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.
1265 строки
40 KiB
C
1265 строки
40 KiB
C
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
|
|
/*
|
|
* Copyright (c) 2009-2013 Oak Ridge National Laboratory. All rights reserved.
|
|
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
|
|
* Copyright (c) 2013-2014 Los Alamos National Security, LLC. All rights
|
|
* reserved.
|
|
* Copyright (c) 2014 Cisco Systems, Inc. All rights reserved.
|
|
* $COPYRIGHT$
|
|
*
|
|
* Additional copyrights may follow
|
|
*
|
|
* $HEADER$
|
|
*/
|
|
|
|
|
|
#ifndef MCA_BCOL_basesmuma_EXPORT_H
|
|
#define MCA_BCOL_basesmuma_EXPORT_H
|
|
|
|
#include "ompi_config.h"
|
|
#include "ompi/mca/bcol/bcol.h"
|
|
#include "ompi/mca/mpool/mpool.h"
|
|
#include "ompi/mca/coll/ml/coll_ml.h"
|
|
#include "ompi/mca/coll/ml/coll_ml_allocation.h"
|
|
#include "ompi/request/request.h"
|
|
#include "ompi/proc/proc.h"
|
|
#include "ompi/patterns/net/netpatterns.h"
|
|
|
|
#include "opal/mca/mca.h"
|
|
#include "opal/util/arch.h"
|
|
#include "opal/util/argv.h"
|
|
#include "opal/datatype/opal_datatype.h"
|
|
#include "opal/util/output.h"
|
|
|
|
#include "bcol_basesmuma_smcm.h"
|
|
BEGIN_C_DECLS
|
|
|
|
struct list_data_t {
|
|
opal_list_item_t super;
|
|
void *data;
|
|
};
|
|
typedef struct list_data_t list_data_t;
|
|
OBJ_CLASS_DECLARATION(list_data_t);
|
|
|
|
/*
|
|
* Macro's for manipulating the 64 bit shared memory control bits.
|
|
* The 64 bit field is devided into 4 bit fields
|
|
*
|
|
* | 48-63: src | 32-47: index | 16-31: flag | 0-15: sequence number |
|
|
*
|
|
* Only the low 16 bits of the sequence number will be put in the header
|
|
* space. We will use the fact that the use of the shared buffers is
|
|
* synchronous, and get the upper 48 bits from the local process space.
|
|
*/
|
|
|
|
#define BASESMUMA_CACHE_LINE_SIZE 128
|
|
|
|
#define SHIFT_UP <<
|
|
#define SHIFT_DOWN >>
|
|
|
|
#define SEQ_WIDTH 16
|
|
#define SEQ_BASE 0
|
|
#define FIELD_SEQ_MASK ( ( 1 SHIFT_UP SEQ_WIDTH ) - 1 )
|
|
#define INPLACE_SEQ_MASK ( (int64_t)FIELD_SEQ_MASK SHIFT_UP SEQ_BASE)
|
|
|
|
#define FLAG_WIDTH 16
|
|
#define FLAG_BASE 16
|
|
#define FIELD_FLAG_MASK ( ( 1 SHIFT_UP FLAG_WIDTH ) - 1 )
|
|
#define INPLACE_FLAG_MASK ( (int64_t)FIELD_FLAG_MASK SHIFT_UP FLAG_BASE)
|
|
|
|
#define INDX_WIDTH 16
|
|
#define INDX_BASE 32
|
|
#define FIELD_INDX_MASK ( ( 1 SHIFT_UP INDX_WIDTH ) - 1 )
|
|
#define INPLACE_INDX_MASK ( (int64_t)FIELD_INDX_MASK SHIFT_UP INDX_BASE)
|
|
|
|
#define SRC_WIDTH 16
|
|
#define SRC_BASE 48
|
|
#define FIELD_SRC_MASK ( ( 1 SHIFT_UP SRC_WIDTH ) - 1 )
|
|
#define INPLACE_SRC_MASK ( (int64_t)FIELD_SRC_MASK SHIFT_UP SRC_BASE)
|
|
/*int64_t INPLACE_SRC_MASK= ((int64_t)FIELD_SRC_MASK SHIFT_UP SRC_BASE); */
|
|
|
|
|
|
#define EXTRACT_FLAG(INPUT, OUTPUT, OUTPUT_TYPE, FIELD_BASE, FIELD_MASK) \
|
|
OUTPUT = (OUTPUT_TYPE) ( (INPUT SHIFT_DOWN FIELD_BASE ) & FIELD_MASK )
|
|
|
|
#define STORE_FLAG(INPUT, OUTPUT, INPUT_TYPE, OUTPUT_TYPE, FIELD_BASE, INPLACE_FIELD_MASK ) \
|
|
OUTPUT = \
|
|
( \
|
|
/* 3 */ \
|
|
( \
|
|
/* 2 */ \
|
|
( \
|
|
/* 1 - shift the input field to the proper location */ \
|
|
(OUTPUT_TYPE)( \
|
|
((OUTPUT_TYPE)((INPUT_TYPE) (INPUT))) \
|
|
SHIFT_UP FIELD_BASE ) \
|
|
/* mask off the extra bits */ \
|
|
& ((OUTPUT_TYPE)INPLACE_FIELD_MASK) \
|
|
) \
|
|
/* store back to the OUTPUT field, w/o destroying other fields */ \
|
|
) | OUTPUT \
|
|
)
|
|
|
|
/**
|
|
* Structure to hold the basic shared memory bcoll component.
|
|
*/
|
|
struct mca_bcol_basesmuma_component_t {
|
|
/** Base coll component */
|
|
mca_bcol_base_component_2_0_0_t super;
|
|
|
|
/* management data for collectives with no user data */
|
|
|
|
/** MCA parameter: number of memory banks */
|
|
int basesmuma_num_mem_banks;
|
|
|
|
/** MCA parameter: number of regions per memory bank */
|
|
int basesmuma_num_regions_per_bank;
|
|
|
|
/** MCA parameter: Number of simultaneous groups supported */
|
|
int n_groups_supported;
|
|
|
|
/* management data for collectives with user data (ud) - the memory
|
|
* is actually obtained at the ML level
|
|
*/
|
|
|
|
/** MCA paramenter: number of polling loops to run while waiting
|
|
* for children or parent to complete their work
|
|
*/
|
|
int n_poll_loops;
|
|
|
|
/* mpool size */
|
|
size_t mpool_size;
|
|
|
|
|
|
/* mpool inited - will use this to test whether or not the
|
|
* shared memory has been inited
|
|
*/
|
|
bool mpool_inited;
|
|
|
|
/* shared memory control buffer - the control structures reside
|
|
* in shared memory */
|
|
bcol_basesmuma_smcm_mmap_t *sm_ctl_structs;
|
|
|
|
/* shared memory payload buffer
|
|
*/
|
|
bcol_basesmuma_smcm_mmap_t *sm_payload_structs;
|
|
|
|
/*
|
|
* list of shared memory control structures
|
|
*/
|
|
opal_list_t ctl_structures;
|
|
|
|
|
|
/** opal list in which the list of peers that I am "connected" to is stored
|
|
*/
|
|
opal_list_t sm_connections_list;
|
|
|
|
/* opal list in which the list of payload peers that I am "connected" to
|
|
* is stored
|
|
*/
|
|
opal_list_t sm_payload_connections_list;
|
|
|
|
/*
|
|
* list of non-blocking admin barriers to progress */
|
|
opal_mutex_t nb_admin_barriers_mutex;
|
|
opal_list_t nb_admin_barriers;
|
|
|
|
/*
|
|
* order of fan-in tree
|
|
*/
|
|
int radix_fanin;
|
|
|
|
/*
|
|
* order of fan-out tree
|
|
*/
|
|
int radix_fanout;
|
|
|
|
/*
|
|
* Order of read tree
|
|
*/
|
|
int radix_read_tree;
|
|
|
|
/*
|
|
* order of reduction fan-out tree
|
|
*/
|
|
int order_reduction_tree;
|
|
|
|
/*
|
|
* K-nomial tree radix
|
|
*/
|
|
int k_nomial_radix;
|
|
|
|
/*
|
|
* K-ary scatter tree radix
|
|
*/
|
|
int scatter_kary_radix;
|
|
|
|
/*
|
|
* number of polling loops
|
|
*/
|
|
int num_to_probe;
|
|
|
|
/*
|
|
* Portals addressing info
|
|
* void*: because wanted to keep portal library dependencies
|
|
* as local as possible
|
|
*/
|
|
void *portals_info;
|
|
bool portals_init;
|
|
|
|
/*
|
|
* verbosity level
|
|
*/
|
|
int verbose;
|
|
|
|
/*
|
|
* control file name base string
|
|
*/
|
|
char *clt_base_fname;
|
|
|
|
/*
|
|
* data file name base string
|
|
*/
|
|
char *payload_base_fname;
|
|
|
|
/*
|
|
* shared memory scratch space. This is mapped at the end of the
|
|
* segement of memory holding the control structures.
|
|
*/
|
|
char *my_scratch_shared_memory;
|
|
|
|
/*
|
|
* size of scratch memory
|
|
*/
|
|
size_t my_scratch_shared_memory_size;
|
|
|
|
/* the offset will be the same for all ranks */
|
|
size_t scratch_offset_from_base_ctl_file;
|
|
};
|
|
|
|
static inline int mca_bcol_basesmuma_err(const char* fmt, ...)
|
|
{
|
|
va_list list;
|
|
int ret;
|
|
|
|
va_start(list, fmt);
|
|
ret = vfprintf(stderr, fmt, list);
|
|
va_end(list);
|
|
return ret;
|
|
}
|
|
|
|
#if OPAL_ENABLE_DEBUG
|
|
#define BASESMUMA_VERBOSE(level, args) \
|
|
do { \
|
|
if(mca_bcol_basesmuma_component.verbose >= level) { \
|
|
mca_bcol_basesmuma_err("[%s]%s[%s:%d:%s] BCOL-BASESMUMA ", \
|
|
ompi_process_info.nodename, \
|
|
OMPI_NAME_PRINT(OMPI_PROC_MY_NAME), \
|
|
__FILE__, __LINE__, __func__); \
|
|
mca_bcol_basesmuma_err args; \
|
|
mca_bcol_basesmuma_err("\n"); \
|
|
} \
|
|
} while(0)
|
|
#else
|
|
#define BASESMUMA_VERBOSE(level, args)
|
|
#endif
|
|
|
|
|
|
/**
|
|
* Convenience typedef */
|
|
typedef struct mca_bcol_basesmuma_component_t mca_bcol_basesmuma_component_t;
|
|
|
|
#if 0
|
|
/*
|
|
* Implemented function index list
|
|
*/
|
|
|
|
/* barrier */
|
|
enum{
|
|
FANIN_FAN_OUT_BARRIER_FN,
|
|
RECURSIVE_DOUBLING_BARRIER_FN,
|
|
N_BARRIER_FNS
|
|
};
|
|
|
|
/* reduce */
|
|
enum{
|
|
FANIN_REDUCE_FN,
|
|
REDUCE_SCATTER_GATHER_FN,
|
|
N_REDUCE_FNS
|
|
};
|
|
enum{
|
|
SHORT_DATA_FN_REDUCE,
|
|
LONG_DATA_FN_REDUCE,
|
|
N_REDUCE_FNS_USED
|
|
};
|
|
|
|
/* all-reduce */
|
|
enum{
|
|
FANIN_FANOUT_ALLREDUCE_FN,
|
|
REDUCE_SCATTER_ALLGATHER_FN,
|
|
N_ALLREDUCE_FNS
|
|
};
|
|
enum{
|
|
SHORT_DATA_FN_ALLREDUCE,
|
|
LONG_DATA_FN_ALLREDUCE,
|
|
N_ALLREDUCE_FNS_USED
|
|
};
|
|
|
|
|
|
/* enum for node type */
|
|
enum{
|
|
ROOT_NODE,
|
|
LEAF_NODE,
|
|
INTERIOR_NODE
|
|
};
|
|
|
|
|
|
/*
|
|
* N-order tree node description
|
|
*/
|
|
struct tree_node_t {
|
|
/* my rank within the group */
|
|
int my_rank;
|
|
/* my node type - root, leaf, or interior */
|
|
int my_node_type;
|
|
/* number of nodes in the tree */
|
|
int tree_size;
|
|
/* number of parents (0/1) */
|
|
int n_parents;
|
|
/* number of children */
|
|
int n_children;
|
|
/* parent rank within the group */
|
|
int parent_rank;
|
|
/* chidren ranks within the group */
|
|
int *children_ranks;
|
|
};
|
|
typedef struct tree_node_t tree_node_t;
|
|
|
|
/*
|
|
* Pair-wise data exchange
|
|
*/
|
|
/* enum for node type */
|
|
enum{
|
|
EXCHANGE_NODE,
|
|
EXTRA_NODE
|
|
};
|
|
|
|
struct pair_exchange_node_t {
|
|
|
|
/* my rank within the group */
|
|
int my_rank;
|
|
|
|
/* number of nodes this node will exchange data with */
|
|
int n_exchanges;
|
|
|
|
/* ranks of nodes involved in data exchnge */
|
|
int *rank_exchanges;
|
|
|
|
/* number of extra sources of data - outside largest power of 2 in
|
|
* this group */
|
|
int n_extra_sources;
|
|
|
|
/* rank of the extra source */
|
|
int rank_extra_source;
|
|
|
|
/* number of tags needed per stripe */
|
|
int n_tags;
|
|
|
|
/* log 2 of largest full power of 2 for this node set */
|
|
int log_2;
|
|
|
|
/* largest power of 2 that fits in this group */
|
|
int n_largest_pow_2;
|
|
|
|
/* node type */
|
|
int node_type;
|
|
|
|
};
|
|
typedef struct pair_exchange_node_t pair_exchange_node_t;
|
|
#endif
|
|
/*
|
|
* descriptor for managing the admin nonblocking barrier routine.
|
|
* This is an sm internal routine, and assumes only 1 outstanding
|
|
* nb-barrier collective call per block.
|
|
*/
|
|
/* forward declarations */
|
|
struct mca_bcol_basesmuma_module_t;
|
|
struct sm_buffer_mgmt;
|
|
|
|
struct sm_nbbar_desc_t {
|
|
/* make sure we can put this on a list */
|
|
opal_list_item_t super;
|
|
|
|
/* phase of the collective operation - needed to know how to continue
|
|
* progressing the nb-barrier */
|
|
int collective_phase;
|
|
|
|
/* iteration to continue at */
|
|
int recursive_dbl_iteration;
|
|
|
|
/* pointer to the collective module this is associated with */
|
|
struct mca_bcol_basesmuma_module_t *sm_module;
|
|
|
|
/* pointer to payload/control structs buffers */
|
|
struct sm_buffer_mgmt *coll_buff;
|
|
|
|
/* pool index */
|
|
int pool_index;
|
|
|
|
/* pointer to the ml_memory_block_desc_t structure
|
|
* that is actually managing this registration.
|
|
* This is meaningful when these control structures
|
|
* are used in conjunction with the user payload
|
|
* data that is allocated at the ml level.
|
|
*/
|
|
void *ml_memory_block_descriptor;
|
|
|
|
};
|
|
typedef struct sm_nbbar_desc_t sm_nbbar_desc_t;
|
|
|
|
/*
|
|
* Barrier request objects
|
|
*/
|
|
|
|
/* shared memory data strucutures */
|
|
struct mca_bcol_basesmuma_nb_request_process_shared_mem_t {
|
|
volatile uint64_t coll_index;
|
|
/* flag used to indicate the status of this memory region */
|
|
volatile uint64_t flag;
|
|
volatile uint64_t index;
|
|
|
|
/* pading */
|
|
/* Note: need to change this so it takes less memory */
|
|
char padding[BASESMUMA_CACHE_LINE_SIZE-3*sizeof(uint64_t)];
|
|
};
|
|
|
|
typedef struct mca_bcol_basesmuma_nb_request_process_shared_mem_t
|
|
mca_bcol_basesmuma_nb_request_process_shared_mem_t;
|
|
|
|
/* enum for phase at which the nb barrier is in */
|
|
enum{
|
|
NB_BARRIER_INACTIVE,
|
|
|
|
/* fan-in/fan-out */
|
|
NB_BARRIER_FAN_IN,
|
|
NB_BARRIER_FAN_OUT,
|
|
|
|
/* recursive doubling */
|
|
NB_PRE_PHASE,
|
|
NB_RECURSIVE_DOUBLING,
|
|
NB_POST_PHASE,
|
|
|
|
/* done and not started are the same for all practicle
|
|
* purposes, as the init funtion always sets this flag
|
|
*/
|
|
NB_BARRIER_DONE
|
|
};
|
|
|
|
|
|
|
|
/* forward declartion */
|
|
struct mca_bcol_basesmuma_module_t;
|
|
|
|
struct mca_basesmuma_ctrl_4_hdl_t {
|
|
int fd;
|
|
bool status;
|
|
volatile char buf[128];
|
|
/*volatile char buf[OPAL_PATH_MAX];*/
|
|
};
|
|
typedef struct mca_basesmuma_ctrl_4_hdl_t mca_basesmuma_ctrl_4_hdl_t;
|
|
|
|
/* control segment for shared memory */
|
|
struct mca_bcol_basesmuma_ctl_struct_t {
|
|
/* collective identifier */
|
|
volatile int64_t sequence_number;
|
|
volatile int64_t flag;
|
|
volatile int64_t index;
|
|
volatile int64_t offset;
|
|
volatile int64_t offset_zip;
|
|
|
|
|
|
/* used for non-blocking algorithms */
|
|
int status;
|
|
int active_requests;
|
|
int iteration;
|
|
|
|
int *src_ptr;
|
|
|
|
int start;
|
|
|
|
/* process private data */
|
|
int starting_flag_value;
|
|
|
|
/* experiment for large data colls */
|
|
int n_sends;
|
|
int length;
|
|
|
|
/* hdl framework control structure*/
|
|
/* no need to pad at this level anymore */
|
|
volatile int64_t data_hdl;
|
|
volatile mca_basesmuma_ctrl_4_hdl_t hdl_ctrl;
|
|
|
|
#ifdef __PORTALS_AVAIL__
|
|
struct mca_bcol_basesmuma_portal_buf_addr_t portals_buf_addr;
|
|
#endif
|
|
/* padding */
|
|
/*char padding[BASESMUMA_CACHE_LINE_SIZE-4*sizeof(uint64_t)-3*sizeof(int)];*/
|
|
char padding[BASESMUMA_CACHE_LINE_SIZE-6*sizeof(int64_t)-5*sizeof(int)];
|
|
};
|
|
typedef struct mca_bcol_basesmuma_ctl_struct_t mca_bcol_basesmuma_ctl_struct_t;
|
|
|
|
|
|
#define SM_BCOLS_MAX 2
|
|
|
|
/* enum for signaling flag bank, when
|
|
* adding to this list, please keep
|
|
* it alphabetical
|
|
*/
|
|
enum {
|
|
ALLGATHER_FLAG,
|
|
ALLREDUCE_FLAG,
|
|
BARRIER_FANIN_FLAG,
|
|
BARRIER_FANOUT_FLAG,
|
|
BARRIER_RKING_FLAG,
|
|
BCAST_FLAG,
|
|
GATHER_FLAG,
|
|
REDUCE_FLAG,
|
|
NUM_SIGNAL_FLAGS
|
|
};
|
|
|
|
|
|
/* control region for colls with user data - shared memory */
|
|
struct mca_bcol_basesmuma_header_t {
|
|
/* collective identifier */
|
|
volatile int64_t sequence_number;
|
|
volatile int8_t flags[NUM_SIGNAL_FLAGS][SM_BCOLS_MAX];
|
|
volatile int32_t src; /* src of bcast data for unknown root,
|
|
bcol id for known root
|
|
*/
|
|
/* starting flag - hierarchies */
|
|
int8_t starting_flag_value[SM_BCOLS_MAX];
|
|
int8_t ready_flag;
