void _XMP_reflect_async_cardinal(_XMP_array_t *a, int async_id)
{
_XMP_async_comm_t *async = _XMP_get_current_async();
MPI_Request *reqs = &async->reqs[async->nreqs];
int nreqs = 0;
_XMP_TSTART(t0);
for (int i = 0; i < a->dim; i++){
_XMP_array_info_t *ai = &(a->info[i]);
if (ai->shadow_type == _XMP_N_SHADOW_NONE){
continue;
}
else if (ai->shadow_type == _XMP_N_SHADOW_NORMAL){
_XMP_reflect_sched_t *reflect = ai->reflect_sched;
if (_xmp_lwidth[i] || _xmp_uwidth[i]){
_XMP_ASSERT(reflect);
if (reflect->is_periodic == -1 /* not set yet */ ||
_xmp_lwidth[i] != reflect->lo_width ||
_xmp_uwidth[i] != reflect->hi_width ||
_xmp_is_periodic[i] != reflect->is_periodic){
reflect->lo_width = _xmp_lwidth[i];
reflect->hi_width = _xmp_uwidth[i];
reflect->is_periodic = _xmp_is_periodic[i];
_XMP_reflect_normal_sched_dim(a, i, _xmp_lwidth[i], _xmp_uwidth[i], _xmp_is_periodic[i]);
}
if (async->nreqs + nreqs + 4 > _XMP_MAX_ASYNC_REQS){
_XMP_fatal("too many arrays in an asynchronous reflect");
}
memcpy(&reqs[nreqs], reflect->req, 4 * sizeof(MPI_Request));
nreqs += 4;
_XMP_TSTART(t0);
if (reflect->req[0] != MPI_REQUEST_NULL) // if req[0] isn't null, any others shouldn't be null.
MPI_Startall(4, reflect->req);
_XMP_TEND2(xmptiming_.t_comm, xmptiming_.tdim_comm[i], t0);
}
}
else { /* _XMP_N_SHADOW_FULL */
_XMP_reflect_shadow_FULL(a->array_addr_p, a, i);
}
}
_XMP_TEND(xmptiming_.t_sched, t0);
async->nreqs += nreqs;
}
static void _XMP_reflect_wait(_XMP_array_t *a, int *lwidth, int *uwidth, int *is_periodic)
{
for (int i = 0; i < a->dim; i++){
if (!lwidth[i] && !uwidth[i]) continue;
_XMP_array_info_t *ai = &(a->info[i]);
if (ai->shadow_type == _XMP_N_SHADOW_NORMAL){
_XMP_reflect_sched_t *reflect = ai->reflect_sched;
_XMP_TSTART(t0);
MPI_Waitall(4, reflect->req, MPI_STATUSES_IGNORE);
_XMP_TEND2(xmptiming_.t_wait, xmptiming_.tdim_wait[i], t0);
}
else if (ai->shadow_type == _XMP_N_SHADOW_FULL){
//_XMP_reflect_shadow_FULL(a->array_addr_p, a, i);
}
}
if (_xmp_reflect_pack_flag){
_XMP_TSTART(t0);
_XMP_reflect_unpack(a, lwidth, uwidth, is_periodic);
_XMP_TEND(xmptiming_.t_copy, t0);
}
}
static void _XMP_reflect_wait(_XMP_array_t *a, int *lwidth, int *uwidth, int *is_periodic)
{
int nrdmas0 = 0, nrdmas1 = 0;
_XMP_TSTART(t0);
for (int i = 0; i < a->dim; i++){
_XMP_reflect_sched_t *reflect = a->info[i].reflect_sched;
if (lwidth[i] && reflect->hi_rank != -1) nrdmas0 += reflect->count;
if (uwidth[i] && reflect->lo_rank != -1) nrdmas1 += reflect->count;
}
while (nrdmas0 || nrdmas1){
while (FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC0, NULL) == FJMPI_RDMA_NOTICE){
nrdmas0--;
}
while (FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC1, NULL) == FJMPI_RDMA_NOTICE){
nrdmas1--;
}
}
xmp_barrier();
_XMP_TEND(xmptiming_.t_wait, t0);
}
static void _XMP_reflect_start(_XMP_array_t *a, int *lwidth, int *uwidth, int *is_periodic,
int tag)
{
_XMP_TSTART(t1);
xmp_barrier();
for (int i = 0; i < a->dim; i++){
_XMP_reflect_sched_t *reflect = a->info[i].reflect_sched;
_XMP_TSTART(t0);
// for lower reflect
if (lwidth[i] && reflect->hi_rank != -1){
for (int j = 0; j < reflect->count; j++){
FJMPI_Rdma_put(reflect->hi_rank, tag,
(uint64_t)reflect->lo_recv_array + j * reflect->stride,
(uint64_t)reflect->lo_send_array + j * reflect->stride,
lwidth[i] * reflect->blocklength,
FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC2);
}
}
// for upper reflect
if (uwidth[i] && reflect->lo_rank != -1){
for (int j = 0; j < reflect->count; j++){
FJMPI_Rdma_put(reflect->lo_rank, tag,
(uint64_t)reflect->hi_recv_array + j * reflect->stride,
(uint64_t)reflect->hi_send_array + j * reflect->stride,
uwidth[i] * reflect->blocklength,
FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC3);
}
}
_XMP_TEND(xmptiming_.tdim_comm[i], t0);
}
_XMP_TEND(xmptiming_.t_comm, t1);
}
void _XMP_reflect_pcopy_sched_dim(_XMP_array_t *adesc, int target_dim,
int lwidth, int uwidth, int is_periodic, int shadow_comm_type){
if (lwidth == 0 && uwidth == 0) return;
_XMP_array_info_t *ai = &(adesc->info[target_dim]);
_XMP_array_info_t *ainfo = adesc->info;
_XMP_ASSERT(ai->align_manner == _XMP_N_ALIGN_BLOCK);
_XMP_ASSERT(ai->is_shadow_comm_member);
if (lwidth > ai->shadow_size_lo || uwidth > ai->shadow_size_hi){
_XMP_fatal("reflect width is larger than shadow width.");
}
_XMP_reflect_sched_t *reflect = ai->reflect_sched;
int target_tdim = ai->align_template_index;
_XMP_nodes_info_t *ni = adesc->align_template->chunk[target_tdim].onto_nodes_info;
if (ni->size == 1 && !is_periodic) return;
int ndims = adesc->dim;
// 0-origin
int my_pos = ni->rank;
int lb_pos = _XMP_get_owner_pos(adesc, target_dim, ai->ser_lower);
int ub_pos = _XMP_get_owner_pos(adesc, target_dim, ai->ser_upper);
int lo_pos = (my_pos == lb_pos) ? ub_pos : my_pos - 1;
int hi_pos = (my_pos == ub_pos) ? lb_pos : my_pos + 1;
MPI_Comm *comm = adesc->align_template->onto_nodes->comm;
int my_rank = adesc->align_template->onto_nodes->comm_rank;
int lo_rank = my_rank + (lo_pos - my_pos) * ni->multiplier;
int hi_rank = my_rank + (hi_pos - my_pos) * ni->multiplier;
int count = 0, blocklength = 0;
long long stride = 0;
int type_size = adesc->type_size;
void *array_addr = adesc->array_addr_p;
void *lo_send_array = NULL, *lo_recv_array = NULL;
void *hi_send_array = NULL, *hi_recv_array = NULL;
void *lo_send_buf = NULL;
void *lo_recv_buf = NULL;
void *hi_send_buf = NULL;
void *hi_recv_buf = NULL;
int lo_buf_size = 0;
int hi_buf_size = 0;
if (reflect->prev_pcopy_sched_type &&
lwidth == reflect->lo_width &&
uwidth == reflect->hi_width &&
is_periodic == reflect->is_periodic){
if ((adesc->order == MPI_ORDER_FORTRAN && target_dim != ndims - 1) ||
(adesc->order == MPI_ORDER_C && target_dim != 0)){
goto init_comm;
}
else if (reflect->prev_pcopy_sched_type != shadow_comm_type){
count = reflect->count;
blocklength = reflect->blocklength;
stride = reflect->stride;
goto alloc_buf;
}
}
//
// setup data_type
//
if (adesc->order == MPI_ORDER_FORTRAN){ /* for XMP/F */
count = 1;
blocklength = type_size;
stride = ainfo[0].alloc_size * type_size;
for (int i = ndims - 2; i >= target_dim; i--){
count *= ainfo[i+1].alloc_size;
}
for (int i = 1; i <= target_dim; i++){
blocklength *= ainfo[i-1].alloc_size;
stride *= ainfo[i].alloc_size;
}
}
else if (adesc->order == MPI_ORDER_C){ /* for XMP/C */
count = 1;
blocklength = type_size;
stride = ainfo[ndims-1].alloc_size * type_size;
for (int i = 1; i <= target_dim; i++){
count *= ainfo[i-1].alloc_size;
}
for (int i = ndims - 2; i >= target_dim; i--){
blocklength *= ainfo[i+1].alloc_size;
stride *= ainfo[i].alloc_size;
}
}
else {
_XMP_fatal("cannot determin the base language.");
}
//
// calculate base address
//
alloc_buf:
// for lower reflect
if (lwidth){
lo_send_array = array_addr;
lo_recv_array = array_addr;
for (int i = 0; i < ndims; i++) {
int lb_send, lb_recv;
unsigned long long dim_acc;
if (i == target_dim) {
lb_send = ainfo[i].local_upper - lwidth + 1;
lb_recv = ainfo[i].shadow_size_lo - lwidth;;
}
else {
// Note: including shadow area
lb_send = 0;
lb_recv = 0;
}
dim_acc = ainfo[i].dim_acc;
lo_send_array = (void *)((char *)lo_send_array + lb_send * dim_acc * type_size);
lo_recv_array = (void *)((char *)lo_recv_array + lb_recv * dim_acc * type_size);
}
}
// for upper reflect
if (uwidth){
hi_send_array = array_addr;
hi_recv_array = array_addr;
for (int i = 0; i < ndims; i++) {
int lb_send, lb_recv;
unsigned long long dim_acc;
if (i == target_dim) {
lb_send = ainfo[i].local_lower;
lb_recv = ainfo[i].local_upper + 1;
}
else {
// Note: including shadow area
lb_send = 0;
lb_recv = 0;
}
dim_acc = ainfo[i].dim_acc;
hi_send_array = (void *)((char *)hi_send_array + lb_send * dim_acc * type_size);
hi_recv_array = (void *)((char *)hi_recv_array + lb_recv * dim_acc * type_size);
}
}
//
// Allocate buffers
