1a021b8f2d
allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
866 строки
42 KiB
C
866 строки
42 KiB
C
/*
|
|
* Copyright (c) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
|
|
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
|
|
* $COPYRIGHT$
|
|
*
|
|
* Additional copyrights may follow
|
|
*
|
|
* $HEADER$
|
|
*/
|
|
|
|
#ifndef MCA_BCOL_PTPCOLL_BCAST_H
|
|
#define MCA_BCOL_PTPCOLL_BCAST_H
|
|
|
|
#include "ompi_config.h"
|
|
#include "bcol_ptpcoll.h"
|
|
#include "bcol_ptpcoll_utils.h"
|
|
|
|
BEGIN_C_DECLS
|
|
|
|
int bcol_ptpcoll_bcast_init(mca_bcol_base_module_t *super);
|
|
|
|
int bcol_ptpcoll_bcast_k_nomial_anyroot (bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
int bcol_ptpcoll_bcast_k_nomial_anyroot_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
int bcol_ptpcoll_bcast_k_nomial_known_root(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
int bcol_ptpcoll_bcast_k_nomial_known_root_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_anyroot(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_anyroot_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_anyroot_extra(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_anyroot_extra_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_known_root(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_known_root_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_known_root_extra(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_known_root_extra_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
|
|
/* macros */
|
|
#define K_NOMIAL_ROOT_BCAST_NB_BINOMIAL_SCATTER( \
|
|
radix_mask_pow, \
|
|
my_group_index, group_size, group_list, \
|
|
data_buffer, segment_size, count, tag, \
|
|
comm, send_requests, num_pending_sends) \
|
|
do { \
|
|
int rc = OMPI_SUCCESS; \
|
|
int dst; \
|
|
int comm_dst; \
|
|
int send_size; \
|
|
int send_offset; \
|
|
int delta; \
|
|
int dst_boundary_rank; \
|
|
int radix_mask = radix_mask_pow >= 0 ? 1 << radix_mask_pow : 0; \
|
|
\
|
|
while(radix_mask > 0) { \
|
|
/* For each level of tree, do sends */ \
|
|
dst = my_group_index ^ radix_mask; \
|
|
comm_dst = group_list[dst]; \
|
|
\
|
|
dst_boundary_rank = dst & ((~(int)0) << (radix_mask_pow)); \
|
|
\
|
|
send_offset = segment_size * dst_boundary_rank; \
|
|
/* Pasha: make sure that we handle the corner cases */ \
|
|
delta = count - send_offset; \
|
|
if (delta <= 0) { \
|
|
send_size = 0; /* we have to send something, other way it will hang */ \
|
|
} else { \
|
|
/* the tail case */ \
|
|
send_size = (int) \
|
|
(delta - (int)segment_size * radix_mask) < 0 ? delta : \
|
|
(int)segment_size * radix_mask; \
|
|
} \
|
|
\
|
|
/* Non blocking send .... */ \
|
|
PTPCOLL_VERBOSE(9 , \
|
|
("Bcast p2s, Isend to %d[%d],count %d,tag %d,addr %p [%p] send_size %d,send_offset %d, radix %d %d",\
|
|
dst, comm_dst, count, tag, \
|
|
data_buffer, (void *)((unsigned char *)data_buffer + (size_t)send_offset), \
|
|
send_size, \
|
|
send_offset, \
|
|
radix_mask, \
|
|
radix_mask_pow \
|
|
)); \
|
|
rc = MCA_PML_CALL(isend((void *)((unsigned char *)data_buffer + (size_t)send_offset), \
|
|
send_size, MPI_BYTE, \
|
|
comm_dst, tag, \
|
|
MCA_PML_BASE_SEND_STANDARD, comm, \
|
|
&(send_requests[*num_pending_sends]))); \
|
|
PTPCOLL_VERBOSE(10, ("send request addr is %p", send_requests[*num_pending_sends])); \
|
|
if( OMPI_SUCCESS != rc ) { \
|
|
