ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
openmpi/ompi/mca/coll/adapt/coll_adapt_ireduce.c
George Bosilca c98e387a53
Many fixes and improvements to ADAPT 
- Add support for fallback to previous coll module on non-commutative operations (#30)
- Replace mutexes by atomic operations.
- Use the correct nbc request type (for both ibcast and ireduce)
  * coll/base: document type casts in ompi_coll_base_retain_*
- add module-wide topology cache
- use standard instead of synchronous send and add mca parameter to control mode of initial send in ireduce/ibcast
- reduce number of memory allocations
- call the default request completion.
  - Remove the requests from the Fortran lookup conversion tables before completing
    and free it.

Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
Signed-off-by: Joseph Schuchart <schuchart@hlrs.de>

Co-authored-by: Joseph Schuchart <schuchart@hlrs.de>
2020-09-18 12:50:17 -04:00

772 строки
37 KiB
C
Исходник Ответственный История

/*
* Copyright (c) 2014-2020 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/communicator/communicator.h"
#include "coll_adapt.h"
#include "coll_adapt_algorithms.h"
#include "coll_adapt_context.h"
#include "coll_adapt_item.h"
#include "coll_adapt_topocache.h"
#include "ompi/constants.h"
#include "ompi/mca/coll/base/coll_base_util.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/coll/base/coll_base_topo.h"
static int ompi_coll_adapt_ireduce_generic(IREDUCE_ARGS,
ompi_coll_tree_t * tree, size_t seg_size);
/* MPI_Reduce and MPI_Ireduce in the ADAPT module only work for commutative operations */
/*
* Set up MCA parameters of MPI_Reduce and MPI_Ireduce
*/
int ompi_coll_adapt_ireduce_register(void)
{
mca_base_component_t *c = &mca_coll_adapt_component.super.collm_version;
mca_coll_adapt_component.adapt_ireduce_algorithm = 1;
mca_base_component_var_register(c, "reduce_algorithm",
"Algorithm of reduce, 1: binomial, 2: in_order_binomial, 3: binary, 4: pipeline, 5: chain, 6: linear", MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5, MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_ireduce_algorithm);
if( (mca_coll_adapt_component.adapt_ireduce_algorithm < 0) ||
(mca_coll_adapt_component.adapt_ireduce_algorithm > OMPI_COLL_ADAPT_ALGORITHM_COUNT) ) {
mca_coll_adapt_component.adapt_ireduce_algorithm = 1;
}
mca_coll_adapt_component.adapt_ireduce_segment_size = 163740;
mca_base_component_var_register(c, "reduce_segment_size",
"Segment size in bytes used by default for reduce algorithms. Only has meaning if algorithm is forced and supports segmenting. 0 bytes means no segmentation.",
MCA_BASE_VAR_TYPE_SIZE_T, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_ireduce_segment_size);
mca_coll_adapt_component.adapt_ireduce_max_send_requests = 2;
mca_base_component_var_register(c, "reduce_max_send_requests",
"Maximum number of send requests",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_ireduce_max_send_requests);
mca_coll_adapt_component.adapt_ireduce_max_recv_requests = 3;
mca_base_component_var_register(c, "reduce_max_recv_requests",
"Maximum number of receive requests per peer",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_ireduce_max_recv_requests);
mca_coll_adapt_component.adapt_inbuf_free_list_min = 10;
mca_base_component_var_register(c, "inbuf_free_list_min",
"Minimum number of segment in inbuf free list",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_inbuf_free_list_min);
mca_coll_adapt_component.adapt_inbuf_free_list_max = 10000;
mca_base_component_var_register(c, "inbuf_free_list_max",
"Maximum number of segment in inbuf free list",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_inbuf_free_list_max);
mca_coll_adapt_component.adapt_inbuf_free_list_inc = 10;
mca_base_component_var_register(c, "inbuf_free_list_inc",
"Number of segments to allocate when growing the inbuf free list",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_inbuf_free_list_inc);
mca_coll_adapt_component.adapt_ireduce_synchronous_send = true;
(void) mca_base_component_var_register(c, "reduce_synchronous_send",
