1
1
openmpi/ompi/mca/coll/fca/coll_fca_ops.c
Mike Dubman e8d7373b14 COLL/FCA: revert to prev barrier if called from finalize
FCA barrier may not complete if FCA progress is not called periodically.
PMI/PMI2 API that can be used in rte barrier has no provision for calling
external progress function.

So it is possible that during finalize some ranks will be stuck
in fca barrier while others are in PMI barrier.
2015-10-04 09:40:19 +03:00

627 строки
22 KiB
C

/**
Copyright (c) 2011 Mellanox Technologies. All rights reserved.
$COPYRIGHT$
Additional copyrights may follow
$HEADER$
*/
#include "ompi_config.h"
#include "ompi/constants.h"
#include "coll_fca.h"
#include "coll_fca_convertor.h"
static mca_coll_fca_dtype_info_t* mca_coll_fca_get_dtype(ompi_datatype_t *dtype)
{
mca_coll_fca_dtype_info_t *dtype_info;
ptrdiff_t lb, extent;
int id = dtype->id;
int fca_dtype;
if (id < 0 || id >= FCA_DT_MAX_PREDEFINED) {
return NULL;
}
/* Different dtype structures may have the same id. In that case, we assume
* they are aliases.
*/
dtype_info = &mca_coll_fca_component.fca_dtypes[id];
if (dtype_info->mpi_dtype->id == id) {
return dtype_info;
}
/* assert we don't overwrite another datatype */
assert(dtype_info->mpi_dtype == MPI_DATATYPE_NULL);
fca_dtype = fca_translate_mpi_dtype(dtype->name);
if (fca_dtype < 0) {
return NULL;
}
FCA_DT_GET_TRUE_EXTENT(dtype, &lb, &extent);
dtype_info->mpi_dtype = dtype;
dtype_info->mpi_dtype_extent = extent;
dtype_info->fca_dtype = fca_dtype;
dtype_info->fca_dtype_extent = fca_get_dtype_size(fca_dtype);
FCA_VERBOSE(2, "Added new dtype[%d]: %s fca id: %d, mpi size: %lu, fca size: %lu",
id, dtype->name, dtype_info->fca_dtype, dtype_info->mpi_dtype_extent,
dtype_info->fca_dtype_extent);
return dtype_info;
}
static mca_coll_fca_op_info_t *mca_coll_fca_get_op(ompi_op_t *op)
{
mca_coll_fca_op_info_t *op_info;
int i, fca_op;
/*
* Find 'op' in the array by exhaustive search. We assume all valid ops are
* in the beginning. If we stumble on a MPI_OP_NULL, we try to resolve the FCA
* operation code and store it in the array.
*/
op_info = mca_coll_fca_component.fca_reduce_ops;
for (i = 0; i < FCA_MAX_OPS; ++i, ++op_info) {
if (op_info->mpi_op == op) {
return op_info;
} else if (op_info->mpi_op == MPI_OP_NULL) {
fca_op = fca_translate_mpi_op(op->o_name);
if (fca_op < 0)
return NULL;
op_info->mpi_op = op;
op_info->fca_op = fca_op;
FCA_VERBOSE(2, "Added new op[%d]: %s fca id: %d", i, op->o_name, fca_op);
return op_info;
}
}
/* assert the array does not overflow */
/*assert(mca_coll_fca_component.fca_reduce_ops[FCA_MAX_OPS - 1] == MPI_OP_NULL);*/
return NULL;
}
/**
* If "datatype" is contiguous when it appears "count" times, return 1 and
* set "*size" to the total buffer size, and *gap to the gap before the data.
* Otherwise return 0.
*/
static inline int mca_coll_fca_array_size(ompi_datatype_t *dtype, int count,
size_t *gap, size_t *size)
{
ptrdiff_t true_lb, true_extent;
if (FCA_DT_IS_CONTIGUOUS_MEMORY_LAYOUT(dtype, count)) {
FCA_DT_GET_TRUE_EXTENT(dtype, &true_lb, &true_extent);
*gap = true_lb;
*size = true_extent * count;
return 1;
} else {
return 0;
}
}
static int mca_coll_fca_fill_reduce_spec(int count, ompi_datatype_t *dtype,
ompi_op_t *op, fca_reduce_spec_t *spec,
int max_fca_payload)
{
mca_coll_fca_dtype_info_t *dtype_info;
mca_coll_fca_op_info_t *op_info;
/* Check dtype */
dtype_info = mca_coll_fca_get_dtype(dtype);
if (!dtype_info) {
FCA_VERBOSE(10, "Unsupported dtype: %s", dtype->name);
return OMPI_ERROR;
}
/* Check FCA size */
if ((int)(dtype_info->fca_dtype_extent * count) > max_fca_payload) {
FCA_VERBOSE(10, "Unsupported buffer size: %lu", dtype_info->fca_dtype_extent * count);
return OMPI_ERROR;
}
/* Check operation */
op_info = mca_coll_fca_get_op(op);
if (!op_info) {
FCA_VERBOSE(10, "Unsupported op: %s", op->o_name);
return OMPI_ERROR;
}
/* Fill spec */
spec->dtype = dtype_info->fca_dtype;
spec->op = op_info->fca_op;
spec->length = count;
spec->buf_size = dtype_info->mpi_dtype_extent * count;
if (MPI_IN_PLACE == spec->sbuf) {
FCA_VERBOSE(10, "Using MPI_IN_PLACE for sbuf");
spec->sbuf = spec->rbuf;
} else if (MPI_IN_PLACE == spec->rbuf) {
FCA_VERBOSE(10, "Using MPI_IN_PLACE for rbuf");
spec->rbuf = spec->sbuf;
}
return OMPI_SUCCESS;
}
/*
* Function: - barrier
* Returns: - MPI_SUCCESS or error code
*/
int mca_coll_fca_barrier(struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
int ret;
FCA_VERBOSE(5,"Using FCA Barrier");
if (OPAL_UNLIKELY(ompi_mpi_finalize_started)) {
FCA_VERBOSE(5, "In finalize, reverting to previous barrier");
goto orig_barrier;
}
ret = fca_do_barrier(fca_module->fca_comm);
if (ret < 0) {
if (ret == -EUSEMPI) {
goto orig_barrier;
}
FCA_ERROR("Barrier failed: %s", fca_strerror(ret));
return OMPI_ERROR;
}
return OMPI_SUCCESS;
orig_barrier:
return fca_module->previous_barrier(comm, fca_module->previous_barrier_module);
}
/*
* Function: - broadcast
* Accepts: - same arguments as MPI_Bcast()
* Returns: - MPI_SUCCESS or error code
*/
int mca_coll_fca_bcast(void *buff, int count, struct ompi_datatype_t *datatype,
int root, struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
MCA_COLL_FCA_DECLARE_CONVERTOR(conv);
fca_bcast_spec_t spec;
size_t gap, size;
int ret;
FCA_VERBOSE(5, "[%d] Calling mca_coll_fca_bcast, root=%d, count=%d",
ompi_comm_rank(comm), root, count);
/* Setup exchange buffer */
spec.root = root;
if (mca_coll_fca_array_size(datatype, count, &gap, &size)) {
spec.buf = (char*)buff + gap;
} else {
mca_coll_fca_convertor_create(&conv, datatype, count, buff,
(root == fca_module->rank)
? MCA_COLL_FCA_CONV_SEND
: MCA_COLL_FCA_CONV_RECV,
&spec.buf, &size);
}
/* Check that operation size does not exceed limit */
spec.size = size;
if (spec.size > fca_module->fca_comm_caps.max_payload) {
FCA_VERBOSE(5, "Unsupported bcast operation size %d, using fallback",
spec.size);
if (spec.buf != buff) {
mca_coll_fca_convertor_destroy(&conv);
}
goto orig_bcast;
}
/* Sender may pack data */
if (spec.buf != buff && root == fca_module->rank) {
mca_coll_fca_convertor_process(&conv, 0);
}
/* Call FCA Bcast */
FCA_VERBOSE(5, "Using FCA Bcast");
ret = fca_do_bcast(fca_module->fca_comm, &spec);
/* Destroy convertor if operation failed */
if (ret < 0) {
mca_coll_fca_convertor_destroy(&conv);
