void Config::_registerModels()
{
// Register distribution helpers on each config run
const bool createDist = _modelDist.empty(); //first run, create distributors
const size_t nModels = _models.size();
LBASSERT( createDist || _modelDist.size() == nModels );
for( size_t i = 0; i < nModels; ++i )
{
const Model* model = _models[i];
ModelDist* modelDist = 0;
if( createDist )
{
modelDist = new ModelDist( model );
_modelDist.push_back( modelDist );
}
else
modelDist = _modelDist[i];
modelDist->registerTree( getClient( ));
LBASSERT( modelDist->isAttached() );
_frameData.setModelID( modelDist->getID( ));
}
LBASSERT( _modelDist.size() == nModels );
if( !_modelDist.empty( ))
{
ModelAssigner assigner( _modelDist );
accept( assigner );
}
}
int main(int argc, char* argv[])
{
diy::mpi::environment env(argc, argv); // diy equivalent of MPI_Init
diy::mpi::communicator world; // diy equivalent of MPI communicator
int size = 8; // total number of MPI processes
int nblocks = 32; // total number of blocks in global domain
diy::ContiguousAssigner assigner(size, nblocks);
Bounds domain; // global data size
domain.min[0] = domain.min[1] = domain.min[2] = 0;
domain.max[0] = domain.max[1] = domain.max[2] = 255;
int rank = world.rank(); // MPI rank of this process
std::cout << "Rank " << rank << ":" << std::endl;
diy::Master master(world,
1, // one thread
-1, // all blocks in memory
&Block::create,
&Block::destroy);
AddBlock addblock(master); // object for adding new blocks to master
// share_face is an n-dim (size 3 in this example) vector of bools
// indicating whether faces are shared in each dimension
// uninitialized values default to false
diy::RegularDecomposer<Bounds>::BoolVector share_face;
share_face.push_back(true);
// wrap is an n-dim (size 3 in this example) vector of bools
// indicating whether boundary conditions are periodic in each dimension
// uninitialized values default to false
diy::RegularDecomposer<Bounds>::BoolVector wrap;
wrap.push_back(true);
wrap.push_back(true);
// ghosts is an n-dim (size 3 in this example) vector of ints
// indicating number of ghost cells per side in each dimension
// uninitialized values default to 0
diy::RegularDecomposer<Bounds>::CoordinateVector ghosts;
ghosts.push_back(1); ghosts.push_back(2);
// either create the regular decomposer and call its decompose function
// (having the decomposer available is useful for its other member functions
diy::RegularDecomposer<Bounds> decomposer(3,
domain,
assigner,
share_face,
wrap,
ghosts);
decomposer.decompose(rank, addblock);
// or combine the two lines above into the following helper function
// but the decomposer gets destroyed afterwards
// diy::decompose(3, rank, domain, assigner, addblock, share_face, wrap, ghosts);
// display the decomposition
master.foreach(&Block::show_link);
}
inline void apply(InputGeometry const& geometry, partitions& state) const
{
// First pass.
// Get min/max (in most cases left / right) points
// This makes use of the geometry::less/greater predicates
// For the left boundary it is important that multiple points
// are sorted from bottom to top. Therefore the less predicate
// does not take the x-only template parameter (this fixes ticket #6019.
// For the right boundary it is not necessary (though also not harmful),
// because points are sorted from bottom to top in a later stage.
// For symmetry and to get often more balanced lower/upper halves
// we keep it.
