/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
+#include <xbt/config.hpp>
+
#include "private.h"
#include "xbt/virtu.h"
#include "mc/mc.h"
#include "simgrid/sg_config.h"
#include "colls/colls.h"
-XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
+#include "src/simix/SynchroComm.hpp"
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
static int match_recv(void* a, void* b, smx_synchro_t ignored) {
MPI_Request ref = (MPI_Request)a;
} else return 0;
}
-
// Methods used to parse and store the values for timing injections in smpi
// These are taken from surf/network.c and generalized to have more values for each factor
typedef struct s_smpi_factor_multival *smpi_os_factor_multival_t;
int nb_values;
double values[4];//arbitrary set to 4
} s_smpi_factor_multival_t;
+
xbt_dynar_t smpi_os_values = NULL;
xbt_dynar_t smpi_or_values = NULL;
xbt_dynar_t smpi_ois_values = NULL;
-double smpi_wtime_sleep = 0.0;
-double smpi_iprobe_sleep = 1e-4;
-double smpi_test_sleep = 1e-4;
+static simgrid::config::Flag<double> smpi_wtime_sleep(
+ "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
+static simgrid::config::Flag<double> smpi_iprobe_sleep(
+ "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
+static simgrid::config::Flag<double> smpi_test_sleep(
+ "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
static int factor_cmp(const void *pa, const void *pb)
{
(((s_smpi_factor_multival_t*)pa)->factor < ((s_smpi_factor_multival_t*)pb)->factor) ? -1 : 0;
}
-
static xbt_dynar_t parse_factor(const char *smpi_coef_string)
{
char *value = NULL;
xbt_dynar_free(&radical_elements2);
}
xbt_dynar_free(&radical_elements);
- iter=0;
xbt_dynar_sort(smpi_factor, &factor_cmp);
xbt_dynar_foreach(smpi_factor, iter, fact) {
XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
static double smpi_os(double size)
{
if (!smpi_os_values) {
- smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
+ smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
}
unsigned int iter = 0;
static double smpi_ois(double size)
{
if (!smpi_ois_values) {
- smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
+ smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
}
unsigned int iter = 0;
s_smpi_factor_multival_t fact;
double current=0.0;
- // Iterate over all the sections that were specified and find the right
- // value. (fact.factor represents the interval sizes; we want to find the
- // section that has fact.factor <= size and no other such fact.factor <= size)
+ // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
+ // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
// Note: parse_factor() (used before) already sorts the dynar we iterate over!
xbt_dynar_foreach(smpi_ois_values, iter, fact) {
- if (size <= fact.factor) { // Values already too large, use the previously
- // computed value of current!
+ if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
return current;
}else{
static double smpi_or(double size)
{
if (!smpi_or_values) {
- smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
+ smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
}
unsigned int iter = 0;
s_smpi_factor_multival_t fact;
double current=0.0;
- // Iterate over all the sections that were specified and find the right
- // value. (fact.factor represents the interval sizes; we want to find the
- // section that has fact.factor <= size and no other such fact.factor <= size)
+ // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
+ // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
// Note: parse_factor() (used before) already sorts the dynar we iterate over!