"Whether to use synchronous send operations during setup of reduce operations",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_adapt_component.adapt_ireduce_synchronous_send);
mca_coll_adapt_component.adapt_ireduce_context_free_list = NULL;
return OMPI_SUCCESS;
}
/*
* Release the free list created in ompi_coll_adapt_ireduce_generic
*/
int ompi_coll_adapt_ireduce_fini(void)
{
if (NULL != mca_coll_adapt_component.adapt_ireduce_context_free_list) {
OBJ_RELEASE(mca_coll_adapt_component.adapt_ireduce_context_free_list);
mca_coll_adapt_component.adapt_ireduce_context_free_list = NULL;
OPAL_OUTPUT_VERBOSE((10, mca_coll_adapt_component.adapt_output, "ireduce fini\n"));
}
return OMPI_SUCCESS;
}
/*
* Functions to access list
*/
static ompi_coll_adapt_item_t *get_next_ready_item(ompi_coll_adapt_constant_reduce_context_t *con, int num_children)
{
ompi_coll_adapt_item_t *item = NULL, *temp_item;
if (opal_list_is_empty(&con->recv_list)) {
return NULL;
}
OPAL_THREAD_LOCK(&con->mutex_recv_list);
OPAL_LIST_FOREACH(temp_item, &con->recv_list, ompi_coll_adapt_item_t) {
if (temp_item->count == num_children) {
item = temp_item;
opal_list_remove_item(&con->recv_list, (opal_list_item_t *) temp_item);
break;
}
}
OPAL_THREAD_UNLOCK(&con->mutex_recv_list);
return item;
}
static int add_to_recv_list(ompi_coll_adapt_constant_reduce_context_t *con, int id)
{
ompi_coll_adapt_item_t *item;
OPAL_THREAD_LOCK(&con->mutex_recv_list);
OPAL_LIST_FOREACH(item, &con->recv_list, ompi_coll_adapt_item_t) {
if (item->id == id) {
(item->count)++;
OPAL_THREAD_UNLOCK(&con->mutex_recv_list);
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output, "add_to_recv_list_return 1\n"));
return 1;
}
}
/* Add a new object to the list with count set to 1 */
item = OBJ_NEW(ompi_coll_adapt_item_t);
item->id = id;
item->count = 1;
opal_list_append(&con->recv_list, (opal_list_item_t *) item);
OPAL_THREAD_UNLOCK(&con->mutex_recv_list);
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output, "add_to_recv_list_return 2\n"));
return 2;
}
/*
* Get the inbuf address
*/
static ompi_coll_adapt_inbuf_t *to_inbuf(char *buf, int distance)
{
return (ompi_coll_adapt_inbuf_t *) (buf - distance);
}
/*
* Finish a ireduce request
*/
static int ireduce_request_fini(ompi_coll_adapt_reduce_context_t * context)
{
/* Return the allocated recourses */
ompi_request_t *temp_req = context->con->request;
if (context->con->accumbuf != NULL) {
if (context->con->rank != context->con->root) {
for (int i = 0; i < context->con->num_segs; i++) {
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: Return accumbuf %d %p\n",
ompi_comm_rank(context->con->comm), i,
(void *) to_inbuf(context->con->accumbuf[i],
context->con->distance)));
opal_free_list_return_st(&context->con->inbuf_list,
(opal_free_list_item_t *) to_inbuf(context->con->accumbuf[i],
context->con->distance));
}
}
free(context->con->accumbuf);
}
for (int i = 0; i < context->con->num_segs; i++) {
OBJ_DESTRUCT(&context->con->mutex_op_list[i]);
}
free(context->con->mutex_op_list);
if (context->con->tree->tree_nextsize > 0) {
free(context->con->next_recv_segs);
}
OBJ_RELEASE(context->con);
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output, "return context_list\n"));
opal_free_list_return(mca_coll_adapt_component.adapt_ireduce_context_free_list,
(opal_free_list_item_t *) context);
/* Complete the request */
ompi_request_complete(temp_req, 1);
return OMPI_SUCCESS;
}
/*
* Callback function of isend
*/
static int send_cb(ompi_request_t * req)
{
ompi_coll_adapt_reduce_context_t *context =
(ompi_coll_adapt_reduce_context_t *) req->req_complete_cb_data;
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: ireduce_send_cb, peer %d, seg_id %d\n", context->con->rank,
context->peer, context->seg_index));
int err;
opal_atomic_sub_fetch_32((opal_atomic_int32_t*)&(context->con->ongoing_send), 1);
/* Send a new segment */
ompi_coll_adapt_item_t *item =
get_next_ready_item(context->con, context->con->tree->tree_nextsize);
if (item != NULL) {
/* Get new context item from free list */
ompi_coll_adapt_reduce_context_t *send_context =
(ompi_coll_adapt_reduce_context_t *) opal_free_list_wait(mca_coll_adapt_component.
adapt_ireduce_context_free_list);
if (context->con->tree->tree_nextsize > 0) {
send_context->buff = context->con->accumbuf[item->id];
} else {
send_context->buff =
context->buff + (item->id - context->seg_index) * context->con->segment_increment;
}
send_context->seg_index = item->id;
send_context->peer = context->peer;
send_context->con = context->con;
opal_atomic_add_fetch_32((opal_atomic_int32_t*)&(context->con->ongoing_send), 1);
int send_count = send_context->con->seg_count;
if (item->id == (send_context->con->num_segs - 1)) {
send_count = send_context->con->count - item->id * send_context->con->seg_count;
}
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In send_cb, create isend to seg %d, peer %d, tag %d\n",
send_context->con->rank, send_context->seg_index, send_context->peer,
send_context->con->ireduce_tag - send_context->seg_index));
ompi_request_t *send_req;
err = MCA_PML_CALL(isend
(send_context->buff, send_count, send_context->con->datatype,
send_context->peer,
context->con->ireduce_tag - send_context->seg_index,
MCA_PML_BASE_SEND_STANDARD, send_context->con->comm, &send_req));
if (MPI_SUCCESS != err) {
return err;
}
/* Release the item */
OBJ_RELEASE(item);
/* Set the send callback */
ompi_request_set_callback(send_req, send_cb, send_context);
}
int32_t num_sent = opal_atomic_add_fetch_32((opal_atomic_int32_t*)&(context->con->num_sent_segs), 1);
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In send_cb, root = %d, num_sent = %d, num_segs = %d\n",
context->con->rank, context->con->tree->tree_root, num_sent,
context->con->num_segs));
/* Check whether signal the condition, non root and sent all the segments */
if (num_sent == context->con->num_segs &&
context->con->num_recv_segs == context->con->num_segs * context->con->tree->tree_nextsize) {
ireduce_request_fini(context);
} else {
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output, "return context_list\n"));
opal_free_list_return(mca_coll_adapt_component.adapt_ireduce_context_free_list,
(opal_free_list_item_t *) context);
}
/* Call back function return 1, which means successful */
req->req_free(&req);
return 1;
}
/*
* Callback function of irecv
*/
static int recv_cb(ompi_request_t * req)
{
ompi_coll_adapt_reduce_context_t *context = (ompi_coll_adapt_reduce_context_t *) req->req_complete_cb_data;
int32_t new_id = opal_atomic_add_fetch_32((opal_atomic_int32_t*)&(context->con->next_recv_segs[context->child_id]), 1);
int err;
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: ireduce_recv_cb, peer %d, seg_id %d\n", context->con->rank,
context->peer, context->seg_index));
/* Did we still need to receive subsequent fragments from this child ? */
if (new_id < context->con->num_segs) {
char *temp_recv_buf = NULL;
ompi_coll_adapt_inbuf_t *inbuf = NULL;
/* Set inbuf, if it it first child, recv on rbuf, else recv on inbuf */
if (context->child_id == 0 && context->con->sbuf != MPI_IN_PLACE
&& context->con->root == context->con->rank) {
temp_recv_buf = (char *) context->con->rbuf +
(ptrdiff_t) new_id *(ptrdiff_t) context->con->segment_increment;
} else {
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In recv_cb, alloc inbuf\n", context->con->rank));
inbuf = (ompi_coll_adapt_inbuf_t *) opal_free_list_wait(&context->con->inbuf_list);
temp_recv_buf = inbuf->buff - context->con->lower_bound;
}
/* Get new context item from free list */
ompi_coll_adapt_reduce_context_t *recv_context =
(ompi_coll_adapt_reduce_context_t *) opal_free_list_wait(mca_coll_adapt_component.
