Re: [Starpu-devel] GPU issue with r12137

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

I tried the new revision (12533) and still got wrong results...

That is strange because this revision completly include the fixing patch

I did a bissection between 12511 + patch to 12533 + patch and the first bad commit is 12533....

There may be a problem with the part of the commit which differ from the patch :

--- a/src/datawizard/coherency.c

+++ b/src/datawizard/coherency.c

@@ -172,7 +172,13 @@ static int worker_supports_direct_access(unsigned node, unsigned handling_node)

                        enum starpu_node_kind kind = starpu_node_get_kind(handling_node);

                        /* GPUs not always allow direct remote access: if CUDA4

                         * is enabled, we allow two CUDA devices to communicate. */

-                       return kind == STARPU_CPU_RAM || kind == STARPU_CUDA_RAM;

+                       return

+#if 0

+                               /* CUDA does not seem very safe with concurrent

+                                * transfer queueing, avoid queueing from CPUs */

+                               kind == STARPU_CPU_RAM ||

+#endif

+                               kind == STARPU_CUDA_RAM;

                }

I runned the test several time because i didn't believe this can be the cause....

Indeed doing :

diff --git a/src/datawizard/coherency.c b/src/datawizard/coherency.c

index 6e67156..8e82163 100644

--- a/src/datawizard/coherency.c

+++ b/src/datawizard/coherency.c

@@ -172,13 +172,13 @@ static int worker_supports_direct_access(unsigned node, unsigned handling_node)

                        enum starpu_node_kind kind = starpu_node_get_kind(handling_node);

                        /* GPUs not always allow direct remote access: if CUDA4

                         * is enabled, we allow two CUDA devices to communicate. */

-                       return

-#if 0

+                       return kind ==

+#if 1

                                /* CUDA does not seem very safe with concurrent

                                 * transfer queueing, avoid queueing from CPUs */

-                               kind == STARPU_CPU_RAM ||

+                               STARPU_CPU_RAM ||

 #endif

-                               kind == STARPU_CUDA_RAM;

+                               STARPU_CUDA_RAM;

                }

 #else

                        /* Direct GPU-GPU transfers are not allowed in general */

solves the issue.

XL.

Le 3 avr. 2014 à 09:05, Xavier Lacoste <xavier.lacoste@inria.fr> a écrit :

Thanks !

XL.

Le 2 avr. 2014 à 19:09, Samuel Thibault <samuel.thibault@ens-lyon.org> a écrit :

Samuel Thibault, le Wed 02 Apr 2014 18:51:34 +0200, a écrit :

Ok, I have reverted r12137/12138, then.

(I mean, the "use non-zero streams on CPU workers" part)

Samuel

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