int mca_coll_ml_reduce_nb(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
int root, struct ompi_communicator_t *comm,
ompi_request_t **req,
mca_coll_base_module_t *module) {
int ret = OMPI_SUCCESS;
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t*)module;
if (OPAL_UNLIKELY(!ompi_op_is_commute(op) || !opal_datatype_is_contiguous_memory_layout(&dtype->super, count))) {
/* coll/ml does not handle non-communative operations at this time. fallback
* on another collective module */
return ml_module->fallback.coll_ireduce (sbuf, rbuf, count, dtype, op, root, comm, req,
ml_module->fallback.coll_ireduce_module);
}
ML_VERBOSE(10,("Calling Ml Reduce "));
ret = parallel_reduce_start(sbuf, rbuf, count, dtype, op,
root, comm, ml_module,
req, ML_SMALL_DATA_REDUCE,
ML_LARGE_DATA_REDUCE);
if (OPAL_UNLIKELY(ret != OMPI_SUCCESS)) {
ML_VERBOSE(10, ("Failed to launch"));
return ret;
}
ML_VERBOSE(10, ("Non-blocking Reduce is done"));
return OMPI_SUCCESS;
}
int MPI_Op_commutative(MPI_Op op, int *commute)
{
OPAL_CR_NOOP_PROGRESS();
/* Error checking */
if (MPI_PARAM_CHECK) {
OMPI_ERR_INIT_FINALIZE(FUNC_NAME);
if (NULL == op || MPI_OP_NULL == op) {
return OMPI_ERRHANDLER_INVOKE(MPI_COMM_WORLD, MPI_ERR_OP,
FUNC_NAME);
}
if (NULL == commute) {
return OMPI_ERRHANDLER_INVOKE(MPI_COMM_WORLD, MPI_ERR_ARG,
FUNC_NAME);
}
}
/* We have a valid op, get the flag */
*commute = ompi_op_is_commute(op);
/* All done */
return MPI_SUCCESS;
}
/*
* reduce_scatter_intra_dec
*
* Function: - seletects reduce_scatter algorithm to use
* Accepts: - same arguments as MPI_Reduce_scatter()
* Returns: - MPI_SUCCESS or error code (passed from
* the reduce scatter implementation)
* Note: If we detect zero valued counts in the rcounts array, we
* fall back to the nonoverlapping algorithm because the other
* algorithms do not currently handle it.
*/
int ompi_coll_tuned_reduce_scatter_intra_dec_fixed( void *sbuf, void *rbuf,
int *rcounts,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int comm_size, i, pow2;
size_t total_message_size, dsize;
const double a = 0.0012;
const double b = 8.0;
const size_t small_message_size = 12 * 1024;
const size_t large_message_size = 256 * 1024;
bool zerocounts = false;
OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_reduce_scatter_intra_dec_fixed"));
comm_size = ompi_comm_size(comm);
/* We need data size for decision function */
ompi_ddt_type_size(dtype, &dsize);
total_message_size = 0;
for (i = 0; i < comm_size; i++) {
total_message_size += rcounts[i];
if (0 == rcounts[i]) {
zerocounts = true;
}
}
if( !ompi_op_is_commute(op) || (zerocounts)) {
return ompi_coll_tuned_reduce_scatter_intra_nonoverlapping (sbuf, rbuf, rcounts,
dtype, op,
comm, module);
}
total_message_size *= dsize;
/* compute the nearest power of 2 */
for (pow2 = 1; pow2 < comm_size; pow2 <<= 1);
if ((total_message_size <= small_message_size) ||
((total_message_size <= large_message_size) && (pow2 == comm_size)) ||
(comm_size >= a * total_message_size + b)) {
return
ompi_coll_tuned_reduce_scatter_intra_basic_recursivehalving(sbuf, rbuf, rcounts,
dtype, op,
comm, module);
}
return ompi_coll_tuned_reduce_scatter_intra_ring(sbuf, rbuf, rcounts,
dtype, op,
comm, module);
}
/*
* allreduce_intra
*
* Function: - allreduce using other MPI collectives
* Accepts: - same as MPI_Allreduce()
* Returns: - MPI_SUCCESS or error code
*/
int
ompi_coll_tuned_allreduce_intra_dec_fixed (void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
size_t dsize, block_dsize;
int comm_size = ompi_comm_size(comm);
const size_t intermediate_message = 10000;
OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_allreduce_intra_dec_fixed"));
/**
* Decision function based on MX results from the Grig cluster at UTK.
*
* Currently, linear, recursive doubling, and nonoverlapping algorithms
* can handle both commutative and non-commutative operations.
* Ring algorithm does not support non-commutative operations.
*/
ompi_ddt_type_size(dtype, &dsize);
block_dsize = dsize * count;
if (block_dsize < intermediate_message) {
return (ompi_coll_tuned_allreduce_intra_recursivedoubling (sbuf, rbuf,
count, dtype,
op, comm, module));
}
if( ompi_op_is_commute(op) && (count > comm_size) ) {
const size_t segment_size = 1 << 20; /* 1 MB */
if ((comm_size * segment_size >= block_dsize)) {
return (ompi_coll_tuned_allreduce_intra_ring (sbuf, rbuf, count, dtype,
op, comm, module));
} else {
return (ompi_coll_tuned_allreduce_intra_ring_segmented (sbuf, rbuf,
count, dtype,
op, comm, module,
segment_size));
}
}
return (ompi_coll_tuned_allreduce_intra_nonoverlapping (sbuf, rbuf, count,
dtype, op, comm, module));
}
/*
* reduce_intra_dec
*
* Function: - seletects reduce algorithm to use
* Accepts: - same arguments as MPI_reduce()
* Returns: - MPI_SUCCESS or error code (passed from the reduce implementation)
*
*/
int ompi_coll_tuned_reduce_intra_dec_fixed( void *sendbuf, void *recvbuf,
int count, struct ompi_datatype_t* datatype,
struct ompi_op_t* op, int root,
struct ompi_communicator_t* comm,
mca_coll_base_module_t *module)
{
int communicator_size, segsize = 0;
size_t message_size, dsize;
const double a1 = 0.6016 / 1024.0; /* [1/B] */
const double b1 = 1.3496;
const double a2 = 0.0410 / 1024.0; /* [1/B] */
const double b2 = 9.7128;
const double a3 = 0.0422 / 1024.0; /* [1/B] */
const double b3 = 1.1614;
const double a4 = 0.0033 / 1024.0; /* [1/B] */
const double b4 = 1.6761;
const int max_requests = 0; /* no limit on # of outstanding requests */
communicator_size = ompi_comm_size(comm);
/* need data size for decision function */
ompi_ddt_type_size(datatype, &dsize);
message_size = dsize * count; /* needed for decision */
/**
* If the operation is non commutative we currently have choice of linear
* or in-order binary tree algorithm.
*/
if( !ompi_op_is_commute(op) ) {
if ((communicator_size < 12) && (message_size < 2048)) {
return ompi_coll_tuned_reduce_intra_basic_linear (sendbuf, recvbuf, count, datatype, op, root, comm, module);
}
return ompi_coll_tuned_reduce_intra_in_order_binary (sendbuf, recvbuf, count, datatype, op, root, comm, module,
0, max_requests);
}
OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_reduce_intra_dec_fixed "
"root %d rank %d com_size %d msg_length %lu",
root, ompi_comm_rank(comm), communicator_size, (unsigned long)message_size));
if ((communicator_size < 8) && (message_size < 512)){
/* Linear_0K */
return ompi_coll_tuned_reduce_intra_basic_linear (sendbuf, recvbuf, count, datatype, op, root, comm, module);
} else if (((communicator_size < 8) && (message_size < 20480)) ||
(message_size < 2048) || (count <= 1)) {
/* Binomial_0K */
segsize = 0;
return ompi_coll_tuned_reduce_intra_binomial(sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, max_requests);
} else if (communicator_size > (a1 * message_size + b1)) {
/* Binomial_1K */
segsize = 1024;
return ompi_coll_tuned_reduce_intra_binomial(sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, max_requests);
} else if (communicator_size > (a2 * message_size + b2)) {
/* Pipeline_1K */
segsize = 1024;
return ompi_coll_tuned_reduce_intra_pipeline (sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, max_requests);
} else if (communicator_size > (a3 * message_size + b3)) {
/* Binary_32K */
segsize = 32*1024;
return ompi_coll_tuned_reduce_intra_pipeline (sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, max_requests);
}
if (communicator_size > (a4 * message_size + b4)) {
/* Pipeline_32K */
segsize = 32*1024;
} else {
/* Pipeline_64K */
segsize = 64*1024;
}
return ompi_coll_tuned_reduce_intra_pipeline (sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, max_requests);
#if 0
/* for small messages use linear algorithm */
if (message_size <= 4096) {
segsize = 0;
fanout = communicator_size - 1;
/* when linear implemented or taken from basic put here, right now using chain as a linear system */
/* it is implemented and I shouldn't be calling a chain with a fanout bigger than MAXTREEFANOUT from topo.h! */
return ompi_coll_tuned_reduce_intra_basic_linear (sendbuf, recvbuf, count, datatype, op, root, comm, module);
/* return ompi_coll_tuned_reduce_intra_chain (sendbuf, recvbuf, count, datatype, op, root, comm, segsize, fanout); */
}
if (message_size < 524288) {
if (message_size <= 65536 ) {
segsize = 32768;
fanout = 8;
} else {
segsize = 1024;
fanout = communicator_size/2;
}
/* later swap this for a binary tree */
/* fanout = 2; */
return ompi_coll_tuned_reduce_intra_chain (sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, fanout, max_requests);
}
segsize = 1024;
return ompi_coll_tuned_reduce_intra_pipeline (sendbuf, recvbuf, count, datatype, op, root, comm, module,
segsize, max_requests);
#endif /* 0 */
}
/*
* reduce_log_intra
*
* Function: - reduction using O(log N) algorithm
* Accepts: - same as MPI_Reduce()
* Returns: - MPI_SUCCESS or error code
*
*
* Performing reduction on each dimension of the hypercube.
* An example for 8 procs (dimensions = 3):
*
* Stage 1, reduce on X dimension, 1 -> 0, 3 -> 2, 5 -> 4, 7 -> 6
*
* 6----<---7 proc_0: 0+1
* /| /| proc_1: 1
* / | / | proc_2: 2+3
* / | / | proc_3: 3
* 4----<---5 | proc_4: 4+5
* | 2--< |---3 proc_5: 5
* | / | / proc_6: 6+7
* | / | / proc_7: 7
* |/ |/
* 0----<---1
*
* Stage 2, reduce on Y dimension, 2 -> 0, 6 -> 4
*
* 6--------7 proc_0: 0+1+2+3
* /| /| proc_1: 1
* v | / | proc_2: 2+3
* / | / | proc_3: 3
* 4--------5 | proc_4: 4+5+6+7
* | 2--- |---3 proc_5: 5
* | / | / proc_6: 6+7
* | v | / proc_7: 7
* |/ |/
* 0--------1
*
* Stage 3, reduce on Z dimension, 4 -> 0
*
* 6--------7 proc_0: 0+1+2+3+4+5+6+7
* /| /| proc_1: 1
* / | / | proc_2: 2+3
* / | / | proc_3: 3
* 4--------5 | proc_4: 4+5+6+7
* | 2--- |---3 proc_5: 5
* v / | / proc_6: 6+7
* | / | / proc_7: 7
* |/ |/
* 0--------1
*
*
*/
int
mca_coll_basic_reduce_log_intra(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
int root, struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, size, rank, vrank;
int err, peer, dim, mask;
ptrdiff_t true_lb, true_extent, lb, extent;
char *free_buffer = NULL;
char *free_rbuf = NULL;
char *pml_buffer = NULL;
char *snd_buffer = NULL;
char *rcv_buffer = (char*)rbuf;
char *inplace_temp = NULL;
/* JMS Codearound for now -- if the operations is not communative,
* just call the linear algorithm. Need to talk to Edgar / George
* about fixing this algorithm here to work with non-communative
* operations. */
if (!ompi_op_is_commute(op)) {
return mca_coll_basic_reduce_lin_intra(sbuf, rbuf, count, dtype,
op, root, comm, module);
}
/* Some variables */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
vrank = ompi_op_is_commute(op) ? (rank - root + size) % size : rank;
dim = comm->c_cube_dim;
/* Allocate the incoming and resulting message buffers. See lengthy
* rationale above. */
ompi_datatype_get_extent(dtype, &lb, &extent);
ompi_datatype_get_true_extent(dtype, &true_lb, &true_extent);
free_buffer = (char*)malloc(true_extent + (count - 1) * extent);
if (NULL == free_buffer) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
pml_buffer = free_buffer - lb;
/* read the comment about commutative operations (few lines down
* the page) */
if (ompi_op_is_commute(op)) {
rcv_buffer = pml_buffer;
}
/* Allocate sendbuf in case the MPI_IN_PLACE option has been used. See lengthy
* rationale above. */
if (MPI_IN_PLACE == sbuf) {
inplace_temp = (char*)malloc(true_extent + (count - 1) * extent);
if (NULL == inplace_temp) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto cleanup_and_return;
}
sbuf = inplace_temp - lb;
err = ompi_datatype_copy_content_same_ddt(dtype, count, (char*)sbuf, (char*)rbuf);
}
snd_buffer = (char*)sbuf;
if (rank != root && 0 == (vrank & 1)) {
/* root is the only one required to provide a valid rbuf.
* Assume rbuf is invalid for all other ranks, so fix it up
* here to be valid on all non-leaf ranks */
free_rbuf = (char*)malloc(true_extent + (count - 1) * extent);
if (NULL == free_rbuf) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto cleanup_and_return;
}
rbuf = free_rbuf - lb;
}
/* Loop over cube dimensions. High processes send to low ones in the
* dimension. */
for (i = 0, mask = 1; i < dim; ++i, mask <<= 1) {
/* A high-proc sends to low-proc and stops. */
if (vrank & mask) {
peer = vrank & ~mask;
if (ompi_op_is_commute(op)) {
peer = (peer + root) % size;
}
err = MCA_PML_CALL(send(snd_buffer, count,
dtype, peer, MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != err) {
goto cleanup_and_return;
}
snd_buffer = (char*)rbuf;
break;
}
/* A low-proc receives, reduces, and moves to a higher
* dimension. */
else {
peer = vrank | mask;
if (peer >= size) {
continue;
}
if (ompi_op_is_commute(op)) {
peer = (peer + root) % size;
}
/* Most of the time (all except the first one for commutative
* operations) we receive in the user provided buffer
* (rbuf). But the exception is here to allow us to dont have
* to copy from the sbuf to a temporary location. If the
* operation is commutative we dont care in which order we
* apply the operation, so for the first time we can receive
* the data in the pml_buffer and then apply to operation
* between this buffer and the user provided data. */
err = MCA_PML_CALL(recv(rcv_buffer, count, dtype, peer,
MCA_COLL_BASE_TAG_REDUCE, comm,
MPI_STATUS_IGNORE));
if (MPI_SUCCESS != err) {
goto cleanup_and_return;
}
/* Perform the operation. The target is always the user
* provided buffer We do the operation only if we receive it
* not in the user buffer */
if (snd_buffer != sbuf) {
/* the target buffer is the locally allocated one */
ompi_op_reduce(op, rcv_buffer, pml_buffer, count, dtype);
} else {
/* If we're commutative, we don't care about the order of
* operations and we can just reduce the operations now.
