Hi Jed, Harshad,
A different solution to the problem of PETSc and a user code stepping on
each other's feet with cudaSetDevice might be to use the CUDA driver API
for device selection rather than the runtime API. If we were to
explicitly manage a PETSc CUDA context using the driver API we can
control what devices are being used without interfering with the
mechanisms used by other parts of a client code for CUDA device
selection (e.g. cudaSetDevice). PETSc's device management would be
completely decoupled from the rest of an application.
Of course this approach can be combined with lazy initialization as
proposed by Karl. Whenever the first device function is called we create
the PETSc CUDA context. The advantages of lazy initialization mentioned
by Karl and Jed ensue (e.g. ability to run on machines without GPUs
provided one is not using GPU functionality).
Another advantage of a solution using the driver API is that device and
context management would be very similar between CUDA and OpenCL backends.
I realize that this proposal might be impractical as a near term
solution since it involves a pretty major refactor of the CUDA context
infrastructure. Furthermore, as far as I can tell, third party libraries
that we rely on (e.g. cusp and cusparse) assume the runtime api. Perhaps
these difficulties can be overcome?
A possible near term solution would be to turn this around and to have
applications with advanced device selection requirements use the driver
API. Harshad, I'm not familiar with your code but would it be possible
for you to use the driver API on your end to avoid conflicts with
cudaSetDevice calls inside PETSc?
Cheers,
Dominic
On 01/14/2014 09:27 AM, Harshad Sahasrabudhe wrote:
Hi Jed,
Sometime back we talked about an interface which could handle other
libraries calling cudaSetDevice simultaneously with PETSc. For
example, in our case 2 different instances of PETSc calling cudaSetDevice.
>Sure, but how will we actually share the device between libraries? What
>if the other library was not PETSc, but something else, and they also
>called cudaSetDevice, but with a different default mapping strategy?
>We need an interface that handles this case.
Do we already have any solution for this? If not, can we start looking
at this case?
Thanks,
Harshad
--
Dominic Meiser
Tech-X Corporation
5621 Arapahoe Avenue
Boulder, CO 80303
USA
Telephone: 303-996-2036
Fax: 303-448-7756
www.txcorp.com
