-#include "colls.h"
+#include "colls_private.h"
static int bcast_NTSL_segment_size_in_byte = 8192;
MPI_Datatype datatype, int root,
MPI_Comm comm)
{
- int tag = 50;
+ int tag = -COLL_TAG_BCAST;
MPI_Status status;
MPI_Request request;
MPI_Request *send_request_array;
int header_index;
int flag_array[MAX_NODE];
int already_sent[MAX_NODE];
-
+ int to_clean[MAX_NODE];
int header_buf[HEADER_SIZE];
char temp_buf[MAX_NODE];
MPI_Aint extent;
- MPI_Type_extent(datatype, &extent);
+ extent = smpi_datatype_get_extent(datatype);
/* destination */
int to;
- MPI_Comm_rank(MPI_COMM_WORLD, &rank);
- MPI_Comm_size(MPI_COMM_WORLD, &size);
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
/* segment is segment size in number of elements (not bytes) */
*/
if (root != 0) {
if (rank == root) {
- MPI_Send(buf, count, datatype, 0, tag, comm);
+ smpi_mpi_send(buf, count, datatype, 0, tag, comm);
} else if (rank == 0) {
- MPI_Recv(buf, count, datatype, root, tag, comm, &status);
+ smpi_mpi_recv(buf, count, datatype, root, tag, comm, &status);
}
}
/* value == 0 means root has not send data (or header) to the node yet */
for (i = 0; i < MAX_NODE; i++) {
already_sent[i] = 0;
+ to_clean[i] = 0;
}
/* when a message is smaller than a block size => no pipeline */
while (sent_count < (size - 1)) {
for (i = 1; i < size; i++) {
- MPI_Iprobe(i, MPI_ANY_TAG, MPI_COMM_WORLD, &flag_array[i],
+ smpi_mpi_iprobe(i, MPI_ANY_TAG, comm, &flag_array[i],
MPI_STATUSES_IGNORE);
}
/* message arrive */
if ((flag_array[i] == 1) && (already_sent[i] == 0)) {
- MPI_Recv(temp_buf, 1, MPI_CHAR, i, tag, MPI_COMM_WORLD, &status);
+ smpi_mpi_recv(temp_buf, 1, MPI_CHAR, i, tag, comm, &status);
header_buf[header_index] = i;
header_index++;
sent_count++;
if (header_index != 0) {
header_buf[header_index] = -1;
to = header_buf[0];
- MPI_Send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
- MPI_Send(buf, count, datatype, to, tag, comm);
+ smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
+ smpi_mpi_send(buf, count, datatype, to, tag, comm);
}
/* randomly MPI_Send to one */
if (already_sent[i] == 0) {
header_buf[0] = i;
header_buf[1] = -1;
- MPI_Send(header_buf, HEADER_SIZE, MPI_INT, i, tag, comm);
- MPI_Send(buf, count, datatype, i, tag, comm);
+ smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, i, tag, comm);
+ smpi_mpi_send(buf, count, datatype, i, tag, comm);
already_sent[i] = 1;
sent_count++;
break;
else {
/* send 1-byte message to root */
- MPI_Send(temp_buf, 1, MPI_CHAR, 0, tag, comm);
+ smpi_mpi_send(temp_buf, 1, MPI_CHAR, 0, tag, comm);
/* wait for header and data, forward when required */
- MPI_Recv(header_buf, HEADER_SIZE, MPI_INT, MPI_ANY_SOURCE, tag, comm,
+ smpi_mpi_recv(header_buf, HEADER_SIZE, MPI_INT, MPI_ANY_SOURCE, tag, comm,
&status);
- MPI_Recv(buf, count, datatype, MPI_ANY_SOURCE, tag, comm, &status);
+ smpi_mpi_recv(buf, count, datatype, MPI_ANY_SOURCE, tag, comm, &status);
/* search for where it is */
int myordering = 0;
/* send header followed by data */
if (header_buf[myordering + 1] != -1) {
- MPI_Send(header_buf, HEADER_SIZE, MPI_INT, header_buf[myordering + 1],
+ smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, header_buf[myordering + 1],
tag, comm);
- MPI_Send(buf, count, datatype, header_buf[myordering + 1], tag, comm);
+ smpi_mpi_send(buf, count, datatype, header_buf[myordering + 1], tag, comm);
}
}
}
/* pipeline bcast */
else {
send_request_array =
- (MPI_Request *) malloc((size + pipe_length) * sizeof(MPI_Request));
+ (MPI_Request *) xbt_malloc((size + pipe_length) * sizeof(MPI_Request));
recv_request_array =
- (MPI_Request *) malloc((size + pipe_length) * sizeof(MPI_Request));
+ (MPI_Request *) xbt_malloc((size + pipe_length) * sizeof(MPI_Request));
send_status_array =
- (MPI_Status *) malloc((size + pipe_length) * sizeof(MPI_Status));
+ (MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
recv_status_array =
- (MPI_Status *) malloc((size + pipe_length) * sizeof(MPI_Status));
+ (MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
if (rank == 0) {
//double start2 = MPI_Wtime();
//iteration++;
//start = MPI_Wtime();
for (i = 1; i < size; i++) {
- MPI_Iprobe(i, MPI_ANY_TAG, MPI_COMM_WORLD, &flag_array[i],
+ smpi_mpi_iprobe(i, MPI_ANY_TAG, comm, &flag_array[i],
&temp_status_array[i]);
}
//total = MPI_Wtime() - start;
for (i = 1; i < size; i++) {
/* message arrive */
if ((flag_array[i] == 1) && (already_sent[i] == 0)) {