* If we are not commutative, we have to copy the send
* buffer into a temp buffer (pml_buffer) and then reduce
* what we just received against it. */
if (!ompi_op_is_commute(op)) {
ompi_datatype_copy_content_same_ddt(dtype, count, pml_buffer,
(char*)sbuf);
ompi_op_reduce(op, rbuf, pml_buffer, count, dtype);
} else {
ompi_op_reduce(op, sbuf, pml_buffer, count, dtype);
}
/* now we have to send the buffer containing the computed data */
snd_buffer = pml_buffer;
/* starting from now we always receive in the user
* provided buffer */
rcv_buffer = (char*)rbuf;
}
}
}
/* Get the result to the root if needed. */
err = MPI_SUCCESS;
if (0 == vrank) {
if (root == rank) {
ompi_datatype_copy_content_same_ddt(dtype, count, (char*)rbuf, snd_buffer);
} else {
err = MCA_PML_CALL(send(snd_buffer, count,
dtype, root, MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
}
} else if (rank == root) {
err = MCA_PML_CALL(recv(rcv_buffer, count, dtype, 0,
MCA_COLL_BASE_TAG_REDUCE,
comm, MPI_STATUS_IGNORE));
if (rcv_buffer != rbuf) {
ompi_op_reduce(op, rcv_buffer, rbuf, count, dtype);
}
}
cleanup_and_return:
if (NULL != inplace_temp) {
free(inplace_temp);
}
if (NULL != free_buffer) {
free(free_buffer);
}
if (NULL != free_rbuf) {
free(free_rbuf);
}
/* All done */
return err;
}
/**
* This is a generic implementation of the reduce protocol. It used the tree
* provided as an argument and execute all operations using a segment of
* count times a datatype.
* For the last communication it will update the count in order to limit
* the number of datatype to the original count (original_count)
*
* Note that for non-commutative operations we cannot save memory copy
* for the first block: thus we must copy sendbuf to accumbuf on intermediate
* to keep the optimized loop happy.
*/
int ompi_coll_tuned_reduce_generic( void* sendbuf, void* recvbuf, int original_count,
ompi_datatype_t* datatype, ompi_op_t* op,
int root, ompi_communicator_t* comm,
mca_coll_base_module_t *module,
ompi_coll_tree_t* tree, int count_by_segment,
int max_outstanding_reqs )
{
char *inbuf[2] = {NULL, NULL}, *inbuf_free[2] = {NULL, NULL};
char *accumbuf = NULL, *accumbuf_free = NULL;
char *local_op_buffer = NULL, *sendtmpbuf = NULL;
ptrdiff_t extent, lower_bound, segment_increment;
size_t typelng;
ompi_request_t* reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
int num_segments, line, ret, segindex, i, rank;
int recvcount, prevcount, inbi;
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_get_extent( datatype, &lower_bound, &extent );
ompi_datatype_type_size( datatype, &typelng );
num_segments = (original_count + count_by_segment - 1) / count_by_segment;
segment_increment = (ptrdiff_t)count_by_segment * extent;
sendtmpbuf = (char*) sendbuf;
if( sendbuf == MPI_IN_PLACE ) {
sendtmpbuf = (char *)recvbuf;
}
OPAL_OUTPUT((ompi_coll_tuned_stream, "coll:tuned:reduce_generic count %d, msg size %ld, segsize %ld, max_requests %d",
original_count, (unsigned long)((ptrdiff_t)num_segments * (ptrdiff_t)segment_increment),
(unsigned long)segment_increment, max_outstanding_reqs));
rank = ompi_comm_rank(comm);
/*printf("Tree of rank %d - ", rank);
printf("Parent : %d - ", tree->tree_prev);
printf("Child : ");
for (i = 0; i < tree->tree_nextsize; i++)
printf("%d ", tree->tree_next[i]);
printf("\n");*/
/* non-leaf nodes - wait for children to send me data & forward up
(if needed) */
if( tree->tree_nextsize > 0 ) {
ptrdiff_t true_lower_bound, true_extent, real_segment_size;
ompi_datatype_get_true_extent( datatype, &true_lower_bound,
&true_extent );
/* handle non existant recv buffer (i.e. its NULL) and
protect the recv buffer on non-root nodes */
accumbuf = (char*)recvbuf;
if( (NULL == accumbuf) || (root != rank) ) {
/* Allocate temporary accumulator buffer. */
accumbuf_free = (char*)malloc(true_extent +