|
|
|
|
/* Manju: Cached array of receive buffer offsets
|
|
*
|
|
* This array stores the receive buffer offsets (rbuf_offsets) of data buffer.
|
|
* In general, we assume that sbuf_offset and rbuf_offset of
|
|
* processes invoking the collective primitive is same. This is
|
|
* true when the order in which processes invoke their hierarchies are
|
|
* same.
|
|
*
|
|
* For some algorithms (like broadcast, reduce) we split the ML buffer
|
|
* and use first half as
|
|
* source and second half as receive buffer. We swap these buffers for
|
|
* each change when we change levels i.e., if first half is source for
|
|
* level 1, in the level 2 of hierarchy it becomes the receive buffer.
|
|
* For reduce algorithm, each process can invoke hierarchies
|
|
* (primitives) in different order. For example, process 1 might have level 1 as SM
|
|
* and level 2 as p2p, and process 2 might have different order where its
|
|
* level 1 is p2p and level 2 SM. In this case, if in basesmuma reduce
|
|
* algorithm, if parent assumes its rbuf_offset as child's rbuf_offset
|
|
* it is wrong. So we cache rbuf_offset of each process so
|
|
* it could be accessed by processes to obtain the data.
|
|
*/
|
|
|
|
volatile int32_t roffsets[SM_BCOLS_MAX];
|
|
|
|
/* Manju Start: Experimental ctl fields and should be removed later;
|
|
* This is used for lmsg reduce for testing
|
|
* during transition to HDL primitives
|
|
*/
|
|
#if 0
|
|
int lmsg_reduce_snd_completes;
|
|
/* There can be atmost 20 ranks in the subgroup. Since this
|
|
* only for testing this should be good enough */
|
|
int lmsg_reduce_peers[20];
|
|
int lmsg_reduce_send_offsets[20];
|
|
/* End: Experimental ctl fields */
|
|
|
|
|
|
/* no need to pad at this level anymore */
|
|
volatile int64_t data_hdl;
|
|
#endif
|
|
};
|
|
typedef struct mca_bcol_basesmuma_header_t mca_bcol_basesmuma_header_t;
|
|
|
|
/* data needed for large messages */
|
|
struct mca_bcol_basesmuma_large_msg_t {
|
|
/* scatter allgather data */
|
|
uint64_t offset;
|
|
uint64_t n_sends;
|
|
uint64_t length;
|
|
|
|
/* portals data */
|
|
|
|
};
|
|
typedef struct mca_bcol_basesmuma_large_msg_t mca_bcol_basesmuma_large_msg_t;
|
|
|
|
/* payload struct */
|
|
struct mca_bcol_basesmuma_payload_t {
|
|
|
|
/* base pointer to shared memory control structure */
|
|
mca_bcol_basesmuma_header_t *ctl_struct;
|
|
void *payload;
|
|
|
|
};
|
|
|
|
typedef struct mca_bcol_basesmuma_payload_t mca_bcol_basesmuma_payload_t;
|
|
|
|
|
|
|
|
|
|
/* memory bank memory management structure */
|
|
struct mem_bank_management_t {
|
|
|
|
/* generation counter */
|
|
uint64_t bank_gen_counter;
|
|
|
|
/* descriptor for the non-blocking barrier. This is
|
|
* used to manage this bank of memory.
|
|
*/
|
|
sm_nbbar_desc_t nb_barrier_desc;
|
|
|
|
/* the number of buffers that are not in use, and are
|
|
* available. The assumption is that the buffers are
|
|
* recycled all at once, so are available for re-use
|
|
* until all buffers have been made available for re-use.
|
|
*/
|
|
volatile int available_buffers;
|
|
|
|
/*
|
|
* number of buffers freed */
|
|
volatile int n_buffs_freed;
|
|
|
|
/* mutex to ensure atomic recycling of resrouces */
|
|
opal_mutex_t mutex;
|
|
|
|
/* number of buffers being managed */
|
|
int number_of_buffers;
|
|
|
|
/* shared memory control structures */
|
|
int index_shared_mem_ctl_structs;
|
|
|
|
|
|
};
|
|
typedef struct mem_bank_management_t mem_bank_management_t;
|
|
|
|
/* data structure for shared buffers */
|
|
struct sm_buffer_mgmt {
|
|
/* number of buffers per process */
|
|
int number_of_buffs;
|
|
|
|
/* size of group */
|
|
int size_of_group;
|
|
|
|
/* number of memory banks */
|
|
int num_mem_banks;
|
|
|
|
/* number of buffers per memory bank */
|
|
int num_buffs_per_mem_bank;
|
|
|
|
/* log base 2 of num_buffs_per_mem_bank */
|
|
int log2_num_buffs_per_mem_bank;
|
|
|
|
/* log base 2 total number of buffers */
|
|
int log2_number_of_buffs;
|
|
|
|
/* mask - masks off the bits corresponding to buffer index */
|
|
int mask;
|
|
|
|
/* control buffers - these point to regions in shared memory */
|
|
/* leading dimension is the group size - all pointers for a given
|
|
* set of buffers appear consecutively in this array
|
|
*/
|
|
volatile void **ctl_buffs;
|
|
|
|
/* management data for the control structures -
|
|
* one per bank of control structures - Will be used for
|
|
* the payload buffers as well.
