Re: [Mjpeg-users] What is causing this error

[Steven M. S. <sm...@2B...>]

On Wed, 14 Jan 2004, Ray Cole wrote:

> I had this issue with the CVS version as well.  I had to make a change to utils/cpu_accel.c to use memalign rather than posix_memalign in bufalloc().  It now looks like this:

	Have you done a 'cvs update' recently?   There was code added to
	check the alignment of the buffer that posix_memalign() returned
	and if the alignment was not correct then to try allocating the
	buffer with memalign().

	Look in cpu_accel.c - you should see something like this:

        if (posix_memalign( &buf, simd_alignment, size))
                buf = memalign(pgsize, size);
        if (buf && ((int)buf & (simd_alignment - 1)))
        {
                free(buf);
                buf = memalign(pgsize, size);
        }

	If you don't see that then it's time for 'cvs update'.  If you do
	see that and it's not working it's a bug and needs to be fixed.

> I would assume this is caused by some glibc bug.  I'm using Red Hat 8.0.  I don't think I've ever upgraded my glib version.  I appear to have versions 1.2.10-8 and 2.3.2-4.80.

	Yes, it is a glibc bug.  glibc-2.2.5 has posix_memalign() but it
	does not honor the alignment request and can return 8 instead of 16
	or 64 byte aligned buffers.   glibc-2.2.4 is even worse - posix_memalign
	exists but always returns an error.

	The bug's never been reported to happen though (that I know of) with
	glibc-2.3.x (I've never seen it happen on Suse 8.2 or 9.0).

	If you still get the message "could not allocate %d bytes aligned ..."
	then neither posix_memalign nor memalign is working correctly and
	poor ol' mpeg2enc has no choice but to bail out.

	Cheers,
	Steven Schultz