PTPCOLL_VERBOSE(10, ("Failed to isend data")); \
|
|
return OMPI_ERROR; \
|
|
} \
|
|
++(*num_pending_sends); \
|
|
radix_mask >>= 1; \
|
|
radix_mask_pow--; \
|
|
} \
|
|
} while(0)
|
|
|
|
#define NARRAY_SCATTER_NB(narray_node, process_shift, group_size, \
|
|
data_buffer, base_block_size, count, tag, comm, send_requests, \
|
|
num_pending_sends) \
|
|
do { \
|
|
int n, rc = OMPI_SUCCESS; \
|
|
int dst; \
|
|
int comm_dst; \
|
|
int offset; \
|
|
int size_count = count; \
|
|
\
|
|
/* Send out data to all relevant childrens */ \
|
|
for (n = 0; n < narray_node->n_children && size_count > 0; n++) { \
|
|
\
|
|
dst = narray_node->children_ranks[n] + process_shift; \
|
|
if (dst >= group_size) { \
|
|
dst -= group_size; \
|
|
} \
|
|
\
|
|
comm_dst = group_list[dst]; \
|
|
offset = n * base_block_size; \
|
|
size_count -= base_block_size; \
|
|
if (OPAL_UNLIKELY(size_count < 0)) { \
|
|
count = base_block_size + size_count; \
|
|
} else { \
|
|
count = base_block_size; \
|
|
} \
|
|
\
|
|
/* Non blocking send .... */ \
|
|
PTPCOLL_VERBOSE(9 , ("Bcast, Isend data to %d[%d], count %d, tag %d, addr %p", \
|
|
dst, comm_dst, count, tag, \
|
|
data_buffer)); \
|
|
rc = MCA_PML_CALL(isend((void *)((char *)data_buffer + (size_t)offset), count, MPI_BYTE,\
|
|
comm_dst, tag, \
|
|
MCA_PML_BASE_SEND_STANDARD, comm, \
|
|
&(send_requests[*num_pending_sends]))); \
|
|
if( OMPI_SUCCESS != rc ) { \
|
|
PTPCOLL_VERBOSE(10, ("Failed to isend data")); \
|
|
return OMPI_ERROR; \
|
|
} \
|
|
++(*num_pending_sends); \
|
|
} \
|
|
} while(0)
|
|
|
|
#define NARRAY_SCATTER_B(narray_node, process_shift, group_size, \
|
|
data_buffer, base_block_size, count, tag, comm, send_requests, \
|
|
num_pending_sends, completed) \
|
|
do { \
|
|
NARRAY_SCATTER_NB(narray_node, process_shift, group_size, \
|
|
data_buffer, base_block_size, count, tag, comm, send_requests, \
|
|
num_pending_sends); \
|
|
if (*num_pending_sends > 0) { \
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(num_pending_sends, send_requests, &rc); \
|
|
if (OMPI_SUCCESS != rc) { \
|
|
return OMPI_ERROR; \
|
|
} \
|
|
} else { \
|
|
completed = 1; \
|
|
} \
|
|
} while (0)
|
|
|
|
#define CHECK_IF_ROOT_OR_VROOT(module, i) \
|
|
(module->pow_2 == module->ml_mem.ml_buf_desc[i].radix_mask_pow)
|
|
|
|
/* inline functions */
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_bcast_binomial_scatter_gatther_send_extra(
|
|
mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
void *data_buffer, int count, int tag,
|
|
int extra_peer, ompi_communicator_t *comm,
|
|
int *active_requests, ompi_request_t **requests)
|
|
{
|
|
int rc = OMPI_SUCCESS;
|
|
int completed = 0;
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
|
|
/* tag is -1 already */
|
|
/* send the all data to your extra peer */
|
|
PTPCOLL_VERBOSE(10, ("bcol_ptpcoll_bcast_binomial_scatter_gatther_send_extra to %d tag %d",
|
|
extra_peer, tag));
|
|
rc = MCA_PML_CALL(isend(data_buffer, count, MPI_BYTE,
|
|
group_list[extra_peer], tag,
|
|
MCA_PML_BASE_SEND_STANDARD, comm,
|
|
&(requests[*active_requests])));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to send data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
|
|
++(*active_requests);
|
|
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
|
|
if (0 == completed) {
|
|
PTPCOLL_VERBOSE(10, ("PR Extra send was not completed"));
|
|
/* we have to store the iteration number somewhere */
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_send_n_extra(mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
void *data_buffer, int count, int tag,
|
|
int *extra_peers, int num_peers, int skip,
|
|
ompi_communicator_t *comm,
|
|
int *active_requests, ompi_request_t **requests)
|
|
{
|
|
int rc = OMPI_SUCCESS;
|
|
int completed = 0;
|
|
int i;
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
|
|
/* send the all data to your extra peer */
|
|
for (i = 0; i < num_peers; i++) {
|
|
PTPCOLL_VERBOSE(10, ("send_n_extra to %d tag %d",
|
|
extra_peers[i], tag));
|
|
if (extra_peers[i] == skip) {
|
|
PTPCOLL_VERBOSE(10, ("SKIP"));
|
|
continue;
|
|
}
|
|
|
|
rc = MCA_PML_CALL(isend(data_buffer, count, MPI_BYTE,
|
|
group_list[extra_peers[i]], tag,
|
|
MCA_PML_BASE_SEND_STANDARD, comm,
|
|
&(requests[*active_requests])));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to send data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
|
|
++(*active_requests);
|
|
}
|
|
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
|
|
if (0 == completed) {
|
|
PTPCOLL_VERBOSE(10, ("PR Extra send was not completed"));
|
|
/* we have to store the iteration number somewhere */
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_bcast_binomial_gather_anyroot(mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
int buffer_index, void *data_buffer, int count, int base_block_size)
|
|
{
|
|
int rc;
|
|
int completed = 0; /* not completed */
|
|
int *active_requests =
|
|
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
|
|
int i;
|
|
int *iteration =
|
|
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
|
|
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
|
|
ompi_request_t **requests =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
|
|
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
|
|
void *curr_data_sbuffer = NULL,
|
|
*curr_data_rbuffer = NULL;
|
|
int radix_mask_pow = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask_pow;
|
|
int delta;
|
|
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
|
|
PTPCOLL_VERBOSE(10, ("bcol_ptpcoll_bcast_binomial_gather_anyroot %d %d %d",
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration,
|
|
ptpcoll_module->pow_2,
|
|
1 << ptpcoll_module->pow_2));
|
|
|
|
/* we assume the iteration #iteration already was completed with probe */
|
|
for (i = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
|
|
i < ptpcoll_module->pow_2; i++) {
|
|
int pow2 = 1 << i;
|
|
int peer_index = my_group_index ^ pow2;
|
|
int comm_rank = group_list[peer_index];
|
|
int slen, rlen,
|
|
send_offset,
|
|
recv_offset;
|
|
|
|
if (i > radix_mask_pow) {
|
|
/* *active_requests = 0; */
|
|
/* send - receive data from the peer */
|
|
slen = rlen = pow2 * base_block_size;
|
|
send_offset = base_block_size * ((my_group_index) & ((~(int)0) << i));
|
|
recv_offset = base_block_size * ((peer_index) & ((~(int)0) << i));
|
|
curr_data_sbuffer = (void *)((unsigned char *)data_buffer + send_offset);
|
|
curr_data_rbuffer = (void *)((unsigned char *)data_buffer + recv_offset);
|
|
|
|
delta = count - recv_offset;
|
|
if (delta > 0) {
|
|
if (delta < rlen) {
|
|
/* recv the tail */
|
|
rlen = delta;
|
|
}
|
|
PTPCOLL_VERBOSE(10, ("[ pow2 %d, radix %d ] recv data %p (offset %d) , len %d , dest %d",
|
|
pow2,
|
|
1 << ptpcoll_module->pow_2,
