Search code examples
cmultithreadingvalgrindlapackatlas

Are valgrind "uninitialized value" warnings false positives in ATLAS multithreaded BLAS routines?


I am using ATLAS for LAPACK and multithreaded BLAS routines, and have noticed that when my matrices get large enough for ATLAS to use the multithreaded versions of BLAS, I get initialization errors from Valgrind. Here is a minimal example from my code:

#include <stdio.h>
#include <stdlib.h>

extern void dgetrf_(int *, int *, double *, int *, int *, int *);
extern void dgetri_(int *, double *, int *, int *, double *, int *, int *);
extern void dgemm_(char *, char *, int *, int *, int *, double *, double *, int *, double *, int *, double *, double *, int *);

int main(void)
{
    double *m1,*m2,*work,*temp;
    int dim = 576;
    int i,j,info;
    int lwork = dim * dim;
    int *ipiv;
    char transA = 'N';
    char transB = 'N';
    double alpha = 1.0;
    double beta = 0.0;

    m1 = malloc(dim*dim*sizeof(double));
    m2 = malloc(dim*dim*sizeof(double));
    temp = malloc(dim*dim*sizeof(double));
    ipiv = malloc(dim*sizeof(int));
    work = malloc(lwork*sizeof(double));

    for(i=0; i<dim; i++)
     {
       for(j=0; j<dim; j++)
        {
          if(i==j)
           {
             m1[i+dim*j] = .25;
             m2[i+dim*j] = .5;
           }
          else
           {
             m1[i+dim*j] = 0.0;
             m2[i+dim*j] = 0.0;
           }
        }
    }

    dgetrf_(&dim, &dim, m1, &dim, ipiv, &info);
    dgetri_(&dim, m1, &dim, ipiv, work, &lwork, &info);

    dgemm_(&transA, &transB, &dim, &dim, &dim, &alpha, m1, &dim, m2, &dim, &beta, temp, &dim);
    for(i=0; i<dim*dim; i++)
        m1[i] = temp[i];

    dgetrf_(&dim, &dim, m1, &dim, ipiv, &info);
    dgetri_(&dim, m1, &dim, ipiv, work, &lwork, &info);

    free(m1);
    free(m2);
    free(ipiv);
    free(work);
    free(temp);

    return 0;
}

(Note: I've checked to make sure the matrices aren't singular and they aren't.)

I compile the program:

gcc -Wall -DATLAS -m64 -g -c fermi.c
gcc -o fermi fermi.o -L/usr/lib64/atlas/ -lm -ltatlas

And run valgrind:

valgrind --leak-check=yes ./fermi

When I do this I get 193 errors from 11 contexts of "Conditional jump or move depends on uninitialised value(s)" when the second instances of dgetrf_ and dgetri_ are encountered.

==24999== Memcheck, a memory error detector
==24999== Copyright (C) 2002-2015, and GNU GPL'd, by Julian Seward et al.
==24999== Using Valgrind-3.12.0 and LibVEX; rerun with -h for copyright info
==24999== Command: ./fermi
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x524C62B: ??? (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C29E3: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x524C66A: ??? (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C29E3: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x524C6BE: ??? (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C29E3: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x51C2A0B: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x51C2A0D: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x51C2A4E: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x51C2A61: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x524C2D7: ATL_daxpy (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x53426BB: ATL_dgerk_axpy (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C2AC7: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x524C751: ??? (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C29E3: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51CD2BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x5210416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400A97: main (fermi.c:52)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x51CD8E5: ATL_dtrtri (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C2EC3: ATL_dgetriC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520EFA5: atl_f77wrap_dgetri_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F684: dgetri_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400AC0: main (fermi.c:53)
==24999== 
==24999== Conditional jump or move depends on uninitialised value(s)
==24999==    at 0x51CD8E7: ATL_dtrtri (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x51C2EC3: ATL_dgetriC (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520EFA5: atl_f77wrap_dgetri_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x520F684: dgetri_ (in /usr/lib64/atlas/libtatlas.so.3.10)
==24999==    by 0x400AC0: main (fermi.c:53)
==24999== 
==24999== 
==24999== HEAP SUMMARY:
==24999==     in use at exit: 0 bytes in 0 blocks
==24999==   total heap usage: 2,024 allocs, 2,024 frees, 54,831,424 bytes allocated
==24999== 
==24999== All heap blocks were freed -- no leaks are possible
==24999== 
==24999== For counts of detected and suppressed errors, rerun with: -v
==24999== Use --track-origins=yes to see where uninitialised values come from
==24999== ERROR SUMMARY: 193 errors from 11 contexts (suppressed: 0 from 0)

I have found some links that suggest this could be a false positive coming from the way the library is doing things, though they aren't related very much to my context.

memory leak in dgemm_

https://www.open-mpi.org/community/lists/users/2007/05/3192.php

So my question: is valgrind giving me false positive errors?


Solution

  • is valgrind giving me false positive errors?

    Looks like no.

    Instead of running valgrind with --leak-check=yes you should have run it with --track-origins=yes to see where uninitialised values come from as suggested by valgrind at the end of the output. Here is what I've got with --track-origins=yes:

    [ ~]$ valgrind --track-origins=yes ./a.out 
    ==17533== Memcheck, a memory error detector
    ==17533== Copyright (C) 2002-2015, and GNU GPL'd, by Julian Seward et al.
    ==17533== Using Valgrind-3.12.0 and LibVEX; rerun with -h for copyright info
    ==17533== Command: ./a.out
    ==17533== 
    ==17533== Conditional jump or move depends on uninitialised value(s)
    ==17533==    at 0x4F4362B: ??? (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4EB99E3: ATL_dgetf2 (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4EC42BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4EC42BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4EC42BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4EC42BF: ATL_dtgetrfC (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4F06538: atl_f77wrap_dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x4F07416: dgetrf_ (in /usr/lib64/atlas/libtatlas.so.3.10)
    ==17533==    by 0x400A29: main (fermi.c:50)
    ==17533==  Uninitialised value was created by a heap allocation
    ==17533==    at 0x4C2DB9D: malloc (vg_replace_malloc.c:299)
    ==17533==    by 0x40080B: main (fermi.c:22)
    

    So the source of uninitialised values is this line of code:

    temp = malloc(dim*dim*sizeof(double));
    

    It is then used to initialise m1 which is passed to dgetrf_() on line 50.

    I'm not familiar with ATLAS library but I guess you should somehow initialise temp variable. For example zero initialising temp with calloc resolves all these valgrind errors:

    temp = calloc(dim*dim,sizeof(double));