if (ret == -EUSEMPI) {
goto orig_bcast;
}
FCA_ERROR("Bcast failed: %s", fca_strerror(ret));
return OMPI_ERROR;
}
/* Unpack data and clean up convertor */
if (mca_coll_fca_convertor_valid(&conv)) {
if (root != fca_module->rank) {
mca_coll_fca_convertor_process(&conv, 0);
}
mca_coll_fca_convertor_destroy(&conv);
}
return OMPI_SUCCESS;
orig_bcast:
return fca_module->previous_bcast(buff, count, datatype, root, comm,
fca_module->previous_bcast_module);
}
/*
* Reduce
*
* Function: - reduce
* Accepts: - same as MPI_Reduce()
* Returns: - MPI_SUCCESS or error code
*/
int mca_coll_fca_reduce(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)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
fca_reduce_spec_t spec;
int ret;
mca_coll_fca_get_reduce_root(root, fca_module->rank, &spec);
spec.sbuf = sbuf;
spec.rbuf = rbuf;
if (mca_coll_fca_fill_reduce_spec(count, dtype, op, &spec,
fca_module->fca_comm_caps.max_payload)
!= OMPI_SUCCESS) {
FCA_VERBOSE(5, "Unsupported reduce operation %s, using fallback\n", op->o_name);
goto orig_reduce;
}
FCA_VERBOSE(5,"Using FCA Reduce");
ret = fca_do_reduce(fca_module->fca_comm, &spec);
if (ret < 0) {
if (ret == -EUSEMPI) {
goto orig_reduce;
}
FCA_ERROR("Reduce failed: %s", fca_strerror(ret));
return OMPI_ERROR;
}
return OMPI_SUCCESS;
orig_reduce:
return fca_module->previous_reduce(sbuf, rbuf, count, dtype, op, root,
comm, fca_module->previous_reduce_module);
}
/*
* Allreduce
*
* Function: - allreduce
* Accepts: - same as MPI_Allreduce()
* Returns: - MPI_SUCCESS or error code
*/
int mca_coll_fca_allreduce(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
fca_reduce_spec_t spec;
int ret;
spec.sbuf = sbuf;
spec.rbuf = rbuf;
if (mca_coll_fca_fill_reduce_spec(count, dtype, op, &spec,
fca_module->fca_comm_caps.max_payload)
!= OMPI_SUCCESS) {
FCA_VERBOSE(5, "Unsupported allreduce operation %s, using fallback\n", op->o_name);
goto orig_allreduce;
}
FCA_VERBOSE(5,"Using FCA Allreduce");
ret = fca_do_all_reduce(fca_module->fca_comm, &spec);
if (ret < 0) {
if (ret == -EUSEMPI) {
goto orig_allreduce;
}
FCA_ERROR("Allreduce failed: %s", fca_strerror(ret));
return OMPI_ERROR;
}
return OMPI_SUCCESS;
orig_allreduce:
return fca_module->previous_allreduce(sbuf, rbuf, count, dtype, op, comm,
fca_module->previous_allreduce_module);
}
/*
* Prepare a send buffer for allgather/allgatherv, handle packing and MPI_IN_PLACE.
*/
static size_t __setup_gather_sendbuf(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
struct mca_coll_fca_convertor *sconv,
void **real_sendbuf)
{
size_t gap, ssize;
if (mca_coll_fca_array_size(sdtype, scount, &gap, &ssize)) {
*real_sendbuf = (char*) sbuf + gap;
} else {
FCA_VERBOSE(5, "Packing send buffer");
mca_coll_fca_convertor_create(sconv, sdtype, scount, sbuf,
MCA_COLL_FCA_CONV_SEND, real_sendbuf,
&ssize);
mca_coll_fca_convertor_process(sconv, 0);
}
return ssize;
}
static size_t __setup_gather_sendbuf_inplace(void *inplace_sbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct mca_coll_fca_convertor *sconv,
void **real_sendbuf)
{
size_t gap, ssize;
if (mca_coll_fca_array_size(rdtype, rcount, &gap, &ssize)) {
*real_sendbuf = (char*) inplace_sbuf + gap;
} else {
FCA_VERBOSE(5, "Packing send buffer");
mca_coll_fca_convertor_create(sconv, rdtype, rcount, inplace_sbuf,
MCA_COLL_FCA_CONV_SEND, real_sendbuf,
&ssize);
mca_coll_fca_convertor_process(sconv, 0);
}
return ssize;
}
/*
* Allgather
*
* Function: - allgather