//
if (reflect->prev_pcopy_sched_type == _XMP_COMM_REFLECT &&
((adesc->order == MPI_ORDER_FORTRAN && target_dim == ndims - 1) ||
(adesc->order == MPI_ORDER_C && target_dim == 0))){
;
}
else {
_XMP_free(reflect->lo_send_buf);
_XMP_free(reflect->lo_recv_buf);
_XMP_free(reflect->hi_send_buf);
_XMP_free(reflect->hi_recv_buf);
}
// for lower reflect
if (lwidth){
lo_buf_size = lwidth * blocklength * count;
if (shadow_comm_type == _XMP_COMM_REFLECT &&
((adesc->order == MPI_ORDER_FORTRAN && target_dim == ndims - 1) ||
(adesc->order == MPI_ORDER_C && target_dim == 0))){
lo_send_buf = lo_send_array;
lo_recv_buf = lo_recv_array;
}
else {
_XMP_TSTART(t0);
lo_send_buf = _XMP_alloc(lo_buf_size);
lo_recv_buf = _XMP_alloc(lo_buf_size);
_XMP_TEND2(xmptiming_.t_mem, xmptiming_.tdim_mem[target_dim], t0);
}
}
// for upper reflect
if (uwidth){
hi_buf_size = uwidth * blocklength * count;
if (shadow_comm_type == _XMP_COMM_REFLECT &&
((adesc->order == MPI_ORDER_FORTRAN && target_dim == ndims - 1) ||
(adesc->order == MPI_ORDER_C && target_dim == 0))){
hi_send_buf = hi_send_array;
hi_recv_buf = hi_recv_array;
}
else {
_XMP_TSTART(t0);
hi_send_buf = _XMP_alloc(hi_buf_size);
hi_recv_buf = _XMP_alloc(hi_buf_size);
_XMP_TEND2(xmptiming_.t_mem, xmptiming_.tdim_mem[target_dim], t0);
}
}
//
// cache schedule
//
reflect->count = count;
reflect->blocklength = blocklength;
reflect->stride = stride;
reflect->lo_send_array = lo_send_array;
reflect->lo_recv_array = lo_recv_array;
reflect->hi_send_array = hi_send_array;
reflect->hi_recv_array = hi_recv_array;
reflect->lo_send_buf = lo_send_buf;
reflect->lo_recv_buf = lo_recv_buf;
reflect->hi_send_buf = hi_send_buf;
reflect->hi_recv_buf = hi_recv_buf;
//
// initialize communication
//
int src, dst;
init_comm:
if (!is_periodic && my_pos == lb_pos){ // no periodic
lo_rank = MPI_PROC_NULL;
}
if (!is_periodic && my_pos == ub_pos){ // no periodic
hi_rank = MPI_PROC_NULL;
}
lo_buf_size = lwidth * reflect->blocklength * reflect->count;
hi_buf_size = uwidth * reflect->blocklength * reflect->count;
// for lower shadow
if (lwidth){
src = lo_rank;
dst = hi_rank;
}
else {
src = MPI_PROC_NULL;
dst = MPI_PROC_NULL;
}
if (shadow_comm_type == _XMP_COMM_REDUCE_SHADOW){
if (reflect->req_reduce[0] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req_reduce[0]);
}
if (reflect->req_reduce[1] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req_reduce[1]);
}
MPI_Send_init(reflect->lo_recv_buf, lo_buf_size, MPI_BYTE, src,
_XMP_N_MPI_TAG_REFLECT_LO, *comm, &reflect->req_reduce[0]);
MPI_Recv_init(reflect->lo_send_buf, lo_buf_size, MPI_BYTE, dst,
_XMP_N_MPI_TAG_REFLECT_LO, *comm, &reflect->req_reduce[1]);
}
else {
if (reflect->req[0] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[0]);
}
if (reflect->req[1] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[1]);
}
MPI_Recv_init(reflect->lo_recv_buf, lo_buf_size, MPI_BYTE, src,
_XMP_N_MPI_TAG_REFLECT_LO, *comm, &reflect->req[0]);
MPI_Send_init(reflect->lo_send_buf, lo_buf_size, MPI_BYTE, dst,
_XMP_N_MPI_TAG_REFLECT_LO, *comm, &reflect->req[1]);
}
// for upper shadow
if (uwidth){
src = hi_rank;
dst = lo_rank;
}
else {
src = MPI_PROC_NULL;
dst = MPI_PROC_NULL;
}
if (shadow_comm_type == _XMP_COMM_REDUCE_SHADOW){
if (reflect->req_reduce[2] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req_reduce[2]);
}
if (reflect->req_reduce[3] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req_reduce[3]);
}
MPI_Send_init(reflect->hi_recv_buf, hi_buf_size, MPI_BYTE, src,
_XMP_N_MPI_TAG_REFLECT_HI, *comm, &reflect->req_reduce[2]);
MPI_Recv_init(reflect->hi_send_buf, hi_buf_size, MPI_BYTE, dst,
_XMP_N_MPI_TAG_REFLECT_HI, *comm, &reflect->req_reduce[3]);
}
else {
if (reflect->req[2] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[2]);
}
if (reflect->req[3] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[3]);
}
MPI_Recv_init(reflect->hi_recv_buf, hi_buf_size, MPI_BYTE, src,
