Re: [Starpu-devel] Example spmv

[Cédric Augonnet <cedric.augonnet@inria.fr>]

Nathalie Furmento wrote:
> The example spmv from starpu/trunk segfaults.
>
> Here what i did (on hannibal):
>
> $ ../trunk/configure --enable-blas-lib=atlas --enable-debug --enable-verbose
> $ make
> $ libtool --mode=execute gdb ./examples/spmv/dw_spmv 
> ...
> (gdb) bt
> #0  0x0000000000401e07 in cpu_spmv (descr=0x7fffe98438f8, arg=0x0) at 
> ../../trunk/examples/spmv/dw_spmv.c:120
> #1  0x00007ffff7bcdeb7 in execute_job_on_cpu (j=0x7fffe98439a0, 
> cpu_args=0x7ffff7ddb5a0) at ../../trunk/src/drivers/cpu/driver_cpu.c:60
> #2  0x00007ffff7bce291 in _starpu_cpu_worker (arg=0x7ffff7ddb5a0) at 
> ../../trunk/src/drivers/cpu/driver_cpu.c:180
> #3  0x00007ffff219473a in start_thread (arg=<value optimized out>) at 
> pthread_create.c:300
> #4  0x00007ffff1f0569d in clone () at 
> ../sysdeps/unix/sysv/linux/x86_64/clone.S:112
> #5  0x0000000000000000 in ?? ()
> (gdb) list
> 115             for (row = 0; row < nrow; row++)
> 116             {
> 117                     float tmp = 0.0f;
> 118                     unsigned index;
> 119     
> 120                     unsigned firstindex = rowptr[row] - firstelem;
> 121                     unsigned lastindex = rowptr[row+1] - firstelem;
> 122     
> 123                     for (index = firstindex; index < lastindex; index++)
> 124                     {
> (gdb) p rowptr
> $1 = (uint32_t *) 0x0
> (gdb) 
>
> By looking at the source code of the application and StarPU, i could not 
> figure out how rowptr could end up
> being NULL. I first thought it was a problem with the csr filter, i looked at 
> it with Samuel, and it seems to be
> correct.
>
> The application needs the --enable-blas-lib=atlas to be compiled, and so is not tested 
> with a regular "make check".
> We should probably change that.
>
> Thanks,
>
> Nathalie
>   
Hi Nathalie,

Thanks for the report, there was actually a bug in the filter 
implementation which should be fixed by commit #1950.

I agree that we should find a way to have a more intensive test suite so 
that we detect such issues ... Similarly there are some problems which 
only appear when some options are disabled (eg. --disable-cuda). Perhaps 
we  should refresh tests/coverage/coverage-test.sh which has gone 
unmaintained for a while ?

Cédric

--
Cédric AUGONNET
PhD student -- Runtime Team
http://runtime.bordeaux.inria.fr/augonnet/
INRIA Bordeaux - LaBRI