xbt_dynar_foreach(smpi_or_values, iter, fact) {
if (size <= fact.factor) { // Values already too large, use the previously
return time;
}
-static MPI_Request build_request(void *buf, int count,
- MPI_Datatype datatype, int src, int dst,
- int tag, MPI_Comm comm, unsigned flags)
+static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+ unsigned flags)
{
MPI_Request request = NULL;
s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
- if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){
+ if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->sizeof_substruct != 0)){
// This part handles the problem of non-contiguous memory
old_buf = buf;
buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
- if ((datatype->has_subtype == 1) && (flags & SEND)) {
+ if ((datatype->sizeof_substruct != 0) && (flags & SEND)) {
subtype->serialize(old_buf, buf, count, datatype->substruct);
}
}
request->buf = buf;
- // This part handles the problem of non-contiguous memory (for the
- // unserialisation at the reception)
+ // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
request->old_buf = old_buf;
request->old_type = datatype;
request->size = smpi_datatype_size(datatype) * count;
+ smpi_datatype_use(datatype);
request->src = src;
request->dst = dst;
request->tag = tag;
request->comm = comm;
+ smpi_comm_use(request->comm);
request->action = NULL;
request->flags = flags;
request->detached = 0;
request->op = MPI_REPLACE;
request->send = 0;
request->recv = 0;
- if (flags & SEND) smpi_datatype_unuse(datatype);
return request;
}
-
void smpi_empty_status(MPI_Status * status)
{
if(status != MPI_STATUS_IGNORE) {
int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, PERSISTENT | SEND | PREPARED);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
return request;
}
int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, PERSISTENT | SSEND | SEND | PREPARED);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
return request;
}
int src, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
- comm, PERSISTENT | RECV | PREPARED);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype,
+ src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
+ smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
return request;
}
void smpi_mpi_start(MPI_Request request)
{
- smx_rdv_t mailbox;
-
+ smx_mailbox_t mailbox;
+
xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
request->flags &= ~PREPARED;
request->flags &= ~FINISHED;
if (request->flags & RECV) {
print_request("New recv", request);
- int async_small_thresh = sg_cfg_get_int("smpi/async_small_thresh");
+ int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
xbt_mutex_t mut = smpi_process_mailboxes_mutex();
if (async_small_thresh != 0 ||request->flags & RMA)
if (async_small_thresh == 0 && !(request->flags & RMA)) {
mailbox = smpi_process_mailbox();
- }
- else if (request->flags & RMA || static_cast<int>(request->size) < async_small_thresh){
- //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
+ } else if (request->flags & RMA || static_cast<int>(request->size) < async_small_thresh){
+ //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
+ //begin with the more appropriate one : the small one.
mailbox = smpi_process_mailbox_small();
XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
simcall_process_sleep(sleeptime);
XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
}
-
+
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
request->real_size=request->size;
- smpi_datatype_use(request->old_type);
- smpi_comm_use(request->comm);
- request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf,
- &request->real_size, &match_recv,
+ request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
!smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
- : &smpi_comm_null_copy_buffer_callback,
- request, -1.0);
+ : &smpi_comm_null_copy_buffer_callback, request, -1.0);
XBT_DEBUG("recv simcall posted");
if (async_small_thresh != 0 || (request->flags & RMA))
xbt_mutex_release(mut);
} else {
-
-
int receiver = request->dst;
int rank = request->src;
TRACE_smpi_send(rank, rank, receiver,request->size);
}
print_request("New send", request);
-
- //if we are giving back the control to the user without waiting for completion, we have to inject timings
+
+ //if we are giving back the control to the user without waiting for completion, we have to inject timings
double sleeptime = 0.0;
if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
//isend and send timings may be different
XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
}
- int async_small_thresh = sg_cfg_get_int("smpi/async_small_thresh");
+ int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
}
void* buf = request->buf;
- if ( (! (request->flags & SSEND)) && (static_cast<int>(request->size) < sg_cfg_get_int("smpi/send_is_detached_thresh"))) {
+ if ( (! (request->flags & SSEND)) && ((request->flags & RMA) ||
+ (static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh")))) {
void *oldbuf = NULL;
request->detached = 1;
XBT_DEBUG("Send request %p is detached", request);
request->refcount++;
- if(request->old_type->has_subtype == 0){
+ if(request->old_type->sizeof_substruct == 0){
oldbuf = request->buf;
if (!smpi_process_get_replaying() && oldbuf && request->size!=0){
if((smpi_privatize_global_variables)
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
request->real_size=request->size;
- smpi_datatype_use(request->old_type);
- smpi_comm_use(request->comm);
- request->action =
- simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
- buf, request->real_size,
- &match_send,
+ request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
+ buf, request->real_size, &match_send,