_XMP_N_MPI_TAG_REFLECT_HI, *comm, &reflect->req[2]);
MPI_Send_init(reflect->hi_send_buf, hi_buf_size, MPI_BYTE, dst,
_XMP_N_MPI_TAG_REFLECT_HI, *comm, &reflect->req[3]);
}
reflect->prev_pcopy_sched_type = shadow_comm_type;
reflect->lo_rank = lo_rank;
reflect->hi_rank = hi_rank;
}
void _XMP_reflect__(_XMP_array_t *a)
{
int is_ordinal = 1;
//_XMP_RETURN_IF_SINGLE;
if (!a->is_allocated){
_xmp_set_reflect_flag = 0;
return;
}
if (!_xmp_set_reflect_flag){
for (int i = 0; i < a->dim; i++){
_XMP_array_info_t *ai = &(a->info[i]);
_xmp_lwidth[i] = ai->shadow_size_lo;
_xmp_uwidth[i] = ai->shadow_size_hi;
_xmp_is_periodic[i] = 0;
}
}
_XMP_TSTART(t0);
for (int i = 0; i < a->dim; i++){
_XMP_array_info_t *ai = &(a->info[i]);
if (ai->shadow_type == _XMP_N_SHADOW_NONE){
continue;
}
else if (ai->shadow_type == _XMP_N_SHADOW_NORMAL){
_XMP_reflect_sched_t *reflect = ai->reflect_sched;
if (_xmp_lwidth[i] || _xmp_uwidth[i]){
_XMP_ASSERT(reflect);
if (reflect->is_periodic == -1 /* not set yet */ ||
_xmp_lwidth[i] != reflect->lo_width ||
_xmp_uwidth[i] != reflect->hi_width ||
_xmp_is_periodic[i] != reflect->is_periodic){
reflect->lo_width = _xmp_lwidth[i];
reflect->hi_width = _xmp_uwidth[i];
reflect->is_periodic = _xmp_is_periodic[i];
_XMP_reflect_rdma_sched_dim(a, i, _xmp_lwidth[i], _xmp_uwidth[i], _xmp_is_periodic[i]);
}
}
}
else { /* _XMP_N_SHADOW_FULL */
;
}
}
_XMP_TEND(xmptiming_.t_sched, t0);
_XMP_reflect_start(a, _xmp_lwidth, _xmp_uwidth, _xmp_is_periodic, 0);
_XMP_reflect_wait(a, _xmp_lwidth, _xmp_uwidth, _xmp_is_periodic);
_xmp_set_reflect_flag = 0;
for (int i = 0; i < a->dim; i++){
_xmp_lwidth[i] = 0;
_xmp_uwidth[i] = 0;
_xmp_is_periodic[i] = 0;
}
}
void _XMP_reflect__(_XMP_array_t *a)
{
int is_ordinal = 1;
//_XMP_RETURN_IF_SINGLE;
if (!a->is_allocated){
_xmp_set_reflect_flag = 0;
return;
}
if (!_xmp_set_reflect_flag){
for (int i = 0; i < a->dim; i++){
_XMP_array_info_t *ai = &(a->info[i]);
_xmp_lwidth[i] = ai->shadow_size_lo;
_xmp_uwidth[i] = ai->shadow_size_hi;
_xmp_is_periodic[i] = 0;
}
}
_XMP_TSTART(t0);
for (int i = 0; i < a->dim; i++){
_XMP_array_info_t *ai = &(a->info[i]);
if (ai->shadow_type == _XMP_N_SHADOW_NONE){
continue;
}
else if (ai->shadow_type == _XMP_N_SHADOW_NORMAL){
_XMP_reflect_sched_t *reflect = ai->reflect_sched;
if (_xmp_lwidth[i] || _xmp_uwidth[i]){
_XMP_ASSERT(reflect);
/* if (!reflect->reflect_is_initialized || */
/* _xmp_lwidth[i] != reflect->lo_width || */
/* _xmp_uwidth[i] != reflect->hi_width || */
/* _xmp_is_periodic[i] != reflect->is_periodic){ */
/* reflect->lo_width = _xmp_lwidth[i]; */
/* reflect->hi_width = _xmp_uwidth[i]; */
/* reflect->is_periodic = _xmp_is_periodic[i]; */
/* if (_xmp_reflect_pack_flag){ */
/* _XMP_reflect_pcopy_sched_dim(a, i, _xmp_lwidth[i], _xmp_uwidth[i], _xmp_is_periodic[i], 0); */
/* } */
/* else { */
/* _XMP_reflect_normal_sched_dim(a, i, _xmp_lwidth[i], _xmp_uwidth[i], _xmp_is_periodic[i]); */
/* } */
/* reflect->reflect_is_initialized = 1; */
/* } */
if (!reflect->reflect_is_initialized ||
_xmp_lwidth[i] != reflect->lo_width ||
_xmp_uwidth[i] != reflect->hi_width ||
_xmp_is_periodic[i] != reflect->is_periodic){
if (_xmp_reflect_pack_flag){
_XMP_reflect_pcopy_sched_dim(a, i, _xmp_lwidth[i], _xmp_uwidth[i], _xmp_is_periodic[i], _XMP_COMM_REFLECT);
}
else {
_XMP_reflect_normal_sched_dim(a, i, _xmp_lwidth[i], _xmp_uwidth[i], _xmp_is_periodic[i]);
}
reflect->reflect_is_initialized = 1;
reflect->lo_width = _xmp_lwidth[i];
reflect->hi_width = _xmp_uwidth[i];
reflect->is_periodic = _xmp_is_periodic[i];
}
if (_xmp_reflect_pack_flag && reflect->req[0] != MPI_REQUEST_NULL){
_XMP_TSTART(t0);
_XMP_reflect_pack_dim(a, i, _xmp_lwidth, _xmp_uwidth, _xmp_is_periodic, _XMP_COMM_REFLECT);
_XMP_TEND(xmptiming_.t_copy, t0);
}
_XMP_TSTART(t0);
if (reflect->req[0] != MPI_REQUEST_NULL) // if req[0] isn't null, any others shouldn't be null.
