isuruf
commented
3 years ago @conda-forge/help-osx-arm64, looks like checking for rlog1p is disabled in cross compiling mode.
Closed yfyang86 closed 4 months ago
isuruf
commented
3 years ago @conda-forge/help-osx-arm64, looks like checking for rlog1p is disabled in cross compiling mode.
znmeb
commented
2 years ago I can reproduce the Rlog1p erroron an NVIDIA Jetson (arm64) using the current mambaforge installer in Docker. I've got time to troubleshoot if it will help.
> install.packages("RcppArmadillo")
trying URL 'https://ftp.osuosl.org/pub/cran/src/contrib/RcppArmadillo_0.10.7.3.0.tar.gz'
Content type 'application/x-gzip' length 1360186 bytes (1.3 MB)
==================================================
downloaded 1.3 MB
* installing *source* package 'RcppArmadillo' ...
** package 'RcppArmadillo' successfully unpacked and MD5 sums checked
** using staged installation
checking whether the C++ compiler works... yes
checking for C++ compiler default output file name... a.out
checking for suffix of executables...
checking whether we are cross compiling... no
checking for suffix of object files... o
checking whether we are using the GNU C++ compiler... yes
checking whether aarch64-conda-linux-gnu-c++ -std=gnu++14 accepts -g... yes
checking how to run the C++ preprocessor... aarch64-conda-linux-gnu-c++ -std=gnu++14 -E
checking whether we are using the GNU C++ compiler... (cached) yes
checking whether aarch64-conda-linux-gnu-c++ -std=gnu++14 accepts -g... (cached) yes
checking whether we have a suitable tempdir... /tmp
checking whether R CMD SHLIB can already compile programs using OpenMP... yes
checking LAPACK_LIBS... system LAPACK found
configure: creating ./config.status
config.status: creating inst/include/RcppArmadilloConfigGenerated.h
config.status: creating src/Makevars
** libs
aarch64-conda-linux-gnu-c++ -std=gnu++11 -I"/home/synth/miniconda3/envs/r-reticulate/lib/R/include" -DNDEBUG -I'/home/synth/miniconda3/envs/r-reticulate/lib/R/library/Rcpp/include' -DNDEBUG -D_FORTIFY_SOURCE=2 -O2 -isystem /home/synth/miniconda3/envs/r-reticulate/include -I/home/synth/miniconda3/envs/r-reticulate/include -Wl,-rpath-link,/home/synth/miniconda3/envs/r-reticulate/lib -I../inst/include -fpic -fvisibility-inlines-hidden -fmessage-length=0 -ftree-vectorize -fPIC -fstack-protector-strong -fno-plt -O3 -pipe -isystem /home/synth/miniconda3/envs/r-reticulate/include -fdebug-prefix-map=/home/conda/feedstock_root/build_artifacts/r-base-split_1638516584669/work=/usr/local/src/conda/r-base-4.1.1 -fdebug-prefix-map=/home/synth/miniconda3/envs/r-reticulate=/usr/local/src/conda-prefix -c RcppArmadillo.cpp -o RcppArmadillo.o
In file included from /home/synth/miniconda3/envs/r-reticulate/lib/R/library/Rcpp/include/RcppCommon.h:74,
from ../inst/include/RcppArmadilloForward.h:25,
from ../inst/include/RcppArmadillo.h:29,
from RcppArmadillo.cpp:21:
../inst/include/armadillo_bits/eop_aux.hpp: In static member function 'static typename arma::arma_integral_only<T>::result arma::eop_aux::Rlog1p(eT)':
/home/synth/miniconda3/envs/r-reticulate/lib/R/include/Rmath.h:75:15: error: 'Rlog1p' is not a member of 'std'; did you mean 'log1p'?
75 | #define log1p Rlog1p
| ^~~~~~
../inst/include/armadillo_bits/eop_aux.hpp:99:124: note: in expansion of macro 'log1p'
99 | template<typename eT> arma_inline static typename arma_integral_only<eT>::result log1p (const eT x) { return eT( std::log1p(double(x)) ); }
| ^~~~~
../inst/include/armadillo_bits/eop_aux.hpp: In static member function 'static typename arma::arma_real_only<T>::result arma::eop_aux::Rlog1p(eT)':
/home/synth/miniconda3/envs/r-reticulate/lib/R/include/Rmath.h:75:15: error: 'Rlog1p' is not a member of 'std'; did you mean 'log1p'?
