ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

Avoid extra lookup for ompi_proc in homogenous build

Просмотр исходного кода 
Signed-off-by: Jithin Jose <jithin.jose@intel.com>
Этот коммит содержится в:
Jithin Jose 2015-05-14 16:11:14 -07:00
родитель 50089977ac
Коммит 07043894bd
3 изменённых файлов: 135 добавлений и 26 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

50
ompi/mca/pml/cm/pml_cm.h
Просмотреть файл

@ -74,7 +74,9 @@ mca_pml_cm_irecv_init(void *addr,
struct ompi_request_t **request) struct ompi_request_t **request)
{ {
mca_pml_cm_hvy_recv_request_t *recvreq; mca_pml_cm_hvy_recv_request_t *recvreq;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc; ompi_proc_t* ompi_proc;
#endif
MCA_PML_CM_HVY_RECV_REQUEST_ALLOC(recvreq); MCA_PML_CM_HVY_RECV_REQUEST_ALLOC(recvreq);
if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE; if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE;
@ -98,7 +100,9 @@ mca_pml_cm_irecv(void *addr,
{ {
int ret; int ret;
mca_pml_cm_thin_recv_request_t *recvreq; mca_pml_cm_thin_recv_request_t *recvreq;
ompi_proc_t* ompi_proc; #if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc = NULL;
#endif
MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq); MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq);
if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE; if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE;
@ -136,7 +140,9 @@ mca_pml_cm_recv(void *addr,
ompi_status_public_t * status) ompi_status_public_t * status)
{ {
int ret; int ret;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t *ompi_proc; ompi_proc_t *ompi_proc;
#endif
opal_convertor_t convertor; opal_convertor_t convertor;
mca_pml_cm_request_t req; mca_pml_cm_request_t req;
mca_mtl_request_t *req_mtl = mca_mtl_request_t *req_mtl =
@ -154,6 +160,7 @@ mca_pml_cm_recv(void *addr,
req.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS; req.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS;
req.req_ompi.req_status._cancelled = 0; req.req_ompi.req_status._cancelled = 0;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
if( MPI_ANY_SOURCE == src ) { if( MPI_ANY_SOURCE == src ) {
ompi_proc = ompi_proc_local_proc; ompi_proc = ompi_proc_local_proc;
} else { } else {
@ -167,6 +174,16 @@ mca_pml_cm_recv(void *addr,
addr, addr,
0, 0,
&convertor ); &convertor );
#else
opal_convertor_copy_and_prepare_for_recv(
ompi_mpi_local_convertor,
&(datatype->super),
count,
addr,
0,
&convertor );
#endif
ret = OMPI_MTL_CALL(irecv(ompi_mtl, ret = OMPI_MTL_CALL(irecv(ompi_mtl,
comm, comm,
src, src,
@ -198,7 +215,9 @@ mca_pml_cm_isend_init(void* buf,
ompi_request_t** request) ompi_request_t** request)
{ {
mca_pml_cm_hvy_send_request_t *sendreq; mca_pml_cm_hvy_send_request_t *sendreq;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc; ompi_proc_t* ompi_proc;
#endif
MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc); MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc);
if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE; if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE;
@ -225,7 +244,9 @@ mca_pml_cm_isend(void* buf,
if(sendmode == MCA_PML_BASE_SEND_BUFFERED ) { if(sendmode == MCA_PML_BASE_SEND_BUFFERED ) {
mca_pml_cm_hvy_send_request_t* sendreq; mca_pml_cm_hvy_send_request_t* sendreq;
ompi_proc_t* ompi_proc; #if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc = NULL;
#endif
MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc); MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc);
if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE; if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE;
@ -248,7 +269,9 @@ mca_pml_cm_isend(void* buf,
} else { } else {
mca_pml_cm_thin_send_request_t* sendreq; mca_pml_cm_thin_send_request_t* sendreq;
ompi_proc_t* ompi_proc; #if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc = NULL;
#endif
MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc); MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc);
if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE; if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE;
@ -288,10 +311,11 @@ mca_pml_cm_send(void *buf,
ompi_communicator_t* comm) ompi_communicator_t* comm)
{ {
int ret = OMPI_ERROR; int ret = OMPI_ERROR;
ompi_proc_t * ompi_proc;
if(sendmode == MCA_PML_BASE_SEND_BUFFERED) { if(sendmode == MCA_PML_BASE_SEND_BUFFERED) {
mca_pml_cm_hvy_send_request_t *sendreq; mca_pml_cm_hvy_send_request_t *sendreq;
ompi_proc_t * ompi_proc;
MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc); MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc);
if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE; if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE;
@ -315,14 +339,13 @@ mca_pml_cm_send(void *buf,
ompi_request_free( (ompi_request_t**)&sendreq ); ompi_request_free( (ompi_request_t**)&sendreq );
} else { } else {
opal_convertor_t convertor; opal_convertor_t convertor;
ompi_proc_t *ompi_proc = ompi_comm_peer_lookup(comm, dst);
#if !(OPAL_ENABLE_HETEROGENEOUS_SUPPORT) #if !(OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
if (opal_datatype_is_contiguous_memory_layout(&datatype->super, count)) { if (opal_datatype_is_contiguous_memory_layout(&datatype->super, count)) {
convertor.remoteArch = ompi_proc->super.proc_convertor->remoteArch; convertor.remoteArch = ompi_mpi_local_convertor->remoteArch;
convertor.flags = ompi_proc->super.proc_convertor->flags; convertor.flags = ompi_mpi_local_convertor->flags;
convertor.master = ompi_proc->super.proc_convertor->master; convertor.master = ompi_mpi_local_convertor->master;
convertor.local_size = count * datatype->super.size; convertor.local_size = count * datatype->super.size;
convertor.pBaseBuf = (unsigned char*)buf; convertor.pBaseBuf = (unsigned char*)buf;
@ -331,6 +354,7 @@ mca_pml_cm_send(void *buf,
} else } else
#endif #endif
{ {
ompi_proc = ompi_comm_peer_lookup(comm, dst);
opal_convertor_copy_and_prepare_for_send( opal_convertor_copy_and_prepare_for_send(
ompi_proc->super.proc_convertor, ompi_proc->super.proc_convertor,
&datatype->super, count, buf, 0, &datatype->super, count, buf, 0,
@ -422,9 +446,10 @@ mca_pml_cm_imrecv(void *buf,
{ {
int ret; int ret;
mca_pml_cm_thin_recv_request_t *recvreq; mca_pml_cm_thin_recv_request_t *recvreq;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc; ompi_proc_t* ompi_proc;
#endif
ompi_communicator_t *comm = (*message)->comm; ompi_communicator_t *comm = (*message)->comm;
int peer = (*message)->peer;
MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq); MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq);
if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE; if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE;
@ -432,7 +457,7 @@ mca_pml_cm_imrecv(void *buf,
MCA_PML_CM_THIN_RECV_REQUEST_INIT(recvreq, MCA_PML_CM_THIN_RECV_REQUEST_INIT(recvreq,
ompi_proc, ompi_proc,
comm, comm,
peer, (*message)->peer,
datatype, datatype,
buf, buf,
count); count);
@ -453,9 +478,10 @@ mca_pml_cm_mrecv(void *buf,
{ {
int ret; int ret;
mca_pml_cm_thin_recv_request_t *recvreq; mca_pml_cm_thin_recv_request_t *recvreq;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
ompi_proc_t* ompi_proc; ompi_proc_t* ompi_proc;
#endif
ompi_communicator_t *comm = (*message)->comm; ompi_communicator_t *comm = (*message)->comm;
int peer = (*message)->peer;
MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq); MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq);
if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE; if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE;
@ -463,7 +489,7 @@ mca_pml_cm_mrecv(void *buf,
MCA_PML_CM_THIN_RECV_REQUEST_INIT(recvreq, MCA_PML_CM_THIN_RECV_REQUEST_INIT(recvreq,
ompi_proc, ompi_proc,
comm, comm,
peer, (*message)->peer,
datatype, datatype,
buf, buf,
count); count);

