Closed
Description
Hi. I have a machine with 80 cores, and a trivial invocation of DGESV goes into an infinite loop with OpenBLAS. There's something about the core count that's making it break; maybe some internal buffer is being overrun or something. This is on Fedora 27, but I suspect it doesn't really matter. Recipe:
- Check out OpenBLAS from git. I see the failure in the master, develop and release-0.2.21 branches.
- Build thusly:
make USE_THREAD=1 USE_OPENMP=0 INTERFACE64=0 BINARY=64 - Save this into /tmp/inv.c: (github is refusing to accept this as an upload)
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
int dgesv_(int *n, int *nrhs, double *a,
int *lda, int *ipiv, double *b, int *ldb, int *info);
int main(int argc, char* argv[])
{
int N = 25;
double* m1 = calloc(N,N*sizeof(double));
double* m2 = calloc(N,N*sizeof(double));
assert(m1);
assert(m2);
for(int i=0; i<N; i++)
m1[i*(N+1)] = m2[i*(N+1)] = 1.0;
int ipiv[N];
int info;
dgesv_( &N, &N, m1, &N, ipiv, m2, &N, &info);
printf("info: %d\n", info);
return 0;
}- Build and run this source against the openblas we just built. I do this in zsh:
gcc -o /tmp/inv /tmp/inv.c libopenblas*.so(.) -Wl,-rpath=$PWD && /tmp/inv
On this 80-core machine this command goes into an infinite loop, and never exits. Apparently setting OPENBLAS_NUM_THREADS=N for N <= 65 makes the issue go away. I tried to find the issue myself, but it wasn't trivial, and hopefully the devs can do this much faster. This failure is very consistent here, so let me know if you need more debugging data. Thanks!
Metadata
Metadata
Assignees
Labels
No labels