&xbt_free_f, // how to free the userdata if a detached send fails
!smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
- : &smpi_comm_null_copy_buffer_callback,
- request,
+ : &smpi_comm_null_copy_buffer_callback, request,
// detach if msg size < eager/rdv switch limit
request->detached);
XBT_DEBUG("send simcall posted");
-
-
/* FIXME: detached sends are not traceable (request->action == NULL) */
if (request->action)
simcall_set_category(request->action, TRACE_internal_smpi_get_category());
if (async_small_thresh != 0 || request->flags & RMA)
xbt_mutex_release(mut);
}
-
}
void smpi_mpi_startall(int count, MPI_Request * requests)
{
- int i;
if(requests==NULL) return;
- for(i = 0; i < count; i++) {
+ for(int i = 0; i < count; i++) {
smpi_mpi_start(requests[i]);
}
}
if((*request)->refcount<0) xbt_die("wrong refcount");
if((*request)->refcount==0){
+ smpi_datatype_unuse((*request)->old_type);
+ smpi_comm_unuse((*request)->comm);
print_request("Destroying", (*request));
xbt_free(*request);
*request = MPI_REQUEST_NULL;
}
}
-
-MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
- int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
+MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+ MPI_Op op)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
if(op==MPI_OP_NULL){
request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
request->op = op;
+// smpi_datatype_use(datatype);
}
return request;
}
-MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
- int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
+MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+ MPI_Op op)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
if(op==MPI_OP_NULL){
request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
request->op = op;
+// smpi_datatype_use(datatype);
}
return request;
}
-
-MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
- int dst, int tag, MPI_Comm comm)
+MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, PERSISTENT | ISEND | SEND | PREPARED);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
return request;
}
-MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
- int dst, int tag, MPI_Comm comm)
+MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, NON_PERSISTENT | ISEND | SEND);
+ request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
smpi_mpi_start(request);
return request;
}
-MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
- int dst, int tag, MPI_Comm comm)
+MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, NON_PERSISTENT | ISEND | SSEND | SEND);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
smpi_mpi_start(request);
return request;
}
-MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
- int src, int tag, MPI_Comm comm)
+MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
+ smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
comm, PERSISTENT | RECV | PREPARED);
return request;
}
-MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
- int src, int tag, MPI_Comm comm)
+MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
- comm, NON_PERSISTENT | RECV);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
+ smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
+ NON_PERSISTENT | RECV);
smpi_mpi_start(request);
return request;
}
-void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
- int tag, MPI_Comm comm, MPI_Status * status)
+void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
request = NULL;
}
-
-
-void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
- int tag, MPI_Comm comm)
+void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, NON_PERSISTENT | SEND);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
smpi_mpi_start(request);
smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
request = NULL;
}
-void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
- int dst, int tag, MPI_Comm comm)
+void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
- comm, NON_PERSISTENT | SSEND | SEND);
+ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(),
+ smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
smpi_mpi_start(request);
smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
request = NULL;
}
-void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
- int dst, int sendtag, void *recvbuf, int recvcount,
- MPI_Datatype recvtype, int src, int recvtag,
+void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
+ void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
MPI_Comm comm, MPI_Status * status)
{
MPI_Request requests[2];
MPI_Status stats[2];
int myid=smpi_process_index();
- if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
- smpi_datatype_copy(sendbuf, sendcount, sendtype,
- recvbuf, recvcount, recvtype);
+ if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
+ smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
return;
}
- requests[0] =
- smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
- requests[1] =
- smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
+ requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
+ requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
smpi_mpi_startall(2, requests);
smpi_mpi_waitall(2, requests, stats);
smpi_mpi_request_free(&requests[0]);
MPI_Request req = *request;
smpi_empty_status(status);
- if(!(req->detached && req->flags & SEND)
- && !(req->flags & PREPARED)){
+ if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){
if(status != MPI_STATUS_IGNORE) {
int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
// this handles the case were size in receive differs from size in send
- // FIXME: really this should just contain the count of receive-type blocks,
- // right?
