/** General ARMCI global operation (reduction). Collective on group.
*
* @param[in] scope Scope in which to perform the GOP (only SCOPE_ALL is supported)
* @param[inout] x Vector of n doubles, contains input and will contain output.
* @param[in] n Length of x
* @param[in] op One of '+', '*', 'max', 'min', 'absmax', 'absmin'
* @param[in] type Data type of x
* @param[in] group Group on which to perform the GOP
*/
void armci_msg_group_gop_scope(int scope, void *x, int n, char *op, int type, ARMCI_Group *group) {
void *out;
MPI_Op mpi_op;
MPI_Datatype mpi_type;
MPI_Comm comm;
int mpi_type_size;
if (scope == SCOPE_ALL || scope == SCOPE_MASTERS)
comm = group->comm;
else
comm = MPI_COMM_SELF;
if (op[0] == '+') {
mpi_op = MPI_SUM;
} else if (op[0] == '*') {
mpi_op = MPI_PROD;
} else if (strncmp(op, "max", 3) == 0) {
mpi_op = MPI_MAX;
} else if (strncmp(op, "min", 3) == 0) {
mpi_op = MPI_MIN;
} else if (strncmp(op, "or", 2) == 0) {
mpi_op = MPI_BOR;
} else if (strncmp(op, "absmax", 6) == 0) {
mpi_op = MPI_ABSMAX_OP;
} else if (strncmp(op, "absmin", 6) == 0) {
mpi_op = MPI_ABSMIN_OP;
} else {
ARMCII_Error("unknown operation \'%s\'", op);
return;
}
switch(type) {
case ARMCI_INT:
mpi_type = MPI_INT;
break;
case ARMCI_LONG:
mpi_type = MPI_LONG;
break;
case ARMCI_LONG_LONG:
mpi_type = MPI_LONG_LONG;
break;
case ARMCI_FLOAT:
mpi_type = MPI_FLOAT;
break;
case ARMCI_DOUBLE:
mpi_type = MPI_DOUBLE;
break;
default:
ARMCII_Error("unknown type (%d)", type);
return;
}
// ABS MAX/MIN are unary as well as binary. We need to also apply abs in the
// single processor case when reduce would normally just be a no-op.
if (group->size == 1 && (mpi_op == MPI_ABSMAX_OP || mpi_op == MPI_ABSMIN_OP)) {
ARMCII_Absv_op(x, x, &n, &mpi_type);
return;
}
MPI_Type_size(mpi_type, &mpi_type_size);
out = malloc(n*mpi_type_size);
ARMCII_Assert(out != NULL);
MPI_Allreduce(x, out, n, mpi_type, mpi_op, group->comm);
ARMCI_Copy(out, x, n*mpi_type_size);
free(out);
}
/** General ARMCI global operation (reduction). Collective on group.
*
* @param[in] scope Scope in which to perform the GOP (only SCOPE_ALL is supported)
* @param[inout] x Vector of n data elements, contains input and will contain output.
* @param[in] n Length of x
* @param[in] op One of '+', '*', 'max', 'min', 'absmax', 'absmin'
* @param[in] type Data type of x (e.g. ARMCI_INT, ...)
* @param[in] group Group on which to perform the GOP
*/
void armci_msg_group_gop_scope(int scope, void *x, int n, char *op, int type, ARMCI_Group *group) {
void *out, **x_buf;
MPI_Op mpi_op;
MPI_Datatype mpi_type;
MPI_Comm comm;
int mpi_type_size;
/* FIXME: scope argument presently ignored */
if (scope == SCOPE_ALL || scope == SCOPE_MASTERS)
comm = group->comm;
else
comm = MPI_COMM_SELF;
if (op[0] == '+') {
mpi_op = MPI_SUM;
} else if (op[0] == '*') {
mpi_op = MPI_PROD;
} else if (strncmp(op, "max", 3) == 0) {
mpi_op = MPI_MAX;
} else if (strncmp(op, "min", 3) == 0) {
mpi_op = MPI_MIN;
} else if (strncmp(op, "or", 2) == 0) {
mpi_op = MPI_BOR;
} else if (strncmp(op, "absmax", 6) == 0) {
mpi_op = ARMCI_MPI_ABSMAX_OP;
} else if (strncmp(op, "absmin", 6) == 0) {
mpi_op = ARMCI_MPI_ABSMIN_OP;
/* The following were added ComEx/ARMCI in 2017. */
/* https://github.com/GlobalArrays/ga/commit/14ef3cfa4ea3ffa7ee721c2a98685669359f7044 */
/* && and || need to be tested before & and | to avoid the latter matching the former. */
} else if ((strncmp(op, "land", 4) == 0) || (strncmp(op, "&&", 2) == 0)) {
mpi_op = MPI_LAND;
} else if ((strncmp(op, "lor", 3) == 0) || (strncmp(op, "||", 2) == 0)) {
mpi_op = MPI_LOR;
} else if ((strncmp(op, "band", 4) == 0) || (strncmp(op, "&", 1) == 0)) {
mpi_op = MPI_BAND;
} else if ((strncmp(op, "bor", 3) == 0) || (strncmp(op, "|", 1) == 0)) {
mpi_op = MPI_BOR;
} else {
ARMCII_Error("unknown operation \'%s\'", op);
return;
}
switch(type) {
case ARMCI_INT:
mpi_type = MPI_INT;
break;
case ARMCI_LONG:
mpi_type = MPI_LONG;
break;
case ARMCI_LONG_LONG:
mpi_type = MPI_LONG_LONG;
break;
case ARMCI_FLOAT:
mpi_type = MPI_FLOAT;
break;
case ARMCI_DOUBLE:
mpi_type = MPI_DOUBLE;
break;
default:
ARMCII_Error("unknown type (%d)", type);
return;
}
MPI_Type_size(mpi_type, &mpi_type_size);
ARMCII_Buf_prepare_read_vec(&x, &x_buf, 1, n*mpi_type_size);
// ABS MAX/MIN are unary as well as binary. We need to also apply abs in the
// single processor case when reduce would normally just be a no-op.
if (group->size == 1 && (mpi_op == ARMCI_MPI_ABSMAX_OP || mpi_op == ARMCI_MPI_ABSMIN_OP)) {
ARMCII_Absv_op(x_buf[0], x_buf[0], &n, &mpi_type);
}
else {
out = malloc(n*mpi_type_size);
ARMCII_Assert(out != NULL);
MPI_Allreduce(x_buf[0], out, n, mpi_type, mpi_op, comm);
ARMCI_Copy(out, x_buf[0], n*mpi_type_size);
free(out);
}
ARMCII_Buf_finish_write_vec(&x, x_buf, 1, n*mpi_type_size);
}
