Re: [Starpu-devel] starpu_data_unregister() failing at pthread_spin_unlock()

[Jonathan Adamczewski <jadamcze@utas.edu.au>]

I managed to reduce it and think I've worked out what the trigger is: the 
length of the .cl file I'm using (or - more likely - the amount of text being 
output by starpu,)

I'm using starpu built with --enable-verbose  and --enable-debug, which 
generates a lot of text that scrolls very slowly in the windows command 
prompt 
(not sure why - it's quite tedious). Additionally, I've added _STARPU_DEBUG() 
calls wherever cl functions are called in the library.

If I use a short .cl file, _starpu_opencl_load_progam_source() has less to 
print and the problem goes away - both variables are correctly unregistered.

I would guess that there is a multithreading timing issue being exposed here 
- 
or that my liberal addition of extra debugging is somehow invalid.

If I redirect the output of my program to a file, I see a log of 
clGetEventInfo() in _starpu_driver_test_request_completion() being called 
thousands of times before exiting with the same error previously reported in 
starpu_spinlock().

Is my use of extra debugging invalid, or is there a bug? Or something else? :)

Thanks

Jonathan.


On Saturday 16 April 2011 17:44, Jonathan Adamczewski wrote:
> I have been unable to extract a simple test case for this - I'll explain as
> best as I can. I'm using svn trunk and using OpenCL only under WinXP.
> 
> I have a function that registers an array of variables through a loop like
> :
> 
>     starpu_data_handle *data_handles = [allocated chunk of memory];
>     uintptr_t varsint = (uintptr_t)[address of vars];
>     for(unsigned int i = 0; i < argnumber; ++i) { // argnumber = 2
>         starpu_variable_data_register(&data_handles[i], 0, varsint,
> var_size); vars += var_size; // var_size = 4
>     }
> 
> These are later added to a task struct:
> 
>     for(unsigned int i = 0; i < argnumber; ++i) { // argnumber = 2
>         task->buffers[i].handle = data_handles[i];
>         task->buffers[i].mode = STARPU_RW;
>     }
> 
> The task is submitted, and the args are set for the cl kernel in the set
> OpenCL function:
> 
>     for(unsigned int i = 0; i < argnumber; ++i) { // argnumber = 2
>         clSetKernelArg(kernel, i, STARPU_VARIABLE_GET_ELEMSIZE(buffer[i]),
>                               &STARPU_VARIABLE_GET_PTR(buffer[i]));
>     }
> 
> The kernel runs and I attempt to unregister the variables to have their
> changed values copied back to the local buffer:
> 
>     for(unsigned int i = 0; i < argnumber; ++i) { // argnumber = 2
>         starpu_data_unregister(data_handles[i]);
>     }
> 
> The first variable is unregistered without problem - I see the correct
> result in memory.
> 
> Unregistering the second variable causes an assert to fail in
> _starpu_spin_unlock() - the assert checking the return value of
> pthread_spin_unlock() fails. According to gdb, lock->lock contains the
> value 0xfeeefeee, which looks rather suspicious.
> 
> Suggestions of things to consider or check for would be appreciated,
> thanks.
> 
> Jonathan.
> 
> _______________________________________________
> Starpu-devel mailing list
> Starpu-devel@lists.gforge.inria.fr
> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel