Open Radar

 
Summary:
The attached file demonstrates the problem.

Calling cblas_sgemv for matrix data that is not aligned at 8 byte boundaries causes a segfault.   This may be due to unprotected use of AVX instructions because running the code on a machine that does not have AVX does not produce the segfault.

See: https://github.com/numpy/numpy/issues/4007 for discussion

Steps to Reproduce:
Compile and run attached c code:

    cc accelerate_aligned.c -lblas
    ./a.outOutput on my 6,2 MBA:

    0 16 0
    Segmentation fault: 11

Expected Results:
No segmentation fault. 

Actual Results:
Output on my 6,2 MBA:

    0 16 0
    Segmentation fault: 11

Version:
XCode version : 5.1.0.0.1.1396320587
OSX version : 10.9.3 (13D65)

Notes:


Configuration:
Does not segfault on two OSX machines without AVX.  However, these are running 10.6 and 10.7.


accelerate_aligned.c

// cc file.c -lblas
#include <stdlib.h>
#include <stdio.h>
#include "Accelerate/Accelerate.h"

int main(int argc, const char *argv[])
{
    long m = 10000;
    long n = 200;
    float * A = calloc(m*n + 4, sizeof(float));
    float * X = calloc(m*n + 4, sizeof(float));
    float * R = calloc(m*n + 4, sizeof(float));
    X +=4;
    printf("%ld %ld %ld", (long)A % 32, (long)X % 32, (long)R % 32);
    puts("");
    cblas_sgemv(CblasRowMajor, CblasNoTrans, m, n, 1.f, A, n, X, 1,
                 0.f, R, 1);
    return 0;
}