typedef typename geometry::detail::range_type<InputGeometry>::type range_type;
typedef typename boost::range_iterator
<
range_type const
>::type range_iterator;
detail::get_extremes
<
range_type,
range_iterator,
geometry::less<point_type>,
geometry::greater<point_type>
> extremes;
geometry::detail::for_each_range(geometry, extremes);
// Bounding left/right points
// Second pass, now that extremes are found, assign all points
// in either lower, either upper
detail::assign_range
<
range_type,
range_iterator,
container_type,
typename strategy::side::services::default_strategy<cs_tag>::type
> assigner(extremes.left, extremes.right);
geometry::detail::for_each_range(geometry, assigner);
// Sort both collections, first on x(, then on y)
detail::sort(assigner.lower_points);
detail::sort(assigner.upper_points);
//std::cout << boost::size(assigner.lower_points) << std::endl;
//std::cout << boost::size(assigner.upper_points) << std::endl;
// And decide which point should be in the final hull
build_half_hull<-1>(assigner.lower_points, state.m_lower_hull,
extremes.left, extremes.right);
build_half_hull<1>(assigner.upper_points, state.m_upper_hull,
extremes.left, extremes.right);
}
static int kl_script_event_dequeue_wrap(lua_State* L)
{
kl_script_context_t sctx = (kl_script_context_t)lua_topointer(L, 1);
kl_script_event_t event;
if(kl_script_event_dequeue(sctx, &event) == KL_SUCCESS)
{
uint32_t context_type = kl_get_script_event_context_type(event.event.id);
int ret = 3;
lua_pushinteger(L, event.event.id);
if(context_type == KL_SCRIPT_CONTEXT_TYPE_ASSIGNER)
{
kl_script_event_context_assigner_fn assigner =
kl_get_script_event_context_assigner(event.event.id);
ret = assigner(L, &event) + 1;
}
else
{
switch(context_type)
{
case 0:
{
lua_pushlightuserdata(L, event.event.context.as_ptr);
break;
}
case LUA_TNUMBER:
{
lua_pushnumber(L, *((lua_Number*)event.event.context.as_ptr));
kl_heap_free(event.event.context.as_ptr);
break;
}
case LUA_TSTRING:
{
lua_pushstring(L, event.event.context.as_ptr);
kl_heap_free(event.event.context.as_ptr);
break;
}
}
lua_pushinteger(L, event.event.arg.as_uint32);
}
return ret;
}
lua_pushnil(L);
return 1;
}
int main(int argc, char* argv[])
{
diy::mpi::environment env(argc, argv);
diy::mpi::communicator world;
diy::Master master(world, 1, -1,
&create_block, // master will take ownership after read_blocks(),
&destroy_block); // so it needs create and destroy functions
diy::ContiguousAssigner assigner(world.size(), 0);
//diy::RoundRobinAssigner assigner(world.size(), 0); // nblocks will be filled by read_blocks()
diy::io::read_blocks("blocks.out", world, assigner, master, &load_block);
master.foreach(&output);
}
int main(int argc, char* argv[])
{
diy::mpi::environment env(argc, argv);
diy::mpi::communicator world;
int nblocks = 4*world.size();
diy::FileStorage storage("./DIY.XXXXXX");
diy::Master master(world,
-1,
2,
&create_block,
&destroy_block,
&storage,
&save_block,
&load_block);
srand(time(NULL));
//diy::ContiguousAssigner assigner(world.size(), nblocks);
diy::RoundRobinAssigner assigner(world.size(), nblocks);
for (int gid = 0; gid < nblocks; ++gid)
if (assigner.rank(gid) == world.rank())
master.add(gid, new Block, new diy::Link);
bool all_done = false;
while (!all_done)
{
master.foreach(&flip_coin);
master.exchange();
all_done = master.proxy(master.loaded_block()).read<bool>();
}
if (world.rank() == 0)
std::cout << "Total iterations: " << master.block<Block>(master.loaded_block())->count << std::endl;
}
/*!
* \param[in] data Data for the current frame.
* \param[in] sel Selection element being evaluated.
* \param[in] g Group for which \p sel should be evaluated.
* \returns 0 on success, a non-zero error code on error.
*
* Finds the part of \p g for which the subexpression
* has not yet been evaluated by comparing \p g to \p sel->u.cgrp.
* If the part is not empty, the child expression is evaluated for this
* part, and the results merged to the old values of the child.
* The value of \p sel itself is undefined after the call.
*
* \todo
* The call to gmx_ana_index_difference() can take quite a lot of unnecessary
* time if the subexpression is evaluated either several times for the same
* group or for completely distinct groups.
* However, in the majority of cases, these situations occur when
* _gmx_sel_evaluate_subexpr_staticeval() can be used, so this should not be a
* major problem.