+ // FIXME: really this should just contain the count of receive-type blocks, right?
status->count = req->real_size;
}
print_request("Finishing", req);
MPI_Datatype datatype = req->old_type;
- if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
+ if((req->flags & ACCUMULATE) || (datatype->sizeof_substruct != 0)){
if (!smpi_process_get_replaying()){
- if( smpi_privatize_global_variables
- && ((char*)req->old_buf >= smpi_start_data_exe)
- && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )
- ){
+ if( smpi_privatize_global_variables && ((char*)req->old_buf >= smpi_start_data_exe)
+ && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
smpi_switch_data_segment(smpi_process_index());
}
}
- if(datatype->has_subtype == 1){
- // This part handles the problem of non-contignous memory
- // the unserialization at the reception
+ if(datatype->sizeof_substruct != 0){
+ // This part handles the problem of non-contignous memory the unserialization at the reception
s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
if(req->flags & RECV)
- subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
- if(req->detached == 0) free(req->buf);
+ subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
+ datatype->substruct, req->op);
+ xbt_free(req->buf);
}else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
int n =req->real_size/smpi_datatype_size(datatype);
smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
+ xbt_free(req->buf);
}
}
- smpi_comm_unuse(req->comm);
- smpi_datatype_unuse(datatype);
-
}
if (TRACE_smpi_view_internals()) {
req->flags |= FINISHED;
smpi_mpi_request_free(request);
-
}
int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
- int flag;
-
//assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
// to avoid deadlocks if used as a break condition, such as
// while (MPI_Test(request, flag, status) && flag) {
// }
- // because the time will not normally advance when only calls to MPI_Test
- // are made -> deadlock
+ // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
// multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
static int nsleeps = 1;
if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
smpi_empty_status(status);
- flag = 1;
+ int flag = 1;
if (!((*request)->flags & PREPARED)) {
if ((*request)->action != NULL)
flag = simcall_comm_test((*request)->action);
return flag;
}
-int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
- MPI_Status * status)
+int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
{
xbt_dynar_t comms;
int i, flag, size;
map = xbt_new(int, count);
size = 0;
for(i = 0; i < count; i++) {
- if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
- !(requests[i]->flags & PREPARED)) {
+ if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
xbt_dynar_push(comms, &requests[i]->action);
map[size] = i;
size++;
return flag;
}
-
-int smpi_mpi_testall(int count, MPI_Request requests[],
- MPI_Status status[])
+int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
{
MPI_Status stat;
MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
- MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
- comm, PERSISTENT | RECV);
+ MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
+ smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
// to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
// (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
static int nsleeps = 1;
if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
// behave like a receive, but don't do it
- smx_rdv_t mailbox;
+ smx_mailbox_t mailbox;
print_request("New iprobe", request);
// We have to test both mailboxes as we don't know if we will receive one one or another
- if (sg_cfg_get_int("smpi/async_small_thresh")>0){
+ if (xbt_cfg_get_int("smpi/async-small-thresh")>0){
mailbox = smpi_process_mailbox_small();
XBT_DEBUG("trying to probe the perm recv mailbox");
request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
}
if (request->action){
- MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
+ simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
+ MPI_Request req = (MPI_Request)sync_comm->src_data;
*flag = 1;
if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
return;
}
- if ((*request)->action != NULL) { // this is not a detached send
+ if ((*request)->action != NULL)
+ // this is not a detached send
simcall_comm_wait((*request)->action, -1.0);
- if((MC_is_active() || MC_record_replay_is_active()) && (*request)->action)
- (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison
- }
-
finish_wait(request, status);
if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
*request = MPI_REQUEST_NULL;
// FIXME for a detached send, finish_wait is not called:
}
-int smpi_mpi_waitany(int count, MPI_Request requests[],
- MPI_Status * status)
+int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
{
xbt_dynar_t comms;
int i, size, index;
size = 0;
XBT_DEBUG("Wait for one of %d", count);
for(i = 0; i < count; i++) {
- if (requests[i] != MPI_REQUEST_NULL
- && !(requests[i]->flags & PREPARED)
- && !(requests[i]->flags & FINISHED)) {
+ if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
if (requests[i]->action != NULL) {
XBT_DEBUG("Waiting any %p ", requests[i]);