|
|
*/
|
|
mem_bank_management_t *ctl_buffs_mgmt;
|
|
|
|
/* data buffers - these point to regions in shared memory */
|
|
/* leading dimension is the group size - all pointers for a given
|
|
* set of buffers appear consecutively in this array
|
|
*/
|
|
|
|
volatile mca_bcol_basesmuma_payload_t *data_buffs;
|
|
|
|
|
|
|
|
};
|
|
typedef struct sm_buffer_mgmt sm_buffer_mgmt;
|
|
|
|
|
|
struct mca_bcol_basesmuma_nb_coll_buff_desc_t {
|
|
void *data_addr;
|
|
uint64_t bank_index;
|
|
uint64_t buffer_index;
|
|
int active_requests;
|
|
ompi_request_t **requests;
|
|
int data_src;
|
|
int radix_mask;
|
|
int radix_mask_pow;
|
|
int iteration;
|
|
int status;
|
|
/* this is for testing */
|
|
int tag;
|
|
|
|
volatile mca_bcol_basesmuma_ctl_struct_t **ctl_structs;
|
|
volatile mca_bcol_basesmuma_ctl_struct_t *my_ctl_pointer;
|
|
volatile mca_bcol_basesmuma_ctl_struct_t *parent_ctl_pointer;
|
|
volatile mca_bcol_basesmuma_ctl_struct_t *extra_partner_ctl_pointer;
|
|
};
|
|
|
|
typedef struct mca_bcol_basesmuma_nb_coll_buff_desc_t mca_bcol_basesmuma_nb_coll_buff_desc_t;
|
|
|
|
struct mca_bcol_basesmuma_local_mlmem_desc_t {
|
|
|
|
uint32_t bank_index_for_release;
|
|
struct ml_memory_block_desc_t *ml_mem_desc;
|
|
uint32_t num_banks;
|
|
uint32_t num_buffers_per_bank;
|
|
uint32_t size_buffer;
|
|
uint32_t *bank_release_counter;
|
|
|
|
/*
|
|
* Number of descriptors allocated is equivalent to number of ml buffers
|
|
* (number of banks * number of buffers per bank)
|
|
*/
|
|
mca_bcol_basesmuma_nb_coll_buff_desc_t *nb_coll_desc;
|
|
};
|
|
|
|
typedef struct mca_bcol_basesmuma_local_mlmem_desc_t mca_bcol_basesmuma_local_mlmem_desc_t;
|
|
|
|
#ifdef __PORTALS_AVAIL__
|
|
#define MAX_SM_GROUP_SIZE 32
|
|
|
|
|
|
struct portals_scatter_allgather_nb_bcast_state_t
|
|
{
|
|
/* local variables */
|
|
uint64_t length;
|
|
int my_rank, src, matched;
|
|
int src_list[MAX_SM_GROUP_SIZE];
|
|
int group_size;
|
|
int64_t ready_flag;
|
|
int pow_2, pow_2_levels;
|
|
int src_list_index;
|
|
uint64_t fragment_size; /* user buffer size */
|
|
|
|
/* Input argument variables */
|
|
void *my_userbuf;
|
|
int64_t sequence_number;
|
|
|
|
/* Extra source variables */
|
|
bool secondary_root;
|
|
int partner , extra_partner;
|
|
|
|
/* Scatter Allgather offsets */
|
|
uint64_t local_sg_offset , global_sg_offset , partner_offset ;
|
|
|
|
/* Portals messaging relevant variables */
|
|
/*
|
|
* ptl_handle_eq_t allgather_eq_h;
|
|
*/
|
|
ptl_handle_eq_t read_eq;
|
|
ptl_event_t allgather_event;
|
|
bool msg_posted;
|
|
|
|
/* OMPI module and component variables */
|
|
mca_bcol_basesmuma_component_t *cs;
|
|
struct mca_bcol_basesmuma_module_t *bcol_module;
|
|
|
|
/* Control structure and payload variables */
|
|
volatile mca_bcol_basesmuma_ctl_struct_t **ctl_structs;
|
|
volatile mca_bcol_basesmuma_ctl_struct_t *my_ctl_pointer;
|
|
volatile mca_bcol_basesmuma_ctl_struct_t *parent_ctl_pointer; /* scatter source */
|
|
volatile mca_bcol_basesmuma_ctl_struct_t *extra_partner_ctl_pointer; /* scatter source */
|
|
|
|
int phase;
|
|
};
|
|
|
|
|
|
typedef struct portals_scatter_allgather_nb_bcast_state_t sg_state_t;
|
|
#endif
|
|
|
|
#define SM_ARRAY_INDEX(LEAD_DIM,BUF_INDEX,PROC_INDEX) \
|
|
((LEAD_DIM)*(BUF_INDEX)+(PROC_INDEX))
|
|
/* debug */
|
|
#define BARRIER_BANK_LIST_SIZE 32
|
|
/* end debug */
|
|
|
|
struct mca_bcol_basesmuma_module_t {
|
|
/* base structure */
|
|
mca_bcol_base_module_t super;
|
|
|
|
/* free list item with the control structures used for
|
|
* the no user data collective operations
|
|
*/
|
|
list_data_t *no_userdata_ctl;
|
|
|
|
/* free list item with the control structures used for
|
|
* the with user data collective operations
|
|
*/
|
|
list_data_t *userdata_ctl;
|
|
|
|
/*
|
|
* information on sm control backing files for the subgroup
|
|
* associated with this module.