MPI_Startall(4, reflect->req);
_XMP_TEND2(xmptiming_.t_comm, xmptiming_.tdim_comm[i], t0);
if (is_ordinal){
_XMP_TSTART(t0);
MPI_Waitall(4, reflect->req, MPI_STATUSES_IGNORE);
_XMP_TEND2(xmptiming_.t_wait, xmptiming_.tdim_wait[i], t0);
if (_xmp_reflect_pack_flag && reflect->req[0] != MPI_REQUEST_NULL){
_XMP_TSTART(t0);
_XMP_reflect_unpack_dim(a, i, _xmp_lwidth, _xmp_uwidth, _xmp_is_periodic);
_XMP_TEND(xmptiming_.t_copy, t0);
}
}
}
}
else { /* _XMP_N_SHADOW_FULL */
_XMP_reflect_shadow_FULL(a->array_addr_p, a, i);
}
}
_XMP_TEND(xmptiming_.t_sched, t0);
// t0 = MPI_Wtime();
if (!is_ordinal)
_XMP_reflect_wait(a, _xmp_lwidth, _xmp_uwidth, _xmp_is_periodic);
// t_wait = t_wait + (MPI_Wtime() - t0);
_xmp_set_reflect_flag = 0;
for (int i = 0; i < a->dim; i++){
_xmp_lwidth[i] = 0;
_xmp_uwidth[i] = 0;
_xmp_is_periodic[i] = 0;
}
}
void _XMP_reflect_async_ordinal(_XMP_array_t *a, int async_id){
int n = a->dim;
_XMP_async_reflect_t *async_reflect;
_Bool reusable_sched = false;
if (!a->async_reflect){
int max_nreqs = (pow(3, n) - 1) * 2;
async_reflect = (_XMP_async_reflect_t *)_XMP_alloc(sizeof(_XMP_async_reflect_t));
async_reflect->datatype = (MPI_Datatype *)_XMP_alloc(sizeof(MPI_Datatype) * max_nreqs);
async_reflect->reqs = (MPI_Request *)_XMP_alloc(sizeof(MPI_Request) * max_nreqs);
for (int i = 0; i < max_nreqs; i++){
async_reflect->datatype[i] = MPI_DATATYPE_NULL;
async_reflect->reqs[i] = MPI_REQUEST_NULL;
}
async_reflect->nreqs = 0;
a->async_reflect = async_reflect;
}
else {
reusable_sched = true;
async_reflect = a->async_reflect;
for (int i = 0; i < n; i++){
if (async_reflect->lwidth[i] != _xmp_lwidth[i] ||
async_reflect->uwidth[i] != _xmp_uwidth[i] ||
async_reflect->is_periodic[i] != _xmp_is_periodic[i]){
reusable_sched = false;
break;
}
}
}
if (!reusable_sched){
int lb[_XMP_N_MAX_DIM] = { 0 };
int ub[_XMP_N_MAX_DIM] = { 0 };
for (int i = 0; i < n; i++){
async_reflect->lwidth[i] = _xmp_lwidth[i];
async_reflect->uwidth[i] = _xmp_uwidth[i];
async_reflect->is_periodic[i] = _xmp_is_periodic[i];
if (_xmp_lwidth[i] > 0) lb[i] = -1;
if (_xmp_uwidth[i] > 0) ub[i] = 1;
}
for (int i = 0; i < async_reflect->nreqs; i++){
if (async_reflect->datatype[i] != MPI_DATATYPE_NULL)
MPI_Type_free(&async_reflect->datatype[i]);
if (async_reflect->reqs[i] != MPI_REQUEST_NULL)
MPI_Request_free(&async_reflect->reqs[i]);
}
async_reflect->nreqs = 0;
int ishadow[_XMP_N_MAX_DIM];
for (ishadow[0] = lb[0]; ishadow[0] <= ub[0]; ishadow[0]++){
for (ishadow[1] = lb[1]; ishadow[1] <= ub[1]; ishadow[1]++){
for (ishadow[2] = lb[2]; ishadow[2] <= ub[2]; ishadow[2]++){
for (ishadow[3] = lb[3]; ishadow[3] <= ub[3]; ishadow[3]++){
for (ishadow[4] = lb[4]; ishadow[4] <= ub[4]; ishadow[4]++){
for (ishadow[5] = lb[5]; ishadow[5] <= ub[5]; ishadow[5]++){
for (ishadow[6] = lb[6]; ishadow[6] <= ub[6]; ishadow[6]++){
// When ishadow > 0, upper shadow is to be updated, and vice versa.
int nnzero = 0;
for (int i = 0; i < n; i++){
if (ishadow[i] != 0) nnzero++;
}
if (nnzero == 0) continue;
_XMP_reflect_sched_dir(a, ishadow, _xmp_lwidth, _xmp_uwidth, _xmp_is_periodic);
}}}}}}}
}
_XMP_async_comm_t *async = _XMP_get_current_async();
MPI_Request *reqs = &async->reqs[async->nreqs];
// copy to async
if (async->nreqs + async_reflect->nreqs > _XMP_MAX_ASYNC_REQS){
_XMP_fatal("too many arrays in an asynchronous reflect");
}
memcpy(reqs, async_reflect->reqs, async_reflect->nreqs * sizeof(MPI_Request));
async->nreqs += async_reflect->nreqs;
_XMP_TSTART(t0);
MPI_Startall(async_reflect->nreqs, reqs);
_XMP_TEND(xmptiming_.t_start, t0);
}
static void _XMP_reflect_pcopy_sched_dim(_XMP_array_t *adesc, int target_dim,
int lwidth, int uwidth, int is_periodic, void *dev_array_addr, int *lwidths, int *uwidths){
//printf("desc=%p, tardim=%d, lw=%d, uw=%d, devp=%p\n", adesc, target_dim, lwidth, uwidth, dev_array_addr);
if (lwidth == 0 && uwidth == 0) return;
_XMP_array_info_t *ai = &(adesc->info[target_dim]);
_XMP_array_info_t *ainfo = adesc->info;
_XMP_ASSERT(ai->align_manner == _XMP_N_ALIGN_BLOCK);
_XMP_ASSERT(ai->is_shadow_comm_member);
if (lwidth > ai->shadow_size_lo || uwidth > ai->shadow_size_hi){
_XMP_fatal("reflect width is larger than shadow width.");
}
_XMP_reflect_sched_t *reflect = ai->reflect_acc_sched;
int target_tdim = ai->align_template_index;
_XMP_nodes_info_t *ni = adesc->align_template->chunk[target_tdim].onto_nodes_info;