|
|
curr_data_rbuffer,
|
|
recv_offset,
|
|
rlen,
|
|
comm_rank));
|
|
rc = MCA_PML_CALL(irecv(curr_data_rbuffer, rlen, MPI_BYTE,
|
|
comm_rank, tag, comm, &requests[*active_requests]));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
++(*active_requests);
|
|
}
|
|
|
|
delta = count - send_offset;
|
|
if (delta > 0) {
|
|
if (delta < slen) {
|
|
/* recv the tail */
|
|
slen = delta;
|
|
}
|
|
PTPCOLL_VERBOSE(10, ("[ pow2 %d, radix %d ] sending data %p (offset %d) , len %d , dest %d",
|
|
pow2,
|
|
1 << ptpcoll_module->pow_2,
|
|
curr_data_sbuffer,
|
|
send_offset,
|
|
slen,
|
|
comm_rank));
|
|
rc = MCA_PML_CALL(isend(curr_data_sbuffer, slen, MPI_BYTE,
|
|
comm_rank, tag,
|
|
MCA_PML_BASE_SEND_STANDARD, comm,
|
|
&(requests[*active_requests])));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to send data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
++(*active_requests);
|
|
}
|
|
|
|
if (*active_requests > 0) {
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
|
|
if (0 == completed) {
|
|
*iteration = i;
|
|
/* we have to store the iteration number somewhere */
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
}
|
|
} else if (i == radix_mask_pow) {
|
|
/* only receive data */
|
|
rlen = pow2 * base_block_size;
|
|
recv_offset = base_block_size * ((peer_index) & ((~(int)0) << i));
|
|
curr_data_rbuffer = (void *)((unsigned char *)data_buffer + recv_offset);
|
|
delta = count - recv_offset;
|
|
if (0 >= delta) {
|
|
/* we have nothing to send, skip the iteration */
|
|
continue;
|
|
}
|
|
if (delta < rlen) {
|
|
/* recv the tail */
|
|
rlen = delta;
|
|
}
|
|
/* receive data from the peer */
|
|
PTPCOLL_VERBOSE(10, ("[ pow2 %d, radix %d ] recv data %p (offset %d) , len %d , dest %d",
|
|
pow2,
|
|
1 << ptpcoll_module->pow_2,
|
|
curr_data_rbuffer,
|
|
recv_offset,
|
|
rlen,
|
|
comm_rank));
|
|
rc = MCA_PML_CALL(irecv(curr_data_rbuffer, rlen, MPI_BYTE,
|
|
comm_rank, tag, comm, &(requests[*active_requests])));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
++(*active_requests);
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
|
|
if (0 == completed) {
|
|
*iteration = i;
|
|
PTPCOLL_VERBOSE(10, ("Recv was not completed"));
|
|
/* we have to store the iteration number somewhere */
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
PTPCOLL_VERBOSE(10, ("Recv was completed"));
|
|
} else if (i < radix_mask_pow) {
|
|
/* Only send data */
|
|
slen = pow2 * base_block_size;
|
|
send_offset = base_block_size * ((my_group_index) & ((~(int)0) << i));
|
|
curr_data_sbuffer = (void *)((unsigned char *)data_buffer + send_offset);
|
|
delta = count - send_offset;
|
|
if (0 >= delta) {
|
|
/* we have nothing to send, skip the iteration */
|
|
continue;
|
|
}
|
|
if (delta < slen) {
|
|
slen = delta;
|
|
}
|
|
PTPCOLL_VERBOSE(10, ("[ pow2 %d, radix %d ] sending data %p (offset %d) , len %d , dest %d",
|
|
pow2,
|
|
1 << ptpcoll_module->pow_2,
|
|
curr_data_sbuffer,
|
|
send_offset,
|
|
slen,
|
|
comm_rank));
|
|
rc = MCA_PML_CALL(isend(curr_data_sbuffer, slen, MPI_BYTE,
|
|
comm_rank, tag, MCA_PML_BASE_SEND_STANDARD, comm,
|
|
&(requests[*active_requests])));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to send data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
++(*active_requests);
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
|
|
if (0 == completed) {
|
|
*iteration = i;
