Skip to content

Commit

Permalink
Fix bug in rounding up indices & sizes for vectorized halo exchanges.
Browse files Browse the repository at this point in the history 
Ver 2.06.02.
  • Loading branch information
@chuckyount
chuckyount committed Apr 28, 2018
1 parent eb8fcf2 commit 6d2cd4f
Show file tree
Hide file tree
Showing 4 changed files with 70 additions and 35 deletions.
2 changes: 1 addition & 1 deletion src/common/common_utils.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ namespace yask {
// for numbers above 9 (at least up to 99).

// Format: "major.minor.patch".
const string version = "2.06.01";
const string version = "2.06.02";

string yask_get_version_string() {
return version;
Expand Down
4 changes: 2 additions & 2 deletions src/kernel/lib/context.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1281,7 +1281,7 @@ namespace yask {
// Vec ok?
// Domain sizes must be ok, and buffer size must be ok
// as calculated when buffers were created.
bool send_vec_ok = vec_ok && sendBuf.has_all_vlen_mults;
bool send_vec_ok = vec_ok && sendBuf.vec_copy_ok;

// Get first and last ranges.
IdxTuple first = sendBuf.begin_pt;
Expand Down Expand Up @@ -1330,7 +1330,7 @@ namespace yask {
MPI_Wait(&grid_recv_reqs[ni], MPI_STATUS_IGNORE);

// Vec ok?
bool recv_vec_ok = vec_ok && recvBuf.has_all_vlen_mults;
bool recv_vec_ok = vec_ok && recvBuf.vec_copy_ok;

// Get first and last ranges.
IdxTuple first = recvBuf.begin_pt;
Expand Down
4 changes: 2 additions & 2 deletions src/kernel/lib/settings.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -633,8 +633,8 @@ namespace yask {
IdxTuple num_pts;

// Whether the number of points is a multiple of the
// vector length in all dims.
bool has_all_vlen_mults = false;
// vector length in all dims and buffer is aligned.
bool vec_copy_ok = false;

// Number of points overall.
idx_t get_size() const {
Expand Down
95 changes: 65 additions & 30 deletions src/kernel/lib/setup.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -371,7 +371,7 @@ namespace yask {
}

// Is vectorized exchange allowed based on domain sizes?
// Both my rank and neighbor rank must have all domain sizes
// Both my rank and neighbor rank must have *all* domain sizes
// of vector multiples.
bool vec_ok = allow_vec_exchange &&
_mpiInfo->has_all_vlen_mults[_mpiInfo->my_neighbor_index] &&
Expand All @@ -390,6 +390,7 @@ namespace yask {
if (!gp || gp->is_scratch() || gp->is_fixed_size())
continue;
auto& gname = gp->get_name();
bool grid_vec_ok = vec_ok;

// Lookup first & last domain indices and calc exchange sizes
// for this grid.
Expand All @@ -402,6 +403,9 @@ namespace yask {

// Only consider domain dims that are used in this grid.
if (gp->is_dim_used(dname)) {
auto vlen = _dims->_fold_pts[dname];
auto lhalo = gp->get_left_halo_size(dname);
auto rhalo = gp->get_right_halo_size(dname);

// Get domain indices for this grid. If there
// are no more ranks in the given direction,
Expand All @@ -413,12 +417,25 @@ namespace yask {
first_inner_idx.addDimBack(dname, fidx);
last_inner_idx.addDimBack(dname, lidx);
if (_opts->is_first_rank(dname))
fidx -= gp->get_left_halo_size(dname);
fidx -= lhalo;
if (_opts->is_last_rank(dname))
lidx += gp->get_right_halo_size(dname);
lidx += rhalo;
first_outer_idx.addDimBack(dname, fidx);
last_outer_idx.addDimBack(dname, lidx);

// Determine if it is possible to round the
// outer indices to vec-multiples. This will
// be required to allow full vec exchanges for
// this grid. We won't do the actual rounding
// yet, because we need to see if it's safe
// in all dims.
fidx = round_down_flr(fidx, vlen);
lidx = round_up_flr(lidx, vlen);
if (fidx < gp->get_first_rank_alloc_index(dname))
grid_vec_ok = false;
if (lidx > gp->get_last_rank_alloc_index(dname))
grid_vec_ok = false;

// Determine size of exchange in this dim. This
// will be the actual halo size plus any
// wave-front shifts. In the current
Expand All @@ -432,15 +449,12 @@ namespace yask {
if (neigh_offsets[dname] == MPIInfo::rank_prev) {
auto ext = wf_shifts[dname];

// my halo on my left.
auto halo_size = gp->get_left_halo_size(dname);
halo_size += ext;
my_halo_sizes.addDimBack(dname, halo_size);
// My halo on my left.
my_halo_sizes.addDimBack(dname, lhalo + ext);

// neighbor halo on their right.
halo_size = gp->get_right_halo_size(dname); // assume their right == my right.
halo_size += ext;
neigh_halo_sizes.addDimBack(dname, halo_size);
// Neighbor halo on their right.
// Assume my right is same as their right.
neigh_halo_sizes.addDimBack(dname, rhalo + ext);