|
|
*/
|
|
bcol_basesmuma_smcm_proc_item_t **ctl_backing_files_info;
|
|
|
|
/*
|
|
* information on sm payload backing files for the subgroup
|
|
* associated with this module.
|
|
*/
|
|
bcol_basesmuma_smcm_proc_item_t **payload_backing_files_info;
|
|
|
|
/*
|
|
* buffers for the collective that do not involve user data -
|
|
* barrier, fanin, fanout.
|
|
*/
|
|
sm_buffer_mgmt colls_no_user_data;
|
|
|
|
/*
|
|
* buffers for the collective with user data.
|
|
*/
|
|
sm_buffer_mgmt colls_with_user_data;
|
|
|
|
/* recursive-doubling tree node */
|
|
netpatterns_pair_exchange_node_t recursive_doubling_tree;
|
|
|
|
/* k-nomial gather/allgather tree */
|
|
netpatterns_k_exchange_node_t knomial_allgather_tree;
|
|
|
|
/* fanin tree node - root is rank 0 */
|
|
netpatterns_tree_node_t fanin_node;
|
|
|
|
/* fanout tree node - root is rank 0 */
|
|
netpatterns_tree_node_t fanout_node;
|
|
|
|
/* index of blocking barrier memory region to use */
|
|
int index_blocking_barrier_memory_bank;
|
|
|
|
/* comm to shared memory map */
|
|
int *comm_to_sm_map;
|
|
|
|
/* reduction fanout tree */
|
|
netpatterns_tree_node_t* reduction_tree;
|
|
|
|
/* broadcast fanout tree */
|
|
netpatterns_tree_node_t* fanout_read_tree;
|
|
|
|
/* scatter - k-ary tree */
|
|
int scatter_kary_radix;
|
|
netpatterns_tree_node_t *scatter_kary_tree;
|
|
|
|
/* Knomial exchange tree */
|
|
/* Currently used for only large message reduce */
|
|
netpatterns_k_exchange_node_t knomial_exchange_tree;
|
|
|
|
/* sequence number offset - want to make sure that we start
|
|
* id'ing collectives with id 0, so we can have simple
|
|
* resource management.
|
|
*/
|
|
int64_t squence_number_offset;
|
|
|
|
/* basesmuma specific header size into ml buffer
|
|
* was calculated at ml level - it is the sum of
|
|
* all headers from all bcols and then aligned to
|
|
* whatever alignment was requested
|
|
*/
|
|
uint32_t total_header_size;
|
|
|
|
/* list of possible sources */
|
|
int *src_list;
|
|
|
|
/* Number of possible sources */
|
|
int src_size;
|
|
|
|
/* smallest power of k that is smaller
|
|
* than or equal in size to the uma group
|
|
*/
|
|
int pow_k_levels;
|
|
|
|
/* size of power-of-k group */
|
|
int pow_k;
|
|
|
|
/* smallest power of 2 that is smaller
|
|
* than or equal to the smuma group size
|
|
*/
|
|
int pow_2_levels;
|
|
|
|
/* size of power-of-2 group */
|
|
int pow_2;
|
|
|
|
/* pointer to the shared memory scratch array of each
|
|
* process in the group.
|
|
*/
|
|
void **shared_memory_scratch_space;
|
|
|
|
/*
|
|
* Caching information for re-entrant collectives
|
|
*/
|
|
mca_bcol_basesmuma_local_mlmem_desc_t ml_mem;
|
|
|
|
/*
|
|
* Cached offsets for lmsg reduce
|
|
*/
|
|
int **reduce_offsets;
|
|
|
|
/*XXX:
|
|
* Starting to explore the beauty of zero-copy for large message
|
|
*/
|
|
struct mca_hdl_base_module_t **hdl_module;
|
|
|
|
#ifdef __PORTALS_AVAIL__
|
|
/*
|
|
* Store state for NB blocking functions
|
|
*/
|
|
sg_state_t sg_state;
|
|
|
|
#endif
|
|
};
|
|
|
|
typedef struct mca_bcol_basesmuma_module_t mca_bcol_basesmuma_module_t;
|
|
OBJ_CLASS_DECLARATION(mca_bcol_basesmuma_module_t);
|
|
|
|
/* shared memory specific arguments for the bcol registration function */
|
|
typedef struct bcol_basesmuma_registration_data_t {
|
|
char *file_name; /* filename for payload */
|
|
void *base_addr; /* base address to be mapped */
|
|
size_t size; /* size of memory block to be "registered" */
|
|
size_t size_ctl_structure;
|
|
size_t data_seg_alignment;
|
|
bcol_basesmuma_smcm_mmap_t *sm_mmap; /* shared memory map struct */
|
|
mca_coll_ml_release_buff_fn_t buff_release_cb; /* buffer release
|
|
call back */
|
|
} bcol_basesmuma_registration_data_t;
|
|
|
|
|
|
enum {
|
|
BUFFER_AVAILABLE,
|
|
STARTED,
|
|
FANIN,
|
|
FANOUT
|
|
};
|
|
|
|
/* enum used for non-blocking large
|
|
* message bcast
|
|
*/
|
|
|
|
enum {
|
|
INIT,
|
|
START,
|
|
NOT_STARTED,
|
|
SCATTER,
|
|
ALLGATHER,
|
|
EXTRA_RANK,
|
|
PROBE,
|
|
SCATTER_ROOT_WAIT,
|
|
SCATTER_EXTRA_ROOT_WAIT,
|
|
SCATTER_PARENT_WAIT,
|
|
FINISHED
|
|
};
|
|
|
|
/**
|
|
* Global component instance
|
|
*/
|
|
OMPI_MODULE_DECLSPEC extern mca_bcol_basesmuma_component_t mca_bcol_basesmuma_component;
|
|
|
|
|
|
/*
|
|
* coll module functions
|
|
*/
|
|
|
|
/* query to see if the component is available for use, and can
|
|
* satisfy the thread and progress requirements
|
|
*/
|
|
int mca_bcol_basesmuma_init_query(bool enable_progress_threads,
|
|
bool enable_mpi_threads);
|
|
|
|
/* query to see if the module is available for use on the given
|
|
* communicator, and if so, what it's priority is.
|
|
*/
|
|
mca_bcol_base_module_t **
|
|
mca_bcol_basesmuma_comm_query(mca_sbgp_base_module_t *module, int *num_modules);
|
|
|
|
|
|
|
|
/* shared memory specific memory registration function - this will be passed into the mpool */
|
|
int mca_bcol_basesmuma_register_sm(void *context_data, void *base, size_t size,
|
|
void **reg);
|
|
|
|
/* shared memory specific memory deregistration function - also needed by the mpool */
|
|
int mca_bcol_basesmuma_deregister_sm(void *context_data, void *reg);
|
|
|
|
/* setup the new k_nomial tree for collectives */
|
|
int bcol_basesmuma_setup_knomial_tree(mca_bcol_base_module_t *super);
|
|
|
|
/* allocate the memory pool for the shared memory control structures */
|
|
int mca_bcol_basesmuma_allocate_pool_memory(mca_bcol_basesmuma_component_t
|
|
*component);
|
|
|
|
/* initialize the internal scratch buffers and control structs that will be
|
|
used by the module */
|
|
int base_bcol_basesmuma_setup_library_buffers(
|
|
mca_bcol_basesmuma_module_t *sm_module,
|
|
mca_bcol_basesmuma_component_t *cs);
|
|
|
|
|
|
/* shared memory recursive doubling initialization */
|
|
int bcol_basesmuma_rd_barrier_init(mca_bcol_base_module_t *module);
|
|
|
|
/* shared memory recusive double barrier */
|
|
int bcol_basesmuma_recursive_double_barrier(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
/* shared memory fanin */
|
|
int bcol_basesmuma_fanin_init(mca_bcol_base_module_t *super);
|
|
|
|
/* shared memory fanout */
|
|
int bcol_basesmuma_fanout_init(mca_bcol_base_module_t *super);
|
|
|
|
/* shared memory recursive k-ing non-blocking barrier */
|
|
int bcol_basesmuma_barrier_init(mca_bcol_base_module_t *super);
|
|
|
|
/* Shared memory broadcast */
|
|
int bcol_basesmuma_bcast_init(mca_bcol_base_module_t *super);
|
|
|
|
int bcol_basesmuma_bcast(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* Shared memory non-blocking broadcast */
|
|
int bcol_basesmuma_bcast_k_nomial_anyroot(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
int bcol_basesmuma_bcast_k_nomial_knownroot(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* Shared memory non-blocking broadcast - Large message anyroot */
|
|
int bcol_basesmuma_binary_scatter_allgather_segment(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
#if 0
|
|
/*FIXME: having fun here*/
|
|
int bcol_basesmuma_hdl_zerocopy_bcast(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
#endif
|
|
|
|
int bcol_basesmuma_lmsg_bcast_k_nomial_anyroot(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
int bcol_basesmuma_lmsg_scatter_allgather_portals_bcast(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
int bcol_basesmuma_lmsg_scatter_allgather_portals_nb_bcast(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
int bcol_basesmuma_lmsg_scatter_allgather_portals_nb_knownroot_bcast(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/*
|
|
* shared memory scatter
|
|
*/
|
|
int bcol_basesmuma_scatter_init(mca_bcol_base_module_t *super);
|
|
|
|
/* shared memory nonblocking scatter - known root */
|
|
int bcol_basesmuma_nb_scatter_k_array_knownroot(
|
|
bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* shared memory non-blocking k-nomial barrier init */
|
|
int bcol_basesmuma_k_nomial_barrier_init(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
/* shared memory non-blocking k-nomial barrier progress */
|
|
int bcol_basesmuma_k_nomial_barrier_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
/*shared memory non-blocking k-nomial allgather init */
|
|
int bcol_basesmuma_k_nomial_allgather_init(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
/* shared memory non-blocking k-nomial allgather progress */
|
|
int bcol_basesmuma_k_nomial_allgather_progress(bcol_function_args_t *input_args,
|
|
struct coll_ml_function_t *const_args);
|
|
|
|
/* shared memory allgather -- selection logic api */
|
|
int bcol_basesmuma_allgather_init(mca_bcol_base_module_t *super);
|
|
|
|
/* shared memory blocking k-nomial gather */
|
|
int bcol_basesmuma_k_nomial_gather(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* shared memory non blocking k-nomial gather */
|
|
int bcol_basesmuma_k_nomial_gather_init(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* shared memory non blocking k-nomial gather progress*/
|
|
int bcol_basesmuma_k_nomial_gather_progress(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* shared memory init */
|
|
int bcol_basesmuma_gather_init(mca_bcol_base_module_t *super);
|
|
|
|
/* allocate shared memory control memory */
|
|
int mca_bcol_basesmuma_allocate_sm_ctl_memory(
|
|
mca_bcol_basesmuma_component_t *cs);
|
|
|
|
/* Shared memory basesmuma reduce */
|
|
int bcol_basesmuma_reduce_init(mca_bcol_base_module_t *super);
|
|
int bcol_basesmuma_reduce_intra_fanin(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
int bcol_basesmuma_reduce_intra_fanin_old(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
int bcol_basesmuma_reduce_intra_reducescatter_gather(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);
|
|
|
|
/* Shared memory basesmuma allreduce */
|
|
int bcol_basesmuma_allreduce_init(mca_bcol_base_module_t *super);
|
|
|
|
int bcol_basesmuma_allreduce_intra_fanin_fanout(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
int bcol_basesmuma_allreduce_intra_recursive_doubling(bcol_function_args_t *input_args,
|
|
coll_ml_function_t *c_input_args);
|
|
|
|
/* initialize non-blocking barrier for recycling the memory buffers.
|
|
* This is not a general purpose nb_barrier, and relies on the
|
|
* fact that we will have only one outstanding nb-barrier per bank
|
|
* at a time.
