OpenMPI
0.1.1
|
This file implements a simple memory pool that is used by the GPU buffer on the sending side. More...
#include "ompi_config.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/mpool/gpusm/mpool_gpusm.h"
#include "ompi/runtime/params.h"
#include "ompi/mca/common/cuda/common_cuda.h"
This file implements a simple memory pool that is used by the GPU buffer on the sending side.
It just gets a memory handle and event handle that can be sent to the remote side which can then use the handles to get access to the memory and the event to determine when it can start accessing the memory. There is no caching of the memory handles as getting new ones is fast. The event handles are cached by the cuda_common code.
void mca_mpool_gpusm_finalize | ( | struct mca_mpool_base_module_t * | mpool | ) |
int mca_mpool_gpusm_find | ( | mca_mpool_base_module_t * | mpool, |
void * | addr, | ||
size_t | size, | ||
mca_mpool_base_registration_t ** | reg | ||
) |
Just go ahead and get a new registration.
find registration for a given block of memory
The find and register functions are the same thing for this memory pool.
References mca_mpool_gpusm_register().
int mca_mpool_gpusm_ft_event | ( | int | state | ) |
Fault Tolerance Event Notification Function.
state | Checkpoint Stae |
|
static |
Called when the registration free list is created.
An event is created for each entry.
|
static |
Called when the program is exiting.
This destroys the events.