64
ompi/mca/pml/cm/pml_cm_recvreq.h
Просмотреть файл

@ -85,6 +85,7 @@ do { \
* @param comm (IN) Communicator. * @param comm (IN) Communicator.
* @param persistent (IN) Is this a ersistent request. * @param persistent (IN) Is this a ersistent request.
*/ */
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
#define MCA_PML_CM_THIN_RECV_REQUEST_INIT( request, \ #define MCA_PML_CM_THIN_RECV_REQUEST_INIT( request, \
ompi_proc, \ ompi_proc, \
comm, \ comm, \
@ -115,7 +116,35 @@ do { \
0, \ 0, \
&(request)->req_base.req_convertor ); \ &(request)->req_base.req_convertor ); \
} while(0) } while(0)
#else
#define MCA_PML_CM_THIN_RECV_REQUEST_INIT( request, \
ompi_proc, \
comm, \
src, \
datatype, \
addr, \
count ) \
do { \
OMPI_REQUEST_INIT(&(request)->req_base.req_ompi, false); \
(request)->req_base.req_ompi.req_mpi_object.comm = comm; \
(request)->req_base.req_pml_complete = false; \
(request)->req_base.req_free_called = false; \
request->req_base.req_comm = comm; \
request->req_base.req_datatype = datatype; \
OBJ_RETAIN(comm); \
OBJ_RETAIN(datatype); \
\
opal_convertor_copy_and_prepare_for_recv( \
ompi_mpi_local_convertor, \
&(datatype->super), \
count, \
addr, \
0, \
&(request)->req_base.req_convertor ); \
} while(0)
#endif
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
#define MCA_PML_CM_HVY_RECV_REQUEST_INIT( request, \ #define MCA_PML_CM_HVY_RECV_REQUEST_INIT( request, \
ompi_proc, \ ompi_proc, \
comm, \ comm, \
@ -152,7 +181,39 @@ do { \
0, \ 0, \
&(request)->req_base.req_convertor ); \ &(request)->req_base.req_convertor ); \
} while(0) } while(0)
#else
#define MCA_PML_CM_HVY_RECV_REQUEST_INIT( request, \
ompi_proc, \
comm, \
tag, \
src, \
datatype, \
addr, \
count, \
persistent) \
do { \
OMPI_REQUEST_INIT(&(request)->req_base.req_ompi, persistent); \
(request)->req_base.req_ompi.req_mpi_object.comm = comm; \
(request)->req_base.req_pml_complete = OPAL_INT_TO_BOOL(persistent); \
(request)->req_base.req_free_called = false; \
request->req_base.req_comm = comm; \
request->req_base.req_datatype = datatype; \
request->req_tag = tag; \
request->req_peer = src; \
request->req_addr = addr; \
request->req_count = count; \
OBJ_RETAIN(comm); \
OBJ_RETAIN(datatype); \
\
opal_convertor_copy_and_prepare_for_recv( \
ompi_mpi_local_convertor, \
&(datatype->super), \
count, \
addr, \
0, \
&(request)->req_base.req_convertor ); \
} while(0)
#endif
/** /**
* Start an initialized request. * Start an initialized request.
@ -315,7 +376,6 @@ do { \
} }
extern void mca_pml_cm_recv_request_completion(struct mca_mtl_request_t *mtl_request); extern void mca_pml_cm_recv_request_completion(struct mca_mtl_request_t *mtl_request);
extern void mca_pml_cm_recv_fast_completion(struct mca_mtl_request_t *mtl_request);
#endif #endif