xbt_dynar_push(comms, &requests[i]->action);
return index;
}
-int smpi_mpi_waitall(int count, MPI_Request requests[],
- MPI_Status status[])
+int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
{
int index, c;
MPI_Status stat;
//tag invalid requests in the set
if (status != MPI_STATUSES_IGNORE) {
for (c = 0; c < count; c++) {
- if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
- (requests[c]->flags & PREPARED)) {
+ if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
smpi_empty_status(&status[c]);
} else if (requests[c]->src == MPI_PROC_NULL) {
smpi_empty_status(&status[c]);
return retvalue;
}
-int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
- MPI_Status status[])
+int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
{
int i, count, index;
MPI_Status stat;
return count;
}
-int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
- MPI_Status status[])
+int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
{
int i, count, count_dead;
MPI_Status stat;
else return count;
}
-void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
- MPI_Comm comm)
+void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
{
smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
}
}
void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
- void *recvbuf, int recvcount, MPI_Datatype recvtype,
- int root, MPI_Comm comm)
+ void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
{
int system_tag = COLL_TAG_GATHER;
int rank, size, src, index;
// FIXME: check for errors
smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from root
- smpi_datatype_copy(sendbuf, sendcount, sendtype,
- (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
+ smpi_datatype_copy(sendbuf, sendcount, sendtype, (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
// Receive buffers from senders
requests = xbt_new(MPI_Request, size - 1);
index = 0;
for(src = 0; src < size; src++) {
if(src != root) {
- requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
- recvcount, recvtype,
+ requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext, recvcount, recvtype,
src, system_tag, comm);
index++;
}
}
}
-
-void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
- MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
+ MPI_Comm comm)
{
int i, size, count;
int *displs;
- int rank = smpi_process_index();
+ int rank = smpi_comm_rank(comm);
void *tmpbuf;
/* arbitrarily choose root as rank 0 */
tmpbuf=(void*)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
- smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
- recvcounts[rank], datatype, 0, comm);
+ smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
xbt_free(displs);
smpi_free_tmp_buffer(tmpbuf);
}
-void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
- void *recvbuf, int *recvcounts, int *displs,
+void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
MPI_Datatype recvtype, int root, MPI_Comm comm)
{
int system_tag = COLL_TAG_GATHERV;
// FIXME: check for errors
smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from root
- smpi_datatype_copy(sendbuf, sendcount, sendtype,
- (char *)recvbuf + displs[root] * recvext,
+ smpi_datatype_copy(sendbuf, sendcount, sendtype, (char *)recvbuf + displs[root] * recvext,
recvcounts[root], recvtype);
// Receive buffers from senders
requests = xbt_new(MPI_Request, size - 1);
index = 0;
for(src = 0; src < size; src++) {
if(src != root) {
- requests[index] =
- smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
+ requests[index] = smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
recvcounts[src], recvtype, src, system_tag, comm);
index++;
}
}
}
-void smpi_mpi_allgather(void *sendbuf, int sendcount,
- MPI_Datatype sendtype, void *recvbuf,
- int recvcount, MPI_Datatype recvtype,
- MPI_Comm comm)
+void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
+ void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
{
int system_tag = COLL_TAG_ALLGATHER;
int rank, size, other, index;
// FIXME: check for errors
smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from self
- smpi_datatype_copy(sendbuf, sendcount, sendtype,
- (char *)recvbuf + rank * recvcount * recvext, recvcount,
- recvtype);
+ smpi_datatype_copy(sendbuf, sendcount, sendtype, (char *)recvbuf + rank * recvcount * recvext, recvcount, recvtype);
// Send/Recv buffers to/from others;
requests = xbt_new(MPI_Request, 2 * (size - 1));
index = 0;
for(other = 0; other < size; other++) {
if(other != rank) {
- requests[index] =
- smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
- comm);
+ requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
index++;
- requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
- recvcount, recvtype, other,
+ requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext, recvcount, recvtype, other,
system_tag, comm);
index++;
}
xbt_free(requests);
}
-void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
- MPI_Datatype sendtype, void *recvbuf,
- int *recvcounts, int *displs,
- MPI_Datatype recvtype, MPI_Comm comm)
+void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
+ int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
{
int system_tag = COLL_TAG_ALLGATHERV;
int rank, size, other, index;
