1
1
openmpi/ompi/mca/pml/cm/pml_cm_sendreq.h
Aravind Gopalakrishnan 2e83cf15ce Add support for GPU buffers for PSM2 MTL
PSM2 enables support for GPU buffers and CUDA managed memory and it can
directly recognize GPU buffers, handle copies between HFIs and GPUs.
Therefore, it is not required for OMPI to handle GPU buffers for pt2pt cases.
In this patch, we allow the PSM2 MTL to specify when
it does not require CUDA convertor support. This allows us to skip CUDA
convertor init phases and lets PSM2 handle the memory transfers.

This translates to improvements in latency.
The patch enables blocking collectives and workloads with GPU contiguous,
GPU non-contiguous memory.

Signed-off-by: Aravind Gopalakrishnan <Aravind.Gopalakrishnan@intel.com>
2017-09-01 16:59:03 -07:00

499 строки
30 KiB
C

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2013 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2006 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2015 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2015 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2017 Intel, Inc. All rights reserved
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef PML_CM_SENDREQ_H
#define PML_CM_SENDREQ_H
#include "pml_cm_request.h"
#include "ompi/mca/pml/base/pml_base_sendreq.h"
#include "ompi/mca/pml/base/pml_base_bsend.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/mtl/mtl.h"
#include "opal/prefetch.h"
struct mca_pml_cm_send_request_t {
mca_pml_cm_request_t req_base;
mca_pml_base_send_mode_t req_send_mode;
};
typedef struct mca_pml_cm_send_request_t mca_pml_cm_send_request_t;
OBJ_CLASS_DECLARATION(mca_pml_cm_send_request_t);
struct mca_pml_cm_thin_send_request_t {
mca_pml_cm_send_request_t req_send;
mca_mtl_request_t req_mtl; /**< the mtl specific memory. This field should be the last in the struct */
};
typedef struct mca_pml_cm_thin_send_request_t mca_pml_cm_thin_send_request_t;
OBJ_CLASS_DECLARATION(mca_pml_cm_thin_send_request_t);
struct mca_pml_cm_hvy_send_request_t {
mca_pml_cm_send_request_t req_send;
const void *req_addr; /**< pointer to application buffer */
size_t req_count; /**< count of user datatype elements */
int32_t req_peer; /**< peer process - rank w/in this communicator */
int32_t req_tag; /**< user defined tag */
void *req_buff; /**< pointer to send buffer - may not be application buffer */
bool req_blocking;
mca_mtl_request_t req_mtl; /**< the mtl specific memory. This field should be the last in the struct */
};
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);
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
#define MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, \
ompi_proc) \
do { \
ompi_proc = ompi_comm_peer_lookup( comm, dst ); \
\
if(OPAL_UNLIKELY(NULL == ompi_proc)) { \
sendreq = NULL; \
} else { \
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)
#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, \
ompi_proc) \
{ \
ompi_proc = ompi_comm_peer_lookup( comm, dst ); \
if(OPAL_UNLIKELY(NULL == ompi_proc)) { \
sendreq = NULL; \
} else { \
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; \
} \
}
#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)
#define MCA_PML_CM_HVY_SEND_REQUEST_INIT_COMMON(req_send, \
ompi_proc, \
comm, \
tag, \
datatype, \
sendmode, \
buf, \
count, \
flags ) \
{ \
OBJ_RETAIN(comm); \
OMPI_DATATYPE_RETAIN(datatype); \
(req_send)->req_base.req_comm = comm; \
(req_send)->req_base.req_datatype = datatype; \
MCA_PML_CM_SWITCH_CUDA_CONVERTOR_OFF(flags, datatype, count); \
opal_convertor_copy_and_prepare_for_send( \
ompi_proc->super.proc_convertor, \