* Accepts: - same as MPI_Allgather()
* Returns: - MPI_SUCCESS or error code
*/
int mca_coll_fca_allgather(void *sbuf, int scount, struct ompi_datatype_t *sdtype,
void *rbuf, int rcount, struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
#if OMPI_FCA_ALLGATHER == 1
MCA_COLL_FCA_DECLARE_CONVERTOR(sconv);
MCA_COLL_FCA_DECLARE_CONVERTOR(rconv);
fca_gather_spec_t spec = {0,};
size_t rgap, rsize;
ptrdiff_t rdtype_extent;
ssize_t total_rcount;
int ret;
/* Setup send buffer */
if(sbuf == MPI_IN_PLACE ) {
FCA_DT_EXTENT(rdtype, &rdtype_extent);
spec.size = __setup_gather_sendbuf_inplace(
(char *)rbuf + rcount * fca_module->rank * rdtype_extent,
rcount, rdtype, &sconv, &spec.sbuf);
} else {
spec.size = __setup_gather_sendbuf(sbuf, scount, sdtype, &sconv, &spec.sbuf);
}
/* Setup recv buffer */
total_rcount = ompi_comm_size(comm) * rcount;
if (mca_coll_fca_array_size(rdtype, total_rcount, &rgap, &rsize) && rgap == 0) {
spec.rbuf = rbuf;
} else {
mca_coll_fca_convertor_create(&rconv, rdtype, total_rcount, rbuf,
MCA_COLL_FCA_CONV_RECV, &spec.rbuf, &rsize);
}
/* Call FCA Allgather */
FCA_VERBOSE(5,"Using FCA Allgather size");
ret = fca_do_allgather(fca_module->fca_comm, &spec);
/* Destroy convertors if operation failed */
if (ret < 0) {
mca_coll_fca_convertor_destroy(&sconv);
mca_coll_fca_convertor_destroy(&rconv);
if (ret == -EUSEMPI) {
goto orig_allgather;
}
FCA_ERROR("Allgather failed: %s", fca_strerror(ret));
return OMPI_ERROR;
}
/* Unpack data and clean up convertor */
mca_coll_fca_convertor_destroy(&sconv);
if (mca_coll_fca_convertor_valid(&rconv)) {
FCA_VERBOSE(5, "Unpacking Allgather receive buffer");
mca_coll_fca_convertor_process(&rconv, 0);
mca_coll_fca_convertor_destroy(&rconv);
}
return OMPI_SUCCESS;
orig_allgather:
#endif
return fca_module->previous_allgather(sbuf, scount, sdtype, rbuf, rcount, rdtype,
comm, fca_module->previous_allgather_module);
}
int mca_coll_fca_allgatherv(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *disps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
#if OMPI_FCA_ALLGATHER == 1
MCA_COLL_FCA_DECLARE_CONVERTOR(sconv);
MCA_COLL_FCA_DECLARE_CONVERTOR(rconv);
fca_gatherv_spec_t spec;
size_t rgap, rsize;
int sum_rcounts;
ptrdiff_t rdtype_extent;
int comm_size;
int relemsize;
size_t displ;
int i, ret;
comm_size = ompi_comm_size(fca_module->comm);
FCA_DT_EXTENT(rdtype, &rdtype_extent);
/* Setup send buffer */
if(sbuf == MPI_IN_PLACE ) {
spec.sendsize = __setup_gather_sendbuf_inplace(
(char *)rbuf + disps[fca_module->rank] * rdtype_extent,
rcounts[fca_module->rank], rdtype, &sconv, &spec.sbuf);
} else {
spec.sendsize = __setup_gather_sendbuf(sbuf, scount, sdtype, &sconv, &spec.sbuf);
}
/* Allocate alternative recvsizes/displs on the stack, which will be in bytes */
spec.recvsizes = alloca(sizeof *spec.recvsizes * comm_size);
spec.displs = alloca(sizeof *spec.displs * comm_size);
/* Calculate the size of receive buffer */
sum_rcounts = 0;
for (i = 0; i < comm_size; ++i) {
sum_rcounts += rcounts[i];
}
/* convert MPI counts which depend on dtype) to FCA sizes (which are in bytes) */
if (mca_coll_fca_array_size(rdtype, sum_rcounts, &rgap, &rsize) && rgap == 0) {
spec.rbuf = rbuf;
for (i = 0; i < comm_size; ++i) {
spec.recvsizes[i] = rcounts[i] * rdtype_extent;
spec.displs[i] = disps[i] * rdtype_extent;
}
} else {
/*
* Reorder and remove gaps in displs - we want to allocate as little memory
* as possible, and we should unpack one-by-one anyway.
