Re: Ubuntu in vmd and avx

2017-05-02 Thread Karl Pettersson
On Mon, Apr 24, 2017 at 11:56:25AM -0700, Mike Larkin wrote:
> On Mon, Apr 24, 2017 at 11:50:07AM -0700, Mike Larkin wrote:
> > On Mon, Apr 24, 2017 at 08:18:17PM +0200, Karl Pettersson wrote:
> > > I run 64-bit Ubuntu as a vmm guest, according to:
> > > https://gist.github.com/reyk/6d369c5c0bd0c76f4906f83933f3bb71
> > > 
> > > It works well for the most part; I have removed cloud-init becaue I do
> > > not need it. However, I used Python 2.7 with numpy 1.12 (installed via 
> > > Pip),
> > > and experienced crashes which seemed to be related to AVX2
> > > instructions, even though the cpu should support this, according to
> > > /proc/cpuinfo. (cf. https://github.com/numpy/numpy/issues/8128). For
> > > example, when I tried to reproduce this simple chart,
> > > https://matplotlib.org/examples/lines_bars_and_markers/fill_demo.html,
> > > this made Python exit with illegal instruction, according to gdb:
> > > 
> > > Program received signal SIGILL, Illegal instruction.  0x73ce5d56
> > > in UBYTE_less_equal_avx2 (args=0x7fffba50, dimensions= > > out>, steps=, __NPY_UNUSED_TAGGEDfunc=) at
> > > numpy/core/src/umath/loops.c.src:899
> > > 
> > > Downgrading to numpy 1.11.1~rc1 (the version available from as an Ubuntu
> > > package) seems to have fixed the issue.
> > >
> > 
> > we filter out avx2 from the guest, for now.  
> > 
> 
> actually after looking at this it appears we don't, and we should. the filter
> mask is only for avx (and not the later ones). I'll add suppression of avx2,
> that may force your software to downgrade to something we do allow. I'll 
> commit
> that tonight.
>

Was this commit ever made? I only find references to AVX in the CVS log
for vmm.c.



Re: Ubuntu in vmd and avx

2017-05-02 Thread Mike Larkin
On Tue, May 02, 2017 at 10:31:41PM +0200, Karl Pettersson wrote:
> On Mon, Apr 24, 2017 at 11:56:25AM -0700, Mike Larkin wrote:
> > On Mon, Apr 24, 2017 at 11:50:07AM -0700, Mike Larkin wrote:
> > > On Mon, Apr 24, 2017 at 08:18:17PM +0200, Karl Pettersson wrote:
> > > > I run 64-bit Ubuntu as a vmm guest, according to:
> > > > https://gist.github.com/reyk/6d369c5c0bd0c76f4906f83933f3bb71
> > > > 
> > > > It works well for the most part; I have removed cloud-init becaue I do
> > > > not need it. However, I used Python 2.7 with numpy 1.12 (installed via 
> > > > Pip),
> > > > and experienced crashes which seemed to be related to AVX2
> > > > instructions, even though the cpu should support this, according to
> > > > /proc/cpuinfo. (cf. https://github.com/numpy/numpy/issues/8128). For
> > > > example, when I tried to reproduce this simple chart,
> > > > https://matplotlib.org/examples/lines_bars_and_markers/fill_demo.html,
> > > > this made Python exit with illegal instruction, according to gdb:
> > > > 
> > > > Program received signal SIGILL, Illegal instruction.  0x73ce5d56
> > > > in UBYTE_less_equal_avx2 (args=0x7fffba50, dimensions= > > > out>, steps=, __NPY_UNUSED_TAGGEDfunc=) at
> > > > numpy/core/src/umath/loops.c.src:899
> > > > 
> > > > Downgrading to numpy 1.11.1~rc1 (the version available from as an Ubuntu
> > > > package) seems to have fixed the issue.
> > > >
> > > 
> > > we filter out avx2 from the guest, for now.  
> > > 
> > 
> > actually after looking at this it appears we don't, and we should. the 
> > filter
> > mask is only for avx (and not the later ones). I'll add suppression of avx2,
> > that may force your software to downgrade to something we do allow. I'll 
> > commit
> > that tonight.
> >
> 
> Was this commit ever made? I only find references to AVX in the CVS log
> for vmm.c.
> 

Thanks for reminding me, I must have fatfingered 'd'elete on that email.

I'll get to this tonight.

-ml



Re: Ubuntu in vmd and avx

2017-04-24 Thread Mike Larkin
On Mon, Apr 24, 2017 at 08:18:17PM +0200, Karl Pettersson wrote:
> I run 64-bit Ubuntu as a vmm guest, according to:
> https://gist.github.com/reyk/6d369c5c0bd0c76f4906f83933f3bb71
> 
> It works well for the most part; I have removed cloud-init becaue I do
> not need it. However, I used Python 2.7 with numpy 1.12 (installed via Pip),
> and experienced crashes which seemed to be related to AVX2
> instructions, even though the cpu should support this, according to
> /proc/cpuinfo. (cf. https://github.com/numpy/numpy/issues/8128). For
> example, when I tried to reproduce this simple chart,
> https://matplotlib.org/examples/lines_bars_and_markers/fill_demo.html,
> this made Python exit with illegal instruction, according to gdb:
> 
> Program received signal SIGILL, Illegal instruction.  0x73ce5d56
> in UBYTE_less_equal_avx2 (args=0x7fffba50, dimensions= out>, steps=, __NPY_UNUSED_TAGGEDfunc=) at
> numpy/core/src/umath/loops.c.src:899
> 
> Downgrading to numpy 1.11.1~rc1 (the version available from as an Ubuntu
> package) seems to have fixed the issue.
>

we filter out avx2 from the guest, for now.  



Re: Ubuntu in vmd and avx

2017-04-24 Thread Mike Larkin
On Mon, Apr 24, 2017 at 11:50:07AM -0700, Mike Larkin wrote:
> On Mon, Apr 24, 2017 at 08:18:17PM +0200, Karl Pettersson wrote:
> > I run 64-bit Ubuntu as a vmm guest, according to:
> > https://gist.github.com/reyk/6d369c5c0bd0c76f4906f83933f3bb71
> > 
> > It works well for the most part; I have removed cloud-init becaue I do
> > not need it. However, I used Python 2.7 with numpy 1.12 (installed via Pip),
> > and experienced crashes which seemed to be related to AVX2
> > instructions, even though the cpu should support this, according to
> > /proc/cpuinfo. (cf. https://github.com/numpy/numpy/issues/8128). For
> > example, when I tried to reproduce this simple chart,
> > https://matplotlib.org/examples/lines_bars_and_markers/fill_demo.html,
> > this made Python exit with illegal instruction, according to gdb:
> > 
> > Program received signal SIGILL, Illegal instruction.  0x73ce5d56
> > in UBYTE_less_equal_avx2 (args=0x7fffba50, dimensions= > out>, steps=, __NPY_UNUSED_TAGGEDfunc=) at
> > numpy/core/src/umath/loops.c.src:899
> > 
> > Downgrading to numpy 1.11.1~rc1 (the version available from as an Ubuntu
> > package) seems to have fixed the issue.
> >
> 
> we filter out avx2 from the guest, for now.  
> 

actually after looking at this it appears we don't, and we should. the filter
mask is only for avx (and not the later ones). I'll add suppression of avx2,
that may force your software to downgrade to something we do allow. I'll commit
that tonight.