// Flag that this grid has a neighbor to left or right.
found_delta = true;
Expand All @@ -450,15 +464,12 @@ namespace yask {
else if (neigh_offsets[dname] == MPIInfo::rank_next) {
auto ext = wf_shifts[dname];

// my halo on my right.
auto halo_size = gp->get_right_halo_size(dname);
halo_size += ext;
my_halo_sizes.addDimBack(dname, halo_size);
// My halo on my right.
my_halo_sizes.addDimBack(dname, rhalo + ext);

// neighbor halo on their left.
halo_size = gp->get_left_halo_size(dname); // assume their left == my left.
halo_size += ext;
neigh_halo_sizes.addDimBack(dname, halo_size);
// Neighbor halo on their left.
// Assume my left is same as their left.
neigh_halo_sizes.addDimBack(dname, lhalo + ext);

// Flag that this grid has a neighbor to left or right.
found_delta = true;
Expand All @@ -470,13 +481,6 @@ namespace yask {
neigh_halo_sizes.addDimBack(dname, 0);
}

// Round up halo sizes if vectorized exchanges allowed.
// TODO: add a heuristic to avoid increasing by a large factor.
if (vec_ok) {
auto vec_size = _dims->_fold_pts[dname];
my_halo_sizes.setVal(dname, ROUND_UP(my_halo_sizes[dname], vec_size));
neigh_halo_sizes.setVal(dname, ROUND_UP(neigh_halo_sizes[dname], vec_size));
}
} // domain dims in this grid.
} // domain dims.

Expand All @@ -491,14 +495,39 @@ namespace yask {
continue; // to next grid.
}

// Round halo sizes if vectorized exchanges allowed.
// Both self and neighbor must be vec-multiples
// and outer indices must be vec-mults or extendable
// to be so.
// TODO: add a heuristic to avoid increasing by a large factor.
if (grid_vec_ok) {
for (auto& dim : _dims->_domain_dims.getDims()) {
auto& dname = dim.getName();
if (gp->is_dim_used(dname)) {
auto vlen = _dims->_fold_pts[dname];

// first index rounded down.
first_outer_idx.setVal(dname, round_down_flr(first_outer_idx[dname], vlen));

// last index rounded up.
last_outer_idx.setVal(dname, round_up_flr(last_outer_idx[dname], vlen));

// sizes rounded up.
my_halo_sizes.setVal(dname, ROUND_UP(my_halo_sizes[dname], vlen));
neigh_halo_sizes.setVal(dname, ROUND_UP(neigh_halo_sizes[dname], vlen));

} // domain dims in this grid.
} // domain dims.
}

// Make a buffer in both directions (send & receive).
for (int bd = 0; bd < MPIBufs::nBufDirs; bd++) {

// Begin/end vars to indicate what part
// of main grid to read from or write to based on
// the current neighbor being processed.
IdxTuple copy_begin = gp->get_allocs();
IdxTuple copy_end = gp->get_allocs();
IdxTuple copy_end = gp->get_allocs(); // one past last!

// Adjust along domain dims in this grid.
for (auto& dim : _dims->_domain_dims.getDims()) {
Expand All @@ -516,13 +545,15 @@ namespace yask {

// Region to read from, i.e., data from inside
// this rank's domain to be put into neighbor's
// halo.
// halo. So, use neighbor's halo sizes when
// calculating buffer size.
if (bd == MPIBufs::bufSend) {

// Neighbor is to the left.
if (neigh_ofs == idx_t(MPIInfo::rank_prev)) {

// Only read slice as wide as halo from beginning.
copy_begin[dname] = first_inner_idx[dname];
copy_end[dname] = first_inner_idx[dname] + neigh_halo_sizes[dname];
}

Expand All @@ -531,13 +562,15 @@ namespace yask {

// Only read slice as wide as halo before end.
copy_begin[dname] = last_inner_idx[dname] + 1 - neigh_halo_sizes[dname];
copy_end[dname] = last_inner_idx[dname] + 1;
}

// Else, this neighbor is in same posn as I am in this dim,
// so we leave the default begin/end settings.
}

// Region to write to, i.e., into this rank's halo.
// So, use my halo sizes when calculating buffer sizes.
else if (bd == MPIBufs::bufRecv) {

// Neighbor is to the left.
Expand Down Expand Up @@ -573,10 +606,12 @@ namespace yask {
if (_dims->_domain_dims.lookup(dname)) {
dsize = copy_end[dname] - copy_begin[dname];

// Check whether size is multiple of vlen.
// Check whether alignment and size are multiple of vlen.
auto vlen = _dims->_fold_pts[dname];
if (dsize % vlen != 0)
vlen_mults = false;
if (imod_flr(copy_begin[dname], vlen) != 0)
vlen_mults = false;
}

// step dim?
Expand Down Expand Up @@ -638,7 +673,7 @@ namespace yask {
buf.last_pt = copy_last;
buf.num_pts = buf_sizes;
buf.name = bufname;
buf.has_all_vlen_mults = vlen_mults;
buf.vec_copy_ok = vlen_mults;

TRACE_MSG("MPI buffer '" << buf.name <<
"' configured for rank at relative offsets " <<
Expand Down

0 comments on commit 6d2cd4f

Please sign in to comment.