(ptrdiff_t)(original_count - 1) * extent);
if (accumbuf_free == NULL) {
line = __LINE__; ret = -1; goto error_hndl;
}
accumbuf = accumbuf_free - lower_bound;
}
/* If this is a non-commutative operation we must copy
sendbuf to the accumbuf, in order to simplfy the loops */
if (!ompi_op_is_commute(op)) {
ompi_datatype_copy_content_same_ddt(datatype, original_count,
(char*)accumbuf,
(char*)sendtmpbuf);
}
/* Allocate two buffers for incoming segments */
real_segment_size = true_extent + (ptrdiff_t)(count_by_segment - 1) * extent;
inbuf_free[0] = (char*) malloc(real_segment_size);
if( inbuf_free[0] == NULL ) {
line = __LINE__; ret = -1; goto error_hndl;
}
inbuf[0] = inbuf_free[0] - lower_bound;
/* if there is chance to overlap communication -
allocate second buffer */
if( (num_segments > 1) || (tree->tree_nextsize > 1) ) {
inbuf_free[1] = (char*) malloc(real_segment_size);
if( inbuf_free[1] == NULL ) {
line = __LINE__; ret = -1; goto error_hndl;
}
inbuf[1] = inbuf_free[1] - lower_bound;
}
/* reset input buffer index and receive count */
inbi = 0;
recvcount = 0;
/* for each segment */
for( segindex = 0; segindex <= num_segments; segindex++ ) {
prevcount = recvcount;
/* recvcount - number of elements in current segment */
recvcount = count_by_segment;
if( segindex == (num_segments-1) )
recvcount = original_count - (ptrdiff_t)count_by_segment * (ptrdiff_t)segindex;
/* for each child */
for( i = 0; i < tree->tree_nextsize; i++ ) {
/**
* We try to overlap communication:
* either with next segment or with the next child
*/
/* post irecv for current segindex on current child */
if( segindex < num_segments ) {
void* local_recvbuf = inbuf[inbi];
if( 0 == i ) {
/* for the first step (1st child per segment) and
* commutative operations we might be able to irecv
* directly into the accumulate buffer so that we can
* reduce(op) this with our sendbuf in one step as
* ompi_op_reduce only has two buffer pointers,
* this avoids an extra memory copy.
*
* BUT if the operation is non-commutative or
* we are root and are USING MPI_IN_PLACE this is wrong!
*/
if( (ompi_op_is_commute(op)) &&
!((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
local_recvbuf = accumbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment;
}
}
ret = MCA_PML_CALL(irecv(local_recvbuf, recvcount, datatype,
tree->tree_next[i],
MCA_COLL_BASE_TAG_REDUCE, comm,
&reqs[inbi]));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl;}
}
/* wait for previous req to complete, if any.
if there are no requests reqs[inbi ^1] will be
MPI_REQUEST_NULL. */
/* wait on data from last child for previous segment */
ret = ompi_request_wait_all( 1, &reqs[inbi ^ 1],
MPI_STATUSES_IGNORE );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
local_op_buffer = inbuf[inbi ^ 1];
if( i > 0 ) {
/* our first operation is to combine our own [sendbuf] data
* with the data we recvd from down stream (but only
* the operation is commutative and if we are not root and
* not using MPI_IN_PLACE)
*/
if( 1 == i ) {
if( (ompi_op_is_commute(op)) &&
!((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
local_op_buffer = sendtmpbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment;
}
}
/* apply operation */
ompi_op_reduce(op, local_op_buffer,
accumbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
recvcount, datatype );
} else if ( segindex > 0 ) {
void* accumulator = accumbuf + (ptrdiff_t)(segindex-1) * (ptrdiff_t)segment_increment;
if( tree->tree_nextsize <= 1 ) {
if( (ompi_op_is_commute(op)) &&
!((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
local_op_buffer = sendtmpbuf + (ptrdiff_t)(segindex-1) * (ptrdiff_t)segment_increment;
}
}
ompi_op_reduce(op, local_op_buffer, accumulator, prevcount,
datatype );
/* all reduced on available data this step (i) complete,
* pass to the next process unless you are the root.