|
|
/* we have to store the iteration number somewhere */
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
}
|
|
}
|
|
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_bcast_binomial_probe_and_scatter_anyroot(mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
int buffer_index, void *data_buffer, int count, int base_block_size)
|
|
{
|
|
mca_bcol_ptpcoll_component_t *cm = &mca_bcol_ptpcoll_component;
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
int rc;
|
|
int completed = 0; /* not completed */
|
|
int comm_root;
|
|
int i;
|
|
int *radix_mask_pow =
|
|
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask_pow);
|
|
int *active_requests =
|
|
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
|
|
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
|
|
ompi_status_public_t status;
|
|
ompi_request_t **requests =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
|
|
int pow2_group_size = ptpcoll_module->pow_2num;
|
|
int pow2_distance;
|
|
int my_left_boundary_rank;
|
|
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
|
|
int group_root_index = 0;
|
|
void *curr_data_buffer = NULL;
|
|
int tag =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag;
|
|
int recv_count = 0;
|
|
int *coll_status =
|
|
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].status;
|
|
|
|
assert(0 == *active_requests);
|
|
|
|
PTPCOLL_VERBOSE(10, ("Running bcol_ptpcoll_bcast_binomial_probe_and_scatter_anyroot"));
|
|
for (i = 0; i < cm->num_to_probe &&
|
|
0 == completed; i++) {
|
|
MCA_PML_CALL(iprobe(MPI_ANY_SOURCE, tag,
|
|
comm, &completed, &status));
|
|
PTPCOLL_VERBOSE(10, ("Bcast, iprobe tag %d",
|
|
tag));
|
|
}
|
|
|
|
/* the function always returns OMPI_SUCCESS, so we don't check return code */
|
|
if (0 == completed) {
|
|
PTPCOLL_VERBOSE(10, ("IPROBE was not matched"));
|
|
/* No data was received, return no match error */
|
|
return BCOL_FN_NOT_STARTED;
|
|
}
|
|
|
|
comm_root = status.MPI_SOURCE;
|
|
|
|
|
|
PTPCOLL_VERBOSE(9, ("IPROBE was matched, root of the data on communicator is %d", comm_root));
|
|
|
|
/* For proxy we have to check if we got something from extra node */
|
|
if (PTPCOLL_PROXY & ptpcoll_module->pow_2type) {
|
|
if (group_list[ptpcoll_module->proxy_extra_index] == comm_root) {
|
|
PTPCOLL_VERBOSE(9, ("IPROBE was matched, root of the data on communicator is extra node %d",
|
|
comm_root));
|
|
/* scatter the data among other peer in the pow2 group */
|
|
*radix_mask_pow = ptpcoll_module->pow_2;
|
|
|
|
pow2_distance = ptpcoll_module->pow_2 - 1;
|
|
curr_data_buffer = data_buffer;
|
|
recv_count = count;
|
|
goto PR_SCATTHER;
|
|
}
|
|
}
|
|
|
|
/* Find group index for communicator root of the data */
|
|
group_root_index = get_group_index_and_distance_for_binomial
|
|
(my_group_index, comm_root, pow2_group_size, group_list, &pow2_distance);
|
|
if (OPAL_UNLIKELY(group_root_index < 0)) {
|
|
PTPCOLL_ERROR(("Fatal error, no group root index found, my id %d, pow2_g_size %d comm_root %d",
|
|
my_group_index, pow2_group_size, comm_root));
|
|
return OMPI_ERROR;
|
|
}
|
|
|
|
PTPCOLL_VERBOSE(10, ("Group root index is %d distance is %d",
|
|
group_root_index, pow2_distance));
|
|
|
|
/* Use group_root_index to calculate the */
|
|
|
|
/* Post receive that will fetch the data */
|
|
/* Pasha: Who is packing data ?
|
|
Should I assume that we get contiguous buffer ?