&(datatype->super), \
count, \
buf, \
flags, \
&(req_send)->req_base.req_convertor ); \
(req_send)->req_base.req_ompi.req_mpi_object.comm = comm; \
(req_send)->req_base.req_ompi.req_status.MPI_SOURCE = \
comm->c_my_rank; \
(req_send)->req_base.req_ompi.req_status.MPI_TAG = tag; \
(req_send)->req_base.req_ompi.req_status._ucount = count; \
(req_send)->req_send_mode = sendmode; \
(req_send)->req_base.req_free_called = false; \
}
#else
#define MCA_PML_CM_HVY_SEND_REQUEST_INIT_COMMON(req_send, \
ompi_proc, \
comm, \
tag, \
datatype, \
sendmode, \
buf, \
count, \
flags ) \
{ \
OBJ_RETAIN(comm); \
OMPI_DATATYPE_RETAIN(datatype); \
(req_send)->req_base.req_comm = comm; \
(req_send)->req_base.req_datatype = datatype; \
MCA_PML_CM_SWITCH_CUDA_CONVERTOR_OFF(flags, datatype, count); \
opal_convertor_copy_and_prepare_for_send( \
ompi_mpi_local_convertor, \
&(datatype->super), \
count, \
buf, \
flags, \
&(req_send)->req_base.req_convertor ); \
(req_send)->req_base.req_ompi.req_mpi_object.comm = comm; \
(req_send)->req_base.req_ompi.req_status.MPI_SOURCE = \
comm->c_my_rank; \
(req_send)->req_base.req_ompi.req_status.MPI_TAG = tag; \
(req_send)->req_base.req_ompi.req_status._ucount = count; \
(req_send)->req_send_mode = sendmode; \
(req_send)->req_base.req_free_called = false; \
}
#endif
#if (OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
#define MCA_PML_CM_SEND_REQUEST_INIT_COMMON(req_send, \
ompi_proc, \
comm, \
tag, \
datatype, \
sendmode, \
buf, \
count, \
flags ) \
{ \
OBJ_RETAIN(comm); \
OMPI_DATATYPE_RETAIN(datatype); \
(req_send)->req_base.req_comm = comm; \
(req_send)->req_base.req_datatype = datatype; \
MCA_PML_CM_SWITCH_CUDA_CONVERTOR_OFF(flags, datatype, count); \
opal_convertor_copy_and_prepare_for_send( \
ompi_proc->super.proc_convertor, \
&(datatype->super), \
count, \
buf, \
flags, \
&(req_send)->req_base.req_convertor ); \
(req_send)->req_base.req_ompi.req_mpi_object.comm = comm; \
(req_send)->req_base.req_ompi.req_status.MPI_SOURCE = \
comm->c_my_rank; \
(req_send)->req_base.req_ompi.req_status.MPI_TAG = tag; \
(req_send)->req_base.req_ompi.req_status._ucount = count; \
(req_send)->req_send_mode = sendmode; \
(req_send)->req_base.req_free_called = false; \
}
#else
#define MCA_PML_CM_SEND_REQUEST_INIT_COMMON(req_send, \
ompi_proc, \
comm, \
tag, \
datatype, \
sendmode, \
buf, \
count, \
flags ) \
{ \
OBJ_RETAIN(comm); \
OMPI_DATATYPE_RETAIN(datatype); \
(req_send)->req_base.req_comm = comm; \
(req_send)->req_base.req_datatype = datatype; \
if (opal_datatype_is_contiguous_memory_layout(&datatype->super, count)) { \
(req_send)->req_base.req_convertor.remoteArch = \
ompi_mpi_local_convertor->remoteArch; \
(req_send)->req_base.req_convertor.flags = \
ompi_mpi_local_convertor->flags; \
(req_send)->req_base.req_convertor.master = \
ompi_mpi_local_convertor->master; \
(req_send)->req_base.req_convertor.local_size = \
count * datatype->super.size; \
(req_send)->req_base.req_convertor.pBaseBuf = \
(unsigned char*)buf + datatype->super.true_lb; \
(req_send)->req_base.req_convertor.count = count; \
(req_send)->req_base.req_convertor.pDesc = &datatype->super; \
} else { \
MCA_PML_CM_SWITCH_CUDA_CONVERTOR_OFF(flags, datatype, count); \
opal_convertor_copy_and_prepare_for_send( \
ompi_mpi_local_convertor, \
&(datatype->super), \
count, \
buf, \
flags, \
&(req_send)->req_base.req_convertor ); \
} \
(req_send)->req_base.req_ompi.req_mpi_object.comm = comm; \
(req_send)->req_base.req_ompi.req_status.MPI_SOURCE = \
comm->c_my_rank; \
(req_send)->req_base.req_ompi.req_status.MPI_TAG = tag; \
