#include "private.h"
+#include "xbt/time.h"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi,
"Logging specific to SMPI (base)");
XBT_LOG_EXTERNAL_CATEGORY(smpi_bench);
XBT_LOG_EXTERNAL_CATEGORY(smpi_kernel);
XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi);
+XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi_dt);
+XBT_LOG_EXTERNAL_CATEGORY(smpi_coll);
XBT_LOG_EXTERNAL_CATEGORY(smpi_receiver);
XBT_LOG_EXTERNAL_CATEGORY(smpi_sender);
XBT_LOG_EXTERNAL_CATEGORY(smpi_util);
smpi_mpi_global_t smpi_mpi_global = NULL;
+
+/**
+ * Operations of MPI_OP : implemented=land,sum,min,max
+ **/
void smpi_mpi_land_func(void *a, void *b, int *length,
MPI_Datatype * datatype);
}
}
+/**
+ * sum two vectors element-wise
+ *
+ * @param a the first vectors
+ * @param b the second vectors
+ * @return the second vector is modified and contains the element-wise sums
+ **/
void smpi_mpi_sum_func(void *a, void *b, int *length,
MPI_Datatype * datatype);
void smpi_mpi_sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
{
int i;
- if (*datatype == smpi_mpi_global->mpi_int) {
+ if ((*datatype == smpi_mpi_global->mpi_byte) || (*datatype == smpi_mpi_global->mpi_char)) {
+ char *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] + y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_int) {
int *x = a, *y = b;
for (i = 0; i < *length; i++) {
y[i] = x[i] + y[i];
}
+ } else if (*datatype == smpi_mpi_global->mpi_float) {
+ float *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] + y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_double) {
+ double *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] + y[i];
+ }
}
}
+/**
+ *i multiply two vectors element-wise
+ *
+ * @param a the first vectors
+ * @param b the second vectors
+ * @return the second vector is modified and contains the element-wise products
+ **/
+void smpi_mpi_prod_func(void *a, void *b, int *length,
+ MPI_Datatype * datatype);
-int smpi_mpi_comm_rank(smpi_mpi_communicator_t comm)
+void smpi_mpi_prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
{
- return comm->index_to_rank_map[smpi_host_index()];
+ int i;
+ if ((*datatype == smpi_mpi_global->mpi_byte) || (*datatype == smpi_mpi_global->mpi_char)) {
+ char *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] * y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_int) {
+ int *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] * y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_float) {
+ float *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] * y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_double) {
+ double *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] * y[i];
+ }
+ }
}
+/**
+ * compute the min of two vectors element-wise
+ **/
+void smpi_mpi_min_func(void *a, void *b, int *length,
+ MPI_Datatype * datatype);
-void smpi_process_init()
+void smpi_mpi_min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
{
- smx_host_t host;
int i;
- smpi_host_data_t hdata;
+ if ((*datatype == smpi_mpi_global->mpi_byte) || (*datatype == smpi_mpi_global->mpi_char)) {
+ char *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] < y[i] ? x[i] : y[i];
+ }
+ } else {
+ if (*datatype == smpi_mpi_global->mpi_int) {
+ int *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] < y[i] ? x[i] : y[i];
+ }
+ } else {
+ if (*datatype == smpi_mpi_global->mpi_float) {
+ float *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] < y[i] ? x[i] : y[i];
+ }
+ } else {
+ if (*datatype == smpi_mpi_global->mpi_double) {
+ double *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] < y[i] ? x[i] : y[i];
+ }
+
+ }
+ }
+ }
+ }
+}
- smpi_global->running_hosts_count++;
+/**
+ * compute the max of two vectors element-wise
+ **/
+void smpi_mpi_max_func(void *a, void *b, int *length,
+ MPI_Datatype * datatype);
- // initialize some local variables
- host = SIMIX_host_self();
+void smpi_mpi_max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
+{
+ int i;
+ if ((*datatype == smpi_mpi_global->mpi_byte) || (*datatype == smpi_mpi_global->mpi_char)) {
+ char *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] > y[i] ? x[i] : y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_int) {
+ int *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] > y[i] ? x[i] : y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_float) {
+ float *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] > y[i] ? x[i] : y[i];
+ }
+ } else if (*datatype == smpi_mpi_global->mpi_double) {
+ double *x = a, *y = b;
+ for (i = 0; i < *length; i++) {
+ y[i] = x[i] > y[i] ? x[i] : y[i];
+ }
- hdata = xbt_new(s_smpi_host_data_t, 1);
+ }
+}
- for (i = 0; i < smpi_global->host_count && host != smpi_global->hosts[i]; i++);