adapt_ireduce_context_free_list);
recv_context->buff = temp_recv_buf;
recv_context->seg_index = new_id;
recv_context->child_id = context->child_id;
recv_context->peer = context->peer;
recv_context->con = context->con;
recv_context->inbuf = inbuf;
int recv_count = recv_context->con->seg_count;
if (new_id == (recv_context->con->num_segs - 1)) {
recv_count = recv_context->con->count - new_id * recv_context->con->seg_count;
}
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In recv_cb, create irecv for seg %d, peer %d, inbuf %p, tag %d\n",
context->con->rank, recv_context->seg_index, recv_context->peer,
(void *) inbuf,
recv_context->con->ireduce_tag - recv_context->seg_index));
ompi_request_t *recv_req;
err = MCA_PML_CALL(irecv(temp_recv_buf, recv_count, recv_context->con->datatype,
recv_context->peer,
recv_context->con->ireduce_tag - recv_context->seg_index,
recv_context->con->comm, &recv_req));
if (MPI_SUCCESS != err) {
return err;
}
/* Set the receive callback */
ompi_request_set_callback(recv_req, recv_cb, recv_context);
}
/* Do the op */
int op_count = context->con->seg_count;
if (context->seg_index == (context->con->num_segs - 1)) {
op_count = context->con->count - context->seg_index * context->con->seg_count;
}
int keep_inbuf = 0;
OPAL_THREAD_LOCK(&context->con->mutex_op_list[context->seg_index]);
if (NULL == context->con->accumbuf[context->seg_index]) {
if (NULL == context->inbuf) {
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: set accumbuf to rbuf\n", context->con->rank));
context->con->accumbuf[context->seg_index] = context->buff;
} else {
keep_inbuf = 1;
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: set accumbuf to inbuf\n", context->con->rank));
context->con->accumbuf[context->seg_index] = context->inbuf->buff - context->con->lower_bound;
}
/* Op sbuf and accmbuf to accumbuf */
ompi_op_reduce(context->con->op,
context->con->sbuf + (ptrdiff_t) context->seg_index * (ptrdiff_t) context->con->segment_increment,
context->con->accumbuf[context->seg_index], op_count, context->con->datatype);
} else {
if (NULL == context->inbuf) {
/* Op rbuf and accumbuf to rbuf */
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: op rbuf and accumbuf to rbuf\n", context->con->rank));
ompi_op_reduce(context->con->op, context->con->accumbuf[context->seg_index],
context->buff, op_count, context->con->datatype);
/* Free old accumbuf */
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: free old accumbuf %p\n", context->con->rank,
(void *) to_inbuf(context->con->accumbuf[context->seg_index],
context->con->distance)));
opal_free_list_return(&context->con->inbuf_list,
(opal_free_list_item_t *) to_inbuf(context->con->accumbuf[context->seg_index],
context->con->distance));
/* Set accumbut to rbuf */
context->con->accumbuf[context->seg_index] = context->buff;
} else {
/* Op inbuf and accmbuf to accumbuf */
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: op inbuf and accmbuf to accumbuf\n", context->con->rank));
ompi_op_reduce(context->con->op, context->inbuf->buff - context->con->lower_bound,
context->con->accumbuf[context->seg_index], op_count,
context->con->datatype);
}
}
OPAL_THREAD_UNLOCK(&context->con->mutex_op_list[context->seg_index]);
/* Set recv list */
if (context->con->rank != context->con->tree->tree_root) {
add_to_recv_list(context->con, context->seg_index);
}
/* Send to parent */
if (context->con->rank != context->con->tree->tree_root
&& context->con->ongoing_send < mca_coll_adapt_component.adapt_ireduce_max_send_requests) {
ompi_coll_adapt_item_t *item = get_next_ready_item(context->con, context->con->tree->tree_nextsize);
if (NULL != item) {
/* Get new context item from free list */
ompi_coll_adapt_reduce_context_t *send_context =
(ompi_coll_adapt_reduce_context_t *) opal_free_list_wait(mca_coll_adapt_component.