|
|
*/
|
|
int bcol_basesmuma_rd_nb_barrier_init_admin(sm_nbbar_desc_t *sm_desc);
|
|
|
|
/* admin nonblocking barrier - progress function */
|
|
int bcol_basesmuma_rd_nb_barrier_progress_admin(sm_nbbar_desc_t *sm_desc);
|
|
|
|
/* Memory syncronization registration function */
|
|
int bcol_basesmuma_memsync_init(mca_bcol_base_module_t *super);
|
|
|
|
/* smcm allgather function used to exchange file offsets. */
|
|
int bcol_basesmuma_smcm_allgather_connection(
|
|
mca_bcol_basesmuma_module_t *sm_bcol_module,
|
|
mca_sbgp_base_module_t *module,
|
|
opal_list_t *peer_list,
|
|
bcol_basesmuma_smcm_proc_item_t ***backing_files,
|
|
ompi_communicator_t *comm,
|
|
bcol_basesmuma_smcm_file_t input, char *base_fname,
|
|
bool map_all);
|
|
|
|
/*
|
|
* this function initializes the internal scratch buffers and control
|
|
* structures that will be used by the module
|
|
*/
|
|
int base_bcol_masesmuma_setup_library_buffers(
|
|
mca_bcol_basesmuma_module_t *sm_bcol_module,
|
|
mca_bcol_basesmuma_component_t *sm_bcol_component);
|
|
|
|
/* get the index of the shared memory buffer to be used */
|
|
int bcol_basesmuma_get_buff_index( sm_buffer_mgmt * buff_block,
|
|
uint64_t buff_id );
|
|
|
|
int bcol_basesmuma_free_buff( sm_buffer_mgmt * buff_block,
|
|
uint64_t buff_id );
|
|
|
|
/* This function does bcol_basesmuma specific memory registration and
|
|
issues call back for ml level bank recycling
|
|
*/
|
|
int bcol_basesmuma_bank_init(struct mca_coll_ml_module_t *ml_module,
|
|
mca_bcol_base_module_t *bcol_module,
|
|
void *reg_data);
|
|
|
|
/* bank init which is used for shared memory optimization, fall back to
|
|
* the bank init above if this causes problems
|
|
*/
|
|
int bcol_basesmuma_bank_init_opti(struct mca_coll_ml_module_t *ml_module,
|
|
mca_bcol_base_module_t *bcol_module,
|
|
void *reg_data);
|
|
|
|
/* used for shared memory offset exchange */
|
|
int base_bcol_basesmuma_exchange_offsets(
|
|
mca_bcol_basesmuma_module_t *sm_bcol_module,
|
|
void **result_array, uint64_t mem_offset, int loop_limit,
|
|
int leading_dim);
|
|
|
|
|
|
/* the progress function to be called from the opal progress function
|
|
*/
|
|
int bcol_basesmuma_progress(void);
|
|
|
|
/* Macro for initializing my shared memory control structure */
|
|
#define BASESMUMA_HEADER_INIT(my_ctl_pointer,ready_flag, seqn, bcol_id) \
|
|
do{ \
|
|
int i,j; \
|
|
int8_t flag_offset = 0; \
|
|
/* setup resource recycling */ \
|
|
if( (my_ctl_pointer)->sequence_number < (seqn) ) { \
|
|
/* Signal arrival */ \
|
|
for( j = 0; j < SM_BCOLS_MAX; j++){ \
|
|
(my_ctl_pointer)->starting_flag_value[j]=0; \
|
|
for( i = 0; i < NUM_SIGNAL_FLAGS; i++){ \
|
|
(my_ctl_pointer)->flags[i][j] = -1; \
|
|
} \
|
|
} \
|
|
} \
|
|
/* increment the starting flag by one and return */ \
|
|
flag_offset = (my_ctl_pointer)->starting_flag_value[(bcol_id)]; \
|
|
(ready_flag) = flag_offset + 1; \
|
|
opal_atomic_wmb(); \
|
|
(my_ctl_pointer)->sequence_number = (seqn); \
|
|
}while(0)
|
|
|
|
/* these are all the same, am using a single macro for all collectives */
|
|
|
|
#define IS_PEER_READY(peer, my_flag, my_sequence_number,flag_index, bcol_id) \
|
|
(((peer)->sequence_number == (my_sequence_number) && \
|
|
(peer)->flags[flag_index][bcol_id] >= (my_flag))? true : false )
|
|
|
|
#if 0
|
|
#define IS_AR_DATA_READY(peer, my_flag, my_sequence_number) \
|
|
(((peer)->sequence_number == (my_sequence_number) && \
|
|
(peer)->flags[ALLREDUCE_FLAG][bcol_id] >= (my_flag) \
|
|
)? true : false )
|
|
|
|
#define IS_GDATA_READY(peer, my_flag, my_sequence_number) \
|
|
(((peer)->sequence_number == (my_sequence_number) && \
|
|
(peer)->flags[GATHER_FLAG][bcol_id] == (my_flag) \
|
|
)? true : false )
|
|
|
|
#define IS_PEER_READY(peer, my_flag, flag_index, my_sequence_number) \
|
|
((((volatile int64_t)(peer)->sequence_number > (my_sequence_number)) || \
|
|
(((volatile int64_t)(peer)->sequence_number == (my_sequence_number)) && \
|
|
((peer)->flags[flag_index][bcol_id] == (my_flag))) \
|
|
)? true : false )
|
|
|
|
#define IS_ALLREDUCE_PEER_READY(peer, my_flag, my_sequence_number) \
|
|
((((volatile int64_t)(peer)->sequence_number == (my_sequence_number)) && \
|
|
(((peer)->flags[ALLREDUCE_FLAG][bcol_id] == (my_flag))||((peer)->flags[ALLREDUCE_FLAG][bcol_id] == (my_flag) + 1)) \
|
|
)? true : false )
|
|
#endif
|
|
|
|
#define IS_LAST_BCOL_FUNC(ml_args) \
|
|
((((ml_args)->n_of_this_type_in_collective == \
|
|
(ml_args)->index_of_this_type_in_collective + 1 ) )? true : false)
|
|
|
|
static inline __opal_attribute_always_inline__
|
|
size_t bcol_basesmuma_data_offset_calc(
|
|
mca_bcol_basesmuma_module_t *basesmuma_module)
|
|
{
|
|
uint32_t offset = basesmuma_module->super.header_size;
|
|
offset = ((offset + BCOL_HEAD_ALIGN - 1) / BCOL_HEAD_ALIGN) * BCOL_HEAD_ALIGN;
|
|
|
|
return (size_t) offset;
|
|
}
|
|
|
|
|
|
END_C_DECLS
|
|
|
|
#endif /* MCA_BCOL_basesmuma_EXPORT_H */
|