Re: [Starpu-devel] [Starpu-commits] r5692 - in trunk/src: core core/dependencies datawizard drivers/cpu drivers/cuda drivers/gordon drivers/opencl

[Xavier Lacoste <xavier.lacoste@inria.fr>]

Hello,

As Samuel told me it should be fixed I tried the new version of StarPU.

Did you do changes on data registration ? I can't register any data.

Program received signal SIGSEGV, Segmentation fault.
starpu_data_register (handleptr=0x7fffffffc620, home_node=0, 
data_interface=0x7fffffffc400, ops=0x7ffff763d2e0) at 
datawizard/interfaces/data_interface.c:296
296                       ((struct starpu_data_copy_methods 
*)ops->copy_methods)->ram_to_cuda_async = NULL;
(gdb) bt
#0  starpu_data_register (handleptr=0x7fffffffc620, home_node=0, 
data_interface=0x7fffffffc400, ops=0x7ffff763d2e0) at 
datawizard/interfaces/data_interface.c:296
#1  0x00007ffff74151ce in starpu_vector_data_register 
(handleptr=0x7fffc3efaf88, home_node=0, ptr=<value optimized out>, nx=<value 
optimized out>, elemsize=40)
    at datawizard/interfaces/vector_interface.c:163
#2  0x000000000047d830 in Dstarpu_submit_tasks (sopalin_data=0x26ed7d8) at 
sopalin/src/starpu_submit_tasks.c:583
#3  0x000000000047c60f in Dsopalin_updo_gmres_thread (m=0x26ec708, 
sopaparam=0x26ec898) at sopalin/src/sopalin3d.c:1522
#4  0x00000000004347dc in pastix_task_sopalin (pastix_data=0x26ec708, 
pastix_comm=0x7a69a0, n=12111, colptr=0x896a00, row=<value optimized out>, 
avals=<value optimized out>, 
    b=0x8a2750, rhsnbr=1, loc2glob=0x0) at sopalin/src/pastix.c:3195
#5  0x0000000000436a35 in pastix (pastix_data=0x7fffffffcc88, 
pastix_comm=0x7a69a0, n=12111, colptr=0x896a00, row=0x7ffff7fd1010, 
avals=0x7fffe6b13010, perm=0x8ba1f0, 
    invp=0x8c5f40, b=0x8a2750, rhs=1, iparm=0x7fffffffcb40, 
dparm=0x7fffffffc940) at sopalin/src/pastix.c:4273
#6  0x0000000000412745 in main (argc=13, argv=0x7fffffffcdc8) at simple.c:204

Note that the size in   starpu_vector_data_register elemsize is 40 whereas it 
should be 4 :

p vector
$1 = {ptr = 140736721846296, dev_handle = 140736721846296, offset = 0, nx = 
39900, elemsize = 4}

I'll run it in valgrind but it's quite long

XL.

Le 14 mars 2012 à 09:56, Xavier Lacoste a écrit :

> 
> Le 13 mars 2012 à 18:56, Cyril Roelandt a écrit :
> 
>> On 02/24/2012 02:55 PM, Xavier Lacoste wrote:
>>> 
>>> Le 24 févr. 2012 à 14:47, Cyril Roelandt a écrit :
>>> 
>>>> On 02/24/2012 11:01 AM, Xavier Lacoste wrote:
>>>>> Hello again,
>>>>> 
>>>>> I did a smaller example creating an empty sparse matrix and doing a 
>>>>> fake factorization using StarPU.
>>>>> The problem is that for the moment it doesn't reproduce the bug.
>>>>> Can you remember me the define I gave you in config.in file in STARPU 
>>>>> section ?
>>>>> 
>>>> 
>>>> Hello,
>>>> 
>>>> Here is the content of the StarPU section in config.in:
>>>> 
>>>> CCPASTIX   := -L/opt/cluster/plafrim-dev/hwloc/1.3.1/lib $(CCPASTIX) 
>>>> `pkg-config libstarpu --cflags` -DWITH_STARPU -DSTARPU_BLOCKTAB_GEMM 
>>>> -DSTARPU_BLOCKTAB_SELFCOPY
>>>> EXTRALIB   := $(EXTRALIB) `pkg-config libstarpu --libs --static`
>>>> 
>>>> 
>>>> Cyril Roelandt.
>>> 
>>> Ok it's what I remembered... Maybe my example (11x11 matrix) is not big 
>>> enough to reproduce the bug... I don't know...
>> 
>> 
>> I see. It's probably quite hard to write a very simple program that 
>> reproduces this issue. Using the exact same code you sent me, do you think 
>> we could use an input file representing a smaller matrix than shipsec5.rsa 
>> ? The thing is, we are trying to run helgrind on your code, but the data 
>> seems big enough to prevent it from completing.
>> 
>> Cyril Roelandt.
> 
> Mis-replied to you instead of starpu-devel :
> 
>> The code is working on an other machine so I tried to update StarPU, 
>> hoping that it was fixed, sadly it's not.
>> I cleaned some of the StarPU code in PaStiX if you want I can give you a 
>> newer version...
>> The problem is that I don't really know how to simply generate a valid 
>> sparse matrix so I couldn't go bigger than 11x11 matrix for the example I 
>> tried to write...
>> For the input file you can use matrix/small.rsa which is smaller 
>> (orsirr.rua does not crash, -mm matrix/young4c.mtc neither) you can also 
>> try with -lap <size> which will create a laplacian of the given size.
>> 
>> 
>> Thanks
>> 
>> XL.
> 
> A little precision, matrix/small.rsa only works with LDLT Factorzation (in 
> the main example if you have API_FACT_LLT, change it to API_FACT_LDLT) If 
> you need any help tell me.
> _______________________________________________
> Starpu-devel mailing list
> Starpu-devel@lists.gforge.inria.fr
> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel