1
1

reorg to support new pml - move ptl/teg specific data

from pml_base_* data structures
*** NOTE ***
this will break gm/ib - I will fix these shortly

This commit was SVN r5655.
Этот коммит содержится в:
Tim Woodall 2005-05-09 19:37:10 +00:00
родитель ac60eac8f3
Коммит 3d3151abb3
65 изменённых файлов: 3296 добавлений и 334 удалений

Просмотреть файл

@ -972,11 +972,6 @@ int ompi_comm_determine_first ( ompi_communicator_t *intercomm, int high )
/********************************************************************************/ /********************************************************************************/
int ompi_comm_dump ( ompi_communicator_t *comm ) int ompi_comm_dump ( ompi_communicator_t *comm )
{ {
mca_pml_ptl_comm_t *pml_comm;
mca_ptl_sequence_t *seq;
ompi_list_t *list;
int i;
ompi_output(0, "Dumping information for comm_cid %d\n", comm->c_contextid); ompi_output(0, "Dumping information for comm_cid %d\n", comm->c_contextid);
ompi_output(0," f2c index:%d cube_dim: %d\n", comm->c_f_to_c_index, ompi_output(0," f2c index:%d cube_dim: %d\n", comm->c_f_to_c_index,
comm->c_cube_dim); comm->c_cube_dim);
@ -997,23 +992,6 @@ int ompi_comm_dump ( ompi_communicator_t *comm )
if (OMPI_COMM_IS_INTER(comm)) { if (OMPI_COMM_IS_INTER(comm)) {
ompi_output(0," Remote group size:%d\n", comm->c_remote_group->grp_proc_count); ompi_output(0," Remote group size:%d\n", comm->c_remote_group->grp_proc_count);
} }
/* Dump the c_pml_comm->c_unexpexted_frags for each process */
pml_comm = (mca_pml_ptl_comm_t *)comm->c_pml_comm;
seq = (mca_ptl_sequence_t *) pml_comm->c_frags_cant_match;
for ( i = 0; i < comm->c_local_group->grp_proc_count; i++ ){
list = (ompi_list_t *)seq+i;
ompi_output(0,"%d: head->list_next:%p head->list_prev:%p"
" tail->list_next:%p tail->list_next:%p\n",
i,
(char*)list->ompi_list_head.ompi_list_next,
(char*)list->ompi_list_head.ompi_list_prev,
(char*)list->ompi_list_tail.ompi_list_next,
(char*)list->ompi_list_tail.ompi_list_prev );
fflush(stdout);
}
return MPI_SUCCESS; return MPI_SUCCESS;
} }
/********************************************************************************/ /********************************************************************************/

Просмотреть файл

@ -272,6 +272,7 @@ static void ompi_comm_construct(ompi_communicator_t* comm)
comm->c_remote_group = NULL; comm->c_remote_group = NULL;
comm->error_handler = NULL; comm->error_handler = NULL;
comm->c_pml_comm = NULL; comm->c_pml_comm = NULL;
comm->c_pml_procs = NULL;
comm->c_topo = NULL; comm->c_topo = NULL;
comm->c_topo_component = NULL; comm->c_topo_component = NULL;
comm->c_topo_comm = NULL; comm->c_topo_comm = NULL;

Просмотреть файл

@ -131,8 +131,8 @@ struct ompi_communicator_t {
ompi_errhandler_type_t errhandler_type; ompi_errhandler_type_t errhandler_type;
/* Hooks for PML to hang things */ /* Hooks for PML to hang things */
struct mca_pml_comm_t *c_pml_comm; struct mca_pml_comm_t *c_pml_comm;
struct mca_pml_proc_t **c_pml_procs;
mca_coll_base_module_1_0_0_t c_coll; mca_coll_base_module_1_0_0_t c_coll;
/**< Selected collective module, saved by value for speed (instead /**< Selected collective module, saved by value for speed (instead

Просмотреть файл

@ -30,8 +30,6 @@ OMPI_DECLSPEC extern ompi_class_t mca_pml_base_recv_request_t_class;
struct mca_pml_base_recv_request_t { struct mca_pml_base_recv_request_t {
mca_pml_base_request_t req_base; /**< base request */ mca_pml_base_request_t req_base; /**< base request */
size_t req_bytes_packed; /**< size of message being received */ size_t req_bytes_packed; /**< size of message being received */
size_t req_bytes_received; /**< number of bytes received from network */
size_t req_bytes_delivered; /**< number of bytes delivered to user */
}; };
typedef struct mca_pml_base_recv_request_t mca_pml_base_recv_request_t; typedef struct mca_pml_base_recv_request_t mca_pml_base_recv_request_t;

Просмотреть файл

@ -34,11 +34,11 @@ static void mca_pml_base_send_request_construct(mca_pml_base_send_request_t* req
{ {
/* no need to reinit for every send -- never changes */ /* no need to reinit for every send -- never changes */
request->req_base.req_type = MCA_PML_REQUEST_SEND; request->req_base.req_type = MCA_PML_REQUEST_SEND;
request->req_cached = false;
OBJ_CONSTRUCT(&request->req_convertor, ompi_convertor_t); OBJ_CONSTRUCT(&request->req_convertor, ompi_convertor_t);
} }
static void mca_pml_base_send_request_destruct(mca_pml_base_send_request_t* req) static void mca_pml_base_send_request_destruct(mca_pml_base_send_request_t* req)
{ {
OBJ_DESTRUCT(&req->req_convertor);
} }

Просмотреть файл

@ -35,21 +35,12 @@ OMPI_DECLSPEC extern ompi_class_t mca_pml_base_send_request_t_class;
*/ */
struct mca_pml_base_send_request_t { struct mca_pml_base_send_request_t {
mca_pml_base_request_t req_base; /** base request type - common data structure for use by wait/test */ mca_pml_base_request_t req_base; /** base request type - common data structure for use by wait/test */
ompi_datatype_t* req_datatype; /**< pointer to datatype */
void *req_addr; /**< pointer to send buffer - may not be application buffer */ void *req_addr; /**< pointer to send buffer - may not be application buffer */
size_t req_count; /**< number of elements in send buffer */ size_t req_count; /**< number of elements in send buffer */
ompi_datatype_t* req_datatype; /**< pointer to datatype */
size_t req_offset; /**< number of bytes that have already been assigned to a fragment */
size_t req_bytes_packed; /**< packed size of a message given the datatype and count */ size_t req_bytes_packed; /**< packed size of a message given the datatype and count */
size_t req_bytes_sent; /**< number of bytes that have been sent */
mca_pml_base_send_mode_t req_send_mode; /**< type of send */ mca_pml_base_send_mode_t req_send_mode; /**< type of send */
struct mca_ptl_base_module_t* req_ptl; /**< PTL that is selected for first fragment */
struct mca_ptl_base_peer_t* req_peer; /**< PTL peer instance that will be used for first fragment */
ompi_ptr_t req_peer_match; /**< matched receive at peer */
ompi_ptr_t req_peer_addr; /**< peers remote buffer address */
uint64_t req_peer_size; /**< size of peers remote buffer */
bool req_cached; /**< has this request been obtained from the ptls cache */
ompi_convertor_t req_convertor; /**< convertor that describes this datatype */ ompi_convertor_t req_convertor; /**< convertor that describes this datatype */
volatile int32_t req_lock; /**< lock used by the scheduler */
}; };
typedef struct mca_pml_base_send_request_t mca_pml_base_send_request_t; typedef struct mca_pml_base_send_request_t mca_pml_base_send_request_t;
@ -137,32 +128,6 @@ typedef struct mca_pml_base_send_request_t mca_pml_base_send_request_t;
OBJ_RELEASE(request->req_base.req_datatype); \ OBJ_RELEASE(request->req_base.req_datatype); \
} while (0) } while (0)
/**
* Test to check if an acknowledgment has been received, with the match.
*
* @param request (IN) Send request.
* return TRUE if an ack/match has been received from peer.
*/
static inline bool mca_pml_base_send_request_matched(
mca_pml_base_send_request_t* request)
{
return (NULL != request->req_peer_match.pval);
}
/**
* Atomically increase the request offset.
*
* @param request (IN) Send request.
* @param offset (IN) Increment.
*/
static inline void mca_pml_base_send_request_offset(
mca_pml_base_send_request_t* request,
size_t offset)
{
OMPI_THREAD_ADD32((int32_t*)(&request->req_offset), offset);
}
#if defined(c_plusplus) || defined(__cplusplus) #if defined(c_plusplus) || defined(__cplusplus)
} }

Просмотреть файл

@ -14,34 +14,28 @@
# $HEADER$ # $HEADER$
# #
# Use the top-level Makefile.options
include $(top_ompi_srcdir)/config/Makefile.options include $(top_ompi_srcdir)/config/Makefile.options
sources = noinst_LTLIBRARIES = libmca_pml_teg.la
include src/Makefile.extra libmca_pml_teg_la_SOURCES = \
pml_teg.c \
# Make the output library in this directory, and name it either pml_teg.h \
# mca_<type>_<name>.la (for DSO builds) or libmca_<type>_<name>.la pml_teg_cancel.c \
# (for static builds). pml_teg_component.c \
pml_teg_component.h \
if OMPI_BUILD_pml_teg_DSO pml_teg_iprobe.c \
lib = pml_teg_irecv.c \
lib_sources = pml_teg_isend.c \
component = mca_pml_teg.la pml_teg_ptl.c \
component_sources = $(sources) pml_teg_ptl.h \
else pml_teg_proc.c \
lib = libmca_pml_teg.la pml_teg_proc.h \
lib_sources = $(sources) pml_teg_progress.c \
component = pml_teg_recvfrag.c \
component_sources = pml_teg_recvfrag.h \
endif pml_teg_recvreq.c \
pml_teg_recvreq.h \
mcacomponentdir = $(libdir)/openmpi pml_teg_sendreq.c \
mcacomponent_LTLIBRARIES = $(component) pml_teg_sendreq.h \
mca_pml_teg_la_SOURCES = $(component_sources) pml_teg_start.c \
mca_pml_teg_la_LDFLAGS = -module -avoid-version pml_ptl_array.c
noinst_LTLIBRARIES = $(lib)
libmca_pml_teg_la_SOURCES = $(lib_sources)
libmca_pml_teg_la_LDFLAGS = -module -avoid-version

Просмотреть файл

@ -17,6 +17,6 @@
# Specific to this module # Specific to this module
PARAM_INIT_FILE=src/pml_teg.c PARAM_INIT_FILE=pml_teg.c
PARAM_CONFIG_HEADER_FILE="src/teg_config.h" PARAM_CONFIG_HEADER_FILE="teg_config.h"
PARAM_CONFIG_FILES="Makefile" PARAM_CONFIG_FILES="Makefile"

61
src/mca/pml/teg/pml_ptl_array.c Обычный файл
Просмотреть файл

@ -0,0 +1,61 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <string.h>
#include "mca/pml/pml.h"
#include "pml_ptl_array.h"
static void mca_ptl_array_construct(mca_ptl_array_t* array)
{
array->ptl_procs = 0;
array->ptl_size = 0;
array->ptl_index = 0;
array->ptl_reserve = 0;
}
static void mca_ptl_array_destruct(mca_ptl_array_t* array)
{
if(array->ptl_procs != 0)
free(array->ptl_procs);
}
OBJ_CLASS_INSTANCE(
mca_pml_teg_ptl_array_t,
ompi_object_t,
mca_ptl_array_construct,
mca_ptl_array_destruct
);
int mca_ptl_array_reserve(mca_ptl_array_t* array, size_t size)
{
mca_ptl_proc_t *procs;
if(array->ptl_reserve >= size)
return OMPI_SUCCESS;
procs = (mca_ptl_proc_t *)realloc(array->ptl_procs, sizeof(mca_ptl_proc_t)*size);
if(NULL == procs)
return OMPI_ERR_OUT_OF_RESOURCE;
array->ptl_procs = procs;
array->ptl_reserve = size;
memset(array->ptl_procs+array->ptl_size, 0, (size-array->ptl_size)*sizeof(mca_ptl_proc_t));
return OMPI_SUCCESS;
}

146
src/mca/pml/teg/pml_ptl_array.h Обычный файл
Просмотреть файл

@ -0,0 +1,146 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PTL_ARRAY_H
#define OMPI_PTL_ARRAY_H
#include "util/output.h"
#include "mca/ptl/ptl.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
extern ompi_class_t mca_pml_teg_ptl_array_t_class;
/**
* A data structure associated with a ompi_proc_t that caches
* addressing/scheduling attributes for a specific PTL instance
* that can be used to reach the process.
*/
struct mca_ptl_proc_t {
int ptl_weight; /**< PTL weight for scheduling */
struct mca_ptl_base_peer_t* ptl_peer; /**< PTL addressing info */
struct mca_pml_base_ptl_t* ptl_base; /**< PML specific PTL info */
mca_ptl_base_module_t *ptl; /**< PTL module */
};
typedef struct mca_ptl_proc_t mca_ptl_proc_t;
/**
* A dynamically growable array of mca_ptl_proc_t instances.
* Maintains an index into the array that is used for round-robin
* scheduling across contents.
*/
struct mca_ptl_array_t {
ompi_object_t super;
mca_ptl_proc_t* ptl_procs; /**< array of ptl procs */
size_t ptl_size; /**< number available */
size_t ptl_reserve; /**< size of allocated ptl_proc array */
size_t ptl_index; /**< last used index*/
};
typedef struct mca_ptl_array_t mca_ptl_array_t;
typedef struct mca_ptl_array_t mca_pml_teg_ptl_array_t;
/**
* If required, reallocate (grow) the array to the indicate size.
*
* @param array (IN)
* @param size (IN)
*/
int mca_ptl_array_reserve(mca_ptl_array_t*, size_t);
static inline size_t mca_ptl_array_get_size(mca_ptl_array_t* array)
{
return array->ptl_size;
}
/**
* Grow the array if required, and set the size.
*
* @param array (IN)
* @param size (IN)
*/
static inline void mca_ptl_array_set_size(mca_ptl_array_t* array, size_t size)
{
if(array->ptl_size > array->ptl_reserve)
mca_ptl_array_reserve(array, size);
array->ptl_size = size;
}
/**
* Grow the array size by one and return the item at that index.
*
* @param array (IN)
*/
static inline mca_ptl_proc_t* mca_ptl_array_insert(mca_ptl_array_t* array)
{
#if OMPI_ENABLE_DEBUG
if(array->ptl_size >= array->ptl_reserve) {
ompi_output(0, "mca_ptl_array_insert: invalid array index %d >= %d",
array->ptl_size, array->ptl_reserve);
return 0;
}
#endif
return &array->ptl_procs[array->ptl_size++];
}
/**
* Return an array item at the specified index.
*
* @param array (IN)
* @param index (IN)
*/
static inline mca_ptl_proc_t* mca_ptl_array_get_index(mca_ptl_array_t* array, size_t index)
{
#if OMPI_ENABLE_DEBUG
if(index >= array->ptl_size) {
ompi_output(0, "mca_ptl_array_get_index: invalid array index %d >= %d",
index, array->ptl_size);
return 0;
}
#endif
return &array->ptl_procs[index];
}
/**
* Return the next LRU index in the array.
*
* @param array (IN)
* @param index (IN)
*/
static inline mca_ptl_proc_t* mca_ptl_array_get_next(mca_ptl_array_t* array)
{
mca_ptl_proc_t* ptl_proc;
#if OMPI_ENABLE_DEBUG
if(array->ptl_size == 0) {
ompi_output(0, "mca_ptl_array_get_next: invalid array size");
return 0;
}
#endif
ptl_proc = &array->ptl_procs[array->ptl_index++];
if(array->ptl_index == array->ptl_size)
array->ptl_index = 0;
return ptl_proc;
}
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

436
src/mca/pml/teg/pml_teg.c Обычный файл
Просмотреть файл

@ -0,0 +1,436 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <stdlib.h>
#include <string.h>
#include "class/ompi_bitmap.h"
#include "mca/pml/pml.h"
#include "mca/ptl/ptl.h"
#include "mca/ptl/base/base.h"
#include "mca/ptl/base/ptl_base_comm.h"
#include "mca/ptl/base/ptl_base_header.h"
#include "mca/ptl/base/ptl_base_recvfrag.h"
#include "mca/ptl/base/ptl_base_sendfrag.h"
#include "pml_teg.h"
#include "pml_teg_component.h"
#include "pml_teg_proc.h"
#include "pml_teg_ptl.h"
#include "pml_teg_recvreq.h"
#include "pml_teg_sendreq.h"
#include "pml_teg_recvfrag.h"
mca_pml_teg_t mca_pml_teg = {
{
mca_pml_teg_add_procs,
mca_pml_teg_del_procs,
mca_pml_teg_add_ptls,
mca_pml_teg_control,
mca_pml_teg_progress,
mca_pml_teg_add_comm,
mca_pml_teg_del_comm,
mca_pml_teg_irecv_init,
mca_pml_teg_irecv,
mca_pml_teg_recv,
mca_pml_teg_isend_init,
mca_pml_teg_isend,
mca_pml_teg_send,
mca_pml_teg_iprobe,
mca_pml_teg_probe,
mca_pml_teg_start
}
};
int mca_pml_teg_add_comm(ompi_communicator_t* comm)
{
/* allocate pml specific comm data */
mca_pml_ptl_comm_t* pml_comm = OBJ_NEW(mca_pml_ptl_comm_t);
if (NULL == pml_comm) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
mca_pml_ptl_comm_init_size(pml_comm, comm->c_remote_group->grp_proc_count);
comm->c_pml_comm = pml_comm;
return OMPI_SUCCESS;
}
int mca_pml_teg_del_comm(ompi_communicator_t* comm)
{
OBJ_RELEASE(comm->c_pml_comm);
comm->c_pml_comm = 0;
return OMPI_SUCCESS;
}
static int ptl_exclusivity_compare(const void* arg1, const void* arg2)
{
mca_ptl_base_module_t* ptl1 = *(struct mca_ptl_base_module_t**)arg1;
mca_ptl_base_module_t* ptl2 = *(struct mca_ptl_base_module_t**)arg2;
if( ptl1->ptl_exclusivity > ptl2->ptl_exclusivity ) {
return -1;
} else if (ptl1->ptl_exclusivity == ptl2->ptl_exclusivity ) {
return 0;
} else {
return 1;
}
}
int mca_pml_teg_add_ptls(ompi_list_t *ptls)
{
/* build an array of ptls and ptl modules */
mca_ptl_base_selected_module_t* selected_ptl;
size_t num_ptls = ompi_list_get_size(ptls);
size_t cache_bytes = 0;
mca_pml_teg.teg_num_ptl_modules = 0;
mca_pml_teg.teg_num_ptl_progress = 0;
mca_pml_teg.teg_num_ptl_components = 0;
mca_pml_teg.teg_ptl_modules = (mca_ptl_base_module_t **)malloc(sizeof(mca_ptl_base_module_t*) * num_ptls);
mca_pml_teg.teg_ptl_progress = (mca_ptl_base_component_progress_fn_t*)malloc(sizeof(mca_ptl_base_component_progress_fn_t) * num_ptls);
mca_pml_teg.teg_ptl_components = (mca_ptl_base_component_t **)malloc(sizeof(mca_ptl_base_component_t*) * num_ptls);
if (NULL == mca_pml_teg.teg_ptl_modules ||
NULL == mca_pml_teg.teg_ptl_progress ||
NULL == mca_pml_teg.teg_ptl_components) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for(selected_ptl = (mca_ptl_base_selected_module_t*)ompi_list_get_first(ptls);
selected_ptl != (mca_ptl_base_selected_module_t*)ompi_list_get_end(ptls);
selected_ptl = (mca_ptl_base_selected_module_t*)ompi_list_get_next(selected_ptl)) {
mca_ptl_base_module_t *ptl = selected_ptl->pbsm_module;
size_t i;
mca_pml_teg.teg_ptl_modules[mca_pml_teg.teg_num_ptl_modules++] = ptl;
for(i=0; i<mca_pml_teg.teg_num_ptl_components; i++) {
if(mca_pml_teg.teg_ptl_components[i] == ptl->ptl_component) {
break;
}
}
if(i == mca_pml_teg.teg_num_ptl_components) {
mca_pml_teg.teg_ptl_components[mca_pml_teg.teg_num_ptl_components++] = ptl->ptl_component;
}
/*
*setup ptl
*/
/* set pointer to fragment matching logic routine, if this
* not already set by the ptl */
if( NULL == ptl->ptl_match)
ptl->ptl_match = mca_pml_teg_recv_frag_match;
ptl->ptl_send_progress = mca_pml_teg_send_request_progress;
ptl->ptl_recv_progress = mca_pml_teg_recv_request_progress;
ptl->ptl_stack = ptl;
ptl->ptl_base = NULL;
/* find maximum required size for cache */
if(ptl->ptl_cache_bytes > cache_bytes) {
cache_bytes = ptl->ptl_cache_bytes;
}
}
/* setup send fragments based on largest required send request */
ompi_free_list_init(
&mca_pml_teg.teg_send_requests,
sizeof(mca_pml_teg_send_request_t) + cache_bytes,
OBJ_CLASS(mca_pml_teg_send_request_t),
mca_pml_teg.teg_free_list_num,
mca_pml_teg.teg_free_list_max,
mca_pml_teg.teg_free_list_inc,
NULL);
/* sort ptl list by exclusivity */
qsort(mca_pml_teg.teg_ptl_modules, mca_pml_teg.teg_num_ptl_modules, sizeof(struct mca_ptl_t*), ptl_exclusivity_compare);
return OMPI_SUCCESS;
}
/*
* Pass control information through to all PTL modules.
*/
int mca_pml_teg_control(int param, void* value, size_t size)
{
size_t i=0;
for(i=0; i<mca_pml_teg.teg_num_ptl_components; i++) {
if(NULL != mca_pml_teg.teg_ptl_components[i]->ptlm_control) {
int rc = mca_pml_teg.teg_ptl_components[i]->ptlm_control(param,value,size);
if(rc != OMPI_SUCCESS)
return rc;
}
}
return OMPI_SUCCESS;
}
/*
* For each proc setup a datastructure that indicates the PTLs
* that can be used to reach the destination.
*
*/
int mca_pml_teg_add_procs(ompi_proc_t** procs, size_t nprocs)
{
size_t p;
ompi_bitmap_t reachable;
struct mca_ptl_base_peer_t** ptl_peers = NULL;
int rc;
size_t p_index;
if(nprocs == 0)
return OMPI_SUCCESS;
OBJ_CONSTRUCT(&reachable, ompi_bitmap_t);
rc = ompi_bitmap_init(&reachable, nprocs);
if(OMPI_SUCCESS != rc)
return rc;
/* iterate through each of the procs and set the peers architecture */
for(p=0; p<nprocs; p++) {
uint32_t* proc_arch;
size_t size = sizeof(uint32_t);
rc = mca_base_modex_recv(&mca_pml_teg_component.pmlm_version, procs[p],
(void**)&proc_arch, &size);
if(rc != OMPI_SUCCESS)
return rc;
if(size != sizeof(uint32_t))
return OMPI_ERROR;
procs[p]->proc_arch = ntohl(*proc_arch);
free(proc_arch);
}
/* attempt to add all procs to each ptl */
ptl_peers = (struct mca_ptl_base_peer_t **)malloc(nprocs * sizeof(struct mca_ptl_base_peer_t*));
for(p_index = 0; p_index < mca_pml_teg.teg_num_ptl_modules; p_index++) {
mca_ptl_base_module_t* ptl = mca_pml_teg.teg_ptl_modules[p_index];
int ptl_inuse = 0;
/* if the ptl can reach the destination proc it sets the
* corresponding bit (proc index) in the reachable bitmap
* and can return addressing information for each proc
* that is passed back to the ptl on data transfer calls
*/
ompi_bitmap_clear_all_bits(&reachable);
memset(ptl_peers, 0, nprocs * sizeof(struct mca_ptl_base_peer_t*));
rc = ptl->ptl_add_procs(ptl, nprocs, procs, ptl_peers, &reachable);
if(OMPI_SUCCESS != rc) {
free(ptl_peers);
return rc;
}
/* for each proc that is reachable - add the ptl to the procs array(s) */
for(p=0; p<nprocs; p++) {
if(ompi_bitmap_is_set_bit(&reachable, p)) {
ompi_proc_t *proc = procs[p];
mca_pml_proc_t* proc_pml = proc->proc_pml;
mca_ptl_proc_t* proc_ptl;
size_t size;
/* this ptl can be used */
ptl_inuse++;
/* initialize each proc */
if(NULL == proc_pml) {
/* allocate pml specific proc data */
proc_pml = OBJ_NEW(mca_pml_teg_proc_t);
if (NULL == proc_pml) {
ompi_output(0, "mca_pml_teg_add_procs: unable to allocate resources");
free(ptl_peers);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* preallocate space in array for max number of ptls */
mca_ptl_array_reserve(&proc_pml->proc_ptl_first, mca_pml_teg.teg_num_ptl_modules);
mca_ptl_array_reserve(&proc_pml->proc_ptl_next, mca_pml_teg.teg_num_ptl_modules);
proc_pml->proc_ompi = proc;
proc->proc_pml = proc_pml;
}
/* dont allow an additional PTL with a lower exclusivity ranking */
size = mca_ptl_array_get_size(&proc_pml->proc_ptl_next);
if(size > 0) {
proc_ptl = mca_ptl_array_get_index(&proc_pml->proc_ptl_next, size-1);
/* skip this ptl if the exclusivity is less than the previous */
if(proc_ptl->ptl->ptl_exclusivity > ptl->ptl_exclusivity) {
if(ptl_peers[p] != NULL) {
ptl->ptl_del_procs(ptl, 1, &proc, &ptl_peers[p]);
}
continue;
}
}
/* cache the ptl on the proc */
proc_ptl = mca_ptl_array_insert(&proc_pml->proc_ptl_next);
proc_ptl->ptl = ptl;
proc_ptl->ptl_peer = ptl_peers[p];
proc_ptl->ptl_weight = 0;
proc_pml->proc_ptl_flags |= ptl->ptl_flags;
}
}
if(ptl_inuse > 0 && NULL != ptl->ptl_component->ptlm_progress) {
size_t p;
bool found = false;
for(p=0; p<mca_pml_teg.teg_num_ptl_progress; p++) {
if(mca_pml_teg.teg_ptl_progress[p] == ptl->ptl_component->ptlm_progress) {
found = true;
break;
}
}
if(found == false) {
mca_pml_teg.teg_ptl_progress[mca_pml_teg.teg_num_ptl_progress] =
ptl->ptl_component->ptlm_progress;
mca_pml_teg.teg_num_ptl_progress++;
}
}
}
free(ptl_peers);
/* iterate back through procs and compute metrics for registered ptls */
for(p=0; p<nprocs; p++) {
ompi_proc_t *proc = procs[p];
mca_pml_proc_t* proc_pml = proc->proc_pml;
double total_bandwidth = 0;
uint32_t latency = 0;
size_t n_index;
size_t n_size;
/* skip over procs w/ no ptls registered */
if(NULL == proc_pml)
continue;
/* (1) determine the total bandwidth available across all ptls
* note that we need to do this here, as we may already have ptls configured
* (2) determine the highest priority ranking for latency
*/
n_size = mca_ptl_array_get_size(&proc_pml->proc_ptl_next);
for(n_index = 0; n_index < n_size; n_index++) {
struct mca_ptl_proc_t* proc_ptl = mca_ptl_array_get_index(&proc_pml->proc_ptl_next, n_index);
struct mca_ptl_base_module_t* ptl = proc_ptl->ptl;
total_bandwidth += proc_ptl->ptl->ptl_bandwidth;
if(ptl->ptl_latency > latency)
latency = ptl->ptl_latency;
}
/* (1) set the weight of each ptl as a percentage of overall bandwidth
* (2) copy all ptl instances at the highest priority ranking into the
* list of ptls used for first fragments
*/
for(n_index = 0; n_index < n_size; n_index++) {
struct mca_ptl_proc_t* proc_ptl = mca_ptl_array_get_index(&proc_pml->proc_ptl_next, n_index);
struct mca_ptl_base_module_t *ptl = proc_ptl->ptl;
double weight;
/* compute weighting factor for this ptl */
if(ptl->ptl_bandwidth)
weight = proc_ptl->ptl->ptl_bandwidth / total_bandwidth;
else
weight = 1.0 / n_size;
proc_ptl->ptl_weight = (int)(weight * 100);
/*
* save/create ptl extension for use by pml
*/
proc_ptl->ptl_base = ptl->ptl_base;
if (NULL == proc_ptl->ptl_base &&
ptl->ptl_cache_bytes > 0 &&
NULL != ptl->ptl_request_init &&
NULL != ptl->ptl_request_fini) {
mca_pml_base_ptl_t* ptl_base = OBJ_NEW(mca_pml_base_ptl_t);
ptl_base->ptl = ptl;
ptl_base->ptl_cache_size = ptl->ptl_cache_size;
proc_ptl->ptl_base = ptl->ptl_base = ptl_base;
}
/* check to see if this ptl is already in the array of ptls used for first
* fragments - if not add it.
*/
if(ptl->ptl_latency == latency) {
struct mca_ptl_proc_t* proc_new = mca_ptl_array_insert(&proc_pml->proc_ptl_first);
*proc_new = *proc_ptl;
}
}
}
return OMPI_SUCCESS;
}
/*
* iterate through each proc and notify any PTLs associated
* with the proc that it is/has gone away
*/
int mca_pml_teg_del_procs(ompi_proc_t** procs, size_t nprocs)
{
size_t p;
int rc;
for(p = 0; p < nprocs; p++) {
ompi_proc_t *proc = procs[p];
mca_pml_proc_t* proc_pml = proc->proc_pml;
size_t f_index, f_size;
size_t n_index, n_size;
/* notify each ptl that the proc is going away */
f_size = mca_ptl_array_get_size(&proc_pml->proc_ptl_first);
for(f_index = 0; f_index < f_size; f_index++) {
mca_ptl_proc_t* ptl_proc = mca_ptl_array_get_index(&proc_pml->proc_ptl_first, f_index);
mca_ptl_base_module_t* ptl = ptl_proc->ptl;
rc = ptl->ptl_del_procs(ptl,1,&proc,&ptl_proc->ptl_peer);
if(OMPI_SUCCESS != rc) {
return rc;
}
/* remove this from next array so that we dont call it twice w/
* the same address pointer
*/
n_size = mca_ptl_array_get_size(&proc_pml->proc_ptl_first);
for(n_index = 0; n_index < n_size; n_index++) {
mca_ptl_proc_t* next_proc = mca_ptl_array_get_index(&proc_pml->proc_ptl_next, n_index);
if(next_proc->ptl == ptl) {
memset(next_proc, 0, sizeof(mca_ptl_proc_t));
break;
}
}
}
/* notify each ptl that was not in the array of ptls for first fragments */
n_size = mca_ptl_array_get_size(&proc_pml->proc_ptl_next);
for(n_index = 0; n_index < n_size; n_index++) {
mca_ptl_proc_t* ptl_proc = mca_ptl_array_get_index(&proc_pml->proc_ptl_first, n_index);
mca_ptl_base_module_t* ptl = ptl_proc->ptl;
if (ptl != 0) {
rc = ptl->ptl_del_procs(ptl,1,&proc,&ptl_proc->ptl_peer);
if(OMPI_SUCCESS != rc)
return rc;
}
}
/* do any required cleanup */
OBJ_RELEASE(proc_pml);
proc->proc_pml = NULL;
}
return OMPI_SUCCESS;
}
int mca_pml_teg_component_fini(void)
{
/* FIX */
return OMPI_SUCCESS;
}

268
src/mca/pml/teg/pml_teg.h Обычный файл
Просмотреть файл

@ -0,0 +1,268 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PML_TEG_H
#define MCA_PML_TEG_H
#include "threads/thread.h"
#include "threads/condition.h"
#include "class/ompi_free_list.h"
#include "util/cmd_line.h"
#include "request/request.h"
#include "mca/pml/pml.h"
#include "mca/pml/base/pml_base_request.h"
#include "mca/pml/base/pml_base_bsend.h"
#include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/ptl/ptl.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* TEG PML module
*/
struct mca_pml_teg_t {
mca_pml_base_module_t super;
mca_ptl_base_component_t **teg_ptl_components;
size_t teg_num_ptl_components;
mca_ptl_base_module_t** teg_ptl_modules;
size_t teg_num_ptl_modules;
mca_ptl_base_component_progress_fn_t* teg_ptl_progress;
size_t teg_num_ptl_progress;
ompi_list_t teg_procs;
ompi_mutex_t teg_lock;
int teg_priority;
int teg_free_list_num; /* initial size of free list */
int teg_free_list_max; /* maximum size of free list */
int teg_free_list_inc; /* number of elements to grow free list */
int teg_poll_iterations; /* number of iterations to poll for completion */
/* free list of requests */
ompi_free_list_t teg_send_requests;
ompi_free_list_t teg_recv_requests;
/* list of pending send requests */
ompi_list_t teg_send_pending;
};
typedef struct mca_pml_teg_t mca_pml_teg_t;
extern mca_pml_teg_t mca_pml_teg;
/*
* PML module functions.
*/
extern int mca_pml_teg_component_open(void);
extern int mca_pml_teg_component_close(void);
extern mca_pml_base_module_t* mca_pml_teg_component_init(
int *priority,
bool enable_progress_threads,
bool enable_mpi_threads
);
extern int mca_pml_teg_component_fini(void);
/*
* PML interface functions.
*/
extern int mca_pml_teg_add_comm(
struct ompi_communicator_t* comm
);
extern int mca_pml_teg_del_comm(
struct ompi_communicator_t* comm
);
extern int mca_pml_teg_add_procs(
struct ompi_proc_t **procs,
size_t nprocs
);
extern int mca_pml_teg_del_procs(
struct ompi_proc_t **procs,
size_t nprocs
);
extern int mca_pml_teg_add_ptls(
ompi_list_t *ptls
);
extern int mca_pml_teg_control(
int param,
void *size,
size_t value
);
extern int mca_pml_teg_progress(void);
extern int mca_pml_teg_iprobe(
int dst,
int tag,
struct ompi_communicator_t* comm,
int *matched,
ompi_status_public_t* status
);
extern int mca_pml_teg_probe(
int dst,
int tag,
struct ompi_communicator_t* comm,
ompi_status_public_t* status
);
extern int mca_pml_teg_cancelled(
ompi_request_t* request,
int *flag
);
extern int mca_pml_teg_isend_init(
void *buf,
size_t count,
ompi_datatype_t *datatype,
int dst,
int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
struct ompi_request_t **request
);
extern int mca_pml_teg_isend(
void *buf,
size_t count,
ompi_datatype_t *datatype,
int dst,
int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
struct ompi_request_t **request
);
extern int mca_pml_teg_send(
void *buf,
size_t count,
ompi_datatype_t *datatype,
int dst,
int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm
);
extern int mca_pml_teg_irecv_init(
void *buf,
size_t count,
ompi_datatype_t *datatype,
int src,
int tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request
);
extern int mca_pml_teg_irecv(
void *buf,
size_t count,
ompi_datatype_t *datatype,
int src,
int tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request
);
extern int mca_pml_teg_recv(
void *buf,
size_t count,
ompi_datatype_t *datatype,
int src,
int tag,
struct ompi_communicator_t* comm,
ompi_status_public_t* status
);
extern int mca_pml_teg_progress(void);
extern int mca_pml_teg_start(
size_t count,
ompi_request_t** requests
);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#define MCA_PML_TEG_FINI(request) \
{ \
mca_pml_base_request_t* pml_request = *(mca_pml_base_request_t**)(request); \
if(pml_request->req_persistent) { \
if(pml_request->req_free_called) { \
MCA_PML_TEG_FREE(request); \
} else { \
pml_request->req_ompi.req_state = OMPI_REQUEST_INACTIVE; \
} \
} else { \
MCA_PML_TEG_FREE(request); \
} \
}
#define MCA_PML_TEG_FREE(request) \
{ \
mca_pml_base_request_t* pml_request = *(mca_pml_base_request_t**)(request); \
pml_request->req_free_called = true; \
if( pml_request->req_pml_complete == true) \
{ \
OMPI_REQUEST_FINI(*(request)); \
switch(pml_request->req_type) { \
case MCA_PML_REQUEST_SEND: \
{ \
mca_pml_teg_send_request_t* sendreq = (mca_pml_teg_send_request_t*)pml_request; \
while(sendreq->req_lock > 0); \
if(sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) { \
mca_pml_base_bsend_request_fini((ompi_request_t*)sendreq); \
} \
MCA_PML_TEG_SEND_REQUEST_RETURN(sendreq); \
break; \
} \
case MCA_PML_REQUEST_RECV: \
{ \
mca_pml_teg_recv_request_t* recvreq = (mca_pml_teg_recv_request_t*)pml_request; \
MCA_PML_TEG_RECV_REQUEST_RETURN(recvreq); \
break; \
} \
default: \
break; \
} \
} \
*(request) = MPI_REQUEST_NULL; \
}
#endif

28
src/mca/pml/teg/pml_teg_cancel.c Обычный файл
Просмотреть файл

@ -0,0 +1,28 @@
/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "pml_teg.h"
int mca_pml_teg_cancelled(ompi_request_t* request, int* flag)
{
if(NULL != flag)
*flag = (true == request->req_status._cancelled ? 1 : 0);
return OMPI_SUCCESS;
}

177
src/mca/pml/teg/pml_teg_component.c Обычный файл
Просмотреть файл

@ -0,0 +1,177 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "event/event.h"
#include "mpi.h"
#include "mca/pml/pml.h"
#include "mca/ptl/ptl.h"
#include "mca/base/mca_base_param.h"
#include "mca/pml/base/pml_base_bsend.h"
#include "pml_teg.h"
#include "pml_teg_proc.h"
#include "pml_teg_sendreq.h"
#include "pml_teg_recvreq.h"
mca_pml_base_component_1_0_0_t mca_pml_teg_component = {
/* First, the mca_base_component_t struct containing meta
information about the component itself */
{
/* Indicate that we are a pml v1.0.0 component (which also implies
a specific MCA version) */
MCA_PML_BASE_VERSION_1_0_0,
"teg", /* MCA component name */
1, /* MCA component major version */
0, /* MCA component minor version */
0, /* MCA component release version */
mca_pml_teg_component_open, /* component open */
mca_pml_teg_component_close /* component close */
},
/* Next the MCA v1.0.0 component meta data */
{
/* Whether the component is checkpointable or not */
false
},
mca_pml_teg_component_init, /* component init */
mca_pml_teg_component_fini /* component finalize */
};
static inline int mca_pml_teg_param_register_int(
const char* param_name,
int default_value)
{
int id = mca_base_param_register_int("pml","teg",param_name,NULL,default_value);
int param_value = default_value;
mca_base_param_lookup_int(id,&param_value);
return param_value;
}
int mca_pml_teg_component_open(void)
{
#ifdef WIN32
WSADATA win_sock_data;
if (WSAStartup(MAKEWORD(2,2), &win_sock_data) != 0) {
ompi_output (0, "mca_oob_tcp_component_init: failed to initialise windows sockets: %d\n", WSAGetLastError());
return OMPI_ERROR;
}
#endif
OBJ_CONSTRUCT(&mca_pml_teg.teg_lock, ompi_mutex_t);
OBJ_CONSTRUCT(&mca_pml_teg.teg_send_requests, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_teg.teg_recv_requests, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_teg.teg_procs, ompi_list_t);
OBJ_CONSTRUCT(&mca_pml_teg.teg_send_pending, ompi_list_t);
mca_pml_teg.teg_ptl_components = NULL;
mca_pml_teg.teg_num_ptl_components = 0;
mca_pml_teg.teg_ptl_modules = NULL;
mca_pml_teg.teg_num_ptl_modules = 0;
mca_pml_teg.teg_ptl_progress = NULL;
mca_pml_teg.teg_num_ptl_progress = 0;
mca_pml_teg.teg_free_list_num =
mca_pml_teg_param_register_int("free_list_num", 256);
mca_pml_teg.teg_free_list_max =
mca_pml_teg_param_register_int("free_list_max", -1);
mca_pml_teg.teg_free_list_inc =
mca_pml_teg_param_register_int("free_list_inc", 256);
mca_pml_teg.teg_poll_iterations =
mca_pml_teg_param_register_int("poll_iterations", 100000);
mca_pml_teg.teg_priority =
mca_pml_teg_param_register_int("priority", 0);
return OMPI_SUCCESS;
}
int mca_pml_teg_component_close(void)
{
#ifdef WIN32
WSACleanup();
#endif
#if OMPI_ENABLE_DEBUG
if (mca_pml_teg.teg_recv_requests.fl_num_allocated !=
mca_pml_teg.teg_recv_requests.super.ompi_list_length) {
ompi_output(0, "teg recv requests: %d allocated %d returned\n",
mca_pml_teg.teg_recv_requests.fl_num_allocated,
mca_pml_teg.teg_recv_requests.super.ompi_list_length);
}
#endif
if(NULL != mca_pml_teg.teg_ptl_components) {
free(mca_pml_teg.teg_ptl_components);
}
if(NULL != mca_pml_teg.teg_ptl_modules) {
free(mca_pml_teg.teg_ptl_modules);
}
if(NULL != mca_pml_teg.teg_ptl_progress) {
free(mca_pml_teg.teg_ptl_progress);
}
OBJ_DESTRUCT(&mca_pml_teg.teg_send_pending);
OBJ_DESTRUCT(&mca_pml_teg.teg_send_requests);
OBJ_DESTRUCT(&mca_pml_teg.teg_recv_requests);
OBJ_DESTRUCT(&mca_pml_teg.teg_procs);
OBJ_DESTRUCT(&mca_pml_teg.teg_lock);
return OMPI_SUCCESS;
}
mca_pml_base_module_t* mca_pml_teg_component_init(int* priority,
bool enable_progress_threads,
bool enable_mpi_threads)
{
uint32_t proc_arch;
int rc;
*priority = mca_pml_teg.teg_priority;
/* recv requests */
ompi_free_list_init(
&mca_pml_teg.teg_recv_requests,
sizeof(mca_pml_teg_recv_request_t),
OBJ_CLASS(mca_pml_teg_recv_request_t),
mca_pml_teg.teg_free_list_num,
mca_pml_teg.teg_free_list_max,
mca_pml_teg.teg_free_list_inc,
NULL);
/* buffered send */
if(OMPI_SUCCESS != mca_pml_base_bsend_init(enable_mpi_threads)) {
ompi_output(0, "mca_pml_teg_component_init: mca_pml_bsend_init failed\n");
return NULL;
}
/* post this processes datatype */
proc_arch = ompi_proc_local()->proc_arch;
proc_arch = htonl(proc_arch);
rc = mca_base_modex_send(&mca_pml_teg_component.pmlm_version, &proc_arch, sizeof(proc_arch));
if(rc != OMPI_SUCCESS)
return NULL;
return &mca_pml_teg.super;
}

27
src/mca/pml/teg/pml_teg_component.h Обычный файл
Просмотреть файл

@ -0,0 +1,27 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PML_TEG_COMPONENT_H
#define MCA_PML_TEG_COMPONENT_H
/*
* PML module functions.
*/
OMPI_COMP_EXPORT extern mca_pml_base_component_1_0_0_t mca_pml_teg_component;
#endif

93
src/mca/pml/teg/pml_teg_iprobe.c Обычный файл
Просмотреть файл

@ -0,0 +1,93 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "request/request.h"
#include "pml_teg_recvreq.h"
int mca_pml_teg_iprobe(int src,
int tag,
struct ompi_communicator_t *comm,
int *matched, ompi_status_public_t * status)
{
int rc;
mca_ptl_base_recv_request_t recvreq;
OBJ_CONSTRUCT( &(recvreq), mca_ptl_base_recv_request_t );
recvreq.req_recv.req_base.req_ompi.req_type = OMPI_REQUEST_PML;
recvreq.req_recv.req_base.req_type = MCA_PML_REQUEST_IPROBE;
MCA_PML_TEG_RECV_REQUEST_INIT(&recvreq, NULL, 0, &ompi_mpi_char, src, tag, comm, true);
*matched = 0;
if ((rc = mca_pml_teg_recv_request_start(&recvreq)) == OMPI_SUCCESS) {
if( recvreq.req_recv.req_base.req_ompi.req_complete == true ) {
if( NULL != status ) {
*status = recvreq.req_recv.req_base.req_ompi.req_status;
}
*matched = 1;
} else {
/* we are supposed to progress ... */
ompi_progress();
}
}
MCA_PML_BASE_RECV_REQUEST_RETURN((&recvreq.req_recv));
return rc;
}
int mca_pml_teg_probe(int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
int rc;
mca_ptl_base_recv_request_t recvreq;
OBJ_CONSTRUCT( &(recvreq), mca_ptl_base_recv_request_t );
recvreq.req_recv.req_base.req_ompi.req_type = OMPI_REQUEST_PML;
recvreq.req_recv.req_base.req_type = MCA_PML_REQUEST_PROBE;
MCA_PML_TEG_RECV_REQUEST_INIT(&recvreq, NULL, 0, &ompi_mpi_char, src, tag, comm, true);
if ((rc = mca_pml_teg_recv_request_start(&recvreq)) != OMPI_SUCCESS) {
MCA_PML_BASE_RECV_REQUEST_RETURN((&recvreq.req_recv));
return rc;
}
if (recvreq.req_recv.req_base.req_ompi.req_complete == false) {
/* give up and sleep until completion */
if (ompi_using_threads()) {
ompi_mutex_lock(&ompi_request_lock);
ompi_request_waiting++;
while (recvreq.req_recv.req_base.req_ompi.req_complete == false)
ompi_condition_wait(&ompi_request_cond, &ompi_request_lock);
ompi_request_waiting--;
ompi_mutex_unlock(&ompi_request_lock);
} else {
ompi_request_waiting++;
while (recvreq.req_recv.req_base.req_ompi.req_complete == false)
ompi_condition_wait(&ompi_request_cond, &ompi_request_lock);
ompi_request_waiting--;
}
}
if (NULL != status) {
*status = recvreq.req_recv.req_base.req_ompi.req_status;
}
MCA_PML_BASE_RECV_REQUEST_RETURN(&recvreq.req_recv);
return OMPI_SUCCESS;
}

117
src/mca/pml/teg/pml_teg_irecv.c Обычный файл
Просмотреть файл

@ -0,0 +1,117 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "request/request.h"
#include "pml_teg_recvreq.h"
int mca_pml_teg_irecv_init(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
int rc;
mca_ptl_base_recv_request_t *recvreq;
MCA_PML_TEG_RECV_REQUEST_ALLOC(recvreq, rc);
if (NULL == recvreq)
return rc;
MCA_PML_TEG_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, true);
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_teg_irecv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
int rc;
mca_ptl_base_recv_request_t *recvreq;
MCA_PML_TEG_RECV_REQUEST_ALLOC(recvreq, rc);
if (NULL == recvreq)
return rc;
MCA_PML_TEG_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, false);
if ((rc = mca_pml_teg_recv_request_start(recvreq)) != OMPI_SUCCESS) {
MCA_PML_TEG_RECV_REQUEST_RETURN(recvreq);
return rc;
}
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_teg_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
int rc;
mca_ptl_base_recv_request_t *recvreq;
MCA_PML_TEG_RECV_REQUEST_ALLOC(recvreq, rc);
if (NULL == recvreq)
return rc;
MCA_PML_TEG_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, false);
if ((rc = mca_pml_teg_recv_request_start(recvreq)) != OMPI_SUCCESS) {
goto recv_finish;
}
if (recvreq->req_recv.req_base.req_ompi.req_complete == false) {
/* give up and sleep until completion */
if (ompi_using_threads()) {
ompi_mutex_lock(&ompi_request_lock);
ompi_request_waiting++;
while (recvreq->req_recv.req_base.req_ompi.req_complete == false)
ompi_condition_wait(&ompi_request_cond, &ompi_request_lock);
ompi_request_waiting--;
ompi_mutex_unlock(&ompi_request_lock);
} else {
ompi_request_waiting++;
while (recvreq->req_recv.req_base.req_ompi.req_complete == false)
ompi_condition_wait(&ompi_request_cond, &ompi_request_lock);
ompi_request_waiting--;
}
}
recv_finish:
if (NULL != status) { /* return status */
*status = recvreq->req_recv.req_base.req_ompi.req_status;
}
MCA_PML_TEG_RECV_REQUEST_RETURN(recvreq);
return recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
}

128
src/mca/pml/teg/pml_teg_isend.c Обычный файл
Просмотреть файл

@ -0,0 +1,128 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "pml_teg.h"
#include "pml_teg_proc.h"
#include "pml_teg_sendreq.h"
#include "pml_teg_recvreq.h"
int mca_pml_teg_isend_init(void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm,
ompi_request_t ** request)
{
int rc;
mca_pml_teg_send_request_t *sendreq;
MCA_PML_TEG_SEND_REQUEST_ALLOC(comm, dst, sendreq, rc);
if (rc != OMPI_SUCCESS)
return rc;
MCA_PML_TEG_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, true);
*request = (ompi_request_t *) sendreq;
return OMPI_SUCCESS;
}
int mca_pml_teg_isend(void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm,
ompi_request_t ** request)
{
int rc;
mca_pml_teg_send_request_t *sendreq;
MCA_PML_TEG_SEND_REQUEST_ALLOC(comm, dst, sendreq, rc);
if (rc != OMPI_SUCCESS)
return rc;
MCA_PML_TEG_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, false);
MCA_PML_TEG_SEND_REQUEST_START(sendreq, rc);
*request = (ompi_request_t *) sendreq;
return rc;
}
int mca_pml_teg_send(void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm)
{
int rc;
mca_pml_teg_send_request_t *sendreq;
MCA_PML_TEG_SEND_REQUEST_ALLOC(comm, dst, sendreq, rc);
if (rc != OMPI_SUCCESS)
return rc;
MCA_PML_TEG_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, false);
MCA_PML_TEG_SEND_REQUEST_START(sendreq, rc);
if (rc != OMPI_SUCCESS) {
MCA_PML_TEG_FREE((ompi_request_t **) & sendreq);
return rc;
}
if (sendreq->req_send.req_base.req_ompi.req_complete == false) {
/* give up and sleep until completion */
if (ompi_using_threads()) {
ompi_mutex_lock(&ompi_request_lock);
ompi_request_waiting++;
while (sendreq->req_send.req_base.req_ompi.req_complete == false)
ompi_condition_wait(&ompi_request_cond, &ompi_request_lock);
ompi_request_waiting--;
ompi_mutex_unlock(&ompi_request_lock);
} else {
ompi_request_waiting++;
while (sendreq->req_send.req_base.req_ompi.req_complete == false)
ompi_condition_wait(&ompi_request_cond, &ompi_request_lock);
ompi_request_waiting--;
}
}
/* return request to pool */
MCA_PML_TEG_FREE((ompi_request_t **) & sendreq);
return OMPI_SUCCESS;
}

56
src/mca/pml/teg/pml_teg_proc.c Обычный файл
Просмотреть файл

@ -0,0 +1,56 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "include/sys/atomic.h"
#include "pml_teg.h"
#include "pml_teg_proc.h"
#include "pml_ptl_array.h"
static void mca_pml_teg_proc_construct(mca_pml_proc_t* proc)
{
proc->proc_ompi = NULL;
proc->proc_ptl_flags = 0;
OBJ_CONSTRUCT(&proc->proc_lock, ompi_mutex_t);
OBJ_CONSTRUCT(&proc->proc_ptl_first, mca_pml_teg_ptl_array_t);
OBJ_CONSTRUCT(&proc->proc_ptl_next, mca_pml_teg_ptl_array_t);
OMPI_THREAD_LOCK(&mca_pml_teg.teg_lock);
ompi_list_append(&mca_pml_teg.teg_procs, (ompi_list_item_t*)proc);
OMPI_THREAD_UNLOCK(&mca_pml_teg.teg_lock);
}
static void mca_pml_teg_proc_destruct(mca_pml_proc_t* proc)
{
OMPI_THREAD_LOCK(&mca_pml_teg.teg_lock);
ompi_list_remove_item(&mca_pml_teg.teg_procs, (ompi_list_item_t*)proc);
OMPI_THREAD_UNLOCK(&mca_pml_teg.teg_lock);
OBJ_DESTRUCT(&proc->proc_lock);
OBJ_DESTRUCT(&proc->proc_ptl_first);
OBJ_DESTRUCT(&proc->proc_ptl_next);
}
OBJ_CLASS_INSTANCE(
mca_pml_teg_proc_t,
ompi_list_item_t,
mca_pml_teg_proc_construct,
mca_pml_teg_proc_destruct
);

107
src/mca/pml/teg/pml_teg_proc.h Обычный файл
Просмотреть файл

@ -0,0 +1,107 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PML_PROC_H
#define MCA_PML_PROC_H
#include "threads/mutex.h"
#include "communicator/communicator.h"
#include "group/group.h"
#include "proc/proc.h"
#include "pml_ptl_array.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* Structure associated w/ ompi_proc_t that contains data specific
* to the PML. Note that this name is not PML specific.
*/
struct mca_pml_proc_t {
ompi_list_item_t super;
ompi_proc_t *proc_ompi; /**< back-pointer to ompi_proc_t */
ompi_mutex_t proc_lock; /**< lock to protect against concurrent access */
mca_ptl_array_t proc_ptl_first; /**< array of ptls to use for first fragments */
mca_ptl_array_t proc_ptl_next; /**< array of ptls to use for remaining fragments */
uint32_t proc_ptl_flags; /**< aggregate ptl flags */
};
typedef struct mca_pml_proc_t mca_pml_proc_t;
OMPI_COMP_EXPORT extern ompi_class_t mca_pml_teg_proc_t_class;
typedef struct mca_pml_proc_t mca_pml_teg_proc_t;
/**
* Return the mca_pml_proc_t instance cached in the communicators local group.
*
* @param comm Communicator
* @param rank Peer rank
* @return mca_pml_proc_t instance
*/
static inline mca_pml_proc_t* mca_pml_teg_proc_lookup_local(ompi_communicator_t* comm, int rank)
{
ompi_proc_t* proc = comm->c_local_group->grp_proc_pointers[rank];
return proc->proc_pml;
}
/**
* Return the mca_pml_proc_t instance cached on the communicators remote group.
*
* @param comm Communicator
* @param rank Peer rank
* @return mca_pml_proc_t instance
*/
static inline mca_pml_proc_t* mca_pml_teg_proc_lookup_remote(ompi_communicator_t* comm, int rank)
{
ompi_proc_t* proc = comm->c_remote_group->grp_proc_pointers[rank];
return proc->proc_pml;
}
/**
* Return the mca_ptl_peer_t instance corresponding to the process/ptl combination.
*
* @param comm Communicator
* @param rank Peer rank
* @return mca_pml_proc_t instance
*/
static inline struct mca_ptl_base_peer_t* mca_pml_teg_proc_lookup_remote_peer(
ompi_communicator_t* comm,
int rank,
struct mca_ptl_base_module_t* ptl)
{
ompi_proc_t* proc = comm->c_remote_group->grp_proc_pointers[rank];
mca_pml_proc_t* proc_pml = proc->proc_pml;
size_t i, size = mca_ptl_array_get_size(&proc_pml->proc_ptl_first);
mca_ptl_proc_t* proc_ptl = proc_pml->proc_ptl_first.ptl_procs;
for(i = 0; i < size; i++) {
if(proc_ptl->ptl == ptl) {
return proc_ptl->ptl_peer;
}
proc_ptl++;
}
return NULL;
}
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

40
src/mca/pml/teg/pml_teg_progress.c Обычный файл
Просмотреть файл

@ -0,0 +1,40 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "pml_teg.h"
#include "pml_teg_sendreq.h"
int mca_pml_teg_progress(void)
{
mca_ptl_tstamp_t tstamp = 0;
size_t i;
int count = 0;
/*
* Progress each of the PTL modules
*/
for(i=0; i<mca_pml_teg.teg_num_ptl_progress; i++) {
int rc = mca_pml_teg.teg_ptl_progress[i](tstamp);
if(rc > 0) {
count += rc;
}
}
return count;
}

43
src/mca/pml/teg/pml_teg_ptl.c Обычный файл
Просмотреть файл

@ -0,0 +1,43 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "pml_teg_ptl.h"
static void mca_pml_base_ptl_construct(mca_pml_base_ptl_t* ptl)
{
OBJ_CONSTRUCT(&ptl->ptl_cache, ompi_list_t);
OBJ_CONSTRUCT(&ptl->ptl_cache_lock, ompi_mutex_t);
ptl->ptl = NULL;
ptl->ptl_cache_size = 0;
ptl->ptl_cache_alloc = 0;
}
static void mca_pml_base_ptl_destruct(mca_pml_base_ptl_t* ptl)
{
OBJ_DESTRUCT(&ptl->ptl_cache);
OBJ_DESTRUCT(&ptl->ptl_cache_lock);
}
OBJ_CLASS_INSTANCE(
mca_pml_base_ptl_t,
ompi_list_t,
mca_pml_base_ptl_construct,
mca_pml_base_ptl_destruct
);

44
src/mca/pml/teg/pml_teg_ptl.h Обычный файл
Просмотреть файл

@ -0,0 +1,44 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef _MCA_PML_BASE_PTL_
#define _MCA_PML_BASE_PTL_
#include "mca/pml/pml.h"
#include "mca/ptl/ptl.h"
#include "threads/condition.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
struct mca_pml_base_ptl_t {
ompi_list_t ptl_cache; /**< cache of send requests */
size_t ptl_cache_size; /**< maximum size of cache */
size_t ptl_cache_alloc; /**< current number of allocated items */
ompi_mutex_t ptl_cache_lock; /**< lock for queue access */
struct mca_ptl_base_module_t* ptl; /**< back pointer to ptl */
};
typedef struct mca_pml_base_ptl_t mca_pml_base_ptl_t;
OBJ_CLASS_DECLARATION(mca_pml_base_ptl_t);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

93
src/mca/pml/teg/pml_teg_recvfrag.c Обычный файл
Просмотреть файл

@ -0,0 +1,93 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#include "ompi_config.h"
#include "mca/pml/pml.h"
#include "pml_teg_recvfrag.h"
#include "pml_teg_proc.h"
OMPI_DECLSPEC extern ompi_class_t mca_ptl_base_recv_frag_t_class;
/**
* Called by the PTL to match attempt a match for new fragments.
*
* @param ptl (IN) The PTL pointer
* @param frag (IN) Receive fragment descriptor.
* @param header (IN) Header corresponding to the receive fragment.
* @return OMPI_SUCCESS or error status on failure.
*/
bool mca_pml_teg_recv_frag_match(
mca_ptl_base_module_t* ptl,
mca_ptl_base_recv_frag_t* frag,
mca_ptl_base_match_header_t* header)
{
bool matched;
bool matches = false;
ompi_list_t matched_frags;
if((matched = mca_ptl_base_match(header, frag, &matched_frags, &matches)) == false) {
frag = (matches ? (mca_ptl_base_recv_frag_t*)ompi_list_remove_first(&matched_frags) : NULL);
}
while(NULL != frag) {
mca_ptl_base_module_t* ptl = frag->frag_base.frag_owner;
mca_ptl_base_recv_request_t *request = frag->frag_request;
mca_ptl_base_match_header_t *header = &frag->frag_base.frag_header.hdr_match;
/*
* Initialize request status.
*/
request->req_recv.req_bytes_packed = header->hdr_msg_length;
request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = header->hdr_src;
request->req_recv.req_base.req_ompi.req_status.MPI_TAG = header->hdr_tag;
/*
* If probe - signal request is complete - but don't notify PTL
*/
if(request->req_recv.req_base.req_type == MCA_PML_REQUEST_PROBE) {
ptl->ptl_recv_progress( ptl,
request,
header->hdr_msg_length,
header->hdr_msg_length );
matched = mca_pml_teg_recv_frag_match( ptl, frag, header );
} else {
/* if required - setup pointer to ptls peer */
if (NULL == frag->frag_base.frag_peer) {
frag->frag_base.frag_peer = mca_pml_teg_proc_lookup_remote_peer(
request->req_recv.req_base.req_comm,header->hdr_src,ptl);
}
/* notify ptl of match */
ptl->ptl_matched(ptl, frag);
};
/* process any additional fragments that arrived out of order */
frag = (matches ? (mca_ptl_base_recv_frag_t*)ompi_list_remove_first(&matched_frags) : NULL);
};
return matched;
}

42
src/mca/pml/teg/pml_teg_recvfrag.h Обычный файл
Просмотреть файл

@ -0,0 +1,42 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PML_TEG_RECVFRAG_H
#define MCA_PML_TEG_RECVFRAG_H
#include "mca/ptl/ptl.h"
#include "mca/pml/base/pml_base_recvreq.h"
#include "mca/ptl/base/ptl_base_recvfrag.h"
/**
* Called by the PTL to match attempt a match for new fragments.
*
* @param ptl (IN) The PTL pointer
* @param frag (IN) Receive fragment descriptor.
* @param header (IN) Header corresponding to the receive fragment.
* @return OMPI_SUCCESS or error status on failure.
*/
bool mca_pml_teg_recv_frag_match(
mca_ptl_base_module_t* ptl,
mca_ptl_base_recv_frag_t* frag,
mca_ptl_base_match_header_t* header
);
#endif

280
src/mca/pml/teg/pml_teg_recvreq.c Обычный файл
Просмотреть файл

@ -0,0 +1,280 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "mca/pml/pml.h"
#include "mca/ptl/ptl.h"
#include "mca/ptl/base/ptl_base_comm.h"
#include "pml_teg_recvreq.h"
#include "pml_teg_sendreq.h"
static mca_ptl_base_recv_frag_t* mca_pml_teg_recv_request_match_specific_proc(
mca_ptl_base_recv_request_t* request, int proc);
static int mca_pml_teg_recv_request_fini(struct ompi_request_t** request)
{
MCA_PML_TEG_FINI(request);
return OMPI_SUCCESS;
}
static int mca_pml_teg_recv_request_free(struct ompi_request_t** request)
{
MCA_PML_TEG_FREE(request);
return OMPI_SUCCESS;
}
static int mca_pml_teg_recv_request_cancel(struct ompi_request_t* request, int complete)
{
mca_pml_base_request_t* teg_request = (mca_pml_base_request_t*)request;
ompi_communicator_t* ompi_comm = teg_request->req_comm;
mca_pml_ptl_comm_t* pml_comm = (mca_pml_ptl_comm_t*)ompi_comm->c_pml_comm;
if( true == request->req_complete ) { /* way to late to cancel this one */
return OMPI_SUCCESS;
}
/* The rest should be protected behind the match logic lock */
OMPI_THREAD_LOCK(&pml_comm->c_matching_lock);
if( OMPI_ANY_TAG == request->req_status.MPI_TAG ) { /* the match have not been already done */
if( teg_request->req_peer == OMPI_ANY_SOURCE ) {
ompi_list_remove_item( &(pml_comm->c_wild_receives),
(ompi_list_item_t*)request );
} else {
ompi_list_remove_item( pml_comm->c_specific_receives + teg_request->req_peer,
(ompi_list_item_t*)request );
}
}
OMPI_THREAD_UNLOCK(&pml_comm->c_matching_lock);
request->req_status._cancelled = true;
request->req_complete = true; /* mark it as completed so all the test/wait functions
* on this particular request will finish */
/* Now we have a problem if we are in a multi-threaded environment. We shou ld
* broadcast the condition on the request in order to allow the other threa ds
* to complete their test/wait functions.
*/
if(ompi_request_waiting) {
ompi_condition_broadcast(&ompi_request_cond);
}
return OMPI_SUCCESS;
}
static void mca_pml_teg_recv_request_construct(mca_ptl_base_recv_request_t* request)
{
request->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
request->req_recv.req_base.req_ompi.req_fini = mca_pml_teg_recv_request_fini;
request->req_recv.req_base.req_ompi.req_free = mca_pml_teg_recv_request_free;
request->req_recv.req_base.req_ompi.req_cancel = mca_pml_teg_recv_request_cancel;
}
static void mca_pml_teg_recv_request_destruct(mca_ptl_base_recv_request_t* request)
{
}
OBJ_CLASS_INSTANCE(
mca_pml_teg_recv_request_t,
mca_ptl_base_recv_request_t,
mca_pml_teg_recv_request_construct,
mca_pml_teg_recv_request_destruct);
/*
* Update the recv request status to reflect the number of bytes
* received and actually delivered to the application.
*/
void mca_pml_teg_recv_request_progress(
struct mca_ptl_base_module_t* ptl,
mca_ptl_base_recv_request_t* req,
size_t bytes_received,
size_t bytes_delivered)
{
OMPI_THREAD_LOCK(&ompi_request_lock);
req->req_bytes_received += bytes_received;
req->req_bytes_delivered += bytes_delivered;
if (req->req_bytes_received >= req->req_recv.req_bytes_packed) {
/* initialize request status */
req->req_recv.req_base.req_ompi.req_status._count = req->req_bytes_delivered;
req->req_recv.req_base.req_pml_complete = true;
req->req_recv.req_base.req_ompi.req_complete = true;
if(ompi_request_waiting) {
ompi_condition_broadcast(&ompi_request_cond);
}
}
OMPI_THREAD_UNLOCK(&ompi_request_lock);
}
/*
* This routine is used to match a posted receive when the source process
* is specified.
*/
void mca_pml_teg_recv_request_match_specific(mca_ptl_base_recv_request_t* request)
{
ompi_communicator_t *comm = request->req_recv.req_base.req_comm;
mca_pml_ptl_comm_t* pml_comm = comm->c_pml_comm;
int req_peer = request->req_recv.req_base.req_peer;
mca_ptl_base_recv_frag_t* frag;
/* check for a specific match */
OMPI_THREAD_LOCK(&pml_comm->c_matching_lock);
/* assign sequence number */
request->req_recv.req_base.req_sequence = pml_comm->c_recv_seq++;
if (ompi_list_get_size(&pml_comm->c_unexpected_frags[req_peer]) > 0 &&
(frag = mca_pml_teg_recv_request_match_specific_proc(request, req_peer)) != NULL) {
mca_ptl_base_module_t* ptl = frag->frag_base.frag_owner;
/* setup pointer to ptls peer */
if(NULL == frag->frag_base.frag_peer)
frag->frag_base.frag_peer = mca_pml_teg_proc_lookup_remote_peer(comm,req_peer,ptl);
OMPI_THREAD_UNLOCK(&pml_comm->c_matching_lock);
if( !((MCA_PML_REQUEST_IPROBE == request->req_recv.req_base.req_type) ||
(MCA_PML_REQUEST_PROBE == request->req_recv.req_base.req_type)) ) {
ptl->ptl_matched(ptl, frag);
}
return; /* match found */
}
/* We didn't find any matches. Record this irecv so we can match
* it when the message comes in.
*/
if(request->req_recv.req_base.req_type != MCA_PML_REQUEST_IPROBE) {
ompi_list_append(pml_comm->c_specific_receives+req_peer, (ompi_list_item_t*)request);
}
OMPI_THREAD_UNLOCK(&pml_comm->c_matching_lock);
}
/*
* this routine is used to try and match a wild posted receive - where
* wild is determined by the value assigned to the source process
*/
void mca_pml_teg_recv_request_match_wild(mca_ptl_base_recv_request_t* request)
{
ompi_communicator_t *comm = request->req_recv.req_base.req_comm;
mca_pml_ptl_comm_t* pml_comm = comm->c_pml_comm;
int proc_count = comm->c_remote_group->grp_proc_count;
int proc;
/*
* Loop over all the outstanding messages to find one that matches.
* There is an outer loop over lists of messages from each
* process, then an inner loop over the messages from the
* process.
*/
OMPI_THREAD_LOCK(&pml_comm->c_matching_lock);
/* assign sequence number */
request->req_recv.req_base.req_sequence = pml_comm->c_recv_seq++;
for (proc = 0; proc < proc_count; proc++) {
mca_ptl_base_recv_frag_t* frag;
/* continue if no frags to match */
if (ompi_list_get_size(&pml_comm->c_unexpected_frags[proc]) == 0)
continue;
/* loop over messages from the current proc */
if ((frag = mca_pml_teg_recv_request_match_specific_proc(request, proc)) != NULL) {
mca_ptl_base_module_t* ptl = frag->frag_base.frag_owner;
/* if required - setup pointer to ptls peer */
if(NULL == frag->frag_base.frag_peer)
frag->frag_base.frag_peer = mca_pml_teg_proc_lookup_remote_peer(comm,proc,ptl);
OMPI_THREAD_UNLOCK(&pml_comm->c_matching_lock);
if( !((MCA_PML_REQUEST_IPROBE == request->req_recv.req_base.req_type) ||
(MCA_PML_REQUEST_PROBE == request->req_recv.req_base.req_type)) ) {
ptl->ptl_matched(ptl, frag);
}
return; /* match found */
}
}
/* We didn't find any matches. Record this irecv so we can match to
* it when the message comes in.
*/
if(request->req_recv.req_base.req_type != MCA_PML_REQUEST_IPROBE)
ompi_list_append(&pml_comm->c_wild_receives, (ompi_list_item_t*)request);
OMPI_THREAD_UNLOCK(&pml_comm->c_matching_lock);
}
/*
* this routine tries to match a posted receive. If a match is found,
* it places the request in the appropriate matched receive list.
*/
static mca_ptl_base_recv_frag_t* mca_pml_teg_recv_request_match_specific_proc(
mca_ptl_base_recv_request_t* request, int proc)
{
mca_pml_ptl_comm_t *pml_comm = request->req_recv.req_base.req_comm->c_pml_comm;
ompi_list_t* unexpected_frags = pml_comm->c_unexpected_frags+proc;
mca_ptl_base_recv_frag_t* frag;
mca_ptl_base_match_header_t* header;
int tag = request->req_recv.req_base.req_tag;
if( OMPI_ANY_TAG == tag ) {
for (frag = (mca_ptl_base_recv_frag_t*)ompi_list_get_first(unexpected_frags);
frag != (mca_ptl_base_recv_frag_t*)ompi_list_get_end(unexpected_frags);
frag = (mca_ptl_base_recv_frag_t*)ompi_list_get_next(frag)) {
header = &(frag->frag_base.frag_header.hdr_match);
/* check first frag - we assume that process matching has been done already */
if( header->hdr_tag >= 0 ) {
goto find_fragment;
}
}
} else {
for (frag = (mca_ptl_base_recv_frag_t*)ompi_list_get_first(unexpected_frags);
frag != (mca_ptl_base_recv_frag_t*)ompi_list_get_end(unexpected_frags);
frag = (mca_ptl_base_recv_frag_t*)ompi_list_get_next(frag)) {
header = &(frag->frag_base.frag_header.hdr_match);
/* check first frag - we assume that process matching has been done already */
if ( tag == header->hdr_tag ) {
/* we assume that the tag is correct from MPI point of view (ie. >= 0 ) */
goto find_fragment;
}
}
}
return NULL;
find_fragment:
request->req_recv.req_bytes_packed = header->hdr_msg_length;
request->req_recv.req_base.req_ompi.req_status.MPI_TAG = header->hdr_tag;
request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = header->hdr_src;
if( !((MCA_PML_REQUEST_IPROBE == request->req_recv.req_base.req_type) ||
(MCA_PML_REQUEST_PROBE == request->req_recv.req_base.req_type)) ) {
ompi_list_remove_item(unexpected_frags, (ompi_list_item_t*)frag);
frag->frag_request = request;
} else {
/* it's a probe, therefore report it's completion */
mca_pml_teg_recv_request_progress( NULL, request, header->hdr_msg_length, header->hdr_msg_length );
}
return frag;
}

152
src/mca/pml/teg/pml_teg_recvreq.h Обычный файл
Просмотреть файл

@ -0,0 +1,152 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PML_TEG_RECV_REQUEST_H
#define OMPI_PML_TEG_RECV_REQUEST_H
#include "pml_teg.h"
#include "pml_teg_proc.h"
#include "mca/ptl/base/ptl_base_recvreq.h"
#include "mca/ptl/base/ptl_base_recvfrag.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
typedef mca_ptl_base_recv_request_t mca_pml_teg_recv_request_t;
OBJ_CLASS_DECLARATION(mca_pml_teg_recv_request_t);
/**
* Allocate a recv request from the modules free list.
*
* @param rc (OUT) OMPI_SUCCESS or error status on failure.
* @return Receive request.
*/
#define MCA_PML_TEG_RECV_REQUEST_ALLOC(recvreq, rc) \
do { \
ompi_list_item_t* item; \
OMPI_FREE_LIST_GET(&mca_pml_teg.teg_recv_requests, item, rc); \
recvreq = (mca_ptl_base_recv_request_t*)item; \
} while(0)
/**
* Initialize a recv request.
*/
#define MCA_PML_TEG_RECV_REQUEST_INIT( \
request, \
addr, \
count, \
datatype, \
src, \
tag, \
comm, \
persistent) \
{ \
MCA_PML_BASE_RECV_REQUEST_INIT( \
(&(request)->req_recv), \
addr, \
count, \
datatype, \
src, \
tag, \
comm, \
persistent \
); \
}
/**
* Return a recv request to the modules free list.
*
* @param request (IN) Receive request.
*/
#define MCA_PML_TEG_RECV_REQUEST_RETURN(request) \
do { \
MCA_PML_BASE_RECV_REQUEST_RETURN( &request->req_recv ); \
OMPI_FREE_LIST_RETURN(&mca_pml_teg.teg_recv_requests, (ompi_list_item_t*)request); \
} while(0)
/**
* Attempt to match the request against the unexpected fragment list
* for all source ranks w/in the communicator.
*
* @param request (IN) Request to match.
*/
void mca_pml_teg_recv_request_match_wild(mca_ptl_base_recv_request_t* request);
/**
* Attempt to match the request against the unexpected fragment list
* for a specific source rank.
*
* @param request (IN) Request to match.
*/
void mca_pml_teg_recv_request_match_specific(mca_ptl_base_recv_request_t* request);
/**
* Start an initialized request.
*
* @param request Receive request.
* @return OMPI_SUCESS or error status on failure.
*/
static inline int mca_pml_teg_recv_request_start(mca_ptl_base_recv_request_t* request)
{
/* init/re-init the request */
request->req_bytes_received = 0;
request->req_bytes_delivered = 0;
request->req_recv.req_base.req_pml_complete = false;
request->req_recv.req_base.req_ompi.req_complete = false;
request->req_recv.req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE;
/* always set the req_status.MPI_TAG to ANY_TAG before starting the request. This field
* is used on the cancel part in order to find out if the request has been matched or not.
*/
request->req_recv.req_base.req_ompi.req_status.MPI_TAG = OMPI_ANY_TAG;
request->req_recv.req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS;
request->req_recv.req_base.req_ompi.req_status._cancelled = 0;
/* attempt to match posted recv */
if(request->req_recv.req_base.req_peer == OMPI_ANY_SOURCE) {
mca_pml_teg_recv_request_match_wild(request);
} else {
mca_pml_teg_recv_request_match_specific(request);
}
return OMPI_SUCCESS;
}
/**
* Update status of a recv request based on the completion status of
* the receive fragment.
*
* @param ptl (IN) The PTL pointer.
* @param request (IN) Receive request.
* @param bytes_received (IN) Bytes received from peer.
* @param bytes_delivered (IN) Bytes delivered to application.
*/
void mca_pml_teg_recv_request_progress(
struct mca_ptl_base_module_t* ptl,
mca_ptl_base_recv_request_t* request,
size_t bytes_received,
size_t bytes_delivered
);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

215
src/mca/pml/teg/pml_teg_sendreq.c Обычный файл
Просмотреть файл

@ -0,0 +1,215 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/*%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%*/
#include "ompi_config.h"
#include "include/constants.h"
#include "mca/pml/pml.h"
#include "mca/ptl/ptl.h"
#include "pml_teg.h"
#include "pml_teg_proc.h"
#include "pml_teg_sendreq.h"
#include "pml_teg_recvreq.h"
static int mca_pml_teg_send_request_fini(struct ompi_request_t** request)
{
MCA_PML_TEG_FINI(request);
return OMPI_SUCCESS;
}
static int mca_pml_teg_send_request_free(struct ompi_request_t** request)
{
MCA_PML_TEG_FREE(request);
return OMPI_SUCCESS;
}
static int mca_pml_teg_send_request_cancel(struct ompi_request_t* request, int complete)
{
/* we dont cancel send requests by now */
return OMPI_SUCCESS;
}
static void mca_pml_teg_send_request_construct(mca_pml_teg_send_request_t* req)
{
req->req_cached = false;
req->req_send.req_base.req_ompi.req_fini = mca_pml_teg_send_request_fini;
req->req_send.req_base.req_ompi.req_free = mca_pml_teg_send_request_free;
req->req_send.req_base.req_ompi.req_cancel = mca_pml_teg_send_request_cancel;
}
static void mca_pml_teg_send_request_destruct(mca_pml_teg_send_request_t* req)
{
}
OBJ_CLASS_INSTANCE(
mca_pml_teg_send_request_t,
mca_ptl_base_send_request_t,
mca_pml_teg_send_request_construct,
mca_pml_teg_send_request_destruct);
/**
* Schedule message delivery across potentially multiple PTLs.
*
* @param request (IN) Request to schedule
* @return status Error status
*
*/
int mca_pml_teg_send_request_schedule(mca_ptl_base_send_request_t* req)
{
ompi_proc_t *proc = ompi_comm_peer_lookup(req->req_send.req_base.req_comm, req->req_send.req_base.req_peer);
mca_pml_proc_t* proc_pml = proc->proc_pml;
int send_count = 0;
size_t bytes_remaining;
size_t num_ptl_avail;
size_t num_ptl;
/*
* Only allow one thread in this routine for a given request.
* However, we cannot block callers on a mutex, so simply keep track
* of the number of times the routine has been called and run through
* the scheduling logic once for every call.
*/
if(OMPI_THREAD_ADD32(&req->req_lock,1) == 1) {
do {
/* allocate remaining bytes to PTLs */
bytes_remaining = req->req_send.req_bytes_packed - req->req_offset;
num_ptl_avail = proc_pml->proc_ptl_next.ptl_size;
num_ptl = 0;
while(bytes_remaining > 0 && num_ptl++ < num_ptl_avail) {
mca_ptl_proc_t* ptl_proc = mca_ptl_array_get_next(&proc_pml->proc_ptl_next);
mca_ptl_base_module_t* ptl = ptl_proc->ptl;
int rc;
/* if this is the last PTL that is available to use, or the number of
* bytes remaining in the message is less than the PTLs minimum fragment
* size, then go ahead and give the rest of the message to this PTL.
*/
size_t bytes_to_frag;
if(num_ptl == num_ptl_avail || bytes_remaining < ptl->ptl_min_frag_size) {
bytes_to_frag = bytes_remaining;
/* otherwise attempt to give the PTL a percentage of the message
* based on a weighting factor. for simplicity calculate this as
* a percentage of the overall message length (regardless of amount
* previously assigned)
*/
} else {
bytes_to_frag = (ptl_proc->ptl_weight * bytes_remaining) / 100;
}
/* makes sure that we don't exceed ptl_max_frag_size */
if(ptl->ptl_max_frag_size != 0 && bytes_to_frag > ptl->ptl_max_frag_size)
bytes_to_frag = ptl->ptl_max_frag_size;
rc = ptl->ptl_put(ptl, ptl_proc->ptl_peer, req, req->req_offset, bytes_to_frag, 0);
if(rc == OMPI_SUCCESS) {
send_count++;
bytes_remaining = req->req_send.req_bytes_packed - req->req_offset;
}
}
/* unable to complete send - queue for later */
if(send_count == 0) {
OMPI_THREAD_LOCK(&mca_pml_teg.teg_lock);
ompi_list_append(&mca_pml_teg.teg_send_pending, (ompi_list_item_t*)req);
OMPI_THREAD_UNLOCK(&mca_pml_teg.teg_lock);
req->req_lock = 0;
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* fragments completed while scheduling - so retry */
} while(OMPI_THREAD_ADD32(&req->req_lock,-1) > 0);
/* free the request if completed while in the scheduler */
if (req->req_send.req_base.req_free_called && req->req_send.req_base.req_pml_complete) {
MCA_PML_TEG_FREE((ompi_request_t**)&req);
}
}
return OMPI_SUCCESS;
}
/**
* Update the status of the send request to reflect the number of bytes
* "actually" sent (and acknowledged). This should be called by the
* lower layer PTL after the fragment is actually delivered and has been
* acknowledged (if required). Note that this routine should NOT be called
* directly by the PTL, a function pointer is setup on the PTL at init to
* enable upcalls into the PML w/out directly linking to a specific PML
* implementation.
*/
void mca_pml_teg_send_request_progress(
struct mca_ptl_base_module_t* ptl,
mca_ptl_base_send_request_t* req,
size_t bytes_sent)
{
bool schedule = false;
OMPI_THREAD_LOCK(&ompi_request_lock);
req->req_bytes_sent += bytes_sent;
if (req->req_bytes_sent >= req->req_send.req_bytes_packed) {
req->req_send.req_base.req_pml_complete = true;
if (req->req_send.req_base.req_ompi.req_complete == false) {
req->req_send.req_base.req_ompi.req_status.MPI_SOURCE = req->req_send.req_base.req_comm->c_my_rank;
req->req_send.req_base.req_ompi.req_status.MPI_TAG = req->req_send.req_base.req_tag;
req->req_send.req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS;
req->req_send.req_base.req_ompi.req_status._count = req->req_bytes_sent;
req->req_send.req_base.req_ompi.req_complete = true;
if(ompi_request_waiting) {
ompi_condition_broadcast(&ompi_request_cond);
}
} else if(req->req_send.req_base.req_free_called) {
/* don't free the request if in the scheduler */
if(req->req_lock == 0) {
MCA_PML_TEG_FREE((ompi_request_t**)&req);
}
} else if (req->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) {
mca_pml_base_bsend_request_fini((ompi_request_t*)req);
}
/* test to see if we have scheduled the entire request */
} else if (req->req_offset < req->req_send.req_bytes_packed) {
schedule = true;
}
OMPI_THREAD_UNLOCK(&ompi_request_lock);
/* schedule remaining fragments of this request */
if(schedule) {
mca_pml_teg_send_request_schedule(req);
}
/* check for pending requests that need to be progressed */
while(ompi_list_get_size(&mca_pml_teg.teg_send_pending) != 0) {
OMPI_THREAD_LOCK(&mca_pml_teg.teg_lock);
req = (mca_ptl_base_send_request_t*)ompi_list_remove_first(&mca_pml_teg.teg_send_pending);
OMPI_THREAD_UNLOCK(&mca_pml_teg.teg_lock);
if(req == NULL)
break;
if(mca_pml_teg_send_request_schedule(req) != OMPI_SUCCESS)
break;
}
}

211
src/mca/pml/teg/pml_teg_sendreq.h Обычный файл
Просмотреть файл

@ -0,0 +1,211 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PML_TEG_SEND_REQUEST_H
#define OMPI_PML_TEG_SEND_REQUEST_H
#include "mca/ptl/ptl.h"
#include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/ptl/base/ptl_base_sendfrag.h"
#include "mca/ptl/base/ptl_base_comm.h"
#include "pml_teg_proc.h"
#include "pml_teg_ptl.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
typedef mca_ptl_base_send_request_t mca_pml_teg_send_request_t;
OBJ_CLASS_DECLARATION(mca_pml_teg_send_request_t);
#define MCA_PML_TEG_SEND_REQUEST_ALLOC( \
comm, \
dst, \
sendreq, \
rc) \
{ \
mca_pml_proc_t *proc = mca_pml_teg_proc_lookup_remote(comm,dst); \
mca_ptl_proc_t* ptl_proc; \
mca_pml_base_ptl_t* ptl_base; \
\
if(NULL == proc) { \
return OMPI_ERR_OUT_OF_RESOURCE; \
} \
OMPI_THREAD_SCOPED_LOCK(&proc->proc_lock, \
(ptl_proc = mca_ptl_array_get_next(&proc->proc_ptl_first))); \
ptl_base = ptl_proc->ptl_base; \
/* \
* check to see if there is a cache of send requests associated with \
* this ptl - if so try the allocation from there. \
*/ \
if(NULL != ptl_base) { \
OMPI_THREAD_LOCK(&ptl_base->ptl_cache_lock); \
sendreq = (mca_pml_teg_send_request_t*) \
ompi_list_remove_first(&ptl_base->ptl_cache); \
if(NULL != sendreq) { \
OMPI_THREAD_UNLOCK(&ptl_base->ptl_cache_lock); \
rc = OMPI_SUCCESS; \
} else if (ptl_base->ptl_cache_alloc < ptl_base->ptl_cache_size) { \
/* \
* allocate an additional request to the cache \
*/ \
mca_ptl_base_module_t* ptl = ptl_base->ptl; \
ompi_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&mca_pml_teg.teg_send_requests, item, rc); \
sendreq = (mca_pml_teg_send_request_t*)item; \
sendreq->req_ptl = ptl; \
if(ptl->ptl_request_init(ptl, sendreq) == OMPI_SUCCESS) { \
sendreq->req_cached = true; \
ptl_base->ptl_cache_alloc++; \
} \
OMPI_THREAD_UNLOCK(&ptl_base->ptl_cache_lock); \
} else { \
/* \
* take a request from the global pool \
*/ \
ompi_list_item_t* item; \
OMPI_THREAD_UNLOCK(&ptl_base->ptl_cache_lock); \
OMPI_FREE_LIST_WAIT(&mca_pml_teg.teg_send_requests, item, rc); \
sendreq = (mca_pml_teg_send_request_t*)item; \
sendreq->req_ptl = ptl_proc->ptl; \
} \
\
/* otherwise - take the allocation from the global list */ \
} else { \
ompi_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&mca_pml_teg.teg_send_requests, item, rc); \
sendreq = (mca_pml_teg_send_request_t*)item; \
sendreq->req_ptl = ptl_proc->ptl; \
} \
/* update request to point to current peer */ \
sendreq->req_peer = ptl_proc->ptl_peer; \
}
#define MCA_PML_TEG_SEND_REQUEST_INIT( request, \
addr, \
count, \
datatype, \
peer, \
tag, \
comm, \
mode, \
persistent) \
{ \
MCA_PML_BASE_SEND_REQUEST_INIT((&request->req_send), \
addr, \
count, \
datatype, \
peer, \
tag, \
comm, \
mode, \
persistent \
); \
}
#define MCA_PML_TEG_SEND_REQUEST_RETURN(sendreq) \
{ \
mca_ptl_base_module_t* ptl = (sendreq)->req_ptl; \
mca_pml_base_ptl_t* ptl_base = ptl->ptl_base; \
\
/* Let the base handle the reference counts */ \
MCA_PML_BASE_SEND_REQUEST_RETURN((&sendreq->req_send)); \
\
/* \
* If there is a cache associated with the ptl - first attempt \
* to return the send descriptor to the cache. \
*/ \
if(NULL != ptl->ptl_base && (sendreq)->req_cached) { \
OMPI_THREAD_LOCK(&ptl_base->ptl_cache_lock); \
ompi_list_prepend(&ptl_base->ptl_cache, \
(ompi_list_item_t*)sendreq); \
OMPI_THREAD_UNLOCK(&ptl_base->ptl_cache_lock); \
} else { \
OMPI_FREE_LIST_RETURN( \
&mca_pml_teg.teg_send_requests, (ompi_list_item_t*)sendreq); \
} \
}
/**
* Start a send request.
*/
#define MCA_PML_TEG_SEND_REQUEST_START(req, rc) \
{ \
mca_ptl_base_module_t* ptl = req->req_ptl; \
size_t first_fragment_size = ptl->ptl_first_frag_size; \
int flags; \
\
req->req_lock = 0; \
req->req_bytes_sent = 0; \
req->req_peer_match.lval = 0; \
req->req_peer_addr.lval = 0; \
req->req_peer_size = 0; \
req->req_offset = 0; \
req->req_send.req_base.req_pml_complete = false; \
req->req_send.req_base.req_ompi.req_complete = false; \
req->req_send.req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE; \
req->req_send.req_base.req_sequence = mca_pml_ptl_comm_send_sequence( \
req->req_send.req_base.req_comm->c_pml_comm, req->req_send.req_base.req_peer); \
\
/* handle buffered send */ \
if(req->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) { \
mca_pml_base_bsend_request_start(&req->req_send.req_base.req_ompi); \
} \
\
/* start the first fragment */ \
if (first_fragment_size == 0 || \
req->req_send.req_bytes_packed <= first_fragment_size) { \
first_fragment_size = req->req_send.req_bytes_packed; \
flags = (req->req_send.req_send_mode == MCA_PML_BASE_SEND_SYNCHRONOUS) ? \
MCA_PTL_FLAGS_ACK : 0; \
} else { \
/* require match for first fragment of a multi-fragment */ \
flags = MCA_PTL_FLAGS_ACK; \
} \
rc = ptl->ptl_send(ptl, req->req_peer, req, 0, first_fragment_size, \
flags); \
}
/**
* Schedule any data that was not delivered in the first fragment
* across the available PTLs.
*/
int mca_pml_teg_send_request_schedule(mca_ptl_base_send_request_t* req);
/**
* Update the request to reflect the number of bytes delivered. If this
* was the first fragment - schedule the rest of the data.
*/
void mca_pml_teg_send_request_progress(
struct mca_ptl_base_module_t* ptl,
mca_ptl_base_send_request_t* send_request,
size_t bytes_sent
);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

122
src/mca/pml/teg/pml_teg_start.c Обычный файл
Просмотреть файл

@ -0,0 +1,122 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "pml_teg.h"
#include "pml_teg_recvreq.h"
#include "pml_teg_sendreq.h"
int mca_pml_teg_start(size_t count, ompi_request_t** requests)
{
int rc;
size_t i;
for(i=0; i<count; i++) {
mca_pml_base_request_t *pml_request = (mca_pml_base_request_t*)requests[i];
if(NULL == pml_request)
continue;
/* If the persistent request is currently active - obtain the
* request lock and verify the status is incomplete. if the
* pml layer has not completed the request - mark the request
* as free called - so that it will be freed when the request
* completes - and create a new request.
*/
switch(pml_request->req_ompi.req_state) {
case OMPI_REQUEST_INACTIVE:
if(pml_request->req_pml_complete == true)
break;
/* otherwise fall through */
case OMPI_REQUEST_ACTIVE: {
ompi_request_t *request;
OMPI_THREAD_LOCK(&ompi_request_lock);
if (pml_request->req_pml_complete == false) {
/* free request after it completes */
pml_request->req_free_called = true;
} else {
/* can reuse the existing request */
OMPI_THREAD_UNLOCK(&ompi_request_lock);
break;
}
/* allocate a new request */
switch(pml_request->req_type) {
case MCA_PML_REQUEST_SEND: {
mca_pml_base_send_mode_t sendmode =
((mca_pml_base_send_request_t*)pml_request)->req_send_mode;
rc = mca_pml_teg_isend_init(
pml_request->req_addr,
pml_request->req_count,
pml_request->req_datatype,
pml_request->req_peer,
pml_request->req_tag,
sendmode,
pml_request->req_comm,
&request);
break;
}
case MCA_PML_REQUEST_RECV:
rc = mca_pml_teg_irecv_init(
pml_request->req_addr,
pml_request->req_count,
pml_request->req_datatype,
pml_request->req_peer,
pml_request->req_tag,
pml_request->req_comm,
&request);
break;
default:
rc = OMPI_ERR_REQUEST;
break;
}
OMPI_THREAD_UNLOCK(&ompi_request_lock);
if(OMPI_SUCCESS != rc)
return rc;
pml_request = (mca_pml_base_request_t*)request;
requests[i] = request;
break;
}
default:
return OMPI_ERR_REQUEST;
}
/* start the request */
switch(pml_request->req_type) {
case MCA_PML_REQUEST_SEND:
{
mca_ptl_base_send_request_t* sendreq = (mca_ptl_base_send_request_t*)pml_request;
MCA_PML_TEG_SEND_REQUEST_START(sendreq, rc);
if(rc != OMPI_SUCCESS)
return rc;
break;
}
case MCA_PML_REQUEST_RECV:
{
mca_ptl_base_recv_request_t* recvreq = (mca_ptl_base_recv_request_t*)pml_request;
if((rc = mca_pml_teg_recv_request_start(recvreq)) != OMPI_SUCCESS)
return rc;
break;
}
default:
return OMPI_ERR_REQUEST;
}
}
return OMPI_SUCCESS;
}

Просмотреть файл

@ -81,10 +81,10 @@ static int mca_pml_teg_recv_request_cancel(struct ompi_request_t* request, int c
static void mca_pml_teg_recv_request_construct(mca_pml_base_recv_request_t* request) static void mca_pml_teg_recv_request_construct(mca_pml_base_recv_request_t* request)
{ {
request->req_base.req_type = MCA_PML_REQUEST_RECV; request->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
request->req_base.req_ompi.req_fini = mca_pml_teg_recv_request_fini; request->req_recv.req_base.req_ompi.req_fini = mca_pml_teg_recv_request_fini;
request->req_base.req_ompi.req_free = mca_pml_teg_recv_request_free; request->req_recv.req_base.req_ompi.req_free = mca_pml_teg_recv_request_free;
request->req_base.req_ompi.req_cancel = mca_pml_teg_recv_request_cancel; request->req_recv.req_base.req_ompi.req_cancel = mca_pml_teg_recv_request_cancel;
} }
static void mca_pml_teg_recv_request_destruct(mca_pml_base_recv_request_t* request) static void mca_pml_teg_recv_request_destruct(mca_pml_base_recv_request_t* request)
@ -93,7 +93,7 @@ static void mca_pml_teg_recv_request_destruct(mca_pml_base_recv_request_t* reque
OBJ_CLASS_INSTANCE( OBJ_CLASS_INSTANCE(
mca_pml_teg_recv_request_t, mca_pml_teg_recv_request_t,
mca_pml_base_recv_request_t, mca_ptl_base_recv_request_t,
mca_pml_teg_recv_request_construct, mca_pml_teg_recv_request_construct,
mca_pml_teg_recv_request_destruct); mca_pml_teg_recv_request_destruct);

Просмотреть файл

@ -21,13 +21,13 @@
#include "pml_teg.h" #include "pml_teg.h"
#include "pml_teg_proc.h" #include "pml_teg_proc.h"
#include "mca/pml/base/pml_base_recvreq.h" #include "mca/ptl/base/ptl_base_recvreq.h"
#include "mca/ptl/base/ptl_base_recvfrag.h" #include "mca/ptl/base/ptl_base_recvfrag.h"
#if defined(c_plusplus) || defined(__cplusplus) #if defined(c_plusplus) || defined(__cplusplus)
extern "C" { extern "C" {
#endif #endif
typedef mca_pml_base_recv_request_t mca_pml_teg_recv_request_t; typedef mca_ptl_base_recv_request_t mca_pml_teg_recv_request_t;
OBJ_CLASS_DECLARATION(mca_pml_teg_recv_request_t); OBJ_CLASS_DECLARATION(mca_pml_teg_recv_request_t);

Просмотреть файл

@ -31,7 +31,9 @@ headers = \
ptl_base_header.h \ ptl_base_header.h \
ptl_base_match.h \ ptl_base_match.h \
ptl_base_recvfrag.h \ ptl_base_recvfrag.h \
ptl_base_sendfrag.h ptl_base_recvreq.h \
ptl_base_sendfrag.h \
ptl_base_sendreq.h
libmca_ptl_base_la_SOURCES = \ libmca_ptl_base_la_SOURCES = \
$(headers) \ $(headers) \
@ -41,8 +43,10 @@ libmca_ptl_base_la_SOURCES = \
ptl_base_match.c \ ptl_base_match.c \
ptl_base_open.c \ ptl_base_open.c \
ptl_base_recvfrag.c \ ptl_base_recvfrag.c \
ptl_base_recvreq.c \
ptl_base_select.c \ ptl_base_select.c \
ptl_base_sendfrag.c ptl_base_sendfrag.c \
ptl_base_sendreq.c
# Conditionally install the header files # Conditionally install the header files

Просмотреть файл

@ -50,7 +50,7 @@
do { \ do { \
/* local parameters */ \ /* local parameters */ \
ompi_list_t* wild_receives = &pml_comm->c_wild_receives; \ ompi_list_t* wild_receives = &pml_comm->c_wild_receives; \
mca_pml_base_recv_request_t *wild_recv; \ mca_ptl_base_recv_request_t *wild_recv; \
int frag_tag,recv_tag; \ int frag_tag,recv_tag; \
\ \
/* initialization */ \ /* initialization */ \
@ -61,14 +61,14 @@ do { \
* locking is protecting from having other threads trying to \ * locking is protecting from having other threads trying to \
* change this list. \ * change this list. \
*/ \ */ \
for(wild_recv = (mca_pml_base_recv_request_t *) \ for(wild_recv = (mca_ptl_base_recv_request_t *) \
ompi_list_get_first(wild_receives); \ ompi_list_get_first(wild_receives); \
wild_recv != (mca_pml_base_recv_request_t *) \ wild_recv != (mca_ptl_base_recv_request_t *) \
ompi_list_get_end(wild_receives); \ ompi_list_get_end(wild_receives); \
wild_recv = (mca_pml_base_recv_request_t *) \ wild_recv = (mca_ptl_base_recv_request_t *) \
((ompi_list_item_t *)wild_recv)->ompi_list_next) { \ ((ompi_list_item_t *)wild_recv)->ompi_list_next) { \
\ \
recv_tag = wild_recv->req_base.req_tag; \ recv_tag = wild_recv->req_recv.req_base.req_tag; \
if ( \ if ( \
/* exact tag match */ \ /* exact tag match */ \
(frag_tag == recv_tag) || \ (frag_tag == recv_tag) || \
@ -112,7 +112,7 @@ do { \
do { \ do { \
/* local variables */ \ /* local variables */ \
ompi_list_t* specific_receives = (pml_comm->c_specific_receives)+frag_src; \ ompi_list_t* specific_receives = (pml_comm->c_specific_receives)+frag_src; \
mca_pml_base_recv_request_t *specific_recv; \ mca_ptl_base_recv_request_t *specific_recv; \
int frag_src,recv_tag,frag_tag; \ int frag_src,recv_tag,frag_tag; \
\ \
/* initialization */ \ /* initialization */ \
@ -122,16 +122,16 @@ do { \
/* \ /* \
* Loop over the specific irecvs. \ * Loop over the specific irecvs. \
*/ \ */ \
for(specific_recv = (mca_pml_base_recv_request_t *) \ for(specific_recv = (mca_ptl_base_recv_request_t *) \
ompi_list_get_first(specific_receives); \ ompi_list_get_first(specific_receives); \
specific_recv != (mca_pml_base_recv_request_t *) \ specific_recv != (mca_ptl_base_recv_request_t *) \
ompi_list_get_end(specific_receives); \ ompi_list_get_end(specific_receives); \
specific_recv = (mca_pml_base_recv_request_t *) \ specific_recv = (mca_ptl_base_recv_request_t *) \
((ompi_list_item_t *)specific_recv)->ompi_list_next) { \ ((ompi_list_item_t *)specific_recv)->ompi_list_next) { \
/* \ /* \
* Check for a match \ * Check for a match \
*/ \ */ \
recv_tag = specific_recv->req_base.req_tag; \ recv_tag = specific_recv->req_recv.req_base.req_tag; \
if ( (frag_tag == recv_tag) || \ if ( (frag_tag == recv_tag) || \
( (recv_tag == OMPI_ANY_TAG) && (0 <= frag_tag) ) ) { \ ( (recv_tag == OMPI_ANY_TAG) && (0 <= frag_tag) ) ) { \
\ \
@ -170,7 +170,7 @@ do { \
frag_header, pml_comm, return_match) \ frag_header, pml_comm, return_match) \
do { \ do { \
/* local variables */ \ /* local variables */ \
mca_pml_base_recv_request_t *specific_recv, *wild_recv; \ mca_ptl_base_recv_request_t *specific_recv, *wild_recv; \
mca_ptl_sequence_t wild_recv_seq, specific_recv_seq; \ mca_ptl_sequence_t wild_recv_seq, specific_recv_seq; \
int frag_src,frag_tag, wild_recv_tag, specific_recv_tag; \ int frag_src,frag_tag, wild_recv_tag, specific_recv_tag; \
\ \
@ -182,13 +182,13 @@ do { \
* We know that when this is called, both specific and wild irecvs \ * We know that when this is called, both specific and wild irecvs \
* have been posted. \ * have been posted. \
*/ \ */ \
specific_recv = (mca_pml_base_recv_request_t *) \ specific_recv = (mca_ptl_base_recv_request_t *) \
ompi_list_get_first((pml_comm->c_specific_receives)+frag_src); \ ompi_list_get_first((pml_comm->c_specific_receives)+frag_src); \
wild_recv = (mca_pml_base_recv_request_t *) \ wild_recv = (mca_ptl_base_recv_request_t *) \
ompi_list_get_first(&(pml_comm->c_wild_receives)); \ ompi_list_get_first(&(pml_comm->c_wild_receives)); \
\ \
specific_recv_seq = specific_recv->req_base.req_sequence; \ specific_recv_seq = specific_recv->req_recv.req_base.req_sequence; \
wild_recv_seq = wild_recv->req_base.req_sequence; \ wild_recv_seq = wild_recv->req_recv.req_base.req_sequence; \
\ \
while (true) { \ while (true) { \
if (wild_recv_seq < specific_recv_seq) { \ if (wild_recv_seq < specific_recv_seq) { \
@ -198,7 +198,7 @@ do { \
/* \ /* \
* try and match \ * try and match \
*/ \ */ \
wild_recv_tag = wild_recv->req_base.req_tag; \ wild_recv_tag = wild_recv->req_recv.req_base.req_tag; \
if ( (frag_tag == wild_recv_tag) || \ if ( (frag_tag == wild_recv_tag) || \
( (wild_recv_tag == OMPI_ANY_TAG) && (0 <= frag_tag) ) ) { \ ( (wild_recv_tag == OMPI_ANY_TAG) && (0 <= frag_tag) ) ) { \
/* \ /* \
@ -215,14 +215,14 @@ do { \
/* \ /* \
* No match, go to the next. \ * No match, go to the next. \
*/ \ */ \
wild_recv=(mca_pml_base_recv_request_t *) \ wild_recv=(mca_ptl_base_recv_request_t *) \
((ompi_list_item_t *)wild_recv)->ompi_list_next; \ ((ompi_list_item_t *)wild_recv)->ompi_list_next; \
\ \
/* \ /* \
* If that was the last wild one, just look at the \ * If that was the last wild one, just look at the \
* rest of the specific ones. \ * rest of the specific ones. \
*/ \ */ \
if (wild_recv == (mca_pml_base_recv_request_t *) \ if (wild_recv == (mca_ptl_base_recv_request_t *) \
ompi_list_get_end(&(pml_comm->c_wild_receives)) ) \ ompi_list_get_end(&(pml_comm->c_wild_receives)) ) \
{ \ { \
MCA_PTL_BASE_CHECK_SPECIFIC_RECEIVES_FOR_MATCH(frag_header, pml_comm, return_match); \ MCA_PTL_BASE_CHECK_SPECIFIC_RECEIVES_FOR_MATCH(frag_header, pml_comm, return_match); \
@ -233,13 +233,13 @@ do { \
* Get the sequence number for this recv, and go \ * Get the sequence number for this recv, and go \
* back to the top of the loop. \ * back to the top of the loop. \
*/ \ */ \
wild_recv_seq = wild_recv->req_base.req_sequence; \ wild_recv_seq = wild_recv->req_recv.req_base.req_sequence; \
\ \
} else { \ } else { \
/* \ /* \
* specific recv is earlier than the wild one. \ * specific recv is earlier than the wild one. \
*/ \ */ \
specific_recv_tag=specific_recv->req_base.req_tag; \ specific_recv_tag=specific_recv->req_recv.req_base.req_tag; \
if ( (frag_tag == specific_recv_tag) || \ if ( (frag_tag == specific_recv_tag) || \
( (specific_recv_tag == OMPI_ANY_TAG) && (0<=frag_tag)) ) \ ( (specific_recv_tag == OMPI_ANY_TAG) && (0<=frag_tag)) ) \
{ \ { \
@ -256,14 +256,14 @@ do { \
/* \ /* \
* No match, go on to the next specific irecv. \ * No match, go on to the next specific irecv. \
*/ \ */ \
specific_recv = (mca_pml_base_recv_request_t *) \ specific_recv = (mca_ptl_base_recv_request_t *) \
((ompi_list_item_t *)specific_recv)->ompi_list_next; \ ((ompi_list_item_t *)specific_recv)->ompi_list_next; \
\ \
/* \ /* \
* If that was the last specific irecv, process the \ * If that was the last specific irecv, process the \
* rest of the wild ones. \ * rest of the wild ones. \
*/ \ */ \
if (specific_recv == (mca_pml_base_recv_request_t *) \ if (specific_recv == (mca_ptl_base_recv_request_t *) \
ompi_list_get_end((pml_comm->c_specific_receives)+frag_src) ) \ ompi_list_get_end((pml_comm->c_specific_receives)+frag_src) ) \
{ \ { \
MCA_PTL_BASE_CHECK_WILD_RECEIVES_FOR_MATCH(frag_header, pml_comm, return_match); \ MCA_PTL_BASE_CHECK_WILD_RECEIVES_FOR_MATCH(frag_header, pml_comm, return_match); \
@ -273,7 +273,7 @@ do { \
* Get the sequence number for this recv, and go \ * Get the sequence number for this recv, and go \
* back to the top of the loop. \ * back to the top of the loop. \
*/ \ */ \
specific_recv_seq = specific_recv->req_base.req_sequence; \ specific_recv_seq = specific_recv->req_recv.req_base.req_sequence; \
} \ } \
} \ } \
} while(0) } while(0)
@ -328,7 +328,7 @@ bool mca_ptl_base_match(
uint16_t next_msg_seq_expected, frag_msg_seq; uint16_t next_msg_seq_expected, frag_msg_seq;
ompi_communicator_t *comm_ptr; ompi_communicator_t *comm_ptr;
mca_pml_base_recv_request_t *matched_receive = NULL; mca_ptl_base_recv_request_t *matched_receive = NULL;
mca_pml_ptl_comm_t *pml_comm; mca_pml_ptl_comm_t *pml_comm;
int frag_src; int frag_src;
bool match_made=false; bool match_made=false;
@ -398,12 +398,12 @@ bool mca_ptl_base_match(
frag_desc->frag_request=matched_receive; frag_desc->frag_request=matched_receive;
/* set lenght of incoming message */ /* set lenght of incoming message */
matched_receive->req_bytes_packed = frag_header->hdr_msg_length; matched_receive->req_recv.req_bytes_packed = frag_header->hdr_msg_length;
/* /*
* update delivered sequence number information, if needed. * update delivered sequence number information, if needed.
*/ */
if( (matched_receive->req_base.req_type == MCA_PML_REQUEST_PROBE) ) { if( (matched_receive->req_recv.req_base.req_type == MCA_PML_REQUEST_PROBE) ) {
/* Match a probe, rollback the next expected sequence number */ /* Match a probe, rollback the next expected sequence number */
(pml_comm->c_next_msg_seq[frag_src])--; (pml_comm->c_next_msg_seq[frag_src])--;
} }
@ -461,7 +461,7 @@ static bool mca_ptl_base_check_cantmatch_for_match(ompi_list_t *additional_match
int match_found; int match_found;
uint16_t next_msg_seq_expected, frag_seq; uint16_t next_msg_seq_expected, frag_seq;
mca_ptl_base_recv_frag_t *frag_desc; mca_ptl_base_recv_frag_t *frag_desc;
mca_pml_base_recv_request_t *matched_receive = NULL; mca_ptl_base_recv_request_t *matched_receive = NULL;
bool match_made = false; bool match_made = false;
/* /*
@ -604,7 +604,7 @@ bool mca_ptl_base_match_in_order_network_delivery(
{ {
/* local variables */ /* local variables */
ompi_communicator_t *comm_ptr; ompi_communicator_t *comm_ptr;
mca_pml_base_recv_request_t *matched_receive = NULL; mca_ptl_base_recv_request_t *matched_receive = NULL;
mca_pml_ptl_comm_t *pml_comm; mca_pml_ptl_comm_t *pml_comm;
int frag_src; int frag_src;
@ -658,7 +658,7 @@ bool mca_ptl_base_match_in_order_network_delivery(
frag_desc->frag_request=matched_receive; frag_desc->frag_request=matched_receive;
/* set lenght of incoming message */ /* set lenght of incoming message */
matched_receive->req_bytes_packed=frag_header->hdr_msg_length; matched_receive->req_recv.req_bytes_packed=frag_header->hdr_msg_length;
} else { } else {
/* if no match found, place on unexpected queue */ /* if no match found, place on unexpected queue */

Просмотреть файл

@ -20,7 +20,7 @@
#define MCA_PTL_BASE_RECVFRAG_H #define MCA_PTL_BASE_RECVFRAG_H
#include "mca/ptl/ptl.h" #include "mca/ptl/ptl.h"
#include "mca/pml/base/pml_base_recvreq.h" #include "mca/ptl/base/ptl_base_recvreq.h"
#include "mca/ptl/base/ptl_base_fragment.h" #include "mca/ptl/base/ptl_base_fragment.h"
#include "mca/ptl/base/ptl_base_match.h" #include "mca/ptl/base/ptl_base_match.h"
@ -34,7 +34,7 @@ OMPI_DECLSPEC extern ompi_class_t mca_ptl_base_recv_frag_t_class;
*/ */
struct mca_ptl_base_recv_frag_t { struct mca_ptl_base_recv_frag_t {
mca_ptl_base_frag_t frag_base; /**< base fragment descriptor */ mca_ptl_base_frag_t frag_base; /**< base fragment descriptor */
mca_pml_base_recv_request_t *frag_request; /**< matched posted receive */ mca_ptl_base_recv_request_t *frag_request; /**< matched posted receive */
bool frag_is_buffered; /**< does fragment need to be unpacked into users buffer */ bool frag_is_buffered; /**< does fragment need to be unpacked into users buffer */
}; };
typedef struct mca_ptl_base_recv_frag_t mca_ptl_base_recv_frag_t; typedef struct mca_ptl_base_recv_frag_t mca_ptl_base_recv_frag_t;

10
src/mca/ptl/base/ptl_base_recvreq.c Обычный файл
Просмотреть файл

@ -0,0 +1,10 @@
#include "ptl_base_recvreq.h"
OBJ_CLASS_INSTANCE(
mca_ptl_base_recv_request_t,
mca_pml_base_recv_request_t,
NULL,
NULL
);

47
src/mca/ptl/base/ptl_base_recvreq.h Обычный файл
Просмотреть файл

@ -0,0 +1,47 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PTL_BASE_RECV_REQUEST_H
#define MCA_PTL_BASE_RECV_REQUEST_H
#include "mca/ptl/ptl.h"
#include "mca/pml/base/pml_base_recvreq.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* Base type for recv requests
*/
struct mca_ptl_base_recv_request_t {
mca_pml_base_recv_request_t req_recv;
size_t req_bytes_received; /**< number of bytes received from network */
size_t req_bytes_delivered; /**< number of bytes delivered to user */
};
typedef struct mca_ptl_base_recv_request_t mca_ptl_base_recv_request_t;
OMPI_DECLSPEC OBJ_CLASS_DECLARATION(mca_ptl_base_recv_request_t);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

Просмотреть файл

@ -32,7 +32,7 @@ OMPI_DECLSPEC extern ompi_class_t mca_ptl_base_send_frag_t_class;
*/ */
struct mca_ptl_base_send_frag_t { struct mca_ptl_base_send_frag_t {
mca_ptl_base_frag_t frag_base; /**< base fragment descriptor */ mca_ptl_base_frag_t frag_base; /**< base fragment descriptor */
struct mca_pml_base_send_request_t *frag_request; /**< pointer to send request */ struct mca_ptl_base_send_request_t *frag_request; /**< pointer to send request */
}; };
typedef struct mca_ptl_base_send_frag_t mca_ptl_base_send_frag_t; typedef struct mca_ptl_base_send_frag_t mca_ptl_base_send_frag_t;

10
src/mca/ptl/base/ptl_base_sendreq.c Обычный файл
Просмотреть файл

@ -0,0 +1,10 @@
#include "ptl_base_sendreq.h"
OBJ_CLASS_INSTANCE(
mca_ptl_base_send_request_t,
mca_pml_base_send_request_t,
NULL,
NULL
);

69
src/mca/ptl/base/ptl_base_sendreq.h Обычный файл
Просмотреть файл

@ -0,0 +1,69 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PTL_BASE_SEND_REQUEST_H
#define MCA_PTL_BASE_SEND_REQUEST_H
#include "mca/ptl/ptl.h"
#include "mca/pml/base/pml_base_sendreq.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* Base type for send requests
*/
struct mca_ptl_base_send_request_t {
mca_pml_base_send_request_t req_send;
size_t req_offset; /**< number of bytes that have been scheduled */
size_t req_bytes_sent; /**< number of bytes that have been sent */
ompi_ptr_t req_peer_match; /**< matched receive at peer */
ompi_ptr_t req_peer_addr; /**< peers remote buffer address */
uint64_t req_peer_size; /**< size of peers remote buffer */
bool req_cached; /**< has this request been obtained from the ptls cache */
volatile int32_t req_lock; /**< lock used by the scheduler */
struct mca_ptl_base_module_t* req_ptl; /**< ptl allocated for first fragment */
struct mca_ptl_base_peer_t* req_peer; /**< peer associated w/ this ptl */
};
typedef struct mca_ptl_base_send_request_t mca_ptl_base_send_request_t;
OMPI_DECLSPEC OBJ_CLASS_DECLARATION(mca_ptl_base_send_request_t);
/**
* Atomically increase the request offset.
*
* @param request (IN) Send request.
* @param offset (IN) Increment.
*/
static inline void mca_ptl_base_send_request_offset(
mca_ptl_base_send_request_t* request,
size_t offset)
{
OMPI_THREAD_ADD32((int32_t*)(&request->req_offset), offset);
}
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

Просмотреть файл

@ -89,7 +89,7 @@ static void *mca_ptl_mx_alloc(size_t *size)
* *
*/ */
int mca_ptl_mx_request_init(struct mca_ptl_base_module_t* ptl, mca_pml_base_send_request_t* request) int mca_ptl_mx_request_init(struct mca_ptl_base_module_t* ptl, mca_ptl_base_send_request_t* request)
{ {
OBJ_CONSTRUCT(request+1, mca_ptl_mx_send_frag_t); OBJ_CONSTRUCT(request+1, mca_ptl_mx_send_frag_t);
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -109,7 +109,7 @@ int mca_ptl_mx_request_init(struct mca_ptl_base_module_t* ptl, mca_pml_base_send
* descriptor by the PTL. * descriptor by the PTL.
*/ */
void mca_ptl_mx_request_fini(struct mca_ptl_base_module_t* ptl, mca_pml_base_send_request_t* request) void mca_ptl_mx_request_fini(struct mca_ptl_base_module_t* ptl, mca_ptl_base_send_request_t* request)
{ {
OBJ_DESTRUCT(request+1); OBJ_DESTRUCT(request+1);
} }
@ -142,7 +142,7 @@ void mca_ptl_mx_request_fini(struct mca_ptl_base_module_t* ptl, mca_pml_base_sen
int mca_ptl_mx_send( int mca_ptl_mx_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t* sendreq, struct mca_ptl_base_send_request_t* sendreq,
size_t offset, size_t offset,
size_t size, size_t size,
int flags) int flags)
@ -177,13 +177,13 @@ int mca_ptl_mx_send(
int rc; int rc;
convertor = &sendfrag->frag_send.frag_base.frag_convertor; convertor = &sendfrag->frag_send.frag_base.frag_convertor;
ompi_convertor_copy(&sendreq->req_convertor, convertor); ompi_convertor_copy(&sendreq->req_send.req_convertor, convertor);
ompi_convertor_init_for_send( ompi_convertor_init_for_send(
convertor, convertor,
0, 0,
sendreq->req_datatype, sendreq->req_send.req_datatype,
sendreq->req_count, sendreq->req_send.req_count,
sendreq->req_addr, sendreq->req_send.req_addr,
offset, offset,
mca_ptl_mx_alloc ); mca_ptl_mx_alloc );
@ -218,12 +218,12 @@ int mca_ptl_mx_send(
/* first fragment - need to try and match at the receiver */ /* first fragment - need to try and match at the receiver */
if(offset == 0) { if(offset == 0) {
hdr->hdr_common.hdr_flags = flags; hdr->hdr_common.hdr_flags = flags;
hdr->hdr_match.hdr_contextid = sendreq->req_base.req_comm->c_contextid; hdr->hdr_match.hdr_contextid = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_dst = sendreq->req_base.req_peer; hdr->hdr_match.hdr_dst = sendreq->req_send.req_base.req_peer;
hdr->hdr_match.hdr_tag = sendreq->req_base.req_tag; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_msg_length = sendreq->req_bytes_packed; hdr->hdr_match.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_match.hdr_msg_seq = sendreq->req_base.req_sequence; hdr->hdr_match.hdr_msg_seq = sendreq->req_send.req_base.req_sequence;
/* for the first 32K - send header for matching + data */ /* for the first 32K - send header for matching + data */
segments = sendfrag->frag_segments; segments = sendfrag->frag_segments;
@ -292,7 +292,7 @@ int mca_ptl_mx_send(
/* must update the offset after actual fragment size is determined /* must update the offset after actual fragment size is determined
* before attempting to send the fragment * before attempting to send the fragment
*/ */
mca_pml_base_send_request_offset(sendreq, mca_ptl_base_send_request_offset(sendreq,
sendfrag->frag_send.frag_base.frag_size); sendfrag->frag_send.frag_base.frag_size);
/* start the fragment */ /* start the fragment */
@ -329,7 +329,7 @@ int mca_ptl_mx_send(
int mca_ptl_mx_send_continue( int mca_ptl_mx_send_continue(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t* sendreq, struct mca_ptl_base_send_request_t* sendreq,
size_t offset, size_t offset,
size_t size, size_t size,
int flags) int flags)
@ -353,13 +353,13 @@ int mca_ptl_mx_send_continue(
/* initialize convertor */ /* initialize convertor */
convertor = &sendfrag->frag_send.frag_base.frag_convertor; convertor = &sendfrag->frag_send.frag_base.frag_convertor;
ompi_convertor_copy(&sendreq->req_convertor, convertor); ompi_convertor_copy(&sendreq->req_send.req_convertor, convertor);
ompi_convertor_init_for_send( ompi_convertor_init_for_send(
convertor, convertor,
0, 0,
sendreq->req_datatype, sendreq->req_send.req_datatype,
sendreq->req_count, sendreq->req_send.req_count,
sendreq->req_addr, sendreq->req_send.req_addr,
offset, offset,
mca_ptl_mx_alloc ); mca_ptl_mx_alloc );
@ -400,7 +400,7 @@ int mca_ptl_mx_send_continue(
/* must update the offset after actual fragment size is determined /* must update the offset after actual fragment size is determined
* before attempting to send the fragment * before attempting to send the fragment
*/ */
mca_pml_base_send_request_offset(sendreq, size); mca_ptl_base_send_request_offset(sendreq, size);
/* start the fragment */ /* start the fragment */
match.sval.uval = sendreq->req_peer_match.ival; match.sval.uval = sendreq->req_peer_match.ival;
@ -446,7 +446,7 @@ void mca_ptl_mx_matched(
mca_ptl_base_recv_frag_t* frag) mca_ptl_base_recv_frag_t* frag)
{ {
mca_ptl_base_header_t* hdr = &frag->frag_base.frag_header; mca_ptl_base_header_t* hdr = &frag->frag_base.frag_header;
mca_pml_base_recv_request_t* request = frag->frag_request; mca_ptl_base_recv_request_t* request = frag->frag_request;
mca_ptl_mx_module_t* mx_ptl = (mca_ptl_mx_module_t*)ptl; mca_ptl_mx_module_t* mx_ptl = (mca_ptl_mx_module_t*)ptl;
mca_ptl_mx_recv_frag_t* mx_frag = (mca_ptl_mx_recv_frag_t*)frag; mca_ptl_mx_recv_frag_t* mx_frag = (mca_ptl_mx_recv_frag_t*)frag;
unsigned int bytes_delivered = mx_frag->frag_size; unsigned int bytes_delivered = mx_frag->frag_size;
@ -493,8 +493,8 @@ void mca_ptl_mx_matched(
struct iovec iov; struct iovec iov;
unsigned int iov_count = 1; unsigned int iov_count = 1;
int free_after = 0; int free_after = 0;
ompi_proc_t *proc = ompi_comm_peer_lookup(request->req_base.req_comm, ompi_proc_t *proc = ompi_comm_peer_lookup(request->req_recv.req_base.req_comm,
request->req_base.req_ompi.req_status.MPI_SOURCE); request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE);
ompi_convertor_t* convertor = &frag->frag_base.frag_convertor; ompi_convertor_t* convertor = &frag->frag_base.frag_convertor;
/* initialize receive convertor */ /* initialize receive convertor */
@ -502,9 +502,9 @@ void mca_ptl_mx_matched(
ompi_convertor_init_for_recv( ompi_convertor_init_for_recv(
convertor, /* convertor */ convertor, /* convertor */
0, /* flags */ 0, /* flags */
request->req_base.req_datatype, /* datatype */ request->req_recv.req_base.req_datatype, /* datatype */
request->req_base.req_count, /* count elements */ request->req_recv.req_base.req_count, /* count elements */
request->req_base.req_addr, /* users buffer */ request->req_recv.req_base.req_addr, /* users buffer */
0, /* offset in bytes into packed buffer */ 0, /* offset in bytes into packed buffer */
NULL ); /* not allocating memory */ NULL ); /* not allocating memory */
/*ompi_convertor_get_packed_size(convertor, &request->req_bytes_packed); */ /*ompi_convertor_get_packed_size(convertor, &request->req_bytes_packed); */

Просмотреть файл

@ -249,7 +249,7 @@ extern int mca_ptl_mx_del_procs(
extern int mca_ptl_mx_request_init( extern int mca_ptl_mx_request_init(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
@ -268,7 +268,7 @@ extern int mca_ptl_mx_request_init(
extern void mca_ptl_mx_request_fini( extern void mca_ptl_mx_request_fini(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
@ -325,7 +325,7 @@ extern void mca_ptl_mx_matched(
extern int mca_ptl_mx_send( extern int mca_ptl_mx_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags
@ -348,7 +348,7 @@ extern int mca_ptl_mx_send(
extern int mca_ptl_mx_send_continue( extern int mca_ptl_mx_send_continue(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags

Просмотреть файл

@ -178,7 +178,7 @@ static void* mca_ptl_mx_thread(ompi_object_t *arg)
static void mca_ptl_mx_match(void* context, uint64_t match_value, int size) static void mca_ptl_mx_match(void* context, uint64_t match_value, int size)
{ {
mca_ptl_mx_module_t* ptl = (mca_ptl_mx_module_t*)context; mca_ptl_mx_module_t* ptl = (mca_ptl_mx_module_t*)context;
mca_pml_base_recv_request_t* request; mca_ptl_base_recv_request_t* request;
mca_ptl_mx_recv_frag_t *frag; mca_ptl_mx_recv_frag_t *frag;
mx_return_t mx_return; mx_return_t mx_return;
ompi_ptr_t match; ompi_ptr_t match;
@ -193,10 +193,10 @@ static void mca_ptl_mx_match(void* context, uint64_t match_value, int size)
/* otherwise extract request pointer and offset */ /* otherwise extract request pointer and offset */
match.lval = match_value; match.lval = match_value;
request = (mca_pml_base_recv_request_t*)match.sval.uval; request = (mca_ptl_base_recv_request_t*)match.sval.uval;
offset = match.sval.lval; offset = match.sval.lval;
proc = ompi_comm_peer_lookup(request->req_base.req_comm, proc = ompi_comm_peer_lookup(request->req_base.req_comm,
request->req_base.req_ompi.req_status.MPI_SOURCE); request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE);
/* allocate a fragment for receive */ /* allocate a fragment for receive */
MCA_PTL_MX_RECV_FRAG_ALLOC(frag, rc); MCA_PTL_MX_RECV_FRAG_ALLOC(frag, rc);
@ -220,15 +220,15 @@ static void mca_ptl_mx_match(void* context, uint64_t match_value, int size)
ompi_convertor_init_for_recv( ompi_convertor_init_for_recv(
convertor, convertor,
0, /* flags */ 0, /* flags */
request->req_base.req_datatype, /* datatype */ request->req_recv.req_base.req_datatype, /* datatype */
request->req_base.req_count, /* count elements */ request->req_recv.req_base.req_count, /* count elements */
request->req_base.req_addr, /* users buffer */ request->req_recv.req_base.req_addr, /* users buffer */
offset, /* offset in bytes into packed buffer */ offset, /* offset in bytes into packed buffer */
mca_ptl_mx_mem_alloc ); /* not allocating memory */ mca_ptl_mx_mem_alloc ); /* not allocating memory */
/* non-contiguous - allocate buffer for receive */ /* non-contiguous - allocate buffer for receive */
if( 1 == ompi_convertor_need_buffers( convertor ) || if( 1 == ompi_convertor_need_buffers( convertor ) ||
request->req_bytes_packed < offset + frag->frag_size ) { request->req_recv.req_bytes_packed < offset + frag->frag_size ) {
/* TODO - use a fixed fragment size for non-contigous and convert /* TODO - use a fixed fragment size for non-contigous and convert
* this to a free-list of buffers. * this to a free-list of buffers.
@ -242,14 +242,14 @@ static void mca_ptl_mx_match(void* context, uint64_t match_value, int size)
} }
/* check for sending more than receiving */ /* check for sending more than receiving */
if( offset > request->req_bytes_packed ) { if( offset > request->req_recv.req_bytes_packed ) {
frag->frag_recv.frag_base.frag_size = 0; frag->frag_recv.frag_base.frag_size = 0;
} else if (offset + frag->frag_size > request->req_bytes_packed ) { } else if (offset + frag->frag_size > request->req_recv.req_bytes_packed ) {
frag->frag_recv.frag_base.frag_size = request->req_bytes_packed - offset; frag->frag_recv.frag_base.frag_size = request->req_recv.req_bytes_packed - offset;
} }
/* calculate offset into users buffer */ /* calculate offset into users buffer */
} else { } else {
frag->frag_recv.frag_base.frag_addr = ((unsigned char*)request->req_base.req_addr) + offset; frag->frag_recv.frag_base.frag_addr = ((unsigned char*)request->req_recv.req_base.req_addr) + offset;
frag->frag_recv.frag_is_buffered = false; frag->frag_recv.frag_is_buffered = false;
} }

Просмотреть файл

@ -110,7 +110,7 @@ do {
case MCA_PTL_HDR_TYPE_ACK: \ case MCA_PTL_HDR_TYPE_ACK: \
{ \ { \
mca_ptl_mx_send_frag_t* sendfrag; \ mca_ptl_mx_send_frag_t* sendfrag; \
mca_pml_base_send_request_t* sendreq; \ mca_ptl_base_send_request_t* sendreq; \
sendfrag = (mca_ptl_mx_send_frag_t*) \ sendfrag = (mca_ptl_mx_send_frag_t*) \
hdr->hdr_ack.hdr_src_ptr.pval; \ hdr->hdr_ack.hdr_src_ptr.pval; \
sendreq = sendfrag->frag_send.frag_request; \ sendreq = sendfrag->frag_send.frag_request; \

Просмотреть файл

@ -21,7 +21,7 @@
#include "ompi_config.h" #include "ompi_config.h"
#include "include/sys/atomic.h" #include "include/sys/atomic.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/ptl/base/ptl_base_sendfrag.h" #include "mca/ptl/base/ptl_base_sendfrag.h"
#include "ptl_mx.h" #include "ptl_mx.h"
#include "ptl_mx_peer.h" #include "ptl_mx_peer.h"
@ -88,7 +88,7 @@ OBJ_CLASS_DECLARATION(mca_ptl_mx_send_frag_t);
#define MCA_PTL_MX_SEND_FRAG_PROGRESS(frag) \ #define MCA_PTL_MX_SEND_FRAG_PROGRESS(frag) \
do { \ do { \
mca_pml_base_send_request_t* request = frag->frag_send.frag_request; \ mca_ptl_base_send_request_t* request = frag->frag_send.frag_request; \
bool frag_ack; \ bool frag_ack; \
\ \
/* if this is an ack - simply return to pool */ \ /* if this is an ack - simply return to pool */ \

Просмотреть файл

@ -105,7 +105,7 @@
* PTLs. * PTLs.
* *
* The PML is responsible for managing the state (allocation, initialization, * The PML is responsible for managing the state (allocation, initialization,
* and release) of send request descriptors (mca_pml_base_send_request_t). * and release) of send request descriptors (mca_ptl_base_send_request_t).
* However, to minimize the latency associated with allocating resources to * However, to minimize the latency associated with allocating resources to
* a request, the PML provides the capability to cache send requests * a request, the PML provides the capability to cache send requests
* descriptors on a per-PTL basis. Each PTL exports two variables * descriptors on a per-PTL basis. Each PTL exports two variables
@ -117,7 +117,7 @@
* is grown by the PML up to the specified limit. The PTL variable, * is grown by the PML up to the specified limit. The PTL variable,
* ptl_cache_bytes, can be used to specify that additional memory should be * ptl_cache_bytes, can be used to specify that additional memory should be
* allocated by the PML in one contigous block along with the base send request * allocated by the PML in one contigous block along with the base send request
* (mca_pml_base_send_request_t) for use by the PTL. The PTLs ptl_request_init() * (mca_ptl_base_send_request_t) for use by the PTL. The PTLs ptl_request_init()
* method is then called to initialize this additional memory and associating * method is then called to initialize this additional memory and associating
* any PTL specific resources with the request. If a request is removed from * any PTL specific resources with the request. If a request is removed from
* the cache, the ptl_request_fini() method will be called to allow the PTL * the cache, the ptl_request_fini() method will be called to allow the PTL
@ -129,7 +129,7 @@
* number of bytes. However, this may not be possible due to resource * number of bytes. However, this may not be possible due to resource
* contraints or datatype alignment/offset. The PTL is responsible for * contraints or datatype alignment/offset. The PTL is responsible for
* updating the number of bytes actually fragmented and queued for delivery * updating the number of bytes actually fragmented and queued for delivery
* on the send request (mca_pml_base_send_request.req_offset) to reflect * on the send request (mca_ptl_base_send_request.req_offset) to reflect
* the current offset into the send buffer. * the current offset into the send buffer.
* *
* If the request is larger than ptl_first_frag_size, the remainder of * If the request is larger than ptl_first_frag_size, the remainder of
@ -258,8 +258,8 @@
struct mca_ptl_base_module_t; struct mca_ptl_base_module_t;
struct mca_ptl_base_peer_t; struct mca_ptl_base_peer_t;
struct mca_ptl_base_fragment_t; struct mca_ptl_base_fragment_t;
struct mca_pml_base_recv_request_t; struct mca_ptl_base_recv_request_t;
struct mca_pml_base_send_request_t; struct mca_ptl_base_send_request_t;
struct mca_ptl_base_recv_frag_t; struct mca_ptl_base_recv_frag_t;
struct mca_ptl_base_send_frag_t; struct mca_ptl_base_send_frag_t;
struct mca_ptl_base_match_header_t; struct mca_ptl_base_match_header_t;
@ -458,7 +458,7 @@ typedef int (*mca_ptl_base_module_del_procs_fn_t)(
*/ */
typedef int (*mca_ptl_base_module_request_init_fn_t)( typedef int (*mca_ptl_base_module_request_init_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* request struct mca_ptl_base_send_request_t* request
); );
@ -477,7 +477,7 @@ typedef int (*mca_ptl_base_module_request_init_fn_t)(
typedef void (*mca_ptl_base_module_request_fini_fn_t)( typedef void (*mca_ptl_base_module_request_fini_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* request struct mca_ptl_base_send_request_t* request
); );
/** /**
@ -506,7 +506,7 @@ typedef void (*mca_ptl_base_module_request_fini_fn_t)(
typedef int (*mca_ptl_base_module_send_fn_t)( typedef int (*mca_ptl_base_module_send_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_base_peer, struct mca_ptl_base_peer_t* ptl_base_peer,
struct mca_pml_base_send_request_t* request, struct mca_ptl_base_send_request_t* request,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags
@ -541,7 +541,7 @@ typedef int (*mca_ptl_base_module_send_fn_t)(
typedef int (*mca_ptl_base_module_put_fn_t)( typedef int (*mca_ptl_base_module_put_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_base_peer, struct mca_ptl_base_peer_t* ptl_base_peer,
struct mca_pml_base_send_request_t* request, struct mca_ptl_base_send_request_t* request,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags
@ -566,7 +566,7 @@ typedef int (*mca_ptl_base_module_put_fn_t)(
typedef int (*mca_ptl_base_module_get_fn_t)( typedef int (*mca_ptl_base_module_get_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_base_peer, struct mca_ptl_base_peer_t* ptl_base_peer,
struct mca_pml_base_recv_request_t* request, struct mca_ptl_base_recv_request_t* request,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags
@ -637,7 +637,7 @@ typedef void (*mca_ptl_base_module_matched_fn_t)(
*/ */
typedef void (*mca_ptl_base_module_recv_progress_fn_t)( typedef void (*mca_ptl_base_module_recv_progress_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_recv_request_t* recv_request, struct mca_ptl_base_recv_request_t* recv_request,
size_t bytes_received, size_t bytes_received,
size_t bytes_delivered size_t bytes_delivered
); );
@ -657,7 +657,7 @@ typedef void (*mca_ptl_base_module_recv_progress_fn_t)(
typedef void (*mca_ptl_base_module_send_progress_fn_t)( typedef void (*mca_ptl_base_module_send_progress_fn_t)(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* send_request, struct mca_ptl_base_send_request_t* send_request,
size_t bytes_sent size_t bytes_sent
); );

Просмотреть файл

@ -101,13 +101,13 @@ int mca_ptl_self_finalize(struct mca_ptl_base_module_t* ptl)
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
int mca_ptl_self_request_init(struct mca_ptl_base_module_t* ptl, mca_pml_base_send_request_t* request) int mca_ptl_self_request_init(struct mca_ptl_base_module_t* ptl, mca_ptl_base_send_request_t* request)
{ {
OBJ_CONSTRUCT(request+1, mca_ptl_base_recv_frag_t); OBJ_CONSTRUCT(request+1, mca_ptl_base_recv_frag_t);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
void mca_ptl_self_request_fini(struct mca_ptl_base_module_t* ptl, mca_pml_base_send_request_t* request) void mca_ptl_self_request_fini(struct mca_ptl_base_module_t* ptl, mca_ptl_base_send_request_t* request)
{ {
OBJ_DESTRUCT(request+1); OBJ_DESTRUCT(request+1);
} }
@ -122,7 +122,7 @@ void mca_ptl_self_request_fini(struct mca_ptl_base_module_t* ptl, mca_pml_base_s
int mca_ptl_self_send( int mca_ptl_self_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_base_peer, struct mca_ptl_base_peer_t* ptl_base_peer,
struct mca_pml_base_send_request_t* request, struct mca_ptl_base_send_request_t* request,
size_t offset, size_t offset,
size_t size, size_t size,
int flags ) int flags )
@ -133,16 +133,16 @@ int mca_ptl_self_send(
hdr->hdr_common.hdr_type = MCA_PTL_HDR_TYPE_MATCH; hdr->hdr_common.hdr_type = MCA_PTL_HDR_TYPE_MATCH;
hdr->hdr_common.hdr_flags = flags; hdr->hdr_common.hdr_flags = flags;
hdr->hdr_match.hdr_contextid = request->req_base.req_comm->c_contextid; hdr->hdr_match.hdr_contextid = request->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = request->req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_src = request->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_dst = request->req_base.req_peer; hdr->hdr_match.hdr_dst = request->req_send.req_base.req_peer;
hdr->hdr_match.hdr_tag = request->req_base.req_tag; hdr->hdr_match.hdr_tag = request->req_send.req_base.req_tag;
hdr->hdr_match.hdr_msg_length = request->req_bytes_packed; hdr->hdr_match.hdr_msg_length = request->req_send.req_bytes_packed;
hdr->hdr_match.hdr_msg_seq = request->req_base.req_sequence; hdr->hdr_match.hdr_msg_seq = request->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_src_ptr.lval = 0; hdr->hdr_rndv.hdr_src_ptr.lval = 0;
hdr->hdr_rndv.hdr_src_ptr.pval = request; hdr->hdr_rndv.hdr_src_ptr.pval = request;
req->req_frag.frag_base.frag_peer = ptl_base_peer; req->req_frag.frag_base.frag_peer = ptl_base_peer;
req->req_frag.frag_base.frag_size = request->req_bytes_packed; req->req_frag.frag_base.frag_size = request->req_send.req_bytes_packed;
req->req_frag.frag_base.frag_owner = &mca_ptl_self_module; req->req_frag.frag_base.frag_owner = &mca_ptl_self_module;
req->req_frag.frag_request = NULL; req->req_frag.frag_request = NULL;
req->req_frag.frag_is_buffered = 0; req->req_frag.frag_is_buffered = 0;
@ -169,10 +169,10 @@ void mca_ptl_self_matched( mca_ptl_base_module_t* ptl,
{ {
mca_ptl_self_send_request_t* sendreq = (mca_ptl_self_send_request_t*) mca_ptl_self_send_request_t* sendreq = (mca_ptl_self_send_request_t*)
frag->frag_base.frag_header.hdr_rndv.hdr_src_ptr.pval; frag->frag_base.frag_header.hdr_rndv.hdr_src_ptr.pval;
mca_pml_base_recv_request_t* recvreq = frag->frag_request; mca_ptl_base_recv_request_t* recvreq = frag->frag_request;
if( (recvreq->req_base.req_count != 0) && if( (recvreq->req_recv.req_base.req_count != 0) &&
(sendreq->req_send.req_base.req_count != 0) ) { (sendreq->req_ptl.req_send.req_base.req_count != 0) ) {
/* Did you have the same datatype or not ? If yes we can use an optimized version /* Did you have the same datatype or not ? If yes we can use an optimized version
* for the copy function, if not we have to use a temporary buffer to pack/unpack * for the copy function, if not we have to use a temporary buffer to pack/unpack
* *
@ -180,12 +180,12 @@ void mca_ptl_self_matched( mca_ptl_base_module_t* ptl,
* a contigous buffer and the convertor on the send request initialized to point * a contigous buffer and the convertor on the send request initialized to point
* into this buffer. * into this buffer.
*/ */
if( sendreq->req_send.req_datatype == recvreq->req_base.req_datatype ) { if( sendreq->req_ptl.req_send.req_datatype == recvreq->req_recv.req_base.req_datatype ) {
ompi_ddt_copy_content_same_ddt( recvreq->req_base.req_datatype, ompi_ddt_copy_content_same_ddt( recvreq->req_recv.req_base.req_datatype,
recvreq->req_base.req_count > sendreq->req_send.req_count ? recvreq->req_recv.req_base.req_count > sendreq->req_ptl.req_send.req_count ?
sendreq->req_send.req_count : recvreq->req_base.req_count, sendreq->req_ptl.req_send.req_count : recvreq->req_recv.req_base.req_count,
(char *)recvreq->req_base.req_addr, (char *)recvreq->req_recv.req_base.req_addr,
(const char *)sendreq->req_send.req_addr ); (const char *)sendreq->req_ptl.req_send.req_addr );
} else { } else {
ompi_convertor_t *pSendConvertor, *pRecvConvertor; ompi_convertor_t *pSendConvertor, *pRecvConvertor;
struct iovec iov[1]; struct iovec iov[1];
@ -199,10 +199,10 @@ void mca_ptl_self_matched( mca_ptl_base_module_t* ptl,
ompi_convertor_init_for_recv( &frag->frag_base.frag_convertor, ompi_convertor_init_for_recv( &frag->frag_base.frag_convertor,
0, 0,
recvreq->req_base.req_datatype, recvreq->req_recv.req_base.req_datatype,
recvreq->req_base.req_count, recvreq->req_recv.req_base.req_count,
recvreq->req_base.req_addr, 0, NULL ); recvreq->req_recv.req_base.req_addr, 0, NULL );
pSendConvertor = &(sendreq->req_send.req_convertor); pSendConvertor = &(sendreq->req_ptl.req_send.req_convertor);
pRecvConvertor = &(frag->frag_base.frag_convertor); pRecvConvertor = &(frag->frag_base.frag_convertor);
completed = 0; completed = 0;
freeAfter = 0; freeAfter = 0;
@ -221,8 +221,8 @@ void mca_ptl_self_matched( mca_ptl_base_module_t* ptl,
free( buf ); free( buf );
} }
} }
ptl->ptl_send_progress( ptl, &sendreq->req_send, ptl->ptl_send_progress( ptl, &sendreq->req_ptl,
sendreq->req_send.req_bytes_packed ); sendreq->req_ptl.req_send.req_bytes_packed );
ptl->ptl_recv_progress( ptl, ptl->ptl_recv_progress( ptl,
recvreq, recvreq,
frag->frag_base.frag_header.hdr_match.hdr_msg_length, frag->frag_base.frag_header.hdr_match.hdr_msg_length,

Просмотреть файл

@ -24,8 +24,8 @@
#include "event/event.h" #include "event/event.h"
#include "mca/pml/pml.h" #include "mca/pml/pml.h"
#include "mca/ptl/ptl.h" #include "mca/ptl/ptl.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/pml/base/pml_base_recvreq.h" #include "mca/ptl/base/ptl_base_recvreq.h"
#include "mca/ptl/base/ptl_base_recvfrag.h" #include "mca/ptl/base/ptl_base_recvfrag.h"
#if defined(c_plusplus) || defined(__cplusplus) #if defined(c_plusplus) || defined(__cplusplus)
extern "C" { extern "C" {
@ -54,7 +54,7 @@ typedef struct mca_ptl_self_component_t mca_ptl_self_component_t;
* base send request, and the base receive fragment which will be used to do the match. * base send request, and the base receive fragment which will be used to do the match.
*/ */
struct mca_ptl_self_send_request_t { struct mca_ptl_self_send_request_t {
mca_pml_base_send_request_t req_send; mca_ptl_base_send_request_t req_ptl;
mca_ptl_base_recv_frag_t req_frag; /* first fragment */ mca_ptl_base_recv_frag_t req_frag; /* first fragment */
}; };
typedef struct mca_ptl_self_send_request_t mca_ptl_self_send_request_t; typedef struct mca_ptl_self_send_request_t mca_ptl_self_send_request_t;
@ -91,9 +91,9 @@ extern mca_ptl_base_module_t** mca_ptl_self_component_init(
int mca_ptl_self_add_proc(struct mca_ptl_base_module_t* ptl, size_t nprocs, struct ompi_proc_t **ompi_proc, struct mca_ptl_base_peer_t** peer_ret, struct ompi_bitmap_t* reachable); int mca_ptl_self_add_proc(struct mca_ptl_base_module_t* ptl, size_t nprocs, struct ompi_proc_t **ompi_proc, struct mca_ptl_base_peer_t** peer_ret, struct ompi_bitmap_t* reachable);
int mca_ptl_self_del_proc(struct mca_ptl_base_module_t* ptl, size_t nprocs, struct ompi_proc_t **proc, struct mca_ptl_base_peer_t** ptl_peer); int mca_ptl_self_del_proc(struct mca_ptl_base_module_t* ptl, size_t nprocs, struct ompi_proc_t **proc, struct mca_ptl_base_peer_t** ptl_peer);
int mca_ptl_self_finalize(struct mca_ptl_base_module_t* ptl); int mca_ptl_self_finalize(struct mca_ptl_base_module_t* ptl);
int mca_ptl_self_request_init(struct mca_ptl_base_module_t* ptl, struct mca_pml_base_send_request_t* request); int mca_ptl_self_request_init(struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_send_request_t* request);
void mca_ptl_self_request_fini(struct mca_ptl_base_module_t* ptl, struct mca_pml_base_send_request_t* request); void mca_ptl_self_request_fini(struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_send_request_t* request);
int mca_ptl_self_send( struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_peer_t* ptl_base_peer, struct mca_pml_base_send_request_t* request, int mca_ptl_self_send( struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_peer_t* ptl_base_peer, struct mca_ptl_base_send_request_t* request,
size_t offset, size_t size, int flags ); size_t offset, size_t size, int flags );
void mca_ptl_self_matched( mca_ptl_base_module_t* ptl, mca_ptl_base_recv_frag_t* frag ); void mca_ptl_self_matched( mca_ptl_base_module_t* ptl, mca_ptl_base_recv_frag_t* frag );

Просмотреть файл

@ -27,8 +27,8 @@
#include "util/if.h" #include "util/if.h"
#include "mca/pml/pml.h" #include "mca/pml/pml.h"
#include "mca/ptl/ptl.h" #include "mca/ptl/ptl.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/pml/base/pml_base_recvreq.h" #include "mca/ptl/base/ptl_base_recvreq.h"
#include "mca/ptl/base/ptl_base_header.h" #include "mca/ptl/base/ptl_base_header.h"
#include "mca/ptl/base/ptl_base_sendfrag.h" #include "mca/ptl/base/ptl_base_sendfrag.h"
#include "mca/ptl/base/ptl_base_recvfrag.h" #include "mca/ptl/base/ptl_base_recvfrag.h"
@ -55,7 +55,7 @@ mca_ptl_sm_t mca_ptl_sm[2] = {
&mca_ptl_sm_component.super, &mca_ptl_sm_component.super,
20, /* number of elements in the send descriptor cache */ 20, /* number of elements in the send descriptor cache */
sizeof(mca_ptl_sm_send_request_t) - sizeof(mca_ptl_sm_send_request_t) -
sizeof(mca_pml_base_send_request_t), /* size of shared memory send sizeof(mca_ptl_base_send_request_t), /* size of shared memory send
descriptor */ descriptor */
1, /* ptl_first_frag_size */ 1, /* ptl_first_frag_size */
0, /* ptl_min_frag_size */ 0, /* ptl_min_frag_size */
@ -80,7 +80,7 @@ mca_ptl_sm_t mca_ptl_sm[2] = {
&mca_ptl_sm_component.super, &mca_ptl_sm_component.super,
20, /* number of elements in the send descriptor cache */ 20, /* number of elements in the send descriptor cache */
sizeof(mca_ptl_sm_send_request_t) - sizeof(mca_ptl_sm_send_request_t) -
sizeof(mca_pml_base_send_request_t), /* size of shared memory sizeof(mca_ptl_base_send_request_t), /* size of shared memory
send descriptor */ send descriptor */
1, /* ptl_first_frag_size */ 1, /* ptl_first_frag_size */
0, /* ptl_min_frag_size */ 0, /* ptl_min_frag_size */
@ -711,7 +711,7 @@ int mca_ptl_sm_finalize(struct mca_ptl_base_module_t* ptl)
} }
int mca_ptl_sm_request_alloc(struct mca_ptl_base_module_t* ptl, struct mca_pml_base_send_request_t* request) int mca_ptl_sm_request_alloc(struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_send_request_t* request)
{ {
mca_ptl_sm_send_request_t *sm_request; mca_ptl_sm_send_request_t *sm_request;
ompi_list_item_t* item; ompi_list_item_t* item;
@ -731,7 +731,7 @@ int mca_ptl_sm_request_alloc(struct mca_ptl_base_module_t* ptl, struct mca_pml_b
} }
void mca_ptl_sm_request_return(struct mca_ptl_base_module_t* ptl, struct mca_pml_base_send_request_t* request) void mca_ptl_sm_request_return(struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_send_request_t* request)
{ {
mca_ptl_sm_send_request_t *sm_request; mca_ptl_sm_send_request_t *sm_request;
ompi_list_item_t* item; ompi_list_item_t* item;
@ -763,7 +763,7 @@ void mca_ptl_sm_request_return(struct mca_ptl_base_module_t* ptl, struct mca_pml
int mca_ptl_sm_send( int mca_ptl_sm_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t* sendreq, struct mca_ptl_base_send_request_t* sendreq,
size_t offset, size_t offset,
size_t size, size_t size,
int flags) int flags)
@ -794,11 +794,11 @@ int mca_ptl_sm_send(
int free_after=0; int free_after=0;
struct iovec address; struct iovec address;
convertor = &sendreq->req_convertor; convertor = &sendreq->req_send.req_convertor;
ompi_convertor_init_for_send( convertor, 0, ompi_convertor_init_for_send( convertor, 0,
sendreq->req_datatype, sendreq->req_send.req_datatype,
sendreq->req_count, sendreq->req_send.req_count,
sendreq->req_addr, sendreq->req_send.req_addr,
offset, NULL); offset, NULL);
sm_data_ptr=sm_request->req_frag->buff; sm_data_ptr=sm_request->req_frag->buff;
@ -822,11 +822,11 @@ int mca_ptl_sm_send(
hdr = &(send_frag->super.frag_base.frag_header); hdr = &(send_frag->super.frag_base.frag_header);
hdr->hdr_common.hdr_type = MCA_PTL_HDR_TYPE_MATCH; hdr->hdr_common.hdr_type = MCA_PTL_HDR_TYPE_MATCH;
hdr->hdr_common.hdr_flags = flags; hdr->hdr_common.hdr_flags = flags;
hdr->hdr_match.hdr_contextid = sendreq->req_base.req_comm->c_contextid; hdr->hdr_match.hdr_contextid = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_dst = sendreq->req_base.req_peer; hdr->hdr_match.hdr_dst = sendreq->req_send.req_base.req_peer;
hdr->hdr_match.hdr_tag = sendreq->req_base.req_tag; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_msg_length = sendreq->req_bytes_packed; hdr->hdr_match.hdr_msg_length = sendreq->req_send.req_bytes_packed;
/* update the offset within the payload */ /* update the offset within the payload */
sendreq->req_offset += size; sendreq->req_offset += size;
@ -879,8 +879,8 @@ int mca_ptl_sm_send(
ompi_atomic_unlock(&send_fifo->head_lock); ompi_atomic_unlock(&send_fifo->head_lock);
/* if this is the entire message - signal request is complete */ /* if this is the entire message - signal request is complete */
if(sendreq->req_bytes_packed == size) { if(sendreq->req_send.req_bytes_packed == size) {
ompi_request_complete( &(sendreq->req_base.req_ompi) ); ompi_request_complete( &(sendreq->req_send.req_base.req_ompi) );
} }
/* return */ /* return */
@ -897,7 +897,7 @@ int mca_ptl_sm_send(
int mca_ptl_sm_send_continue( int mca_ptl_sm_send_continue(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t* sendreq, struct mca_ptl_base_send_request_t* sendreq,
size_t offset, size_t offset,
size_t size, size_t size,
int flags) int flags)
@ -928,11 +928,11 @@ int mca_ptl_sm_send_continue(
} }
/* pack data in payload buffer */ /* pack data in payload buffer */
convertor = &sendreq->req_convertor; convertor = &sendreq->req_send.req_convertor;
ompi_convertor_init_for_send( convertor, 0, ompi_convertor_init_for_send( convertor, 0,
sendreq->req_datatype, sendreq->req_send.req_datatype,
sendreq->req_count, sendreq->req_send.req_count,
sendreq->req_addr, sendreq->req_send.req_addr,
offset, NULL); offset, NULL);
sm_data_ptr=send_frag->buff; sm_data_ptr=send_frag->buff;

Просмотреть файл

@ -274,7 +274,7 @@ extern int mca_ptl_sm_del_procs(
*/ */
extern int mca_ptl_sm_request_alloc( extern int mca_ptl_sm_request_alloc(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
/** /**
@ -286,7 +286,7 @@ extern int mca_ptl_sm_request_alloc(
*/ */
extern void mca_ptl_sm_request_return( extern void mca_ptl_sm_request_return(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
/** /**
@ -330,7 +330,7 @@ extern int mca_ptl_sm_del_procs(
*/ */
extern int mca_ptl_sm_request_alloc( extern int mca_ptl_sm_request_alloc(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
/** /**
@ -342,7 +342,7 @@ extern int mca_ptl_sm_request_alloc(
*/ */
extern void mca_ptl_sm_request_return( extern void mca_ptl_sm_request_return(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
/** /**
@ -388,7 +388,7 @@ extern void mca_ptl_sm_matched_common_base_addr(
extern int mca_ptl_sm_send( extern int mca_ptl_sm_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags
@ -407,7 +407,7 @@ extern int mca_ptl_sm_send(
extern int mca_ptl_sm_send_continue( extern int mca_ptl_sm_send_continue(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags

Просмотреть файл

@ -293,7 +293,7 @@ mca_ptl_base_module_t** mca_ptl_sm_component_init(
for( i=0 ; i < 2 ; i++ ) { for( i=0 ; i < 2 ; i++ ) {
mca_ptl_sm[i].super.ptl_cache_size=mca_ptl_sm_component.sm_first_frag_free_list_max; mca_ptl_sm[i].super.ptl_cache_size=mca_ptl_sm_component.sm_first_frag_free_list_max;
mca_ptl_sm[i].super.ptl_cache_bytes=sizeof(mca_ptl_sm_send_request_t) - mca_ptl_sm[i].super.ptl_cache_bytes=sizeof(mca_ptl_sm_send_request_t) -
sizeof(mca_pml_base_send_request_t); sizeof(mca_ptl_base_send_request_t);
mca_ptl_sm[i].super.ptl_first_frag_size=mca_ptl_sm_component.first_fragment_size; mca_ptl_sm[i].super.ptl_first_frag_size=mca_ptl_sm_component.first_fragment_size;
mca_ptl_sm[i].super.ptl_min_frag_size=mca_ptl_sm_component.max_fragment_size; mca_ptl_sm[i].super.ptl_min_frag_size=mca_ptl_sm_component.max_fragment_size;
mca_ptl_sm[i].super.ptl_max_frag_size=mca_ptl_sm_component.max_fragment_size; mca_ptl_sm[i].super.ptl_max_frag_size=mca_ptl_sm_component.max_fragment_size;
@ -363,7 +363,7 @@ int mca_ptl_sm_component_progress(mca_ptl_tstamp_t tstamp)
ompi_fifo_t *send_fifo = NULL; ompi_fifo_t *send_fifo = NULL;
bool frag_matched; bool frag_matched;
mca_ptl_base_match_header_t *matching_header; mca_ptl_base_match_header_t *matching_header;
mca_pml_base_send_request_t *base_send_req; mca_ptl_base_send_request_t *base_send_req;
ompi_list_item_t *item; ompi_list_item_t *item;
int return_status = 0; int return_status = 0;
@ -434,10 +434,10 @@ int mca_ptl_sm_component_progress(mca_ptl_tstamp_t tstamp)
if( frag_matched ) { if( frag_matched ) {
/* deliver data, and ack */ /* deliver data, and ack */
mca_ptl_base_match_header_t *hdr = &header_ptr->super.frag_base.frag_header.hdr_match; mca_ptl_base_match_header_t *hdr = &header_ptr->super.frag_base.frag_header.hdr_match;
mca_pml_base_recv_request_t *request = header_ptr->super.frag_request; mca_ptl_base_recv_request_t *request = header_ptr->super.frag_request;
request->req_bytes_packed = hdr->hdr_msg_length; request->req_recv.req_bytes_packed = hdr->hdr_msg_length;
request->req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_src; request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_src;
request->req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_tag; request->req_recv.req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_tag;
mca_ptl_sm_matched_same_base_addr( mca_ptl_sm_matched_same_base_addr(
(mca_ptl_base_module_t *)&mca_ptl_sm, (mca_ptl_base_module_t *)&mca_ptl_sm,
(mca_ptl_base_recv_frag_t *)header_ptr); (mca_ptl_base_recv_frag_t *)header_ptr);
@ -548,10 +548,10 @@ int mca_ptl_sm_component_progress(mca_ptl_tstamp_t tstamp)
if( frag_matched ) { if( frag_matched ) {
/* deliver data, and ack */ /* deliver data, and ack */
mca_ptl_base_match_header_t *hdr = &header_ptr->super.frag_base.frag_header.hdr_match; mca_ptl_base_match_header_t *hdr = &header_ptr->super.frag_base.frag_header.hdr_match;
mca_pml_base_recv_request_t *request = header_ptr->super.frag_request; mca_ptl_base_recv_request_t *request = header_ptr->super.frag_request;
request->req_bytes_packed = hdr->hdr_msg_length; request->req_recv.req_bytes_packed = hdr->hdr_msg_length;
request->req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_src; request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_src;
request->req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_tag; request->req_recv.req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_tag;
mca_ptl_sm_matched((mca_ptl_base_module_t *)&mca_ptl_sm, mca_ptl_sm_matched((mca_ptl_base_module_t *)&mca_ptl_sm,
(mca_ptl_base_recv_frag_t *)header_ptr); (mca_ptl_base_recv_frag_t *)header_ptr);
} }

Просмотреть файл

@ -47,7 +47,7 @@ struct mca_ptl_sm_frag_t {
int queue_index; /**< local process index, cached for fast int queue_index; /**< local process index, cached for fast
acking */ acking */
struct mca_ptl_base_module_t* send_ptl; /**< PTL that is selected for first fragment */ struct mca_ptl_base_module_t* send_ptl; /**< PTL that is selected for first fragment */
struct mca_pml_base_send_request_t* send_req; struct mca_ptl_base_send_request_t* send_req;
size_t send_offset; size_t send_offset;
void *buff; /**< pointer to buffer */ void *buff; /**< pointer to buffer */
}; };

Просмотреть файл

@ -67,7 +67,7 @@ void mca_ptl_sm_matched(
mca_ptl_base_module_t* ptl, mca_ptl_base_module_t* ptl,
mca_ptl_base_recv_frag_t* frag) mca_ptl_base_recv_frag_t* frag)
{ {
mca_pml_base_recv_request_t* recv_desc; mca_ptl_base_recv_request_t* recv_desc;
mca_ptl_sm_frag_t *sm_frag_desc; mca_ptl_sm_frag_t *sm_frag_desc;
struct iovec iov; struct iovec iov;
ompi_convertor_t frag_convertor; ompi_convertor_t frag_convertor;
@ -92,7 +92,7 @@ void mca_ptl_sm_matched(
* Initialize convertor and use it to unpack data * Initialize convertor and use it to unpack data
*/ */
OBJ_CONSTRUCT(&frag_convertor, ompi_convertor_t); OBJ_CONSTRUCT(&frag_convertor, ompi_convertor_t);
proc = ompi_comm_peer_lookup(recv_desc->req_base.req_comm, proc = ompi_comm_peer_lookup(recv_desc->req_recv.req_base.req_comm,
frag->frag_base.frag_header.hdr_match.hdr_src); frag->frag_base.frag_header.hdr_match.hdr_src);
/* write over converter set on the send side */ /* write over converter set on the send side */
ompi_convertor_copy(proc->proc_convertor, ompi_convertor_copy(proc->proc_convertor,
@ -100,9 +100,9 @@ void mca_ptl_sm_matched(
ompi_convertor_init_for_recv( ompi_convertor_init_for_recv(
&frag_convertor, /* convertor */ &frag_convertor, /* convertor */
0, /* flags */ 0, /* flags */
recv_desc->req_base.req_datatype, /* datatype */ recv_desc->req_recv.req_base.req_datatype, /* datatype */
recv_desc->req_base.req_count, /* count elements */ recv_desc->req_recv.req_base.req_count, /* count elements */
recv_desc->req_base.req_addr, /* users buffer */ recv_desc->req_recv.req_base.req_addr, /* users buffer */
sm_frag_desc->send_offset, /* offset in bytes into packed buffer */ sm_frag_desc->send_offset, /* offset in bytes into packed buffer */
NULL ); /* dont allocate memory */ NULL ); /* dont allocate memory */

Просмотреть файл

@ -21,7 +21,7 @@
#include <sys/types.h> #include <sys/types.h>
#include "ompi_config.h" #include "ompi_config.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/ptl/base/ptl_base_sendfrag.h" #include "mca/ptl/base/ptl_base_sendfrag.h"
#include "ptl_sm.h" #include "ptl_sm.h"
@ -56,7 +56,7 @@ typedef struct mca_ptl_sm_send_frag_t mca_ptl_sm_send_frag_t;
int mca_ptl_sm_send_frag_init( int mca_ptl_sm_send_frag_init(
mca_ptl_sm_send_frag_t*, mca_ptl_sm_send_frag_t*,
struct mca_ptl_base_peer_t*, struct mca_ptl_base_peer_t*,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t* size, size_t* size,
int flags); int flags);

Просмотреть файл

@ -30,7 +30,7 @@ static void mca_ptl_sm_send_request_destruct(mca_ptl_sm_send_request_t*);
OBJ_CLASS_INSTANCE( OBJ_CLASS_INSTANCE(
mca_ptl_sm_send_request_t, mca_ptl_sm_send_request_t,
mca_pml_base_send_request_t, mca_ptl_base_send_request_t,
mca_ptl_sm_send_request_construct, mca_ptl_sm_send_request_construct,
mca_ptl_sm_send_request_destruct mca_ptl_sm_send_request_destruct
); );
@ -54,7 +54,7 @@ void mca_ptl_sm_send_request_destruct(mca_ptl_sm_send_request_t* request)
* to allocate fragment descriptor and payload memory * to allocate fragment descriptor and payload memory
*/ */
int mca_ptl_sm_send_request_init(struct mca_ptl_base_module_t* ptl, int mca_ptl_sm_send_request_init(struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* request) struct mca_ptl_base_send_request_t* request)
{ {
mca_ptl_sm_send_request_t *sm_request; mca_ptl_sm_send_request_t *sm_request;

Просмотреть файл

@ -22,7 +22,7 @@
#include <sys/types.h> #include <sys/types.h>
#include "ompi_config.h" #include "ompi_config.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "ptl_sm_frag.h" #include "ptl_sm_frag.h"
#if defined(c_plusplus) || defined(__cplusplus) #if defined(c_plusplus) || defined(__cplusplus)
@ -38,7 +38,7 @@ OBJ_CLASS_DECLARATION(mca_ptl_sm_send_request_t);
struct mca_ptl_sm_send_request_t { struct mca_ptl_sm_send_request_t {
/* base send descriptor */ /* base send descriptor */
mca_pml_base_send_request_t super; mca_ptl_base_send_request_t super;
/* pointer to first fragment descriptor */ /* pointer to first fragment descriptor */
mca_ptl_sm_frag_t *req_frag; mca_ptl_sm_frag_t *req_frag;
@ -50,7 +50,7 @@ typedef struct mca_ptl_sm_send_request_t mca_ptl_sm_send_request_t;
* memory send request is initialized. * memory send request is initialized.
*/ */
int mca_ptl_sm_send_request_init(struct mca_ptl_base_module_t* ptl, int mca_ptl_sm_send_request_init(struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* request); struct mca_ptl_base_send_request_t* request);
#if defined(c_plusplus) || defined(__cplusplus) #if defined(c_plusplus) || defined(__cplusplus)
} }

Просмотреть файл

@ -22,9 +22,9 @@
#include "mca/pml/pml.h" #include "mca/pml/pml.h"
#include "mca/ptl/ptl.h" #include "mca/ptl/ptl.h"
#include "mca/ptl/base/ptl_base_header.h" #include "mca/ptl/base/ptl_base_header.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/ptl/base/ptl_base_sendfrag.h" #include "mca/ptl/base/ptl_base_sendfrag.h"
#include "mca/pml/base/pml_base_recvreq.h" #include "mca/ptl/base/ptl_base_recvreq.h"
#include "mca/ptl/base/ptl_base_recvfrag.h" #include "mca/ptl/base/ptl_base_recvfrag.h"
#include "mca/base/mca_base_module_exchange.h" #include "mca/base/mca_base_module_exchange.h"
#include "ptl_tcp.h" #include "ptl_tcp.h"
@ -39,7 +39,7 @@ mca_ptl_tcp_module_t mca_ptl_tcp_module = {
{ {
&mca_ptl_tcp_component.super, &mca_ptl_tcp_component.super,
16, /* max size of request cache */ 16, /* max size of request cache */
sizeof(mca_ptl_tcp_send_request_t) - sizeof(mca_pml_base_send_request_t), /* bytes required by ptl for a request */ sizeof(mca_ptl_tcp_send_request_t) - sizeof(mca_ptl_base_send_request_t), /* bytes required by ptl for a request */
0, /* max size of first fragment */ 0, /* max size of first fragment */
0, /* min fragment size */ 0, /* min fragment size */
0, /* max fragment size */ 0, /* max fragment size */
@ -179,7 +179,7 @@ int mca_ptl_tcp_finalize(struct mca_ptl_base_module_t* ptl)
* along w/ the ptl to cache the first fragment control information. * along w/ the ptl to cache the first fragment control information.
*/ */
int mca_ptl_tcp_request_init(struct mca_ptl_base_module_t* ptl, struct mca_pml_base_send_request_t* request) int mca_ptl_tcp_request_init(struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_send_request_t* request)
{ {
OBJ_CONSTRUCT(request+1, mca_ptl_tcp_send_frag_t); OBJ_CONSTRUCT(request+1, mca_ptl_tcp_send_frag_t);
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -190,7 +190,7 @@ int mca_ptl_tcp_request_init(struct mca_ptl_base_module_t* ptl, struct mca_pml_b
* Cleanup any resources cached along w/ the request. * Cleanup any resources cached along w/ the request.
*/ */
void mca_ptl_tcp_request_fini(struct mca_ptl_base_module_t* ptl, struct mca_pml_base_send_request_t* request) void mca_ptl_tcp_request_fini(struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_send_request_t* request)
{ {
OBJ_DESTRUCT(request+1); OBJ_DESTRUCT(request+1);
} }
@ -240,7 +240,7 @@ void mca_ptl_tcp_send_frag_return(struct mca_ptl_base_module_t* ptl, struct mca_
int mca_ptl_tcp_send( int mca_ptl_tcp_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t* sendreq, struct mca_ptl_base_send_request_t* sendreq,
size_t offset, size_t offset,
size_t size, size_t size,
int flags) int flags)
@ -261,7 +261,7 @@ int mca_ptl_tcp_send(
/* must update the offset after actual fragment size is determined -- and very important -- /* must update the offset after actual fragment size is determined -- and very important --
* before attempting to send the fragment * before attempting to send the fragment
*/ */
mca_pml_base_send_request_offset(sendreq, size); mca_ptl_base_send_request_offset(sendreq, size);
return mca_ptl_tcp_peer_send(ptl_peer, sendfrag, offset); return mca_ptl_tcp_peer_send(ptl_peer, sendfrag, offset);
} }

Просмотреть файл

@ -197,7 +197,7 @@ extern int mca_ptl_tcp_del_procs(
*/ */
extern int mca_ptl_tcp_request_init( extern int mca_ptl_tcp_request_init(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
/** /**
@ -209,7 +209,7 @@ extern int mca_ptl_tcp_request_init(
*/ */
extern void mca_ptl_tcp_request_fini( extern void mca_ptl_tcp_request_fini(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_pml_base_send_request_t* struct mca_ptl_base_send_request_t*
); );
/** /**
@ -237,7 +237,7 @@ extern void mca_ptl_tcp_matched(
extern int mca_ptl_tcp_send( extern int mca_ptl_tcp_send(
struct mca_ptl_base_module_t* ptl, struct mca_ptl_base_module_t* ptl,
struct mca_ptl_base_peer_t* ptl_peer, struct mca_ptl_base_peer_t* ptl_peer,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t size, size_t size,
int flags int flags

Просмотреть файл

@ -21,7 +21,7 @@
#include <sys/types.h> #include <sys/types.h>
#endif #endif
#include "include/ompi_socket_errno.h" #include "include/ompi_socket_errno.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "ptl_tcp.h" #include "ptl_tcp.h"
#include "ptl_tcp_peer.h" #include "ptl_tcp_peer.h"
#include "ptl_tcp_recvfrag.h" #include "ptl_tcp_recvfrag.h"
@ -184,7 +184,7 @@ static bool mca_ptl_tcp_recv_frag_header(mca_ptl_tcp_recv_frag_t* frag, int sd,
static bool mca_ptl_tcp_recv_frag_ack(mca_ptl_tcp_recv_frag_t* frag, int sd) static bool mca_ptl_tcp_recv_frag_ack(mca_ptl_tcp_recv_frag_t* frag, int sd)
{ {
mca_ptl_tcp_send_frag_t* sendfrag; mca_ptl_tcp_send_frag_t* sendfrag;
mca_pml_base_send_request_t* sendreq; mca_ptl_base_send_request_t* sendreq;
sendfrag = (mca_ptl_tcp_send_frag_t*)frag->frag_recv.frag_base.frag_header.hdr_ack.hdr_src_ptr.pval; sendfrag = (mca_ptl_tcp_send_frag_t*)frag->frag_recv.frag_base.frag_header.hdr_ack.hdr_src_ptr.pval;
sendreq = sendfrag->frag_send.frag_request; sendreq = sendfrag->frag_send.frag_request;
sendreq->req_peer_match = frag->frag_recv.frag_base.frag_header.hdr_ack.hdr_dst_match; sendreq->req_peer_match = frag->frag_recv.frag_base.frag_header.hdr_ack.hdr_dst_match;
@ -227,7 +227,7 @@ static bool mca_ptl_tcp_recv_frag_frag(mca_ptl_tcp_recv_frag_t* frag, int sd)
{ {
/* get request from header */ /* get request from header */
if(frag->frag_msg_cnt == 0) { if(frag->frag_msg_cnt == 0) {
frag->frag_recv.frag_request = (mca_pml_base_recv_request_t *)frag->frag_recv.frag_base.frag_header.hdr_frag.hdr_dst_ptr.pval; frag->frag_recv.frag_request = (mca_ptl_base_recv_request_t *)frag->frag_recv.frag_base.frag_header.hdr_frag.hdr_dst_ptr.pval;
mca_ptl_tcp_recv_frag_matched(frag, mca_ptl_tcp_recv_frag_matched(frag,
frag->frag_recv.frag_base.frag_header.hdr_frag.hdr_frag_offset, frag->frag_recv.frag_base.frag_header.hdr_frag.hdr_frag_offset,
frag->frag_recv.frag_base.frag_header.hdr_frag.hdr_frag_length); frag->frag_recv.frag_base.frag_header.hdr_frag.hdr_frag_length);

Просмотреть файл

@ -89,21 +89,21 @@ static inline void mca_ptl_tcp_recv_frag_matched(
size_t frag_offset, size_t frag_offset,
size_t frag_length) size_t frag_length)
{ {
mca_pml_base_recv_request_t* request = frag->frag_recv.frag_request; mca_ptl_base_recv_request_t* request = frag->frag_recv.frag_request;
/* if there is data associated with the fragment -- setup to receive */ /* if there is data associated with the fragment -- setup to receive */
if(frag_length > 0) { if(frag_length > 0) {
/* initialize receive convertor */ /* initialize receive convertor */
ompi_proc_t *proc = ompi_proc_t *proc =
ompi_comm_peer_lookup(request->req_base.req_comm, ompi_comm_peer_lookup(request->req_recv.req_base.req_comm,
request->req_base.req_ompi.req_status.MPI_SOURCE); request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE);
ompi_convertor_copy(proc->proc_convertor, &frag->frag_recv.frag_base.frag_convertor); ompi_convertor_copy(proc->proc_convertor, &frag->frag_recv.frag_base.frag_convertor);
ompi_convertor_init_for_recv( ompi_convertor_init_for_recv(
&frag->frag_recv.frag_base.frag_convertor, /* convertor */ &frag->frag_recv.frag_base.frag_convertor, /* convertor */
0, /* flags */ 0, /* flags */
request->req_base.req_datatype, /* datatype */ request->req_recv.req_base.req_datatype, /* datatype */
request->req_base.req_count, /* count elements */ request->req_recv.req_base.req_count, /* count elements */
request->req_base.req_addr, /* users buffer */ request->req_recv.req_base.req_addr, /* users buffer */
0, /* offset in bytes into packed buffer */ 0, /* offset in bytes into packed buffer */
mca_ptl_tcp_memalloc ); /* not allocating memory */ mca_ptl_tcp_memalloc ); /* not allocating memory */
@ -115,16 +115,16 @@ static inline void mca_ptl_tcp_recv_frag_matched(
} else { } else {
long true_lb, true_extent; long true_lb, true_extent;
ompi_ddt_get_true_extent( request->req_base.req_datatype, &true_lb, &true_extent ); ompi_ddt_get_true_extent( request->req_recv.req_base.req_datatype, &true_lb, &true_extent );
frag->frag_recv.frag_base.frag_addr = ((unsigned char*)request->req_base.req_addr) + frag->frag_recv.frag_base.frag_addr = ((unsigned char*)request->req_recv.req_base.req_addr) +
frag_offset + true_lb; frag_offset + true_lb;
} }
frag->frag_size = frag_length; frag->frag_size = frag_length;
if(frag_offset + frag_length > request->req_bytes_packed) { if(frag_offset + frag_length > request->req_recv.req_bytes_packed) {
if(frag_offset > request->req_bytes_packed) if(frag_offset > request->req_recv.req_bytes_packed)
frag->frag_recv.frag_base.frag_size = 0; frag->frag_recv.frag_base.frag_size = 0;
else else
frag->frag_recv.frag_base.frag_size = request->req_bytes_packed - frag_offset; frag->frag_recv.frag_base.frag_size = request->req_recv.req_bytes_packed - frag_offset;
} else { } else {
frag->frag_recv.frag_base.frag_size = frag_length; frag->frag_recv.frag_base.frag_size = frag_length;
} }
@ -151,7 +151,7 @@ static inline void mca_ptl_tcp_recv_frag_progress(mca_ptl_tcp_recv_frag_t* frag)
(frag->frag_recv.frag_base.frag_header.hdr_common.hdr_type != MCA_PTL_HDR_TYPE_FRAG && (frag->frag_recv.frag_base.frag_header.hdr_common.hdr_type != MCA_PTL_HDR_TYPE_FRAG &&
frag_progressed == 2 )) { frag_progressed == 2 )) {
mca_pml_base_recv_request_t* request = frag->frag_recv.frag_request; mca_ptl_base_recv_request_t* request = frag->frag_recv.frag_request;
if(frag->frag_recv.frag_is_buffered) { if(frag->frag_recv.frag_is_buffered) {
mca_ptl_base_frag_header_t* header = &(frag)->frag_recv.frag_base.frag_header.hdr_frag; mca_ptl_base_frag_header_t* header = &(frag)->frag_recv.frag_base.frag_header.hdr_frag;
size_t offset = (header->hdr_common.hdr_type == MCA_PTL_HDR_TYPE_FRAG) ? size_t offset = (header->hdr_common.hdr_type == MCA_PTL_HDR_TYPE_FRAG) ?
@ -162,15 +162,15 @@ static inline void mca_ptl_tcp_recv_frag_progress(mca_ptl_tcp_recv_frag_t* frag)
*/ */
struct iovec iov; struct iovec iov;
ompi_proc_t *proc = ompi_proc_t *proc =
ompi_comm_peer_lookup(request->req_base.req_comm, ompi_comm_peer_lookup(request->req_recv.req_base.req_comm,
request->req_base.req_ompi.req_status.MPI_SOURCE); request->req_recv.req_base.req_ompi.req_status.MPI_SOURCE);
ompi_convertor_copy(proc->proc_convertor, &frag->frag_recv.frag_base.frag_convertor); ompi_convertor_copy(proc->proc_convertor, &frag->frag_recv.frag_base.frag_convertor);
ompi_convertor_init_for_recv( ompi_convertor_init_for_recv(
&frag->frag_recv.frag_base.frag_convertor, /* convertor */ &frag->frag_recv.frag_base.frag_convertor, /* convertor */
0, /* flags */ 0, /* flags */
request->req_base.req_datatype, /* datatype */ request->req_recv.req_base.req_datatype, /* datatype */
request->req_base.req_count, /* count elements */ request->req_recv.req_base.req_count, /* count elements */
request->req_base.req_addr, /* users buffer */ request->req_recv.req_base.req_addr, /* users buffer */
offset, /* offset in bytes into packed buffer */ offset, /* offset in bytes into packed buffer */
NULL ); /* dont allocate memory */ NULL ); /* dont allocate memory */

Просмотреть файл

@ -78,7 +78,7 @@ void* mca_ptl_tcp_memalloc( size_t* length )
int mca_ptl_tcp_send_frag_init( int mca_ptl_tcp_send_frag_init(
mca_ptl_tcp_send_frag_t* sendfrag, mca_ptl_tcp_send_frag_t* sendfrag,
mca_ptl_base_peer_t* ptl_peer, mca_ptl_base_peer_t* ptl_peer,
mca_pml_base_send_request_t* sendreq, mca_ptl_base_send_request_t* sendreq,
size_t offset, size_t offset,
size_t* size, size_t* size,
int flags) int flags)
@ -96,13 +96,13 @@ int mca_ptl_tcp_send_frag_init(
int rc; int rc;
convertor = &sendfrag->frag_convertor; convertor = &sendfrag->frag_convertor;
ompi_convertor_copy(&sendreq->req_convertor, convertor); ompi_convertor_copy(&sendreq->req_send.req_convertor, convertor);
ompi_convertor_init_for_send( ompi_convertor_init_for_send(
convertor, convertor,
0, 0,
sendreq->req_datatype, sendreq->req_send.req_datatype,
sendreq->req_count, sendreq->req_send.req_count,
sendreq->req_addr, sendreq->req_send.req_addr,
offset, offset,
mca_ptl_tcp_memalloc ); mca_ptl_tcp_memalloc );
@ -130,12 +130,12 @@ int mca_ptl_tcp_send_frag_init(
if(offset == 0) { if(offset == 0) {
hdr->hdr_common.hdr_type = (flags & MCA_PTL_FLAGS_ACK) ? MCA_PTL_HDR_TYPE_MATCH : MCA_PTL_HDR_TYPE_RNDV; hdr->hdr_common.hdr_type = (flags & MCA_PTL_FLAGS_ACK) ? MCA_PTL_HDR_TYPE_MATCH : MCA_PTL_HDR_TYPE_RNDV;
hdr->hdr_common.hdr_flags = flags; hdr->hdr_common.hdr_flags = flags;
hdr->hdr_match.hdr_contextid = sendreq->req_base.req_comm->c_contextid; hdr->hdr_match.hdr_contextid = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_dst = sendreq->req_base.req_peer; hdr->hdr_match.hdr_dst = sendreq->req_send.req_base.req_peer;
hdr->hdr_match.hdr_tag = sendreq->req_base.req_tag; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_msg_length = sendreq->req_bytes_packed; hdr->hdr_match.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_match.hdr_msg_seq = sendreq->req_base.req_sequence; hdr->hdr_match.hdr_msg_seq = sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_frag_length = size_out; hdr->hdr_rndv.hdr_frag_length = size_out;
hdr->hdr_rndv.hdr_src_ptr.lval = 0; /* for VALGRIND/PURIFY - REPLACE WITH MACRO */ hdr->hdr_rndv.hdr_src_ptr.lval = 0; /* for VALGRIND/PURIFY - REPLACE WITH MACRO */
hdr->hdr_rndv.hdr_src_ptr.pval = sendfrag; hdr->hdr_rndv.hdr_src_ptr.pval = sendfrag;

Просмотреть файл

@ -30,7 +30,7 @@
#include <netinet/in.h> #include <netinet/in.h>
#endif #endif
#include "include/sys/atomic.h" #include "include/sys/atomic.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "mca/ptl/base/ptl_base_sendfrag.h" #include "mca/ptl/base/ptl_base_sendfrag.h"
#include "ptl_tcp.h" #include "ptl_tcp.h"
#include "ptl_tcp_recvfrag.h" #include "ptl_tcp_recvfrag.h"
@ -77,7 +77,7 @@ bool mca_ptl_tcp_send_frag_handler(mca_ptl_tcp_send_frag_t*, int sd);
int mca_ptl_tcp_send_frag_init( int mca_ptl_tcp_send_frag_init(
mca_ptl_tcp_send_frag_t*, mca_ptl_tcp_send_frag_t*,
struct mca_ptl_base_peer_t*, struct mca_ptl_base_peer_t*,
struct mca_pml_base_send_request_t*, struct mca_ptl_base_send_request_t*,
size_t offset, size_t offset,
size_t* size, size_t* size,
int flags); int flags);
@ -92,7 +92,7 @@ int mca_ptl_tcp_send_frag_init(
static inline void mca_ptl_tcp_send_frag_progress(mca_ptl_tcp_send_frag_t* frag) static inline void mca_ptl_tcp_send_frag_progress(mca_ptl_tcp_send_frag_t* frag)
{ {
mca_pml_base_send_request_t* request = frag->frag_send.frag_request; mca_ptl_base_send_request_t* request = frag->frag_send.frag_request;
bool frag_ack; bool frag_ack;
/* if this is an ack - simply return to pool */ /* if this is an ack - simply return to pool */
@ -133,15 +133,15 @@ static inline void mca_ptl_tcp_send_frag_init_ack(
mca_ptl_tcp_recv_frag_t* frag) mca_ptl_tcp_recv_frag_t* frag)
{ {
mca_ptl_base_header_t* hdr = &ack->frag_send.frag_base.frag_header; mca_ptl_base_header_t* hdr = &ack->frag_send.frag_base.frag_header;
mca_pml_base_recv_request_t* request = frag->frag_recv.frag_request; mca_ptl_base_recv_request_t* request = frag->frag_recv.frag_request;
hdr->hdr_common.hdr_type = MCA_PTL_HDR_TYPE_ACK; hdr->hdr_common.hdr_type = MCA_PTL_HDR_TYPE_ACK;
hdr->hdr_common.hdr_flags = 0; hdr->hdr_common.hdr_flags = 0;
hdr->hdr_ack.hdr_src_ptr = frag->frag_recv.frag_base.frag_header.hdr_rndv.hdr_src_ptr; hdr->hdr_ack.hdr_src_ptr = frag->frag_recv.frag_base.frag_header.hdr_rndv.hdr_src_ptr;
hdr->hdr_ack.hdr_dst_match.lval = 0; /* for VALGRIND/PURIFY - REPLACE WITH MACRO */ hdr->hdr_ack.hdr_dst_match.lval = 0; /* for VALGRIND/PURIFY - REPLACE WITH MACRO */
hdr->hdr_ack.hdr_dst_match.pval = request; hdr->hdr_ack.hdr_dst_match.pval = request;
hdr->hdr_ack.hdr_dst_addr.lval = 0; /* for VALGRIND/PURIFY - REPLACE WITH MACRO */ hdr->hdr_ack.hdr_dst_addr.lval = 0; /* for VALGRIND/PURIFY - REPLACE WITH MACRO */
hdr->hdr_ack.hdr_dst_addr.pval = request->req_base.req_addr; hdr->hdr_ack.hdr_dst_addr.pval = request->req_recv.req_base.req_addr;
hdr->hdr_ack.hdr_dst_size = request->req_bytes_packed; hdr->hdr_ack.hdr_dst_size = request->req_recv.req_bytes_packed;
ack->frag_send.frag_request = 0; ack->frag_send.frag_request = 0;
ack->frag_send.frag_base.frag_peer = ptl_peer; ack->frag_send.frag_base.frag_peer = ptl_peer;
ack->frag_send.frag_base.frag_owner = ptl; ack->frag_send.frag_base.frag_owner = ptl;

Просмотреть файл

@ -22,7 +22,7 @@
#endif #endif
#include "include/ompi_socket_errno.h" #include "include/ompi_socket_errno.h"
#include "include/types.h" #include "include/types.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "ptl_tcp.h" #include "ptl_tcp.h"
#include "ptl_tcp_sendreq.h" #include "ptl_tcp_sendreq.h"
@ -33,7 +33,7 @@ static void mca_ptl_tcp_send_request_destruct(mca_ptl_tcp_send_request_t*);
ompi_class_t mca_ptl_tcp_send_request_t_class = { ompi_class_t mca_ptl_tcp_send_request_t_class = {
"mca_ptl_tcp_send_request_t", "mca_ptl_tcp_send_request_t",
OBJ_CLASS(mca_pml_base_send_request_t), OBJ_CLASS(mca_ptl_base_send_request_t),
(ompi_construct_t)mca_ptl_tcp_send_request_construct, (ompi_construct_t)mca_ptl_tcp_send_request_construct,
(ompi_destruct_t)mca_ptl_tcp_send_request_destruct (ompi_destruct_t)mca_ptl_tcp_send_request_destruct
}; };

Просмотреть файл

@ -30,7 +30,7 @@
#include <netinet/in.h> #include <netinet/in.h>
#endif #endif
#include "ompi_config.h" #include "ompi_config.h"
#include "mca/pml/base/pml_base_sendreq.h" #include "mca/ptl/base/ptl_base_sendreq.h"
#include "ptl_tcp_sendfrag.h" #include "ptl_tcp_sendfrag.h"
#if defined(c_plusplus) || defined(__cplusplus) #if defined(c_plusplus) || defined(__cplusplus)
extern "C" { extern "C" {
@ -43,7 +43,7 @@ extern ompi_class_t mca_ptl_tcp_send_request_t_class;
* fragment on every send request. * fragment on every send request.
*/ */
struct mca_ptl_tcp_send_request_t { struct mca_ptl_tcp_send_request_t {
mca_pml_base_send_request_t super; mca_ptl_base_send_request_t super;
mca_ptl_tcp_send_frag_t req_frag; /* first fragment */ mca_ptl_tcp_send_frag_t req_frag; /* first fragment */
}; };
typedef struct mca_ptl_tcp_send_request_t mca_ptl_tcp_send_request_t; typedef struct mca_ptl_tcp_send_request_t mca_ptl_tcp_send_request_t;