CUDA releases before 4.0 didn't support this attribute, the #ifdef cannot work anymore on recent CUDA releases, I'll fix that, thanks. Interesting to know that NUMAScale machines use PCI domains. Brice
Le 04/06/2015 14:13, Imre Kerr a écrit : > Hi, > Never mind, I figured it out. hwloc_cudart_get_device_pci_ids was > giving the PCI Domain as being 0 for all GPUs. Looking at > hwloc/cudart.h, I found some strange conditional. Removing it solved > the problem. Is CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID a holdover from some > older CUDA version? > > diff --git a/include/hwloc/cudart.h b/include/hwloc/cudart.h > index 243e0fd..e2fd74d 100644 > --- a/include/hwloc/cudart.h > +++ b/include/hwloc/cudart.h > @@ -56,12 +56,7 @@ hwloc_cudart_get_device_pci_ids(hwloc_topology_t > topology __hwloc_attribute_unus > return -1; > } > > -#ifdef CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID > *domain = prop.pciDomainID; > -#else > - *domain = 0; > -#endif > - > *bus = prop.pciBusID; > *dev = prop.pciDeviceID; > > Imre > _______________________________________________ > hwloc-users mailing list > hwloc-us...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/hwloc-users > Link to this post: > http://www.open-mpi.org/community/lists/hwloc-users/2015/06/1181.php