On Fri, Oct 09, 2015 at 13:58:32 +0200, Bernd Schmidt wrote:
> One oddity I noticed in target.c is that there are two different num_devices
> variables:
>
> /* Total number of available devices. */
> static int num_devices;
>
> /* Number of GOMP_OFFLOAD_CAP_OPENMP_400 devices. */
> static int num_devices_openmp;
>
> Confusingly, the get_num_devices function returns num_devices_openmp. That
> function includes a pthread_once call to gomp_target_init, which sets up
> these variables. References to num_devices_openmp through get_num_devices
> are thereforce guaranteed to be initialized. However, there are direct
> references to num_devices, in GOMP_offload_register_ver and
> GOMP_offload_unregister_ver, and they don't seem to enforce any kind of
> initialization:
>
> /* Load image to all initialized devices. */
> for (i = 0; i < num_devices; i++)
> {
> struct gomp_device_descr *devicep = &devices[i];
> gomp_mutex_lock (&devicep->lock);
> if (devicep->type == target_type && devicep->is_initialized)
> gomp_load_image_to_device (devicep, version,
> host_table, target_data, true);
> gomp_mutex_unlock (&devicep->lock);
> }
>
> I'm guessing this only triggers when dlopening something with an offload
> image after devices have been initialized already, and it looks like we have
> symmetrical code in gomp_init_device.
Right, this code offloads given image to all initialized devices, and similar
code in gomp_init_device offloads all registered images to a given device.
> Wouldn't it be possible/better to
> force a gomp_target_init before referencing num_devices, and then relying on
> the code I quoted and deleting the image loading from gomp_init_device?
gomp_target_init only loads plugins and sets num_devices/num_devices_openmp, but
it doesn't call gomp_init_device, because we wanted to defer device
initialization as much as possible. So, gomp_init_device is called immediately
before usage of that device.
-- Ilya