(req_send)->req_base.req_ompi.req_status._ucount = count; \
(req_send)->req_send_mode = sendmode; \
(req_send)->req_base.req_free_called = false; \
}
#endif
#define MCA_PML_CM_HVY_SEND_REQUEST_INIT( sendreq, \
ompi_proc, \
comm, \
tag, \
dst, \
datatype, \
sendmode, \
persistent, \
blocking, \
buf, \
count, \
flags ) \
do { \
OMPI_REQUEST_INIT(&(sendreq->req_send.req_base.req_ompi), \
persistent); \
sendreq->req_tag = tag; \
sendreq->req_peer = dst; \
sendreq->req_addr = buf; \
sendreq->req_count = count; \
MCA_PML_CM_HVY_SEND_REQUEST_INIT_COMMON( (&sendreq->req_send), \
ompi_proc, \
comm, \
tag, \
datatype, \
sendmode, \
buf, \
count, \
flags ) \
opal_convertor_get_packed_size( \
&sendreq->req_send.req_base.req_convertor, \
&sendreq->req_count ); \
\
sendreq->req_blocking = blocking; \
sendreq->req_send.req_base.req_pml_complete = \
(persistent ? true:false); \
} while(0)
#define MCA_PML_CM_THIN_SEND_REQUEST_INIT( sendreq, \
ompi_proc, \
comm, \
tag, \
dst, \
datatype, \
sendmode, \
buf, \
count, \
flags ) \
do { \
OMPI_REQUEST_INIT(&(sendreq->req_send.req_base.req_ompi), \
false); \
MCA_PML_CM_SEND_REQUEST_INIT_COMMON( (&sendreq->req_send), \
ompi_proc, \
comm, \
tag, \
datatype, \
sendmode, \
buf, \
count, \
flags); \
sendreq->req_send.req_base.req_pml_complete = false; \
} while(0)
#define MCA_PML_CM_SEND_REQUEST_START_SETUP(req_send) \
do { \
(req_send)->req_base.req_pml_complete = false; \
(req_send)->req_base.req_ompi.req_complete = REQUEST_PENDING; \
(req_send)->req_base.req_ompi.req_state = \
OMPI_REQUEST_ACTIVE; \
(req_send)->req_base.req_ompi.req_status._cancelled = 0; \
} while (0)
#define MCA_PML_CM_THIN_SEND_REQUEST_START(sendreq, \
comm, \
tag, \
dst, \
sendmode, \
blocking, \
ret) \
do { \
MCA_PML_CM_SEND_REQUEST_START_SETUP(&(sendreq)->req_send); \
ret = OMPI_MTL_CALL(isend(ompi_mtl, \
comm, \
dst, \
tag, \
&sendreq->req_send.req_base.req_convertor, \
sendmode, \
blocking, \
&sendreq->req_mtl)); \
} while (0)
#define MCA_PML_CM_HVY_SEND_REQUEST_BSEND_ALLOC(sendreq, ret) \
do { \
struct iovec iov; \
unsigned int iov_count; \
size_t max_data; \
\
if(sendreq->req_count > 0) { \
sendreq->req_buff = \
mca_pml_base_bsend_request_alloc_buf(sendreq->req_count); \
if (NULL == sendreq->req_buff) { \
ret = MPI_ERR_BUFFER; \
} else { \
iov.iov_base = (IOVBASE_TYPE*)sendreq->req_buff; \
max_data = iov.iov_len = sendreq->req_count; \
iov_count = 1; \
opal_convertor_pack( &sendreq->req_send.req_base.req_convertor, \
&iov, \
&iov_count, \
&max_data ); \
opal_convertor_prepare_for_send( &sendreq->req_send.req_base.req_convertor, \
&(ompi_mpi_packed.dt.super), \
max_data, sendreq->req_buff ); \
} \
} \
} while(0);
#define MCA_PML_CM_HVY_SEND_REQUEST_START(sendreq, ret) \
do { \
ret = OMPI_SUCCESS; \
MCA_PML_CM_SEND_REQUEST_START_SETUP(&(sendreq)->req_send); \
if (sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) { \
MCA_PML_CM_HVY_SEND_REQUEST_BSEND_ALLOC(sendreq, ret); \
} \
if (OMPI_SUCCESS == ret) { \
ret = OMPI_MTL_CALL(isend(ompi_mtl, \
sendreq->req_send.req_base.req_comm, \
sendreq->req_peer, \
sendreq->req_tag, \
&sendreq->req_send.req_base.req_convertor, \
sendreq->req_send.req_send_mode, \
sendreq->req_blocking, \
&sendreq->req_mtl)); \
if(OMPI_SUCCESS == ret && \
sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) { \
sendreq->req_send.req_base.req_ompi.req_status.MPI_ERROR = 0; \
ompi_request_complete(&(sendreq)->req_send.req_base.req_ompi, true); \
} \
} \
} while (0)
/*
* The PML has completed a send request. Note that this request
* may have been orphaned by the user or have already completed
* at the MPI level.
