[Rcpp-devel] segmentation fault when wrapping a big Eigen 2-d array

[Jack Wasey]

On Mon, Apr 09, 2018 at 07:09:55AM -0500, Dirk Eddelbuettel wrote:
> 
> On 9 April 2018 at 07:22, Wasey, Jack O wrote:
> | Dear Rcpp developers,
> | 
> | I'm having trouble with a reproducible error when wrapping a big Eigen dense matrix.
> | 
> | Rcpp::LogicalMatrix mat_out_bool = Rcpp::wrap(result);
> | 
> | Thread 1 "R" received signal SIGSEGV, Segmentation fault.
> | coerceToLogical (v=0x7fffa1a1c010) at coerce.c:441
> | 
> | Where 'result' is an Eigen::MatrixXi of size 472132x30, which is the result of a sparse x dense matrix multiplication.
> | DenseMap result = visit_codes_sparse * map;
> | 
> | The backtrace begins:
> | #0  coerceToLogical (v=0x7fffa1a1c010) at coerce.c:441
> | #1  Rf_coerceVector (v=<optimized out>, type=<optimized out>) at coerce.c:1243
> | #2  0x00007fffdc660515 in Rcpp::internal::basic_cast<10> (x=0x7fffa1a1c010) at /usr/local/lib/R/site-library/Rcpp/include/Rcpp/r_cast.h:66
> | #3  0x00007fffdc65edc5 in Rcpp::internal::r_true_cast<10> (x=0x7fffa1a1c010) at /usr/local/lib/R/site-library/Rcpp/include/Rcpp/r_cast.h:95
> | #4  0x00007fffdc664a9d in Rcpp::r_cast<10> (x=0x7fffa1a1c010) at /usr/local/lib/R/site-library/Rcpp/include/Rcpp/r_cast.h:163
> | #5  0x00007fffdc666f99 in Rcpp::Matrix<10, Rcpp::PreserveStorage>::Matrix (this=0x7fffffff23a0, x=0x7fffa1a1c010)
> |     at /usr/local/lib/R/site-library/Rcpp/include/Rcpp/vector/Matrix.h:53
> | #6  0x00007fffdc68603e in icd9Comorbid_alt_MatMul (icd9df=..., icd9Mapping=..., visitId="visit_id", icd9Field="code", threads=8, chunk_size=256, omp_chunk_size=1)
> |     at comorbid_alt_MatMul.cpp:271
> | #7  0x00007fffdc64a597 in _icd_icd9Comorbid_alt_MatMul_try (icd9dfSEXP=0x555558590cd8, icd9MappingSEXP=0x55555839d9d0, visitIdSEXP=0x55555a25e2b8, icd9FieldSEXP=0x55555c066298,
> |     threadsSEXP=0x5555623279f8, chunk_sizeSEXP=0x5555586122e8, omp_chunk_sizeSEXP=0x555558612258) at RcppExports.cpp:347
> | 
> | I would love to give a small reproducible example, but at this point, I can only trigger the problem with a big data set (which makes valgrind unreasonably slow), and only by repeating the computation, as sometimes it does succeed. The code is in the eigen-sparse branch of https://github.com/jackwasey/icd . I do not expect anyone to download and debug my code (which FWIW compiles without any warnings), but I would like to know whether, in principle, I am doing something wrong with the integer to logical 'cast' when wrapping, or if this is possibly an Rcpp or RcppEigen bug. I can induce the crash on a server with huge amounts of RAM, and on Mac and Linux platforms, with and without OpenMP enabled.
> 
> From the "well then don't do it" school:  can you cast _before_ you interact
> with R / Rcpp / RcppEigen?   What happens when you try the "big" operation
> entirely in C++ / Eigen and then "merely" transfer a known object (ie an int
> vector/matrix) ?  If that works, try bool and then see if that maps to Logical.
> 
> Some code paths are simply less well trodden.  Int would be first route.
> 
> I may of course be entirely off the page. Early morning  ...
> 
> Dirk
> 
> -- 
> http://dirk.eddelbuettel.com | @eddelbuettel | edd at debian.org

I seem to tread in unusual code paths often, which is
why I appreciate so much your guidance. I can't reproduce the error now
I've converted from integer Eigen matrix to Rcpp::IntegerMatrix, then
within Rcpp, Integer to Logical matrices. 

I naively assumed that if the compiler let me do it, it was a valid
cast. Does this mean I unearthered a bug somewhere down the line of
dependencies?

Thanks,
Jack