|
|
Or should I pack by myself
|
|
===================================================================================================
|
|
=== On this stage I assume that data is contiguous. So I use MPI_BYTE datatype and COUNT = size ===
|
|
===================================================================================================
|
|
*/
|
|
|
|
recv_count = base_block_size * (1 << pow2_distance); /* we may receive larger data */
|
|
|
|
my_left_boundary_rank = my_group_index & ((~(int)0) << pow2_distance );
|
|
|
|
curr_data_buffer = (void *)((unsigned char *)data_buffer +
|
|
(size_t) base_block_size * my_left_boundary_rank);
|
|
|
|
*radix_mask_pow = pow2_distance;
|
|
|
|
pow2_distance--;
|
|
|
|
PR_SCATTHER:
|
|
PTPCOLL_VERBOSE(10, ("Bcast, receive data from %d[%d], "
|
|
"recv_count %d, tag %d, addr %p, offset %d, pow2_distace %d",
|
|
comm_root, group_root_index, recv_count,
|
|
tag, curr_data_buffer,
|
|
my_group_index * base_block_size, pow2_distance));
|
|
|
|
rc = MCA_PML_CALL(recv(curr_data_buffer, recv_count, MPI_BYTE,
|
|
comm_root, tag, comm, MPI_STATUS_IGNORE));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
|
|
PTPCOLL_VERBOSE(10, ("Bcast, Data was received"));
|
|
|
|
/* Sending forward the data over K-nomial tree */
|
|
*coll_status = PTPCOLL_SCATTER_STARTED;
|
|
K_NOMIAL_ROOT_BCAST_NB_BINOMIAL_SCATTER(
|
|
pow2_distance,
|
|
my_group_index, group_size, group_list,
|
|
data_buffer, base_block_size,
|
|
count, tag, comm, requests,
|
|
active_requests);
|
|
|
|
/* Since the next step (gather) does not really require
|
|
completion on scatter , we may return complete */
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_binomial_root_to_src(int group_root, int my_rank,
|
|
int pow2_size, int group_size, int *distance)
|
|
{
|
|
int root, relative_rank, src,
|
|
pow2_distance = 0, i;
|
|
|
|
if (group_root < pow2_size) {
|
|
root = group_root;
|
|
} else {
|
|
/* the source of the data is extra node,
|
|
the real root it represented by some rank from
|
|
pow2 group */
|
|
root = group_root - pow2_size;
|
|
/* shortcut for the case when my rank is root for the group */
|
|
if (my_rank == root) {
|
|
*distance = -1;
|
|
return group_root;
|
|
}
|
|
}
|
|
|
|
relative_rank = (my_rank - root) < 0 ? my_rank - root + pow2_size :
|
|
my_rank - root;
|
|
|
|
for (i = 1; i < pow2_size; i<<=1, pow2_distance++) {
|
|
if (relative_rank & i) {
|
|
src = my_rank ^ i;
|
|
if (src >= pow2_size)
|
|
src -= pow2_size;
|
|
|
|
*distance = pow2_distance;
|
|
return src;
|
|
}
|
|
}
|
|
|
|
/* error case */
|
|
*distance = -1;
|
|
return -1;
|
|
}
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_bcast_binomial_test_and_scatter_known_root(mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
int buffer_index, void *data_buffer, int count, int base_block_size)
|
|
{
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
int rc;
|
|
int *active_requests =
|
|
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
|
|
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
|
|
ompi_request_t **requests =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
|
|
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
|
|
int tmp_radix_mask_pow =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask_pow - 1;
|
|
int tag =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag;
|
|
int *status =
|
|
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].status;
|
|
|
|
PTPCOLL_VERBOSE(10, ("Running bcol_ptpcoll_bcast_binomial_probe_and_scatter_anyroot"));
|
|
|
|
if (0 == mca_bcol_ptpcoll_test_all_for_match(active_requests,
|
|
requests, &rc)) {
|
|
PTPCOLL_VERBOSE(10, ("Test was not matched - %d", rc));
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
|
|
PTPCOLL_VERBOSE(10, ("Bcast, Data was received"));
|
|
|
|
/* Sending forward the data over binimial nomial tree */
|
|
*status = PTPCOLL_SCATTER_STARTED;
|
|
K_NOMIAL_ROOT_BCAST_NB_BINOMIAL_SCATTER(
|
|
tmp_radix_mask_pow,
|
|
my_group_index, group_size, group_list,
|
|
data_buffer, base_block_size,
|
|