// FIXME: check for errors
smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from self
- smpi_datatype_copy(sendbuf, sendcount, sendtype,
- (char *)recvbuf + displs[rank] * recvext,
- recvcounts[rank], recvtype);
+ smpi_datatype_copy(sendbuf, sendcount, sendtype, (char *)recvbuf + displs[rank] * recvext,recvcounts[rank], recvtype);
// Send buffers to others;
requests = xbt_new(MPI_Request, 2 * (size - 1));
index = 0;
for(other = 0; other < size; other++) {
if(other != rank) {
requests[index] =
- smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
- comm);
+ smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
index++;
- requests[index] =
- smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
- recvtype, other, system_tag, comm);
+ requests[index] = smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
+ recvtype, other, system_tag, comm);
index++;
}
}
}
void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
- void *recvbuf, int recvcount, MPI_Datatype recvtype,
- int root, MPI_Comm comm)
+ void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
{
int system_tag = COLL_TAG_SCATTER;
int rank, size, dst, index;
size = smpi_comm_size(comm);
if(rank != root) {
// Recv buffer from root
- smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
- MPI_STATUS_IGNORE);
+ smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
} else {
// FIXME: check for errors
smpi_datatype_extent(sendtype, &lb, &sendext);
index = 0;
for(dst = 0; dst < size; dst++) {
if(dst != root) {
- requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
- sendcount, sendtype, dst,
+ requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext, sendcount, sendtype, dst,
system_tag, comm);
index++;
}
}
}
-void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
- MPI_Datatype sendtype, void *recvbuf, int recvcount,
+void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
MPI_Datatype recvtype, int root, MPI_Comm comm)
{
int system_tag = COLL_TAG_SCATTERV;
size = smpi_comm_size(comm);
if(rank != root) {
// Recv buffer from root
- smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
- MPI_STATUS_IGNORE);
+ smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
} else {
// FIXME: check for errors
smpi_datatype_extent(sendtype, &lb, &sendext);
index = 0;
for(dst = 0; dst < size; dst++) {
if(dst != root) {
- requests[index] =
- smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
- sendtype, dst, system_tag, comm);
+ requests[index] = smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
+ sendtype, dst, system_tag, comm);
index++;
}
}
}
}
-void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
- MPI_Datatype datatype, MPI_Op op, int root,
+void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
MPI_Comm comm)
{
int system_tag = COLL_TAG_REDUCE;
MPI_Request *requests;
void **tmpbufs;
-
char* sendtmpbuf = (char*) sendbuf;
- if( sendbuf == MPI_IN_PLACE ) {
- sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
- smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
- }
+
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
//non commutative case, use a working algo from openmpi
if(!smpi_op_is_commute(op)){
- smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
- datatype, op, root, comm);
+ smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
return;
}
+
+ if( sendbuf == MPI_IN_PLACE ) {
+ sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
+ smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
+ }
if(rank != root) {
// Send buffer to root
else
tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
requests[index] =
- smpi_irecv_init(tmpbufs[index], count, datatype, src,
- system_tag, comm);
+ smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
index++;
}
}
xbt_free(tmpbufs);
xbt_free(requests);
- if( sendbuf == MPI_IN_PLACE ) {
- smpi_free_tmp_buffer(sendtmpbuf);
- }
+ }
+ if( sendbuf == MPI_IN_PLACE ) {
+ smpi_free_tmp_buffer(sendtmpbuf);
}
}
-void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
- MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
}
-void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
- MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
int system_tag = -888;
int rank, size, other, index;
// FIXME: possibly overkill we we have contiguous/noncontiguous data
// mapping...
tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
- requests[index] =
- smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
- comm);
+ requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
index++;
}
for(other = rank + 1; other < size; other++) {
- requests[index] =
- smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
+ requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
index++;
}
// Wait for completion of all comms.
xbt_free(requests);
}
-void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
- MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
int system_tag = -888;
int rank, size, other, index;
// mapping...
tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
requests[index] =
- smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
- comm);
+ smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
index++;
}
for(other = rank + 1; other < size; other++) {
