[Starpu-devel] Fwd: Re: [Starpu-commits] r6977 - in trunk: . src/core/perfmodel src/drivers/cuda

[Cyril Roelandt <cyril.roelandt@inria.fr>]

-------- Original Message --------
Subject: Re: [Starpu-commits] r6977 - in trunk: . src/core/perfmodel 
src/drivers/cuda
Date: Fri, 20 Jul 2012 14:27:58 +0200
From: Cyril Roelandt <cyril.roelandt@inria.fr>
To: starpu-commits@lists.gforge.inria.fr

This commit breaks tools/starpu_calibrate_bus.

On hannibal:

oops in initialize_cuda (core/perfmodel/perfmodel_bus.c:101)... 36:
setting the device when a process is active is not allowed
lt-starpu_calibrate_bus: drivers/cuda/driver_cuda.c:507:
starpu_cuda_report_error: Assertion `0' failed.
Aborted (core dumped)

On attila:
[starpu][benchmark_all_gpu_devices] Benchmarking the speed of the bus is
done.
[starpu][_starpu_deinitialize_registered_performance_models] FREE MODELS !
[starpu][_starpu_terminate_workers] wait for worker 0
(gdb) bt
#0  0x00007fffef6aaf07 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#1  0x00007fffef68bf89 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#2  0x00007fffef66cfe5 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#3  0x00007ffff6a3ec39 in ?? () from
/usr/lib/x86_64-linux-gnu/libcudart.so.4
#4  0x00007ffff6a670a9 in cudaStreamDestroy () from
/usr/lib/x86_64-linux-gnu/libcudart.so.4
#5  0x00007ffff7b67ecd in deinit_context (workerid=0, devid=0) at
/home/nfurmento/work/starpu/trunk/src/drivers/cuda/driver_cuda.c:199
#6  0x00007ffff7b68bbf in _starpu_cuda_driver_deinit (d=0x7fffed770ea0)
at /home/nfurmento/work/starpu/trunk/src/drivers/cuda/driver_cuda.c:446
#7  0x00007ffff7b68c2a in _starpu_cuda_worker (arg=0x7ffff7d9d920) at
/home/nfurmento/work/starpu/trunk/src/drivers/cuda/driver_cuda.c:464
#8  0x00007ffff7255b50 in start_thread (arg=<optimized out>) at
pthread_create.c:304
#9  0x00007ffff6fa06dd in clone () at
../sysdeps/unix/sysv/linux/x86_64/clone.S:112
#10 0x0000000000000000 in ?? ()
(gdb)

About the commit log: can we call starpu_worker_get_count() and
_starpu_get_cuda_device_count() without calling starpu_init() first ? In
private, we talked about removing the calls to starpu_init() and
starpu_shutdown() from tools/starpu_calibrate_bus.c, so I'm not sure
that we will be able to call the *_count() functions in the bus
calibration code.


Cyril.

_______________________________________________
Starpu-commits mailing list
Starpu-commits@lists.gforge.inria.fr
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-commits