Re: [Discuss-gnuradio] Segmentation fault in volk_32f_s32f_multiply_32f_a_sse

[Tom Rondeau]

On Sat, Apr 7, 2012 at 1:08 PM, Alexandru Csete <address@hidden> wrote:
> On Sat, Apr 7, 2012 at 7:07 PM, Alexandru Csete <address@hidden> wrote:
>> On Sat, Apr 7, 2012 at 5:35 PM, Tom Rondeau <address@hidden> wrote:
>>> On Sat, Apr 7, 2012 at 10:19 AM, Alexandru Csete <address@hidden> wrote:
>>>> On Sat, Apr 7, 2012 at 4:13 PM, Tom Rondeau <address@hidden> wrote:
>>>>> On Sat, Apr 7, 2012 at 9:48 AM, Alexandru Csete <address@hidden> wrote:
>>>>>> Greetings,
>>>>>>
>>>>>> I'm getting segmentation faults in a C++ application when I restart
>>>>>> the flow graph, i.e by doing:
>>>>>>
>>>>>> tb->start()
>>>>>>  ...
>>>>>> tb->stop()
>>>>>> tb->wait()
>>>>>> ...
>>>>>> tb->start()
>>>>>> (crashes)
>>>>>>
>>>>>> Running the application through a debugger suggests it is
>>>>>> volk_32f_s32f_multiply_32f_a_sse which is used by a gr_multiply_ff()
>>>>>> block. I end up with the attached assembly dump with a marker
>>>>>> highlighting line 19.
>>>>>>
>>>>>> I have attached a simple python flow graph, which also crashes for me.
>>>>>> Note that the python code was originally generated by GRC and it used
>>>>>> gr.multiply_const_vff which didn't crash. I manually changed it to
>>>>>> gr.multiply_const_ff to provoke the crash.
>>>>>>
>>>>>> This is happening on Ubuntu 11.04 64 bit and intel i7 processor. I'm
>>>>>> also seeing the message "Using Volk machine: sse4_2_64" when I start
>>>>>> the graph. GNU Radio is built using cmake.
>>>>>>
>>>>>> Alex
>>>>>
>>>>> What version of GNU Radio are you using? This sounds like a bug that
>>>>> was fixed a little while ago and is not a release version yet (will be
>>>>> in 3.5.3).
>>>>>
>>>>
>>>> Hi Tom,
>>>>
>>>> I pulled from origin/master earlier today. Git describe says
>>>> v3.5.2.1-103-gfaf1d33
>>>> Is there another branch I can try?
>>>>
>>>> Alex
>>>
>>>
>>> Alex,
>>>
>>> Can you try applying the attached patch? It should work, but I've only
>>> tested it on one of my machines. This makes sure all buffers are
>>> aligned and all blocks know if after a start and restart.
>>>
>>> Thanks,
>>> Tom
>>
>> The patch fixed my problem. Thanks for the q
>
> It was supposed to be "thanks for the quick fix" :)
>
> Alex

And thanks for being able to check it quickly!

The fix has been merged into the code. Just in time for 3.5.3, too (whew).

Tom