int ndims = adesc->dim;
// 0-origin
int my_pos = ni->rank;
int lb_pos = _XMP_get_owner_pos(adesc, target_dim, ai->ser_lower);
int ub_pos = _XMP_get_owner_pos(adesc, target_dim, ai->ser_upper);
int lo_pos = (my_pos == lb_pos) ? ub_pos : my_pos - 1;
int hi_pos = (my_pos == ub_pos) ? lb_pos : my_pos + 1;
MPI_Comm *comm = adesc->align_template->onto_nodes->comm;
int my_rank = adesc->align_template->onto_nodes->comm_rank;
int lo_rank = my_rank + (lo_pos - my_pos) * ni->multiplier;
int hi_rank = my_rank + (hi_pos - my_pos) * ni->multiplier;
int type_size = adesc->type_size;
//void *array_addr = adesc->array_addr_p;
void *lo_send_array = NULL;
void *lo_recv_array = NULL;
void *hi_send_array = NULL;
void *hi_recv_array = NULL;
void *lo_send_dev_buf = NULL;
void *lo_recv_dev_buf = NULL;
void *hi_send_dev_buf = NULL;
void *hi_recv_dev_buf = NULL;
void *lo_send_host_buf = NULL;
void *lo_recv_host_buf = NULL;
void *hi_send_host_buf = NULL;
void *hi_recv_host_buf = NULL;
void *mpi_lo_send_buf = NULL;
void *mpi_lo_recv_buf = NULL;
void *mpi_hi_send_buf = NULL;
void *mpi_hi_recv_buf = NULL;
int lo_buf_size = 0;
int hi_buf_size = 0;
//
// setup data_type
//
int count = 0, blocklength = 0;
long long stride = 0;
// int count_offset = 0;
if (_XMPF_running && !_XMPC_running){ /* for XMP/F */
count = 1;
blocklength = type_size;
stride = ainfo[0].alloc_size * type_size;
for (int i = ndims - 2; i >= target_dim; i--){
count *= ainfo[i+1].alloc_size;
}
for (int i = 1; i <= target_dim; i++){
blocklength *= ainfo[i-1].alloc_size;
stride *= ainfo[i].alloc_size;
}
}
else if (!_XMPF_running && _XMPC_running){ /* for XMP/C */
count = 1;
blocklength = type_size;
stride = ainfo[ndims-1].alloc_size * type_size;
/* if(target_dim > 0){ */
/* count *= ainfo[0].par_size; */
/* count_offset = ainfo[0].shadow_size_lo; */
/* } */
/* for (int i = 1; i < target_dim; i++){ */
/* count *= ainfo[i].alloc_size; */
/* } */
/* for (int i = ndims - 2; i >= target_dim; i--){ */
/* blocklength *= ainfo[i+1].alloc_size; */
/* stride *= ainfo[i].alloc_size; */
/* } */
if(target_dim == 0){
count *= 1;
if(ndims >= 2){
blocklength *= (ainfo[1].par_size + lwidths[1] + uwidths[1]);
}
}else{
count *= (ainfo[0].par_size + lwidths[0] + uwidths[0]);
for(int i = 1; i < target_dim; i++){
count *= ainfo[i].alloc_size;
}
blocklength *= ainfo[target_dim+1].alloc_size;
stride *= ainfo[target_dim].alloc_size;
}
for(int i = target_dim+2; i < ndims; i++){
blocklength *= ainfo[i].alloc_size;
}
for(int i = target_dim+1 ; i < ndims - 1; i++){
stride *= ainfo[i].alloc_size;
}
/* mod_4 */
count = 1;
blocklength = 1;
stride = 1;
for(int i = 0; i < ndims; i++){
int fact = (i == target_dim)? 1 : (ainfo[i].par_size + lwidths[i] + uwidths[i]);
int alloc_size = ainfo[i].alloc_size;
if(blocklength == 1 || fact == alloc_size){
blocklength *= fact;
stride *= alloc_size;
}else if(count == 1 && target_dim != 0){ //to be contiguous if target_dim==0
count = blocklength;
blocklength = fact;
stride = alloc_size;
}else{
blocklength *= alloc_size;
stride *= alloc_size;
}
//printf("tar=%d, i=%d, fact=%d, allocsize=%d, (%d,%d,%lld)\n", target_dim, i, fact, alloc_size, count , blocklength, stride);
}
blocklength *= type_size;
stride *= type_size;
/* mod_4 end */
/* it used at 150717
for (int i = 1; i <= target_dim; i++){
count *= ainfo[i-1].alloc_size;
}
for (int i = ndims - 2; i >= target_dim; i--){
blocklength *= ainfo[i+1].alloc_size;
stride *= ainfo[i].alloc_size;
}
*/
/* for (int i = target_dim + 1; i < ndims; i++){ */
/* blocklength *= ainfo[i].alloc_size; */
/* } */
/* for (int i = target_dim; i < ndims - 1; i++){ */
/* stride *= ainfo[i].alloc_size; */
/* } */
// printf("count =%d, blength=%d, stride=%lld\n", count ,blocklength, stride);
// printf("ainfo[0].par_size=%d\n", ainfo[0].par_size);
// printf("count_ofset=%d,\n", count_offset);
}
else {
_XMP_fatal("cannot determin the base language.");
}
//
// calculate base address
//
// for lower reflect
if (lwidth){
lo_send_array = lo_recv_array = (void *)((char*)dev_array_addr + /*count_offset*/0 * stride);
for (int i = 0; i < ndims; i++) {
int lb_send, lb_recv;
unsigned long long dim_acc;
if (i == target_dim) {
//printf("ainfo[%d].local_upper=%d\n",i,ainfo[i].local_upper);
lb_send = ainfo[i].local_upper - lwidth + 1;