count, tag, comm, requests,
|
|
active_requests);
|
|
|
|
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
#define NARRAY_BLOCK_SIZE(size, module, level_size) \
|
|
((size + (module)->full_narray_tree_num_leafs - 1) / \
|
|
(module)->full_narray_tree_num_leafs) * \
|
|
((module)->full_narray_tree_num_leafs / \
|
|
((0 == level_size) ? \
|
|
mca_bcol_ptpcoll_component.narray_knomial_radix : \
|
|
level_size))
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_bcast_narray_test_and_scatter_known_root(mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
int buffer_index, void *data_buffer, int count, int process_shift,
|
|
int relative_group_index)
|
|
{
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
int rc;
|
|
int *active_requests =
|
|
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
|
|
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
|
|
ompi_request_t **requests =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
|
|
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag;
|
|
int *status =
|
|
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].status;
|
|
int scatter_count = 0;
|
|
int offset = 0;
|
|
int base_block_size = 0;
|
|
void *curr_data_buffer = NULL;
|
|
|
|
PTPCOLL_VERBOSE(10, ("Running bcol_ptpcoll_bcast_narray_test_and_scatter_known_root"));
|
|
|
|
if (0 == mca_bcol_ptpcoll_test_all_for_match(active_requests,
|
|
requests, &rc)) {
|
|
PTPCOLL_VERBOSE(10, ("Test was not matched - %d", rc));
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
|
|
/* Sending forward the data over binimial nomial tree */
|
|
*status = PTPCOLL_SCATTER_STARTED;
|
|
if(0 == relative_group_index) {
|
|
scatter_count = count;
|
|
} else {
|
|
scatter_count = NARRAY_BLOCK_SIZE(count, ptpcoll_module,
|
|
ptpcoll_module->narray_knomial_node[relative_group_index].level_size);
|
|
}
|
|
|
|
offset = scatter_count *
|
|
ptpcoll_module->narray_knomial_node[relative_group_index].rank_on_level;
|
|
|
|
/* make sure that we do not overun memory */
|
|
if (OPAL_UNLIKELY(offset + scatter_count > count)) {
|
|
scatter_count = count - offset;
|
|
}
|
|
|
|
PTPCOLL_VERBOSE(10, ("Bcast, Data was received %d %d %d",
|
|
scatter_count,
|
|
ptpcoll_module->narray_knomial_node[relative_group_index].level_size,
|
|
ptpcoll_module->narray_knomial_node[relative_group_index].rank_on_level));
|
|
|
|
|
|
curr_data_buffer = (void *)((unsigned char *)data_buffer + (size_t)offset);
|
|
|
|
/* calculating scatter block size for next level of tree */
|
|
base_block_size = NARRAY_BLOCK_SIZE(count, ptpcoll_module,
|
|
ptpcoll_module->narray_knomial_node[relative_group_index].level_size *
|
|
mca_bcol_ptpcoll_component.narray_knomial_radix);
|
|
|
|
PTPCOLL_VERBOSE(10, ("scatter_known_rootaaa %d %d %d %d %d",scatter_count, offset, base_block_size,
|
|
ptpcoll_module->narray_knomial_node[relative_group_index].level_size /mca_bcol_ptpcoll_component.narray_knomial_radix,
|
|
ptpcoll_module->full_narray_tree_num_leafs));
|
|
|
|
NARRAY_SCATTER_NB((&ptpcoll_module->narray_knomial_node[relative_group_index]),
|
|
process_shift, ptpcoll_module->full_narray_tree_size,
|
|
curr_data_buffer, base_block_size, scatter_count, tag, comm,
|
|
requests, active_requests);
|
|
|
|
/* Bummer, I tried to prevent this, special case for virtual root */
|
|
if(0 == relative_group_index) {
|
|
if (0 == mca_bcol_ptpcoll_test_all_for_match(active_requests,
|
|
requests, &rc)) {
|
|
PTPCOLL_VERBOSE(10, ("Test was not matched - %d", rc));
|
|
*status = PTPCOLL_ROOT_SEND_STARTED;
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
}
|
|
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
int bcol_ptpcoll_bcast_narray_knomial_gather(mca_bcol_ptpcoll_module_t *ptpcoll_module,
|
|
const int buffer_index, void *data_buffer, const int count,
|
|
const int relative_group_index)
|
|
{
|
|
int completed = 0; /* not completed */
|
|
int *active_requests =
|
|
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
|
|
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
|
|