- hdata->index = i;
- hdata->mutex = SIMIX_mutex_init();
- hdata->cond = SIMIX_cond_init();
- SIMIX_host_set_data(host, hdata);
- return;
+/**
+ * tell the MPI rank of the calling process (from its SIMIX process id)
+ **/
+int smpi_mpi_comm_rank(smpi_mpi_communicator_t comm)
+{
+ return comm->index_to_rank_map[smpi_process_index()];
}
-void smpi_process_finalize()
+void smpi_process_init(int *argc, char ***argv)
{
- int i;
-
- i = --smpi_global->running_hosts_count;
+ smpi_process_data_t pdata;
- SIMIX_mutex_destroy(smpi_host_mutex());
- SIMIX_cond_destroy(smpi_host_cond());
-
- if (0 >= i) {
-
- // wake up senders/receivers
- for (i = 0; i < smpi_global->host_count; i++) {
- if (SIMIX_process_is_suspended(smpi_global->sender_processes[i])) {
- SIMIX_process_resume(smpi_global->sender_processes[i]);
- }
- if (SIMIX_process_is_suspended(smpi_global->receiver_processes[i])) {
- SIMIX_process_resume(smpi_global->receiver_processes[i]);
- }
- }
-
- SIMIX_mutex_destroy(smpi_mpi_global->mpi_comm_world->barrier_mutex);
- SIMIX_cond_destroy(smpi_mpi_global->mpi_comm_world->barrier_cond);
- xbt_free(smpi_mpi_global->mpi_comm_world);
+ // initialize some local variables
- xbt_free(smpi_mpi_global->mpi_byte);
- xbt_free(smpi_mpi_global->mpi_int);
- xbt_free(smpi_mpi_global->mpi_double);
+ pdata = xbt_new(s_smpi_process_data_t, 1);
+ SIMIX_process_set_data(SIMIX_process_self(), pdata);
- xbt_free(smpi_mpi_global->mpi_land);
- xbt_free(smpi_mpi_global->mpi_sum);
+ /* get rank from command line, and remove it from argv */
+ pdata->index = atoi((*argv)[1]);
+ DEBUG1("I'm rank <%d>", pdata->index);
+ if (*argc > 2) {
+ memmove((*argv)[1], (*argv)[2], sizeof(char *) * (*argc - 2));
+ (*argv)[(*argc) - 1] = NULL;
+ }
+ (*argc)--;
+
+ pdata->mutex = SIMIX_mutex_init();
+ pdata->cond = SIMIX_cond_init();
+ pdata->finalize = 0;
+
+ pdata->pending_recv_request_queue = xbt_fifo_new();
+ pdata->pending_send_request_queue = xbt_fifo_new();
+ pdata->received_message_queue = xbt_fifo_new();
+
+ pdata->main = SIMIX_process_self();
+ pdata->sender = SIMIX_process_create("smpi_sender",
+ smpi_sender, pdata,
+ SIMIX_host_get_name(SIMIX_host_self()),
+ 0, NULL,
+ /*props */ NULL);
+ pdata->receiver = SIMIX_process_create("smpi_receiver",
+ smpi_receiver, pdata,
+ SIMIX_host_get_name(SIMIX_host_self
+ ()), 0, NULL,
+ /*props */ NULL);
+
+ smpi_global->main_processes[pdata->index] = SIMIX_process_self();
+ return;
+}
- xbt_free(smpi_mpi_global);
+void smpi_process_finalize()
+{
+ smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self());
+ pdata->finalize = 2; /* Tell sender and receiver to quit */
+ SIMIX_process_resume(pdata->sender);
+ SIMIX_process_resume(pdata->receiver);
+ while (pdata->finalize > 0) { /* wait until it's done */
+ SIMIX_cond_wait(pdata->cond, pdata->mutex);
}
+ SIMIX_mutex_destroy(pdata->mutex);
+ SIMIX_cond_destroy(pdata->cond);
+ xbt_fifo_free(pdata->pending_recv_request_queue);
+ xbt_fifo_free(pdata->pending_send_request_queue);
+ xbt_fifo_free(pdata->received_message_queue);
+ xbt_free(pdata);
}
-int smpi_mpi_barrier(smpi_mpi_communicator_t comm)
+
+/*int smpi_mpi_barrier(smpi_mpi_communicator_t comm)
{
SIMIX_mutex_lock(comm->barrier_mutex);
return MPI_SUCCESS;
}
+*/
int smpi_mpi_isend(smpi_mpi_request_t request)
{
+ smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self());
int retval = MPI_SUCCESS;
- int index = smpi_host_index();
if (NULL == request) {
retval = MPI_ERR_INTERN;
} else {
- xbt_fifo_push(smpi_global->pending_send_request_queues[index], request);
-
- if (SIMIX_process_is_suspended(smpi_global->sender_processes[index])) {
- SIMIX_process_resume(smpi_global->sender_processes[index]);
- }
+ xbt_fifo_push(pdata->pending_send_request_queue, request);
+ SIMIX_process_resume(pdata->sender);
}
return retval;
int smpi_mpi_irecv(smpi_mpi_request_t request)
{
int retval = MPI_SUCCESS;
- int index = smpi_host_index();
+ smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self());
if (NULL == request) {
retval = MPI_ERR_INTERN;
} else {
- xbt_fifo_push(smpi_global->pending_recv_request_queues[index], request);
+ xbt_fifo_push(pdata->pending_recv_request_queue, request);
- if (SIMIX_process_is_suspended(smpi_global->receiver_processes[index])) {
- SIMIX_process_resume(smpi_global->receiver_processes[index]);
+ if (SIMIX_process_is_suspended(pdata->receiver)) {
+ SIMIX_process_resume(pdata->receiver);
}
}
return retval;
}
+void print_req( smpi_mpi_request_t r );
+void print_req( smpi_mpi_request_t r ) {
+ fprintf(stderr,"***req %p-> src=%d dst=%d tag=%d completed=0x%x consumed=0x%x\n",r,r->src,r->dst,r->tag,r->completed,r->consumed);
+}
+
+
+/**
+ * wait and friends ...