75 | #define log1p Rlog1p
| ^~~~~~
../inst/include/armadillo_bits/eop_aux.hpp:100:120: note: in expansion of macro 'log1p'
100 | template<typename eT> arma_inline static typename arma_real_only<eT>::result log1p (const eT x) { return std::log1p(x);
}
| ^~~~~
../inst/include/armadillo_bits/fn_misc.hpp: In function 'typename arma::arma_real_only<T>::result arma::log_add_exp(eT, eT)':
/home/synth/miniconda3/envs/r-reticulate/lib/R/include/Rmath.h:75:15: error: 'Rlog1p' is not a member of 'std'; did you mean 'log1p'?
75 | #define log1p Rlog1p
| ^~~~~~
../inst/include/armadillo_bits/fn_misc.hpp:172:26: note: in expansion of macro 'log1p'
172 | return (log_a + std::log1p(std::exp(negdelta)));
| ^~~~~
make: *** [/home/synth/miniconda3/envs/r-reticulate/lib/R/etc/Makeconf:177: RcppArmadillo.o] Error 1
ERROR: compilation failed for package 'RcppArmadillo'
* removing '/home/synth/miniconda3/envs/r-reticulate/lib/R/library/RcppArmadillo'
The downloaded source packages are in
'/tmp/RtmpJaPZdK/downloaded_packages'
Updating HTML index of packages in '.Library'
Making 'packages.html' ... done
Warning message:
In install.packages("RcppArmadillo") :
installation of package 'RcppArmadillo' had non-zero exit status
>
yfyang86
commented
2 years ago @znmeb This suggestion may help. It would be great if you could share the Dockerfile (and docker configuration) to reproduce the issue. I only tried triton alike on Jetson Kit, but no experience on R (on Jetson) yet.
znmeb
commented
2 years ago @yfyang86 OK ... let me strip all the extraneous stuff out - the image is 6 GB at the moment and I can probably make it happen on the default Ubuntu 18.04 arm64 image.
znmeb
commented
2 years ago @yfyang86 Done. Repository is https://github.com/AlgoCompSynth/RcppArmadillo-crash-dockerfile. In the process of building it I "accidentally" discovered that RcppArmadillo installs successfully with R 3.6.3 but not with R 4.1.1. One thing you might want to check is the build processes for the Rmath library. https://cran.r-project.org/doc/manuals/r-release/R-admin.html#The-standalone-Rmath-library. In the failure that led me to this issue, Rmath was prominent in the error messages.
I'll try my issue (package warbleR) tomorrow, although dropping back to R 3.6.3 isn't an option.
akrherz
commented
2 years ago FWIW, I found with conda-forge/r-sf-feedstock#66 that adding the following in recipe/build.sh was sufficient to get a successful build.
export PKG_CPPFLAGS="-DHAVE_WORKING_LOG1P"I did not attempt testing if the compiled package worked or not.
znmeb
commented
2 years ago FWIW, I found with conda-forge/r-sf-feedstock#66 that adding the following in
recipe/build.shwas sufficient to get a successful build.export PKG_CPPFLAGS="-DHAVE_WORKING_LOG1P"I did not attempt testing if the compiled package worked or not.
I'll try that with warbleR - that's the package I was having trouble with.
znmeb
commented
2 years ago Yep - warbleR compiles with this definition in place.
