Accelerate framework bug with fork

bug.c

/*
 * This small snippet shows an issue when calling dgemm for square matrices above a certain size inside a forked child process.
 *
 * To reproduce the issue:
 *  - compile it ("gcc -framework Accelerate bug.c -o bug")
 *  - ./bug 64 (64x64 matrix) will show the output for d matrix for both parents and child:
 
       computing for size 64
       c[0][0] is 64.000000
       parent d[0][0] is 64.000000
       child d[0][0] is 64.000000

 *  - ./bug 65 (65x65 matrix) will not show the output for child process
 
       computing for size 65
       c[0][0] is 65.000000
       parent d[0][0] is 65.000000

 * sgemm does not have the issue, and not calling cblas_dgemm before fork does not either. 
 * We also attached the valgrind ouptut, which suggests an issue with Grand Central.
 */

#include <string.h>
#include <stdio.h>
#include <unistd.h>

#include <Accelerate/Accelerate.h>

void* xmalloc(size_t n)
{
	void* tmp;

	tmp = malloc(n);
	if (tmp == NULL) {
		fprintf(stderr, "You are about to die\n");
		exit(EXIT_FAILURE);
	} else {
		return tmp;
	}
}

int main(int argc, char *argv[])
{
	int n, lda, ldb, ldc;
	int i;

	double *a, *b, *c, *d;
	size_t n_bytes;

	pid_t p;

	if (argc != 2) {
		fprintf(stderr, "USAGE: %s size\n", argv[0]);
		exit(EXIT_FAILURE);
	} else {
		n = atoi(argv[1]);
		printf("computing for size %d\n", n);
	}

	lda = n;
	ldb = n;
	ldc = n;

	n_bytes = sizeof(*a) * n * n;

	a = xmalloc(n_bytes);
	b = xmalloc(n_bytes);
	c = xmalloc(n_bytes);
	d = xmalloc(n_bytes);

	for(i = 0; i < n * n; ++i) {
		a[i] = 1;
		b[i] = 1;
	}

	/* If this call is disabled, the crash disappears */
#if 1
	cblas_dgemm(CblasRowMajor,
		CblasNoTrans, CblasNoTrans, n, n, n,
		1.0, a, lda, b, ldb, 0.0, c, ldc);
	printf("c[0][0] is %f\n", c[0]);
#endif

	p = fork();
	if (p == 0) {
		cblas_dgemm(CblasRowMajor,
				CblasNoTrans, CblasNoTrans, n, n, n,
				1.0, a, lda, b, ldb, 0.0, d, ldc);
		printf("child d[0][0] is %f\n", d[0]);
	} else {
		cblas_dgemm(CblasRowMajor,
			CblasNoTrans, CblasNoTrans, n, n, n,
			1.0, a, lda, b, ldb, 0.0, d, ldc);
		printf("parent d[0][0] is %f\n", d[0]);
	}

	return 0;
}

gistfile1.txt

==39502== Memcheck, a memory error detector
==39502== Copyright (C) 2002-2011, and GNU GPL'd, by Julian Seward et al.
==39502== Using Valgrind-3.7.0 and LibVEX; rerun with -h for copyright info
==39502== Command: ./a.out 65
==39502== 
--39502-- ./a.out:
--39502-- dSYM directory is missing; consider using --dsymutil=yes
computing for size 65
c[0][0] is 65.000000
UNKNOWN task message [id 3403, to mach_task_self(), reply 0x2e03]
parent d[0][0] is 65.000000
==39502== 
==39502== HEAP SUMMARY:
==39502==     in use at exit: 143,723 bytes in 43 blocks
==39502==   total heap usage: 65 allocs, 22 frees, 1,459,867 bytes allocated
==39502== 
==39503== 
==39503== Process terminating with default action of signal 10 (SIGBUS)
==39503==  Non-existent physical address at address 0x108
==39503==    at 0x373324: dispatch_group_async_f (in /usr/lib/system/libdispatch.dylib)
==39503==    by 0xFD565F: dgemmGCD (in /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib)
==39503==    by 0xB62A47: cblas_dgemm (in /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib)
==39503==    by 0x100000CF9: main (in ./a.out)
==39503== 
==39503== HEAP SUMMARY:
==39503==     in use at exit: 146,067 bytes in 47 blocks
==39503==   total heap usage: 58 allocs, 11 frees, 804,139 bytes allocated
==39503== 
==39502== LEAK SUMMARY:
==39502==    definitely lost: 101,440 bytes in 5 blocks
==39502==    indirectly lost: 0 bytes in 0 blocks
==39502==      possibly lost: 33,800 bytes in 1 blocks
==39502==    still reachable: 8,483 bytes in 37 blocks
==39502==         suppressed: 0 bytes in 0 blocks
==39502== Rerun with --leak-check=full to see details of leaked memory
==39502== 
==39502== For counts of detected and suppressed errors, rerun with: -v
==39502== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 1 from 1)
==39503== LEAK SUMMARY:
==39503==    definitely lost: 40 bytes in 2 blocks
==39503==    indirectly lost: 0 bytes in 0 blocks
==39503==      possibly lost: 0 bytes in 0 blocks
==39503==    still reachable: 146,027 bytes in 45 blocks
==39503==         suppressed: 0 bytes in 0 blocks
==39503== Rerun with --leak-check=full to see details of leaked memory
==39503== 
==39503== For counts of detected and suppressed errors, rerun with: -v
==39503== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 1 from 1)