+ **/
int smpi_mpi_wait(smpi_mpi_request_t request, smpi_mpi_status_t * status)
{
int retval = MPI_SUCCESS;
if (NULL == request) {
retval = MPI_ERR_INTERN;
} else {
+
+ DEBUG3("entered smpi_mpi_wait() for req_src=%d,req_dst=%d,req_tag=%d",
+ request->src,request->dst,request->tag);
SIMIX_mutex_lock(request->mutex);
+//#define DEBUG_STEPH
+#ifdef DEBUG_STEPH
+ print_req( request ); //@@
+#endif
while (!request->completed) {
SIMIX_cond_wait(request->cond, request->mutex);
}
return retval;
}
+
+/**
+ * waitall
+ **/
+int smpi_mpi_waitall(int count, smpi_mpi_request_t requests[],
+ smpi_mpi_status_t status[])
+{
+ int cpt;
+ int index;
+ int retval;
+ smpi_mpi_status_t stat;
+
+ for (cpt = 0; cpt < count; cpt++) {
+ retval = smpi_mpi_waitany(count, requests, &index, &stat);
+ if (retval != MPI_SUCCESS)
+ return retval;
+ if (MPI_STATUS_IGNORE != status)
+ memcpy(&(status[index]), &stat, sizeof(stat));
+ }
+ return MPI_SUCCESS;
+}
+
+/**
+ * waitany
+ **/
+int smpi_mpi_waitany(int count, smpi_mpi_request_t * requests, int *index,
+ smpi_mpi_status_t * status)
+{
+ int cpt;
+
+ DEBUG0("entering smpi_wait_any() ...");
+ *index = MPI_UNDEFINED;
+ if (NULL == requests) {
+ return MPI_ERR_INTERN;
+ }
+ /* First check if one of them is already done */
+ for (cpt = 0; cpt < count; cpt++) {
+ DEBUG2(" exam req[%d] of msg from <%d>",cpt,requests[cpt]->src);
+ if (requests[cpt]->completed && !requests[cpt]->consumed) { /* got ya */
+ DEBUG2("smpi_wait_any() found match req[%d] of msg from <%d>",cpt,requests[cpt]->src);
+ *index = cpt;
+ goto found_request;
+ }
+ }
+ /* If none found, block */
+ /* FIXME: should use a SIMIX_cond_waitany, when implemented. For now, block on the first one */
+ while (1) {
+ for (cpt = 0; cpt < count; cpt++) {
+
+#ifdef DEBUG_STEPH
+ print_req( requests[cpt] );
+#endif
+ if (!requests[cpt]->completed) { /* this one is not done, wait on it */
+ DEBUG3("smpi_waitany() blocked waiting a msg <%d> -> <%d>, tag=%d",requests[cpt]->src,requests[cpt]->dst,requests[cpt]->tag);
+ while (!requests[cpt]->completed)
+ SIMIX_cond_wait(requests[cpt]->cond, requests[cpt]->mutex);
+
+ *index = cpt;
+ goto found_request;
+ }
+ }
+ if (cpt == count) /* they are all done. Damn user */
+ return MPI_ERR_REQUEST;
+ }
+
+found_request:
+#ifdef DEBUG_STEPH
+ print_req( requests[cpt] );
+#endif
+ requests[*index]->consumed = 1;
+#ifdef DEBUG_STEPH
+ print_req( requests[cpt] );
+#endif
+ DEBUG2("smpi_waitany() request %p unblocked ... mark *req[%d]->consumed",requests[*index],cpt);
+ if (NULL != status) {
+ status->MPI_SOURCE = requests[*index]->src;
+ status->MPI_TAG = requests[*index]->tag;
+ status->MPI_ERROR = MPI_SUCCESS;
+ }
+ return MPI_SUCCESS;
+
+}