int blocks_in_step =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask;
|
|
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
|
|
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
|
|
int group_size = ptpcoll_module->full_narray_tree_size;
|
|
int i, k,
|
|
rc,
|
|
len, slen, rlen,
|
|
peer, group_peer;
|
|
size_t s_offset,
|
|
r_offset;
|
|
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
|
|
ompi_request_t **requests =
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
|
|
netpatterns_narray_knomial_tree_node_t *narray_node =
|
|
&ptpcoll_module->narray_knomial_node[relative_group_index];
|
|
netpatterns_k_exchange_node_t *k_node =
|
|
&narray_node->k_node;
|
|
mca_bcol_ptpcoll_component_t *cm =
|
|
&mca_bcol_ptpcoll_component;
|
|
size_t base_block_size =
|
|
NARRAY_BLOCK_SIZE(count, ptpcoll_module, narray_node->level_size);
|
|
|
|
PTPCOLL_VERBOSE(10, ("bcol_ptpcoll_bcast_narray_knomial_gather %d %d %d %d %d %d %d",
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration,
|
|
base_block_size, count, narray_node->level_size,
|
|
relative_group_index, k_node->n_exchanges, tag));
|
|
|
|
/* we assume the iteration #iteration already was completed with probe */
|
|
for (i = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
|
|
i < k_node->n_exchanges; i++, blocks_in_step *= cm->narray_knomial_radix) {
|
|
|
|
len = base_block_size * blocks_in_step;
|
|
|
|
for (k = 0; k < cm->narray_knomial_radix - 1; k++) {
|
|
group_peer = my_group_index +
|
|
(k_node->rank_exchanges[i][k] - narray_node->rank_on_level);
|
|
if (group_peer >= group_size) {
|
|
group_peer -= group_size;
|
|
} else if (group_peer < 0) {
|
|
group_peer += group_size;
|
|
}
|
|
peer = group_list[group_peer];
|
|
|
|
r_offset = (size_t)k_node->rank_exchanges[i][k] / blocks_in_step *
|
|
len;
|
|
|
|
/* check that we do not run out of message boundary */
|
|
if (OPAL_UNLIKELY(r_offset + len > (size_t)count)) {
|
|
rlen = count - r_offset;
|
|
if (OPAL_UNLIKELY(rlen <= 0)) {
|
|
continue;
|
|
}
|
|
} else {
|
|
rlen = len;
|
|
}
|
|
PTPCOLL_VERBOSE(10, ("Recv data from %d, addr %p offset %d len %d %d %d tag %d",
|
|
peer, data_buffer, r_offset, rlen, len, blocks_in_step, tag));
|
|
rc = MCA_PML_CALL(irecv((void *)((unsigned char *)data_buffer + r_offset),
|
|
rlen, MPI_BYTE,
|
|
peer, tag, comm, &requests[*active_requests]));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
++(*active_requests);
|
|
}
|
|
|
|
for (k = 0; k < cm->narray_knomial_radix - 1; k++) {
|
|
group_peer = my_group_index +
|
|
(k_node->rank_exchanges[i][k] - narray_node->rank_on_level);
|
|
if (group_peer >= group_size) {
|
|
group_peer -= group_size;
|
|
} else if (group_peer < 0) {
|
|
group_peer += group_size;
|
|
}
|
|
peer = group_list[group_peer];
|
|
|
|
s_offset = (size_t)narray_node->rank_on_level / blocks_in_step *
|
|
len;
|
|
|
|
/* check that we do not run out of message boundary */
|
|
if (OPAL_UNLIKELY(s_offset + len > (size_t)count)) {
|
|
slen = count - s_offset;
|
|
if (OPAL_UNLIKELY(slen <= 0)) {
|
|
continue;
|
|
}
|
|
} else {
|
|
slen = len;
|
|
}
|
|
|
|
PTPCOLL_VERBOSE(10, ("Send data from %d, addr %p offset %d len %d %d %d tag %d",
|
|
peer, data_buffer, s_offset, slen, len, blocks_in_step, tag));
|
|
rc = MCA_PML_CALL(isend((void *)((unsigned char *)data_buffer + s_offset),
|
|
slen, MPI_BYTE,
|
|
peer, tag, MCA_PML_BASE_SEND_STANDARD, comm,
|
|
&(requests[*active_requests])));
|
|
if( OMPI_SUCCESS != rc ) {
|
|
PTPCOLL_VERBOSE(10, ("Failed to send data"));
|
|
return OMPI_ERROR;
|
|
}
|
|
++(*active_requests);
|
|
}
|
|
|
|
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
|
|
if (0 == completed) {
|
|
/* cache data for next iteration */
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration =
|
|
i; /* why not to store step for next iteration ?! */
|
|
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask =
|
|
blocks_in_step * cm->narray_knomial_radix;
|
|
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
|
|
}
|
|
}
|
|
|
|
return BCOL_FN_COMPLETE;
|
|
}
|
|
|
|
END_C_DECLS
|
|
|
|
#endif
|