isuruf
commented
2 years ago We need to set some variables. Here's a list thanks to @erykoff,
r_cv_size_max, ac_cv_c_stack_direction, r_cv_putenv_unset, r_cv_putenv_unset2, r_cv_func_calloc_works, r_cv_func_isfinite_works, r_cv_func_log1p_works, r_cv_working_ftell, r_cv_func_sigaction_works, r_cv_mktime_errno, r_cv_working_mktime, r_cv_func_ctanh_works, r_cv_icu, r_cv_header_zlib_h, r_cv_have_bzlib, r_cv_have_lzma, r_cv_have_pcre2utf, r_cv_have_pcre832, r_cv_have_curl728, r_cv_have_curl_https, r_cv_kern_usrstack, r_cv_libc_stack_end, ac_cv_func_mmap_fixed_mapped, r_cv_openmp_simdred
genecell
commented
2 years ago FWIW, I found with conda-forge/r-sf-feedstock#66 that adding the following in
recipe/build.shwas sufficient to get a successful build.export PKG_CPPFLAGS="-DHAVE_WORKING_LOG1P"I did not attempt testing if the compiled package worked or not.
Thanks! this works for me on installing the R package sctransform in M1 Pro Max:
This is the error I met:
error: no member named 'Rlog1p' in namespace 'std'; did you mean simply 'Rlog1p'?I first run export PKG_CPPFLAGS="-DHAVE_WORKING_LOG1P" in the terminal and then install the package in R.
mfansler
commented
1 year ago In case anyone else finds exporting isn't sufficient, here's another example situation. For r-ade, the package's src/Makevars already defined PKG_CPPFLAGS, so the export directive gets ignored. Instead we insert the flag via sed:
sed -ie 's/PKG_CPPFLAGS =/PKG_CPPFLAGS = -DHAVE_WORKING_LOG1P/' src/MakevarsFor anyone doing an interactive install in R (as above), the withr package has functions for similar on the fly modifications to the Makevars.
Br1anChou
commented
1 year ago This is due to the ARM processor, you need to download the arm64 packages from CRAN or Bioconductor and compile it manually, like R CMD INSTALL XXX.tgz
yfyang86
commented
1 year ago This issue is too old and is solved according to the above discussions. I will close it.
skullcap4704
commented
1 year ago FWIW, I found with conda-forge/r-sf-feedstock#66 that adding the following in
recipe/build.shwas sufficient to get a successful build.export PKG_CPPFLAGS="-DHAVE_WORKING_LOG1P"I did not attempt testing if the compiled package worked or not.
Thanks! this works for me on installing the R package
sctransformin M1 Pro Max:This is the error I met:
error: no member named 'Rlog1p' in namespace 'std'; did you mean simply 'Rlog1p'?I first run
export PKG_CPPFLAGS="-DHAVE_WORKING_LOG1P"in the terminal and then install the package in R.
It doesn't work properly. So I modified CPPFLAGS on $R_HOME/etc/Makeconf, It worked!
mfansler
commented
1 year ago @skullcap4704 interesting. That might be a better solution for Conda Forge, i.e., we might want to consider adding the flag generally by having r-base deliver a modified lib/R/etc/Makeconf with the flag.
Right now we're multiplying our efforts by having to add the workaround to each R package feedstock. @conda-forge/r thoughts?
(@yfyang86 I'm reopening as I think this still merits discussion)
wolfv
commented
4 months ago I have dug a little bit and compiled r-base with conda-build locally on a osx-arm64 machine. I think the main problem lies in the generated Rmath.h file. This file looks like the following on the natively compiled:
#ifndef HAVE_WORKING_LOG1P
# define HAVE_WORKING_LOG1P 1
#endif
#if !defined(HAVE_WORKING_LOG1P)
/* remap to avoid problems with getting the right entry point */
double Rlog1p(double);
#define log1p Rlog1p
#endifAnd on the cross-compiled, conda-forge version it looks like this:
#ifndef HAVE_WORKING_LOG1P
# undef HAVE_WORKING_LOG1P
#endif
...I do see the following in the local autotools logs:
checking for working isfinite... yes
checking for working log1p... yes
checking whether ftell works correctly on files opened for append... yes
checking for working sigaction... yesSo I am not sure why the Rconfig.h file does not include HAVE_WORKING_LOG1P.
