplified structure where available). ARMA_CRIPPLED_LAPACK should rule out these routines:

https://github.com/conradsnicta/armadillo-code/blob/unstable/include/armadillo_bits/auxlib_meat.hpp#L3755

My guess is that the configure script saw an R version > 3.3.0 and did not set ARMA_CRIPPLED_LAPACK, so the band matrix routines were not removed.

[baptiste]

This is also my understanding, as was discussed at length on the mailing list. Thank you for the work, and I believe it should fix such issues much more reliably than testing for OS and R version, which (especially on CRAN) are not the relevant variables to enforce the crippled flag. My package was archived by CRAN following this failure and will not automatically rebuild; I will not have time in the near future to work on this and push a new version but I appreciate everyone's efforts to improve the global situation. The R-devel discussion also suggests that a full Lapack could potentially be included, especially if someone were to submit a patch; it seems to be more of a case of not having resources to provide a careful implementation for the remaining wrappers than any technical or philosophical reservations. Regards, baptiste

…

On 30 March 2018 at 02:02, kthohr ***@***.***> wrote: From @baptiste <https://github.com/baptiste> 's description, the Linux build on CRAN failed due to a generic solve() call when the inputs were complex-valued. This is also what my unit test was doing. From my understanding of the Armadillo code, solve() will alternate between different solvers depending on the properties of the input matrices, which includes the band matrix case, but this special case isn't required to return a solution (it's just used to exploit a simplified structure where available). ARMA_CRIPPLED_LAPACK should rule out these routines: https://github.com/conradsnicta/armadillo-code/blob/unstable/include/ armadillo_bits/auxlib_meat.hpp#L3755 My guess is that the configure script saw an R version > 3.3.0 and did not set ARMA_CRIPPLED_LAPACK, so the band matrix routines were not removed. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#213 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AABKGnG2hBy1ixUng3-uGQZ6M9X_DS9lks5tjNtPgaJpZM4TAHfg> .