adapt_ireduce_context_free_list);
send_context->buff = context->con->accumbuf[context->seg_index];
send_context->seg_index = item->id;
send_context->peer = context->con->tree->tree_prev;
send_context->con = context->con;
opal_atomic_add_fetch_32((opal_atomic_int32_t*)&(context->con->ongoing_send), 1);
int send_count = send_context->con->seg_count;
if (item->id == (send_context->con->num_segs - 1)) {
send_count = send_context->con->count - item->id * send_context->con->seg_count;
}
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In recv_cb, create isend to seg %d, peer %d, tag %d\n",
send_context->con->rank, send_context->seg_index, send_context->peer,
send_context->con->ireduce_tag - send_context->seg_index));
ompi_request_t *send_req;
err = MCA_PML_CALL(isend(send_context->buff, send_count, send_context->con->datatype,
send_context->peer,
send_context->con->ireduce_tag - send_context->seg_index,
MCA_PML_BASE_SEND_STANDARD, send_context->con->comm, &send_req));
if (MPI_SUCCESS != err) {
return err;
}
OBJ_RELEASE(item);
/* Set the send callback */
ompi_request_set_callback(send_req, send_cb, send_context);
}
}
int32_t num_recv_segs = opal_atomic_add_fetch_32((opal_atomic_int32_t*)&(context->con->num_recv_segs), 1);
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In recv_cb, tree = %p, root = %d, num_recv = %d, num_segs = %d, num_child = %d\n",
context->con->rank, (void *) context->con->tree,
context->con->tree->tree_root, num_recv_segs, context->con->num_segs,
context->con->tree->tree_nextsize));
/* Prepare for releasing all acquired resources */
if (!keep_inbuf && NULL != context->inbuf) {
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: root free context inbuf %p", context->con->rank,
(void *) context->inbuf));
opal_free_list_return(&context->con->inbuf_list,
(opal_free_list_item_t *) context->inbuf);
}
/* If this is root and has received all the segments */
if (num_recv_segs == context->con->num_segs * context->con->tree->tree_nextsize &&
(context->con->tree->tree_root == context->con->rank || context->con->num_sent_segs == context->con->num_segs)) {
ireduce_request_fini(context);
} else {
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output, "[%d]: return context_list",
context->con->rank));
opal_free_list_return(mca_coll_adapt_component.adapt_ireduce_context_free_list,
(opal_free_list_item_t *) context);
}
req->req_free(&req);
return 1;
}
int ompi_coll_adapt_ireduce(const 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 ** request, mca_coll_base_module_t * module)
{
/* Fall-back if operation is commutative */
if (!ompi_op_is_commute(op)){
mca_coll_adapt_module_t *adapt_module = (mca_coll_adapt_module_t *) module;
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"ADAPT cannot handle reduce with this (non-commutative) operation. It needs to fall back on another component\n"));
return adapt_module->previous_ireduce(sbuf, rbuf, count, dtype, op, root,
comm, request,
adapt_module->previous_reduce_module);
}
OPAL_OUTPUT_VERBOSE((10, mca_coll_adapt_component.adapt_output,
"ireduce root %d, algorithm %d, coll_adapt_ireduce_segment_size %zu, coll_adapt_ireduce_max_send_requests %d, coll_adapt_ireduce_max_recv_requests %d\n",
root, mca_coll_adapt_component.adapt_ireduce_algorithm,
mca_coll_adapt_component.adapt_ireduce_segment_size,
mca_coll_adapt_component.adapt_ireduce_max_send_requests,
mca_coll_adapt_component.adapt_ireduce_max_recv_requests));
if (OMPI_COLL_ADAPT_ALGORITHM_TUNED == mca_coll_adapt_component.adapt_ireduce_algorithm) {