*/
if (rank != tree->tree_root) {
/* send combined/accumulated data to parent */
ret = MCA_PML_CALL( send( accumulator, prevcount,
datatype, tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD,
comm) );
if (ret != MPI_SUCCESS) {
line = __LINE__; goto error_hndl;
}
}
/* we stop when segindex = number of segments
(i.e. we do num_segment+1 steps for pipelining */
if (segindex == num_segments) break;
}
/* update input buffer index */
inbi = inbi ^ 1;
} /* end of for each child */
} /* end of for each segment */
/* clean up */
if( inbuf_free[0] != NULL) free(inbuf_free[0]);
if( inbuf_free[1] != NULL) free(inbuf_free[1]);
if( accumbuf_free != NULL ) free(accumbuf_free);
}
/* leaf nodes
Depending on the value of max_outstanding_reqs and
the number of segments we have two options:
- send all segments using blocking send to the parent, or
- avoid overflooding the parent nodes by limiting the number of
outstanding requests to max_oustanding_reqs.
TODO/POSSIBLE IMPROVEMENT: If there is a way to determine the eager size
for the current communication, synchronization should be used only
when the message/segment size is smaller than the eager size.
*/
else {
/* If the number of segments is less than a maximum number of oustanding
requests or there is no limit on the maximum number of outstanding
requests, we send data to the parent using blocking send */
if ((0 == max_outstanding_reqs) ||
(num_segments <= max_outstanding_reqs)) {
segindex = 0;
while ( original_count > 0) {
if (original_count < count_by_segment) {
count_by_segment = original_count;
}
ret = MCA_PML_CALL( send((char*)sendbuf +
(ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
count_by_segment, datatype,
tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD,
comm) );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
segindex++;
original_count -= count_by_segment;
}
}
/* Otherwise, introduce flow control:
- post max_outstanding_reqs non-blocking synchronous send,
- for remaining segments
- wait for a ssend to complete, and post the next one.
- wait for all outstanding sends to complete.
*/
else {
int creq = 0;
ompi_request_t **sreq = NULL;
sreq = (ompi_request_t**) calloc( max_outstanding_reqs,
sizeof(ompi_request_t*) );
if (NULL == sreq) { line = __LINE__; ret = -1; goto error_hndl; }
/* post first group of requests */
for (segindex = 0; segindex < max_outstanding_reqs; segindex++) {
ret = MCA_PML_CALL( isend((char*)sendbuf +
(ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
count_by_segment, datatype,
tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_SYNCHRONOUS, comm,
&sreq[segindex]) );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
original_count -= count_by_segment;
}
creq = 0;
while ( original_count > 0 ) {
/* wait on a posted request to complete */
ret = ompi_request_wait(&sreq[creq], MPI_STATUS_IGNORE);
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
sreq[creq] = MPI_REQUEST_NULL;
if( original_count < count_by_segment ) {
count_by_segment = original_count;
}
ret = MCA_PML_CALL( isend((char*)sendbuf +
(ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
count_by_segment, datatype,
tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_SYNCHRONOUS, comm,
&sreq[creq]) );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
creq = (creq + 1) % max_outstanding_reqs;
segindex++;
original_count -= count_by_segment;
}
/* Wait on the remaining request to complete */