The check in autotools does something like this:
## R_FUNC_LOG1P
## ------------
## Suggested by Nelson H. F. Beebe <beebe@math.utah.edu> to deal with
## inaccuracies on at least NetBSD 1.6 and OpenBSD 3.2.
## However, don't test all the way into denormalized x (he had k > -1074)
## and at x = 2^-54 (d - x)/x is around 3e-17.
AC_DEFUN([R_FUNC_LOG1P],
[AC_CACHE_CHECK([for working log1p], [r_cv_func_log1p_works],
[AC_RUN_IFELSE([AC_LANG_SOURCE([[
#include <math.h>
#include <stdlib.h>
#include "confdefs.h"
int main (void) {
#ifdef HAVE_LOG1P
int k;
double d;
double x = 1.0;
for(k = 0; k < 53; k++) x /= 2.0;
/* log(1+x) = x - (1/2)x^2 + (1/3)x^3 - (1/4)x^4 ... */
/* = x for x sufficiently small */
for(k = -54; k > -1022; --k) {
x /= 2.0;
if(x == 0.0)
exit(0); /* OK: reached underflow limit */
d = log1p(x);
if(d == 0.0)
exit(1); /* ERROR: inaccurate log1p() */
/* for large k, ((1/2)x^2)/x might appear in the guard digits */
if(k < -80 && d != x)
exit(1); /* ERROR: inaccurate log1p() */
}
exit(0);
#else
exit(1);
#endif
}
]])],
[r_cv_func_log1p_works=yes],
[r_cv_func_log1p_works=no],
[r_cv_func_log1p_works=no])])
if test "x${r_cv_func_log1p_works}" = xyes; then
AC_DEFINE(HAVE_WORKING_LOG1P, 1,
[Define if log1p() exists and is accurate enough.])
RMATH_HAVE_WORKING_LOG1P="# define HAVE_WORKING_LOG1P 1"
else
RMATH_HAVE_WORKING_LOG1P="# undef HAVE_WORKING_LOG1P"
fi
AC_SUBST(RMATH_HAVE_WORKING_LOG1P)
])# R_FUNC_LOG1PInterestingly the Windows Makefile also seems to replace the value (hardcoded):
Rmath.h0: Rmath.h0.in $(R_HOME)/VERSION Makefile.win
@$(SED) \
-e 's/@RMATH_HAVE_WORKING_LOG1P@/# define HAVE_WORKING_LOG1P 1/' \
-e "s/@PACKAGE_VERSION@/`sed 's/\([^ ]*\).*/\1/' < $(R_HOME)/VERSION`/" $< > Rmath.h0So I think it should be easy to follow this patch on Windows and just replace @RMATH_HAVE_WORKING_LOG1P@ with the correct value in Rmath.h0.in.
I don't have a lot of autotools-fu so I am not sure if there would also be a good way to override that check.
Maybe @isuruf, since we briefly discussed this today, you have some more ideas? Otherwise I can send a PR with the sed line tomorrow :)
isuruf
commented
4 months ago wolfv
commented
4 months ago OK, I made the fixes as requested in #314. Curious if that's going to work :)
mbargull
commented
4 months ago Builds from gh-314 should fix this for macOS.
We would need the same for the Linux cross-compiled builds acc. to https://github.com/conda-forge/r-base-feedstock/issues/163#issuecomment-988442868 .
The configure logs from gh-314 also show checking for working log1p... no for linux-aarch64/linux-ppc64le.
Might make sense to compare with the configure logs from native builds of those platforms to gauge what's applicable.
isuruf
commented
4 months ago Hmm, that's true. We had native builds until recently where this issue didn't come up.
isuruf
commented
4 months ago
Issue:
Environment (
conda list):Details about
condaand system (conda info):Issue 1 (see solution later):
When I start the R console, it shows
Issue 2 (Details RcppArmadillo330)
When using
Rcpp, it will fail onstd::log1p, i.e.HAVE_WORKING_LOG1Pis not detected in R configuration stage.Possible Solutions
ISSUE1/ISSUE2
a.
___kmpc_for_static_finiis related to openMP, hence for the current environment,--disable-openmpmight work;b. There is no reason to use OpenBLAS since some benchmarks illustrate that vecLib outperforms OpenBLAS alike. Besides, I came across the same
FFLAGerrors when compiling R--with-BLAS-shlib.I recompiled R with the following options (without X, and veclib should be replaced by
-framework Accelerate):Both issues vanish.