OPAL_OUTPUT_VERBOSE((10, mca_coll_adapt_component.adapt_output, "tuned not implemented\n"));
return OMPI_ERR_NOT_IMPLEMENTED;
}
return ompi_coll_adapt_ireduce_generic(sbuf, rbuf, count, dtype, op, root, comm, request, module,
adapt_module_cached_topology(module, comm, root, mca_coll_adapt_component.adapt_ireduce_algorithm),
mca_coll_adapt_component.adapt_ireduce_segment_size);
}
int ompi_coll_adapt_ireduce_generic(const 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 ** request,
mca_coll_base_module_t * module, ompi_coll_tree_t * tree,
size_t seg_size)
{
ptrdiff_t extent, lower_bound, segment_increment;
ptrdiff_t true_lower_bound, true_extent, real_seg_size;
size_t typelng;
int seg_count = count, num_segs, rank, recv_count, send_count, err, min;
/* Used to store the accumuate result, pointer to every segment */
char **accumbuf = NULL;
opal_mutex_t *mutex_op_list;
/* A list to store the segments need to be sent */
opal_list_t *recv_list;
mca_pml_base_send_mode_t sendmode = (mca_coll_adapt_component.adapt_ireduce_synchronous_send)
? MCA_PML_BASE_SEND_SYNCHRONOUS : MCA_PML_BASE_SEND_STANDARD;
/* Determine number of segments and number of elements sent per operation */
rank = ompi_comm_rank(comm);
ompi_datatype_get_extent(dtype, &lower_bound, &extent);
ompi_datatype_type_size(dtype, &typelng);
COLL_BASE_COMPUTED_SEGCOUNT(seg_size, typelng, seg_count);
num_segs = (count + seg_count - 1) / seg_count;
segment_increment = (ptrdiff_t) seg_count *extent;
ompi_datatype_get_true_extent(dtype, &true_lower_bound, &true_extent);
real_seg_size = true_extent + (ptrdiff_t) (seg_count - 1) * extent;
/* Atomically set up free list */
if (NULL == mca_coll_adapt_component.adapt_ireduce_context_free_list) {
opal_free_list_t* fl = OBJ_NEW(opal_free_list_t);
opal_free_list_init(fl,
sizeof(ompi_coll_adapt_reduce_context_t),
opal_cache_line_size,
OBJ_CLASS(ompi_coll_adapt_reduce_context_t),
0, opal_cache_line_size,
mca_coll_adapt_component.adapt_context_free_list_min,
mca_coll_adapt_component.adapt_context_free_list_max,
mca_coll_adapt_component.adapt_context_free_list_inc,
NULL, 0, NULL, NULL, NULL);
if( !OPAL_ATOMIC_COMPARE_EXCHANGE_STRONG_PTR((opal_atomic_intptr_t *)&mca_coll_adapt_component.adapt_ireduce_context_free_list,
&(intptr_t){0}, fl) ) {
OBJ_RELEASE(fl);
}
}
ompi_coll_base_nbc_request_t *temp_request = NULL;
/* Set up request */
temp_request = OBJ_NEW(ompi_coll_base_nbc_request_t);
OMPI_REQUEST_INIT(&temp_request->super, false);
temp_request->super.req_state = OMPI_REQUEST_ACTIVE;
temp_request->super.req_type = OMPI_REQUEST_COLL;
temp_request->super.req_free = ompi_coll_adapt_request_free;
temp_request->super.req_status.MPI_SOURCE = 0;
temp_request->super.req_status.MPI_TAG = 0;
temp_request->super.req_status.MPI_ERROR = 0;
temp_request->super.req_status._cancelled = 0;
temp_request->super.req_status._ucount = 0;
*request = (ompi_request_t*)temp_request;
/* Set up mutex */
mutex_op_list = (opal_mutex_t *) malloc(sizeof(opal_mutex_t) * num_segs);
for (int32_t i = 0; i < num_segs; i++) {
OBJ_CONSTRUCT(&mutex_op_list[i], opal_mutex_t);
}
/* Set constant context for send and recv call back */
ompi_coll_adapt_constant_reduce_context_t *con =
OBJ_NEW(ompi_coll_adapt_constant_reduce_context_t);
con->count = count;
con->seg_count = seg_count;
con->datatype = dtype;
con->comm = comm;
con->segment_increment = segment_increment;
con->num_segs = num_segs;
con->request = (ompi_request_t*)temp_request;
con->rank = rank;
con->num_recv_segs = 0;
con->num_sent_segs = 0;