*/
FCA_VERBOSE(5, "Reordering AllgatherV displacements");
mca_coll_fca_convertor_create(&rconv, rdtype, sum_rcounts, rbuf,
MCA_COLL_FCA_CONV_RECV, &spec.rbuf, &rsize);
assert(rsize % sum_rcounts == 0);
relemsize = rsize / sum_rcounts;
displ = 0;
for (i = 0; i < comm_size; ++i) {
spec.recvsizes[i] = rcounts[i] * relemsize;
spec.displs[i] = displ;
displ += spec.recvsizes[i];
}
assert(displ == rsize);
}
/* Call FCA AllgatherV */
FCA_VERBOSE(5,"Using FCA Allgatherv");
ret = fca_do_allgatherv(fca_module->fca_comm, &spec);
/* Destroy convertors if operation failed */
if (ret < 0) {
mca_coll_fca_convertor_destroy(&sconv);
mca_coll_fca_convertor_destroy(&rconv);
if (ret == -EUSEMPI) {
goto orig_allgatherv;
}
FCA_ERROR("Allgatherv failed: %s", fca_strerror(ret));
return OMPI_ERROR;
}
/* Unpack data and clean up convertor */
mca_coll_fca_convertor_destroy(&sconv);
if (mca_coll_fca_convertor_valid(&rconv)) {
FCA_VERBOSE(5, "Unpacking AllgatherV receive buffer rdtype_extent=%ld",
rdtype_extent);
for (i = 0; i < comm_size; ++i) {
mca_coll_fca_convertor_set(&rconv, rdtype,
(char*)rbuf + disps[i] * rdtype_extent,
rcounts[i]);
mca_coll_fca_convertor_process(&rconv, spec.displs[i]);
}
mca_coll_fca_convertor_destroy(&rconv);
}
return OMPI_SUCCESS;
orig_allgatherv:
#endif
return fca_module->previous_allgatherv(sbuf, scount, sdtype, rbuf, rcounts,
disps, rdtype, comm,
fca_module->previous_allgatherv_module);
}
int mca_coll_fca_alltoall(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
/* not implemented yet */
return fca_module->previous_alltoall(sbuf, scount, sdtype, rbuf, rcount, rdtype,
comm, fca_module->previous_alltoall_module);
}
int mca_coll_fca_alltoallv(void *sbuf, int *scounts, int *sdisps,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *rdisps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
/* not implemented yet */
return fca_module->previous_alltoallv(sbuf, scounts, sdisps, sdtype, rbuf, rcounts, rdisps, rdtype,
comm, fca_module->previous_alltoallv_module);
}
int mca_coll_fca_alltoallw(void *sbuf, int *scounts, int *sdisps,
struct ompi_datatype_t **sdtypes,
void *rbuf, int *rcounts, int *rdisps,
struct ompi_datatype_t **rdtypes,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
/* not implemented yet */
return fca_module->previous_alltoallw(sbuf, scounts, sdisps, sdtypes, rbuf, rcounts, rdisps, rdtypes,
comm, fca_module->previous_alltoallw_module);
}
int mca_coll_fca_gather(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
int root, struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
/* not implemented yet */
return fca_module->previous_gather(sbuf, scount, sdtype, rbuf, rcount, rdtype, root,
comm, fca_module->previous_gather_module);
}
int mca_coll_fca_gatherv(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *disps,
struct ompi_datatype_t *rdtype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
/* not implemented yet */
return fca_module->previous_gatherv(sbuf, scount, sdtype, rbuf, rcounts, disps, rdtype, root,
comm, fca_module->previous_gatherv_module);
}
int mca_coll_fca_reduce_scatter(void *sbuf, void *rbuf, int *rcounts,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_fca_module_t *fca_module = (mca_coll_fca_module_t*)module;
/* not implemented yet */
return fca_module->previous_reduce_scatter(sbuf, rbuf, rcounts, dtype, op,
comm, fca_module->previous_reduce_scatter_module);
}