33
ompi/mca/pml/cm/pml_cm_sendreq.h
Просмотреть файл

@ -59,6 +59,7 @@ typedef struct mca_pml_cm_hvy_send_request_t mca_pml_cm_hvy_send_request_t;
OBJ_CLASS_DECLARATION(mca_pml_cm_hvy_send_request_t); OBJ_CLASS_DECLARATION(mca_pml_cm_hvy_send_request_t);
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
#define MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, \ #define MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, \
ompi_proc) \ ompi_proc) \
do { \ do { \
@ -74,8 +75,20 @@ do { \
sendreq->req_mtl.completion_callback = mca_pml_cm_send_request_completion; \ sendreq->req_mtl.completion_callback = mca_pml_cm_send_request_completion; \
} \ } \
} while(0) } while(0)
#else
#define MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, \
ompi_proc) \
do { \
sendreq = (mca_pml_cm_thin_send_request_t*) \
opal_free_list_wait (&mca_pml_base_send_requests); \
sendreq->req_send.req_base.req_pml_type = MCA_PML_CM_REQUEST_SEND_THIN; \
sendreq->req_mtl.ompi_req = (ompi_request_t*) sendreq; \
sendreq->req_mtl.completion_callback = mca_pml_cm_send_request_completion; \
} while(0)
#endif
#if (OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
#define MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, \ #define MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, \
ompi_proc) \ ompi_proc) \
{ \ { \
@ -90,7 +103,17 @@ do { \
sendreq->req_mtl.completion_callback = mca_pml_cm_send_request_completion; \ sendreq->req_mtl.completion_callback = mca_pml_cm_send_request_completion; \
} \ } \
} }
#else
#define MCA_PML_CM_HVY_SEND_REQUEST_ALLOC(sendreq, comm, dst, \
ompi_proc) \
{ \
sendreq = (mca_pml_cm_hvy_send_request_t*) \
opal_free_list_wait (&mca_pml_base_send_requests); \
sendreq->req_send.req_base.req_pml_type = MCA_PML_CM_REQUEST_SEND_HEAVY; \
sendreq->req_mtl.ompi_req = (ompi_request_t*) sendreq; \
sendreq->req_mtl.completion_callback = mca_pml_cm_send_request_completion; \
}
#endif
#if (OPAL_ENABLE_HETEROGENEOUS_SUPPORT) #if (OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
#define MCA_PML_CM_SEND_REQUEST_INIT_COMMON(req_send, \ #define MCA_PML_CM_SEND_REQUEST_INIT_COMMON(req_send, \
@ -138,11 +161,11 @@ do { \
(req_send)->req_base.req_datatype = datatype; \ (req_send)->req_base.req_datatype = datatype; \
if (opal_datatype_is_contiguous_memory_layout(&datatype->super, count)) { \ if (opal_datatype_is_contiguous_memory_layout(&datatype->super, count)) { \
(req_send)->req_base.req_convertor.remoteArch = \ (req_send)->req_base.req_convertor.remoteArch = \
ompi_proc->super.proc_convertor->remoteArch; \ ompi_mpi_local_convertor->remoteArch; \
(req_send)->req_base.req_convertor.flags = \ (req_send)->req_base.req_convertor.flags = \
ompi_proc->super.proc_convertor->flags; \ ompi_mpi_local_convertor->flags; \
(req_send)->req_base.req_convertor.master = \ (req_send)->req_base.req_convertor.master = \
ompi_proc->super.proc_convertor->master; \ ompi_mpi_local_convertor->master; \
(req_send)->req_base.req_convertor.local_size = \ (req_send)->req_base.req_convertor.local_size = \
count * datatype->super.size; \ count * datatype->super.size; \
(req_send)->req_base.req_convertor.pBaseBuf = (unsigned char*)buf; \ (req_send)->req_base.req_convertor.pBaseBuf = (unsigned char*)buf; \
@ -150,7 +173,7 @@ do { \
(req_send)->req_base.req_convertor.pDesc = &datatype->super; \ (req_send)->req_base.req_convertor.pDesc = &datatype->super; \
} else { \ } else { \
opal_convertor_copy_and_prepare_for_send( \ opal_convertor_copy_and_prepare_for_send( \
ompi_proc->super.proc_convertor, \ ompi_mpi_local_convertor, \
&(datatype->super), \ &(datatype->super), \
count, \ count, \
buf, \ buf, \