* This macro will never be called directly from the upper level, as it should
* only be an internal call to the PML.
*/
#define MCA_PML_CM_HVY_SEND_REQUEST_PML_COMPLETE(sendreq) \
do { \
assert( false == sendreq->req_send.req_base.req_pml_complete ); \
\
if (sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED && \
sendreq->req_count > 0 ) { \
mca_pml_base_bsend_request_free(sendreq->req_buff); \
} \
\
if( !REQUEST_COMPLETE(&sendreq->req_send.req_base.req_ompi)) { \
/* Should only be called for long messages (maybe synchronous) */ \
ompi_request_complete(&(sendreq->req_send.req_base.req_ompi), true); \
} \
sendreq->req_send.req_base.req_pml_complete = true; \
\
if( sendreq->req_send.req_base.req_free_called ) { \
MCA_PML_CM_HVY_SEND_REQUEST_RETURN( sendreq ); \
} else { \
if(sendreq->req_send.req_base.req_ompi.req_persistent) { \
/* rewind convertor */ \
size_t offset = 0; \
opal_convertor_set_position(&sendreq->req_send.req_base.req_convertor, \
&offset); \
} \
} \
} while (0)
/*
* Release resources associated with a request
*/
#define MCA_PML_CM_HVY_SEND_REQUEST_RETURN(sendreq) \
{ \
/* Let the base handle the reference counts */ \
OMPI_DATATYPE_RETAIN(sendreq->req_send.req_base.req_datatype); \
OBJ_RELEASE(sendreq->req_send.req_base.req_comm); \
OMPI_REQUEST_FINI(&sendreq->req_send.req_base.req_ompi); \
opal_convertor_cleanup( &(sendreq->req_send.req_base.req_convertor) ); \
opal_free_list_return ( &mca_pml_base_send_requests, \
(opal_free_list_item_t*)sendreq); \
}
/*
* The PML has completed a send request. Note that this request
* may have been orphaned by the user or have already completed
* at the MPI level.
* This macro will never be called directly from the upper level, as it should
* only be an internal call to the PML.
*/
#define MCA_PML_CM_THIN_SEND_REQUEST_PML_COMPLETE(sendreq) \
do { \
assert( false == sendreq->req_send.req_base.req_pml_complete ); \
\
if( !REQUEST_COMPLETE(&sendreq->req_send.req_base.req_ompi)) { \
/* Should only be called for long messages (maybe synchronous) */ \
ompi_request_complete(&(sendreq->req_send.req_base.req_ompi), true); \
} \
sendreq->req_send.req_base.req_pml_complete = true; \
\
if( sendreq->req_send.req_base.req_free_called ) { \
MCA_PML_CM_THIN_SEND_REQUEST_RETURN( sendreq ); \
} \
} while (0)
/*
* Release resources associated with a request
*/
#define MCA_PML_CM_THIN_SEND_REQUEST_RETURN(sendreq) \
{ \
/* Let the base handle the reference counts */ \
OMPI_DATATYPE_RETAIN(sendreq->req_send.req_base.req_datatype); \
OBJ_RELEASE(sendreq->req_send.req_base.req_comm); \
OMPI_REQUEST_FINI(&sendreq->req_send.req_base.req_ompi); \
opal_convertor_cleanup( &(sendreq->req_send.req_base.req_convertor) ); \
opal_free_list_return ( &mca_pml_base_send_requests, \
(opal_free_list_item_t*)sendreq); \
}
extern void
mca_pml_cm_send_request_completion(struct mca_mtl_request_t *mtl_request);
#endif