[Rd] hand compile; link to MKL fails at BLAS zdotu

[Ivan Krylov]

В Sat, 30 Mar 2024 10:55:48 +0000
Ramón Fallon <ramonfallon using gmail.com> пишет:

> In contrast to Dirk's solution, I've found R's configure script
> doesn't recognise the update-alternatives system on debian/ubuntu, if
> it's MKL.

It ought to work if configured with --with-blas=-lblas
--with-lapack=-llapack, but, as you found out (and I can confirm), if
libblas.so and liblapack.so already point to MKL, ./configure somehow
fails the test for zdotu and falls back to bundled Rblas and Rlapack.

If you'd like the built R to work with the update-alternatives system,
the workaround seems to help is to temporarily switch the alternatives
to reference BLAS & LAPACK, configure and build R, and then switch the
alternatives back to MKL.

> appending "-lmkl_gf_lp64" to the --with-blas option does not help
> (that's suggested by several posts out there).

MKL has an official "link line advisor" at
<https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-link-line-advisor.html>,
which may suggest a completely different set of linker options
depending on what it is told. Here's how R's zdotu test always fails
when linking directly with MKL:

# pre-configure some variables
echo '#define HAVE_F77_UNDERSCORE 1' > confdefs.h
FC=gfortran
FFLAGS='-g -Og'
CC=gcc
CFLAGS='-g -Og'
CPPFLAGS=-I/usr/local/include
MAIN_LDFLAGS='-Wl,--export-dynamic -fopenmp'
LDFLAGS='-L/usr/local/lib'
LIBM=-lm
FLIBS=' -lgfortran -lm -lquadmath'
# copied & pasted from the Intel web page
BLAS_LIBS='-lmkl_rt -Wl,--no-as-needed -lpthread -lm -ldl'

# R prepares to call zdotu from Fortran...
cat > conftestf.f <<EOF
c Goto's BLAS at least needs a XERBLA
      subroutine xerbla(srname, info)
      character*6 srname
      integer info
      end

      subroutine test1(iflag)
      double complex zx(2), ztemp, zres, zdotu
      integer iflag
      zx(1) = (3.1d0,1.7d0)
      zx(2) = (1.6d0,-0.6d0)
      zres = zdotu(2, zx, 1, zx, 1)
      ztemp = (0.0d0,0.0d0)
      do 10 i = 1,2
         ztemp = ztemp + zx(i)*zx(i)
 10      continue
      if(abs(zres - ztemp) > 1.0d-10) then
        iflag = 1
      else
        iflag = 0
      endif
      end
EOF
${FC} ${FFLAGS} -c conftestf.f

# and then call the Fortran subroutine from the C runner...
cat > conftest.c <<EOF
#include <stdlib.h>
#include "confdefs.h"
#ifdef HAVE_F77_UNDERSCORE
# define F77_SYMBOL(x)   x ## _
#else
# define F77_SYMBOL(x)   x
#endif
extern void F77_SYMBOL(test1)(int *iflag);

int main () {
  int iflag;
  F77_SYMBOL(test1)(&iflag);
  exit(iflag);
}
EOF
${CC} ${CPPFLAGS} ${CFLAGS} -c conftest.c

# and then finally link and execute the program
${CC} ${CPPFLAGS} ${CFLAGS} ${LDFLAGS} ${MAIN_LDFLAGS} \
 -o conftest conftest.o conftestf.o \
 ${BLAS_LIBS} ${FLIBS} ${LIBM}
./conftest

It seems to crash inside MKL!

rax=ffffffffcccccccd rbx=00005590ee102008 rcx=00007ffdab2ddb20 rdx=00005590ee102008 
rsi=00007ffdab2ddb18 rdi=00005590ee10200c rbp=00007ffdab2dd910 rsp=00007ffdab2db600 
 r8=00005590ee102008  r9=00007ffdab2ddb28 r10=00007f4086a99178 r11=00007f4086e02490 
r12=00005590ee10200c r13=00007ffdab2ddb20 r14=00005590ee102008 r15=00007ffdab2ddb28 
ip = 7f4086e02a60, sp = 7ffdab2db600 [mkl_blas_zdotu+1488]
ip = 7f4085dc5250, sp = 7ffdab2dd920 [zdotu+256]
ip = 5590ee1011cc, sp = 7ffdab2ddb40 [test1_+91]
ip = 5590ee101167, sp = 7ffdab2ddb70 [main+14]

It's especially strange that R does seem to work if you just
update-alternatives after linking it with the reference BLAS, but
./conftest starts crashing again in the same place. This is with
Debian's MKL version 2020.4.304-4, by the way.

-- 
Best regards,
Ivan