- MPI_Recv(&temp_buf[i], 1, MPI_CHAR, i, tag, MPI_COMM_WORLD,
+ smpi_mpi_recv(&temp_buf[i], 1, MPI_CHAR, i, tag, comm,
&status);
header_buf[header_index] = i;
header_index++;
//start = MPI_Wtime();
/* send header */
- MPI_Send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
+ smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
//total = MPI_Wtime() - start;
//total *= 1000;
if (0 == 1) {
//if (header_index == 1) {
- MPI_Send(buf, count, datatype, to, tag, comm);
+ smpi_mpi_send(buf, count, datatype, to, tag, comm);
}
/* send data - pipeline */
else {
for (i = 0; i < pipe_length; i++) {
- MPI_Send((char *)buf + (i * increment), segment, datatype, to, tag, comm);
+ smpi_mpi_send((char *)buf + (i * increment), segment, datatype, to, tag, comm);
}
- //MPI_Waitall((pipe_length), send_request_array, send_status_array);
+ //smpi_mpi_waitall((pipe_length), send_request_array, send_status_array);
}
//total = MPI_Wtime() - start;
//total *= 1000;
to = i;
//start = MPI_Wtime();
- MPI_Send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
+ smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
/* still need to chop data so that we can use the same non-root code */
for (j = 0; j < pipe_length; j++) {
- MPI_Send((char *)buf + (j * increment), segment, datatype, to, tag,
+ smpi_mpi_send((char *)buf + (j * increment), segment, datatype, to, tag,
comm);
}
- //MPI_Send(buf,count,datatype,to,tag,comm);
- //MPI_Wait(&request,MPI_STATUS_IGNORE);
+ //smpi_mpi_send(buf,count,datatype,to,tag,comm);
+ //smpi_mpi_wait(&request,MPI_STATUS_IGNORE);
//total = MPI_Wtime() - start;
//total *= 1000;
already_sent[i] = 1;
+ to_clean[i]=1;
sent_count++;
break;
}
} /* while loop */
+ for(i=0; i<size; i++)
+ if(to_clean[i]!=0)smpi_mpi_recv(&temp_buf[i], 1, MPI_CHAR, i, tag, comm,
+ &status);
//total = MPI_Wtime() - start2;
//total *= 1000;
//printf("Node zero iter = %d time = %.2f\n",iteration,total);
/* none root */
else {
/* send 1-byte message to root */
- MPI_Send(temp_buf, 1, MPI_CHAR, 0, tag, comm);
+ smpi_mpi_send(temp_buf, 1, MPI_CHAR, 0, tag, comm);
/* wait for header forward when required */
- MPI_Irecv(header_buf, HEADER_SIZE, MPI_INT, MPI_ANY_SOURCE, tag, comm,
- &request);
- MPI_Wait(&request, MPI_STATUS_IGNORE);
+ request = smpi_mpi_irecv(header_buf, HEADER_SIZE, MPI_INT, MPI_ANY_SOURCE, tag, comm);
+ smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
/* search for where it is */
int myordering = 0;
/* send header when required */
if (header_buf[myordering + 1] != -1) {
- MPI_Send(header_buf, HEADER_SIZE, MPI_INT, header_buf[myordering + 1],
+ smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, header_buf[myordering + 1],
tag, comm);
}
if (0 == -1) {
//if (header_buf[1] == -1) {
- MPI_Irecv(buf, count, datatype, 0, tag, comm, &request);
- MPI_Wait(&request, MPI_STATUS_IGNORE);
+ request = smpi_mpi_irecv(buf, count, datatype, 0, tag, comm);
+ smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
//printf("\t\tnode %d ordering = %d receive data from root\n",rank,myordering);
} else {
for (i = 0; i < pipe_length; i++) {
- MPI_Irecv((char *)buf + (i * increment), segment, datatype, MPI_ANY_SOURCE,
- tag, comm, &recv_request_array[i]);
+ recv_request_array[i] = smpi_mpi_irecv((char *)buf + (i * increment), segment, datatype, MPI_ANY_SOURCE,
+ tag, comm);
}
}
/* send data */
if (header_buf[myordering + 1] != -1) {
for (i = 0; i < pipe_length; i++) {
- MPI_Wait(&recv_request_array[i], MPI_STATUS_IGNORE);
- MPI_Isend((char *)buf + (i * increment), segment, datatype,
- header_buf[myordering + 1], tag, comm,
- &send_request_array[i]);
+ smpi_mpi_wait(&recv_request_array[i], MPI_STATUS_IGNORE);
+ send_request_array[i] = smpi_mpi_isend((char *)buf + (i * increment), segment, datatype,
+ header_buf[myordering + 1], tag, comm);
}
- MPI_Waitall((pipe_length), send_request_array, send_status_array);
- }
-
+ smpi_mpi_waitall((pipe_length), send_request_array, send_status_array);
+ }else{
+ smpi_mpi_waitall(pipe_length, recv_request_array, recv_status_array);
+ }
+
}
free(send_request_array);
/* when count is not divisible by block size, use default BCAST for the remainder */
if ((remainder != 0) && (count > segment)) {
- MPI_Bcast((char *)buf + (pipe_length * increment), remainder, datatype, root, comm);
+ XBT_WARN("MPI_bcast_arrival_pattern_aware use default MPI_bcast.");
+ smpi_mpi_bcast((char *)buf + (pipe_length * increment), remainder, datatype, root, comm);
}
return MPI_SUCCESS;