ret = ompi_request_wait_all( max_outstanding_reqs, sreq,
MPI_STATUSES_IGNORE );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
/* free requests */
free(sreq);
}
}
return OMPI_SUCCESS;
error_hndl: /* error handler */
OPAL_OUTPUT (( ompi_coll_tuned_stream,
"ERROR_HNDL: node %d file %s line %d error %d\n",
rank, __FILE__, line, ret ));
if( inbuf_free[0] != NULL ) free(inbuf_free[0]);
if( inbuf_free[1] != NULL ) free(inbuf_free[1]);
if( accumbuf_free != NULL ) free(accumbuf);
return ret;
}
/*
* exscan_intra
*
* Function: - basic exscan operation
* Accepts: - same arguments as MPI_Exscan()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_exscan_intra(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int size, rank, err;
ptrdiff_t true_lb, true_extent, lb, extent;
char *free_buffer = NULL;
char *reduce_buffer = NULL;
char *source;
MPI_Request req = MPI_REQUEST_NULL;
/* Initialize. */
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
/* If we're rank 0, then we send our sbuf to the next rank */
if (0 == rank) {
return MCA_PML_CALL(send(sbuf, count, dtype, rank + 1,
MCA_COLL_BASE_TAG_EXSCAN,
MCA_PML_BASE_SEND_STANDARD, comm));
}
/* If we're the last rank, then just receive the result from the
* prior rank */
else if ((size - 1) == rank) {
return MCA_PML_CALL(recv(rbuf, count, dtype, rank - 1,
MCA_COLL_BASE_TAG_EXSCAN, comm,
MPI_STATUS_IGNORE));
}
/* Otherwise, get the result from the prior rank, combine it with my
* data, and send it to the next rank */
/* Start the receive for the prior rank's answer */
err = MCA_PML_CALL(irecv(rbuf, count, dtype, rank - 1,
MCA_COLL_BASE_TAG_EXSCAN, comm, &req));
if (MPI_SUCCESS != err) {
goto error;
}
/* Get a temporary buffer to perform the reduction into. Rationale
* for malloc'ing this size is provided in coll_basic_reduce.c. */
ompi_ddt_get_extent(dtype, &lb, &extent);
ompi_ddt_get_true_extent(dtype, &true_lb, &true_extent);
free_buffer = (char*)malloc(true_extent + (count - 1) * extent);
if (NULL == free_buffer) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
reduce_buffer = free_buffer - lb;
if (ompi_op_is_commute(op)) {
/* If we're commutative, we can copy my sbuf into the reduction
* buffer before the receive completes */
err = ompi_ddt_copy_content_same_ddt(dtype, count,
reduce_buffer, (char*)sbuf);
if (MPI_SUCCESS != err) {
goto error;
}
/* Now setup the reduction */
source = (char*)rbuf;
/* Finally, wait for the receive to complete (so that we can do
* the reduction). */
err = ompi_request_wait(&req, MPI_STATUS_IGNORE);
if (MPI_SUCCESS != err) {
goto error;
}
} else {
/* Setup the reduction */
source = (char*)sbuf;
/* If we're not commutative, we have to wait for the receive to
* complete and then copy it into the reduce buffer */
err = ompi_request_wait(&req, MPI_STATUS_IGNORE);
if (MPI_SUCCESS != err) {
goto error;
}
err = ompi_ddt_copy_content_same_ddt(dtype, count,
reduce_buffer, (char*)rbuf);
if (MPI_SUCCESS != err) {
goto error;
}
}
/* Now reduce the received answer with my source into the answer
* that we send off to the next rank */
ompi_op_reduce(op, source, reduce_buffer, count, dtype);
/* Send my result off to the next rank */
err = MCA_PML_CALL(send(reduce_buffer, count, dtype, rank + 1,
MCA_COLL_BASE_TAG_EXSCAN,
MCA_PML_BASE_SEND_STANDARD, comm));
/* Error */
error:
free(free_buffer);
if (MPI_REQUEST_NULL != req) {
ompi_request_cancel(req);
ompi_request_wait(&req, MPI_STATUS_IGNORE);
}
/* All done */
return err;
}