lb_recv = ainfo[i].shadow_size_lo - lwidth; ////ainfo[i].local_lower - lwidth;
} else {
// Note: including shadow area
lb_send = 0; //// ainfo[i].local_lower - ainfo[i].shadow_size_lo;
lb_recv = 0; //// ainfo[i].local_lower - ainfo[i].shadow_size_lo;
}
dim_acc = ainfo[i].dim_acc;
lo_send_array = (void *)((char *)lo_send_array + lb_send * dim_acc * type_size);
lo_recv_array = (void *)((char *)lo_recv_array + lb_recv * dim_acc * type_size);
}
}
// for upper reflect
if (uwidth){
hi_send_array = hi_recv_array = (void *)((char*)dev_array_addr + /*count_offset*/0 * stride);
for (int i = 0; i < ndims; i++) {
int lb_send, lb_recv;
unsigned long long dim_acc;
if (i == target_dim) {
lb_send = ainfo[i].local_lower;
lb_recv = ainfo[i].local_upper + 1;
} else {
// Note: including shadow area
lb_send = 0; //ainfo[i].local_lower - ainfo[i].shadow_size_lo;
lb_recv = 0; //ainfo[i].local_lower - ainfo[i].shadow_size_lo;
}
dim_acc = ainfo[i].dim_acc;
hi_send_array = (void *)((char *)hi_send_array + lb_send * dim_acc * type_size);
hi_recv_array = (void *)((char *)hi_recv_array + lb_recv * dim_acc * type_size);
}
}
// for lower reflect
if (reflect->datatype_lo != MPI_DATATYPE_NULL){
MPI_Type_free(&reflect->datatype_lo);
}
if(packVector || count == 1){
MPI_Type_contiguous(blocklength * lwidth * count, MPI_BYTE, &reflect->datatype_lo);
// MPI_Type_contiguous(blocklength * lwidth * count / type_size, MPI_FLOAT, &reflect->datatype_lo);
fprintf(stderr, "dim=%d, send elements lo = %d\n", target_dim, blocklength * lwidth * count / type_size);
//fprintf(stderr, "useHostBuf=%c , packVector=%c\n", useHostBuffer, packVector);
// if(useHostBuffer){ fprintf(stderr,"using host buffer\n"); }
// if(packVector){ fprintf(stderr, "using pack vector\n"); }
}else{
MPI_Type_vector(count, blocklength * lwidth, stride, MPI_BYTE, &reflect->datatype_lo);
}
MPI_Type_commit(&reflect->datatype_lo);
// for upper reflect
if (reflect->datatype_hi != MPI_DATATYPE_NULL){
MPI_Type_free(&reflect->datatype_hi);
}
if(packVector || count == 1){
MPI_Type_contiguous(blocklength * uwidth * count, MPI_BYTE, &reflect->datatype_hi);
// MPI_Type_contiguous(blocklength * uwidth * count / type_size, MPI_FLOAT, &reflect->datatype_hi);
fprintf(stderr, "dim=%d, send elements hi = %d\n", target_dim, blocklength * uwidth * count / type_size);
}else{
MPI_Type_vector(count, blocklength * uwidth, stride, MPI_BYTE, &reflect->datatype_hi);
}
MPI_Type_commit(&reflect->datatype_hi);
//
// Allocate buffers
//
if(useHostBuffer){
CUDA_SAFE_CALL(cudaFreeHost(reflect->lo_send_host_buf));
CUDA_SAFE_CALL(cudaFreeHost(reflect->lo_recv_host_buf));
}
if ((_XMPF_running && target_dim != ndims - 1) ||
(_XMPC_running && target_dim != 0)){
if(packVector){
CUDA_SAFE_CALL(cudaFree(reflect->lo_send_buf));
CUDA_SAFE_CALL(cudaFree(reflect->lo_recv_buf));
}
}
if ((_XMPF_running && target_dim == ndims - 1) ||
(_XMPC_running && target_dim == 0)){
//
}
// for lower reflect
if (lwidth){
lo_buf_size = lwidth * blocklength * count;
hi_buf_size = uwidth * blocklength * count;
if ((_XMPF_running && target_dim == ndims - 1) ||
(_XMPC_running && target_dim == 0)){
lo_send_dev_buf = lo_send_array;
lo_recv_dev_buf = lo_recv_array;
hi_send_dev_buf = hi_send_array;
hi_recv_dev_buf = hi_recv_array;
} else {
_XMP_TSTART(t0);
if(packVector){
CUDA_SAFE_CALL(cudaMalloc((void **)&lo_send_dev_buf, lo_buf_size + hi_buf_size));
hi_send_dev_buf = (char*)lo_send_dev_buf + lo_buf_size;
CUDA_SAFE_CALL(cudaMalloc((void **)&lo_recv_dev_buf, lo_buf_size + hi_buf_size));
hi_recv_dev_buf = (char*)lo_recv_dev_buf + lo_buf_size;
}else{
lo_send_dev_buf = lo_send_array;
lo_recv_dev_buf = lo_recv_array;
hi_send_dev_buf = hi_send_array;
hi_recv_dev_buf = hi_recv_array;
}
_XMP_TEND2(xmptiming_.t_mem, xmptiming_.tdim_mem[target_dim], t0);
}
if(useHostBuffer){
CUDA_SAFE_CALL(cudaMallocHost((void**)&lo_send_host_buf, lo_buf_size + hi_buf_size));
hi_send_host_buf = (char*)lo_send_host_buf + lo_buf_size;
CUDA_SAFE_CALL(cudaMallocHost((void**)&lo_recv_host_buf, lo_buf_size + hi_buf_size));
hi_recv_host_buf = (char*)lo_recv_host_buf + lo_buf_size;
mpi_lo_send_buf = lo_send_host_buf;
mpi_lo_recv_buf = lo_recv_host_buf;
mpi_hi_send_buf = hi_send_host_buf;
mpi_hi_recv_buf = hi_recv_host_buf;
}else{
mpi_lo_send_buf = lo_send_dev_buf;
mpi_lo_recv_buf = lo_recv_dev_buf;
mpi_hi_send_buf = hi_send_dev_buf;
