nvlink error when defining acc routine information

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hello,

I had posted this before about calls to libraries in a loop:
www.pgroup.com/userforum/viewtopic.php?p=19304

I was advised to try inlining or to try defining routine information. I went with the latter as it generally seemed to work, but then I ran into a problem. The two methods I had defined are mentioned as an undefined reference with nvlink.

I did the following to define the routines:

#pragma acc routine seq
extern __GMP_DECLSPEC void mpz_init_set_si (mpz_ptr, signed long int);

#pragma acc routine seq
extern void mpz_clear (mpz_ptr);

and then when I run:

pgc++ -fast -ta=tesla,cc30 -Minfo=all -Minline=levels:3 -lgmpxx -lgmp main.cpp

I get the following output:

main:
     17, Generating copy(vals[:])
     19, Generating copy(t)
     21, Loop is parallelizable
         Accelerator kernel generated
         Generating Tesla code
         21, #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */
     25, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::__gmp_expr(int) inlined, size=2 (inline) file main.cpp (1423)
        1423, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_si(long) inlined, size=7, file main.cpp (1389)
             1391, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4 (inline) file main.cpp (1382)
             1394, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4 (inline) file main.cpp (1382)
             1395, __gmpz_neg inlined, size=4 (inline) file main.cpp (1801)
     26, Loop is parallelizable
     30, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::~__gmp_expr() inlined, size=2 (inline) file main.cpp (1444)
__gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_si(long):
   1391, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4 (inline) file main.cpp (1382)
   1394, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4 (inline) file main.cpp (1382)
   1395, __gmpz_neg inlined, size=4 (inline) file main.cpp (1801)
__gmp_expr<__mpz_struct [1], __mpz_struct [1]>::__gmp_expr(int):
   1423, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_si(long) inlined, size=7 (inline) file main.cpp (1389)
        1391, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4, file main.cpp (1382)
        1394, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4, file main.cpp (1382)
        1395, __gmpz_neg inlined, size=4, file main.cpp (1801)
__gmp_expr<__mpz_struct [1], __mpz_struct [1]>::__gmp_expr(int) [subobject]:
      0, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::__gmp_expr(int) inlined, size=2 (inline) file main.cpp (1423)
        1423, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_si(long) inlined, size=7, file main.cpp (1389)
             1391, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4 (inline) file main.cpp (1382)
             1394, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::init_ui(unsigned long) inlined, size=4 (inline) file main.cpp (1382)
             1395, __gmpz_neg inlined, size=4 (inline) file main.cpp (1801)
__gmp_expr<__mpz_struct [1], __mpz_struct [1]>::~__gmp_expr() [subobject]:
      0, __gmp_expr<__mpz_struct [1], __mpz_struct [1]>::~__gmp_expr() inlined, size=2 (inline) file main.cpp (1444)
nvlink error   : Undefined reference to '__gmpz_init_set_si' in '/tmp/pgc++MZZfYrbrP8To.o'
nvlink error   : Undefined reference to '__gmpz_clear' in '/tmp/pgc++MZZfYrbrP8To.o'
pgacclnk: child process exit status 2: /opt/pgi/linux86-64/15.7/bin/pgnvd

As you can see, there is some sort of nvlink error. I can’t quite understand how to go about resolving it. Any ideas?

It should be noted that __gmpz_init_set_si is defined as mpz_init_set_si in the gmp header file. Use of either one as the routine definition allows the compiler to get up to this point. If I don’t define a routine for either, the compiler says the loop is not parallelizable due to undefined routine information.

Thanks,

2

Hi jrobb61773,

In addition to the prototype, “acc routine” needs to be added to the definition as well.

Do you have access to the GMPxx source?

  • Mat
3

I do! However, I can’t get the GMP to compile with pgc++. Other suggestions?

4

I do! However, I can’t get the GMP to compile with pgc++. Other suggestions?

That’s a problem. You either need to have a device callable version of these routines available (i.e. add the “routine” directive) or inline them. For inlining you’d need to use pgc++ as well.

What’s the issue with pgc++ compiling GMP?

  • Mat