communicator_stack.cpp

Go to the documentation of this file.

#include <communicator_quda.h>
#include <map>
#include <array>
#include <lattice_field.h>
 
int Communicator::gpuid = -1;
 
static std::map<quda::CommKey, Communicator> communicator_stack;
 
static quda::CommKey current_key = {-1, -1, -1, -1};
 
void init_communicator_stack(int ndim, const int *dims, QudaCommsMap rank_from_coords, void *map_data,
                             bool user_set_comm_handle, void *user_comm)
{
  communicator_stack.emplace(
    std::piecewise_construct, std::forward_as_tuple(default_comm_key),
    std::forward_as_tuple(ndim, dims, rank_from_coords, map_data, user_set_comm_handle, user_comm));
 
  current_key = default_comm_key;
}
 
void finalize_communicator_stack() { communicator_stack.clear(); }
 
static Communicator &get_default_communicator()
{
  auto search = communicator_stack.find(default_comm_key);
  if (search == communicator_stack.end()) { errorQuda("Default communicator can't be found."); }
  return search->second;
}
 
Communicator &get_current_communicator()
{
  auto search = communicator_stack.find(current_key);
  if (search == communicator_stack.end()) { errorQuda("Current communicator can't be found."); }
  return search->second;
}
 
void push_communicator(const quda::CommKey &split_key)
{
  if (comm_nvshmem_enabled())
    errorQuda(
      "Split-grid is currently not supported with NVSHMEM. Please set QUDA_ENABLE_NVSHMEM=0 to disable NVSHMEM.");
  auto search = communicator_stack.find(split_key);
  if (search == communicator_stack.end()) {
    communicator_stack.emplace(std::piecewise_construct, std::forward_as_tuple(split_key),
                               std::forward_as_tuple(get_default_communicator(), split_key.data()));
  }
 
  quda::LatticeField::freeGhostBuffer(); // Destroy the (IPC) Comm buffers with the old communicator.
 
  current_key = split_key;
}
 
int comm_neighbor_rank(int dir, int dim) { return get_current_communicator().comm_neighbor_rank(dir, dim); }
 
int comm_dim(int dim) { return get_current_communicator().comm_dim(dim); }
 
int comm_coord(int dim) { return get_current_communicator().comm_coord(dim); }
 
int comm_rank_from_coords(const int *coords) { return get_current_communicator().comm_rank_from_coords(coords); }
 
void comm_init(int ndim, const int *dims, QudaCommsMap rank_from_coords, void *map_data, bool user_set_comm_handle,
               void *user_comm)
{
  init_communicator_stack(ndim, dims, rank_from_coords, map_data, user_set_comm_handle, user_comm);
}
 
void comm_finalize() { finalize_communicator_stack(); }
 
void comm_dim_partitioned_set(int dim) { get_current_communicator().comm_dim_partitioned_set(dim); }
 
void comm_dim_partitioned_reset() { get_current_communicator().comm_dim_partitioned_reset(); }
 
int comm_dim_partitioned(int dim) { return get_current_communicator().comm_dim_partitioned(dim); }
 
int comm_partitioned() { return get_current_communicator().comm_partitioned(); }
 
const char *comm_dim_topology_string() { return get_current_communicator().topology_string; }
 
const char *comm_config_string() { return get_current_communicator().comm_config_string(); }
 
const char *comm_dim_partitioned_string(const int *comm_dim_override)
{
  return get_current_communicator().comm_dim_partitioned_string(comm_dim_override);
}
 
int comm_rank(void) { return get_current_communicator().comm_rank(); }
 
int comm_rank_global(void) { return Communicator::comm_rank_global(); }
 
int comm_size(void) { return get_current_communicator().comm_size(); }
 
// XXX:
// Note here we are always using the **default** communicator.
// We might need to have a better approach.
int comm_gpuid(void) { return Communicator::comm_gpuid(); }
 
bool comm_deterministic_reduce() { return get_current_communicator().comm_deterministic_reduce(); }
 
void comm_gather_hostname(char *hostname_recv_buf)
{
  get_current_communicator().comm_gather_hostname(hostname_recv_buf);
}
 
void comm_gather_gpuid(int *gpuid_recv_buf) { get_current_communicator().comm_gather_gpuid(gpuid_recv_buf); }
 
void comm_peer2peer_init(const char *hostname_recv_buf)
{
  get_current_communicator().comm_peer2peer_init(hostname_recv_buf);
}
 
bool comm_peer2peer_present() { return get_current_communicator().comm_peer2peer_present(); }
 
int comm_peer2peer_enabled_global() { return get_current_communicator().comm_peer2peer_enabled_global(); }
 
bool comm_peer2peer_enabled(int dir, int dim) { return get_current_communicator().comm_peer2peer_enabled(dir, dim); }
 
void comm_enable_peer2peer(bool enable) { get_current_communicator().comm_enable_peer2peer(enable); }
 
bool comm_intranode_enabled(int dir, int dim) { return get_current_communicator().comm_intranode_enabled(dir, dim); }
 
void comm_enable_intranode(bool enable) { get_current_communicator().comm_enable_intranode(enable); }
 
bool comm_gdr_enabled() { return get_current_communicator().comm_gdr_enabled(); }
 
bool comm_gdr_blacklist() { return get_current_communicator().comm_gdr_blacklist(); }
 
bool comm_nvshmem_enabled() { return get_current_communicator().comm_nvshmem_enabled(); }
 