*/
void
_gmx_sel_evaluate_subexpr(gmx_sel_evaluate_t *data, t_selelem *sel, gmx_ana_index_t *g)
{
gmx_ana_index_t gmiss;
MempoolGroupReserver gmissreserver(data->mp);
if (sel->u.cgrp.isize == 0)
{
{
SelelemTemporaryValueAssigner assigner(sel->child, sel);
sel->child->evaluate(data, sel->child, g);
}
/* We need to keep the name for the cgrp across the copy to avoid
* problems if g has a name set. */
char *name = sel->u.cgrp.name;
gmx_ana_index_copy(&sel->u.cgrp, g, false);
sel->u.cgrp.name = name;
gmiss.isize = 0;
}
else
{
gmissreserver.reserve(&gmiss, g->isize);
gmx_ana_index_difference(&gmiss, g, &sel->u.cgrp);
}
if (gmiss.isize > 0)
{
MempoolSelelemReserver reserver(sel->child, gmiss.isize);
/* Evaluate the missing values for the child */
sel->child->evaluate(data, sel->child, &gmiss);
/* Merge the missing values to the existing ones. */
if (sel->v.type == GROUP_VALUE)
{
gmx_ana_index_merge(sel->v.u.g, sel->child->v.u.g, sel->v.u.g);
}
else
{
int i, j, k;
i = sel->u.cgrp.isize - 1;
j = gmiss.isize - 1;
/* TODO: This switch is kind of ugly, but it may be difficult to
* do this portably without C++ templates. */
switch (sel->v.type)
{
case INT_VALUE:
for (k = sel->u.cgrp.isize + gmiss.isize - 1; k >= 0; k--)
{
if (i < 0 || (j >= 0 && sel->u.cgrp.index[i] < gmiss.index[j]))
{
sel->v.u.i[k] = sel->v.u.i[j--];
}
else
{
sel->v.u.i[k] = sel->child->v.u.i[i--];
}
}
break;
case REAL_VALUE:
for (k = sel->u.cgrp.isize + gmiss.isize - 1; k >= 0; k--)
{
if (i < 0 || (j >= 0 && sel->u.cgrp.index[i] < gmiss.index[j]))
{
sel->v.u.r[k] = sel->v.u.r[j--];
}
else
{
sel->v.u.r[k] = sel->child->v.u.r[i--];
}
}
break;
case STR_VALUE:
for (k = sel->u.cgrp.isize + gmiss.isize - 1; k >= 0; k--)
{
if (i < 0 || (j >= 0 && sel->u.cgrp.index[i] < gmiss.index[j]))
{
sel->v.u.s[k] = sel->v.u.s[j--];
}
else
{
sel->v.u.s[k] = sel->child->v.u.s[i--];
}
}
break;
case POS_VALUE:
/* TODO: Implement this */
GMX_THROW(gmx::NotImplementedError("position subexpressions not implemented properly"));
case NO_VALUE:
case GROUP_VALUE:
GMX_THROW(gmx::InternalError("Invalid subexpression type"));
}
}
gmx_ana_index_merge(&sel->u.cgrp, &sel->u.cgrp, &gmiss);
}
}
int main(int argc, char *argv[])
{
int tot_blocks; // total number of blocks in the domain
int mem_blocks; // max blocks in memory
int dsize[3]; // domain grid size
float jitter; // max amount to randomly displace particles
float minvol, maxvol; // volume range, -1.0 = unused
double times[TESS_MAX_TIMES]; // timing
int wrap; // wraparound neighbors flag
int walls; // apply walls to simulation (wrap must be off)
char outfile[256]; // output file name
int num_threads = 1; // threads diy can use
// init MPI
MPI_Comm comm = MPI_COMM_WORLD;
MPI_Init(&argc, &argv);
GetArgs(argc, argv, tot_blocks, mem_blocks, dsize, &jitter, &minvol, &maxvol, &wrap, &walls,
outfile);
// data extents
typedef diy::ContinuousBounds Bounds;
Bounds domain { 3 };
for(int i = 0; i < 3; i++)
{
domain.min[i] = 0;
domain.max[i] = dsize[i] - 1.0;
}
// init diy
diy::mpi::communicator world(comm);
diy::FileStorage storage("./DIY.XXXXXX");
diy::Master master(world,
num_threads,
mem_blocks,
&create_block,
&destroy_block,
&storage,
&save_block,
&load_block);
diy::RoundRobinAssigner assigner(world.size(), tot_blocks);
AddAndGenerate create(master, jitter);
// decompose
std::vector<int> my_gids;
assigner.local_gids(world.rank(), my_gids);
diy::RegularDecomposer<Bounds>::BoolVector wraps;
diy::RegularDecomposer<Bounds>::BoolVector share_face;
diy::RegularDecomposer<Bounds>::CoordinateVector ghosts;
if (wrap)
wraps.assign(3, true);
diy::decompose(3, world.rank(), domain, assigner, create, share_face, wraps, ghosts);
// tessellate
quants_t quants;
timing(times, -1, -1, world);
timing(times, TOT_TIME, -1, world);
tess(master, quants, times);
// output
tess_save(master, outfile, times);
timing(times, -1, TOT_TIME, world);
tess_stats(master, quants, times);
MPI_Finalize();
return 0;
}