mpi_hi_recv_buf = hi_recv_dev_buf;
}
}
// for upper reflect
//
// initialize communication
//
int src, dst;
if (!is_periodic && my_pos == lb_pos){ // no periodic
lo_rank = MPI_PROC_NULL;
}
if (!is_periodic && my_pos == ub_pos){ // no periodic
hi_rank = MPI_PROC_NULL;
}
// for lower shadow
if (lwidth){
src = lo_rank;
dst = hi_rank;
} else {
src = MPI_PROC_NULL;
dst = MPI_PROC_NULL;
}
// fprintf(stderr, "dim=%d, lo_src=%d, lo_dst=%d\n", target_dim, src, dst);
if (reflect->req[0] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[0]);
}
if (reflect->req[1] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[1]);
}
MPI_Recv_init(mpi_lo_recv_buf, 1, reflect->datatype_lo, src,
_XMP_N_MPI_TAG_REFLECT_LO, *comm, &reflect->req[0]);
MPI_Send_init(mpi_lo_send_buf, 1, reflect->datatype_lo, dst,
_XMP_N_MPI_TAG_REFLECT_LO, *comm, &reflect->req[1]);
// for upper shadow
if (uwidth){
src = hi_rank;
dst = lo_rank;
} else {
src = MPI_PROC_NULL;
dst = MPI_PROC_NULL;
}
// fprintf(stderr, "dim=%d, hi_src=%d, hi_dst=%d\n", target_dim, src, dst);
if (reflect->req[2] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[2]);
}
if (reflect->req[3] != MPI_REQUEST_NULL){
MPI_Request_free(&reflect->req[3]);
}
MPI_Recv_init(mpi_hi_recv_buf, 1, reflect->datatype_hi, src,
_XMP_N_MPI_TAG_REFLECT_HI, *comm, &reflect->req[2]);
MPI_Send_init(mpi_hi_send_buf, 1, reflect->datatype_hi, dst,
_XMP_N_MPI_TAG_REFLECT_HI, *comm, &reflect->req[3]);
//
// cache schedule
//
reflect->count = count;
reflect->blocklength = blocklength;
reflect->stride = stride;
reflect->lo_send_array = lo_send_array;
reflect->lo_recv_array = lo_recv_array;
reflect->hi_send_array = hi_send_array;
reflect->hi_recv_array = hi_recv_array;
if(packVector){
reflect->lo_send_buf = lo_send_dev_buf;
reflect->lo_recv_buf = lo_recv_dev_buf;
reflect->hi_send_buf = hi_send_dev_buf;
reflect->hi_recv_buf = hi_recv_dev_buf;
}
if(useHostBuffer){
reflect->lo_send_host_buf = lo_send_host_buf;
reflect->lo_recv_host_buf = lo_recv_host_buf;
reflect->hi_send_host_buf = hi_send_host_buf;
reflect->hi_recv_host_buf = hi_recv_host_buf;
}
reflect->lo_rank = lo_rank;
reflect->hi_rank = hi_rank;
// gpu async
reflect->lo_async_id = _XMP_alloc(sizeof(cudaStream_t));
CUDA_SAFE_CALL(cudaStreamCreate(reflect->lo_async_id));
if(target_dim != 0 &&
(!useHostBuffer || (lo_rank != MPI_PROC_NULL && hi_rank != MPI_PROC_NULL && (lo_buf_size / type_size) <= useSingleStreamLimit)) ){
reflect->hi_async_id = NULL;
}else{
cudaStream_t *hi_stream = (cudaStream_t*)_XMP_alloc(sizeof(cudaStream_t));
CUDA_SAFE_CALL(cudaStreamCreate(hi_stream));
reflect->hi_async_id = (void*)hi_stream;
}
reflect->event = _XMP_alloc(sizeof(cudaEvent_t));
CUDA_SAFE_CALL(cudaEventCreateWithFlags(reflect->event, cudaEventDisableTiming));
}
static void _XMP_reflect_sched(_XMP_array_t *a, int *lwidth, int *uwidth,
int *is_periodic, int is_async, void *dev_addr)
{
_XMP_TSTART(t0);
for (int i = 0; i < a->dim; i++){
_XMP_array_info_t *ai = &(a->info[i]);
if (ai->shadow_type == _XMP_N_SHADOW_NONE){
continue;
}
else if (ai->shadow_type == _XMP_N_SHADOW_NORMAL){
_XMP_reflect_sched_t *reflect = ai->reflect_acc_sched;
if(reflect == NULL){
reflect = _XMP_alloc(sizeof(_XMP_reflect_sched_t));
reflect->is_periodic = -1; /* not used yet */
reflect->datatype_lo = MPI_DATATYPE_NULL;
reflect->datatype_hi = MPI_DATATYPE_NULL;
for (int j = 0; j < 4; j++) reflect->req[j] = MPI_REQUEST_NULL;
reflect->lo_send_buf = NULL;
reflect->lo_recv_buf = NULL;
reflect->hi_send_buf = NULL;
reflect->hi_recv_buf = NULL;
reflect->lo_send_host_buf = NULL;
reflect->lo_recv_host_buf = NULL;
reflect->hi_send_host_buf = NULL;
reflect->hi_recv_host_buf = NULL;
ai->reflect_acc_sched = reflect;
}else{
//
}
if (1/*lwidth[i] || uwidth[i]*/){
_XMP_ASSERT(reflect);
if (reflect->is_periodic == -1 /* not set yet */ ||
lwidth[i] != reflect->lo_width ||
uwidth[i] != reflect->hi_width ||
is_periodic[i] != reflect->is_periodic){
reflect->lo_width = lwidth[i];
reflect->hi_width = uwidth[i];
reflect->is_periodic = is_periodic[i];
if (/*_xmp_reflect_pack_flag && !is_async*/ 1){
_XMP_reflect_pcopy_sched_dim(a, i, lwidth[i], uwidth[i], is_periodic[i], dev_addr, lwidth, uwidth);
}
else {
//_XMP_reflect_normal_sched_dim(a, i, lwidth[i], uwidth[i], is_periodic[i]);
}
}
}
}
else { /* _XMP_N_SHADOW_FULL */
;
}
}
_XMP_TEND(xmptiming_.t_sched, t0);
}