con->ongoing_send = 0;
con->mutex_op_list = mutex_op_list;
con->op = op;
con->tree = tree;
con->lower_bound = lower_bound;
con->sbuf = (char *) sbuf;
con->rbuf = (char *) rbuf;
con->root = root;
con->distance = 0;
con->ireduce_tag = ompi_coll_base_nbc_reserve_tags(comm, num_segs);
con->real_seg_size = real_seg_size;
/* If the current process is not leaf */
if (tree->tree_nextsize > 0) {
size_t num_allocate_elems = mca_coll_adapt_component.adapt_inbuf_free_list_min;
if (tree->tree_nextsize * num_segs < num_allocate_elems) {
num_allocate_elems = tree->tree_nextsize * num_segs;
}
opal_free_list_init(&con->inbuf_list,
sizeof(ompi_coll_adapt_inbuf_t) + real_seg_size,
opal_cache_line_size,
OBJ_CLASS(ompi_coll_adapt_inbuf_t),
0, opal_cache_line_size,
num_allocate_elems,
mca_coll_adapt_component.adapt_inbuf_free_list_max,
mca_coll_adapt_component.adapt_inbuf_free_list_inc,
NULL, 0, NULL, NULL, NULL);
/* Set up next_recv_segs */
con->next_recv_segs = (int32_t *) malloc(sizeof(opal_atomic_int32_t) * tree->tree_nextsize);
ompi_coll_adapt_inbuf_t *temp_inbuf =
(ompi_coll_adapt_inbuf_t *) opal_free_list_wait_st(&con->inbuf_list);
con->distance = (char *) temp_inbuf->buff - lower_bound - (char *) temp_inbuf; //address of inbuf->buff to address of inbuf
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: distance %d, inbuf %p, inbuf->buff %p, inbuf->buff-lb %p, to_inbuf %p, inbuf_list %p\n",
rank, con->distance, (void *) temp_inbuf, (void *) temp_inbuf->buff,
(char *) temp_inbuf->buff - lower_bound,
(void *) to_inbuf((char *) temp_inbuf->buff - lower_bound, con->distance),
(void *) &con->inbuf_list));
opal_free_list_return_st(&con->inbuf_list, (opal_free_list_item_t *) temp_inbuf);
} else {
con->next_recv_segs = NULL;
}
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: start ireduce root %d tag %d\n", rank, tree->tree_root,
con->ireduce_tag));
/* If the current process is not leaf node */
if (tree->tree_nextsize > 0) {
/* Set up accumbuf */
accumbuf = (char **) malloc(sizeof(char *) * num_segs);
if (root == rank && sbuf == MPI_IN_PLACE) {
for (int32_t i = 0; i < num_segs; i++) {
accumbuf[i] = (char *) rbuf + (ptrdiff_t) i *(ptrdiff_t) segment_increment;
}
} else {
for (int32_t i = 0; i < num_segs; i++) {
accumbuf[i] = NULL;
}
}
con->accumbuf = accumbuf;
/* For the first batch of segments */
min = mca_coll_adapt_component.adapt_ireduce_max_recv_requests;
if (num_segs < mca_coll_adapt_component.adapt_ireduce_max_recv_requests) {
min = num_segs;
}
for (int32_t i = 0; i < tree->tree_nextsize; i++) {
con->next_recv_segs[i] = min - 1;
}
int num_recvs = 0;
for (int32_t seg_index = 0; seg_index < min; seg_index++)
{
/* For each child */
for (int32_t i = 0; i < tree->tree_nextsize; i++) {
recv_count = seg_count;
if (seg_index == (num_segs - 1)) {
recv_count = count - (ptrdiff_t) seg_count *(ptrdiff_t) seg_index;
}
char *temp_recv_buf = NULL;
ompi_coll_adapt_inbuf_t *inbuf = NULL;
/* Set inbuf, if it it first child, recv on rbuf, else recv on inbuf */
if (i == 0 && sbuf != MPI_IN_PLACE && root == rank) {
temp_recv_buf = (char *) rbuf + (ptrdiff_t) seg_index *(ptrdiff_t) segment_increment;
} else {
inbuf = (ompi_coll_adapt_inbuf_t *) opal_free_list_wait(&con->inbuf_list);
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In ireduce, alloc inbuf %p\n", rank,
(void *) inbuf));
temp_recv_buf = inbuf->buff - lower_bound;
}
/* Get context */
ompi_coll_adapt_reduce_context_t *context =
(ompi_coll_adapt_reduce_context_t *)opal_free_list_wait(mca_coll_adapt_component.