MsgHandle *comm_declare_send_rank(void *buffer, int rank, int tag, size_t nbytes)
{
  return get_current_communicator().comm_declare_send_rank(buffer, rank, tag, nbytes);
}
 
MsgHandle *comm_declare_recv_rank(void *buffer, int rank, int tag, size_t nbytes)
{
  return get_current_communicator().comm_declare_recv_rank(buffer, rank, tag, nbytes);
}
 
MsgHandle *comm_declare_send_displaced(void *buffer, const int displacement[], size_t nbytes)
{
  return get_current_communicator().comm_declare_send_displaced(buffer, displacement, nbytes);
}
 
MsgHandle *comm_declare_receive_displaced(void *buffer, const int displacement[], size_t nbytes)
{
  return get_current_communicator().comm_declare_receive_displaced(buffer, displacement, nbytes);
}
 
MsgHandle *comm_declare_strided_send_displaced(void *buffer, const int displacement[], size_t blksize, int nblocks,
                                               size_t stride)
{
  return get_current_communicator().comm_declare_strided_send_displaced(buffer, displacement, blksize, nblocks, stride);
}
 
MsgHandle *comm_declare_strided_receive_displaced(void *buffer, const int displacement[], size_t blksize, int nblocks,
                                                  size_t stride)
{
  return get_current_communicator().comm_declare_strided_receive_displaced(buffer, displacement, blksize, nblocks,
                                                                           stride);
}
 
void comm_free(MsgHandle *&mh) { get_current_communicator().comm_free(mh); }
 
void comm_start(MsgHandle *mh) { get_current_communicator().comm_start(mh); }
 
void comm_wait(MsgHandle *mh) { get_current_communicator().comm_wait(mh); }
 
int comm_query(MsgHandle *mh) { return get_current_communicator().comm_query(mh); }
 
void comm_allreduce(double *data) { get_current_communicator().comm_allreduce(data); }
 
void comm_allreduce_max(double *data) { get_current_communicator().comm_allreduce_max(data); }
 
void comm_allreduce_min(double *data) { get_current_communicator().comm_allreduce_min(data); }
 
void comm_allreduce_array(double *data, size_t size) { get_current_communicator().comm_allreduce_array(data, size); }
 
void comm_allreduce_max_array(double *data, size_t size)
{
  get_current_communicator().comm_allreduce_max_array(data, size);
}
 
void comm_allreduce_int(int *data) { get_current_communicator().comm_allreduce_int(data); }
 
void comm_allreduce_xor(uint64_t *data) { get_current_communicator().comm_allreduce_xor(data); }
 
void comm_broadcast(void *data, size_t nbytes) { get_current_communicator().comm_broadcast(data, nbytes); }
 
void comm_broadcast_global(void *data, size_t nbytes) { get_default_communicator().comm_broadcast(data, nbytes); }
 
void comm_barrier(void) { get_current_communicator().comm_barrier(); }
 
void comm_abort_(int status) { Communicator::comm_abort_(status); };
 
void reduceMaxDouble(double &max) { get_current_communicator().reduceMaxDouble(max); }
 
void reduceDouble(double &sum) { get_current_communicator().reduceDouble(sum); }
 
void reduceDoubleArray(double *max, const int len) { get_current_communicator().reduceDoubleArray(max, len); }
 
int commDim(int dim) { return get_current_communicator().commDim(dim); }
 
int commCoords(int dim) { return get_current_communicator().commCoords(dim); }
 
int commDimPartitioned(int dir) { return get_current_communicator().commDimPartitioned(dir); }
 
void commDimPartitionedSet(int dir) { get_current_communicator().commDimPartitionedSet(dir); }
 
void commDimPartitionedReset() { get_current_communicator().comm_dim_partitioned_reset(); }
 
bool commGlobalReduction() { return get_current_communicator().commGlobalReduction(); }
 
void commGlobalReductionSet(bool global_reduce) { get_current_communicator().commGlobalReductionSet(global_reduce); }
 
bool commAsyncReduction() { return get_current_communicator().commAsyncReduction(); }
 
void commAsyncReductionSet(bool global_reduce) { get_current_communicator().commAsyncReductionSet(global_reduce); }
 
int get_enable_p2p_max_access_rank() { return get_current_communicator().enable_p2p_max_access_rank; }