adapt_ireduce_context_free_list);
context->buff = temp_recv_buf;
context->seg_index = seg_index;
context->child_id = i; //the id of peer in in the tree
context->peer = tree->tree_next[i]; //the actual rank of the peer
context->con = con;
context->inbuf = inbuf;
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In ireduce, create irecv for seg %d, peer %d, recv_count %d, inbuf %p tag %d\n",
context->con->rank, context->seg_index, context->peer,
recv_count, (void *) inbuf,
con->ireduce_tag - seg_index));
/* Create a recv request */
ompi_request_t *recv_req;
err = MCA_PML_CALL(irecv
(temp_recv_buf, recv_count, dtype, tree->tree_next[i],
con->ireduce_tag - seg_index, comm, &recv_req));
if (MPI_SUCCESS != err) {
return err;
}
/* Set the recv callback */
ompi_request_set_callback(recv_req, recv_cb, context);
++num_recvs;
}
}
}
/* Leaf nodes */
else {
/* Set up recv_list */
min = mca_coll_adapt_component.adapt_ireduce_max_send_requests;
if (num_segs <= mca_coll_adapt_component.adapt_ireduce_max_send_requests) {
min = num_segs;
}
/* put all items into the recv_list that won't be sent immediately */
for (int32_t seg_index = min; seg_index < num_segs; seg_index++) {
ompi_coll_adapt_item_t *item;
item = OBJ_NEW(ompi_coll_adapt_item_t);
item->id = seg_index;
item->count = tree->tree_nextsize;
opal_list_append(&con->recv_list, (opal_list_item_t *) item);
}
con->accumbuf = accumbuf;
con->ongoing_send = min;
for (int32_t seg_index = 0; seg_index < min; seg_index++) {
send_count = seg_count;
if (seg_index == (num_segs - 1)) {
send_count = count - (ptrdiff_t) seg_count *(ptrdiff_t) seg_index;
}
ompi_coll_adapt_reduce_context_t *context =
(ompi_coll_adapt_reduce_context_t *)opal_free_list_wait(mca_coll_adapt_component.adapt_ireduce_context_free_list);
context->buff = (char *) sbuf + (ptrdiff_t) seg_index * (ptrdiff_t) segment_increment;
context->seg_index = seg_index;
/* Actural rank of the peer */
context->peer = tree->tree_prev;
context->con = con;
context->inbuf = NULL;
OPAL_OUTPUT_VERBOSE((30, mca_coll_adapt_component.adapt_output,
"[%d]: In ireduce, create isend to seg %d, peer %d, send_count %d tag %d\n",
context->con->rank, context->seg_index, context->peer,
send_count, con->ireduce_tag - context->seg_index));
/* Create send request */
ompi_request_t *send_req;
err = MCA_PML_CALL(isend
(context->buff, send_count, dtype, tree->tree_prev,
con->ireduce_tag - context->seg_index,
sendmode, comm, &send_req));
if (MPI_SUCCESS != err) {
return err;
}
/* Set the send callback */
ompi_request_set_callback(send_req, send_cb, context);
}
}
return MPI_SUCCESS;
}
Показать git blame Копировать постоянную ссылку