40 size_t blksize,
int nblocks,
size_t stride)
44 size_t blksize,
int nblocks,
size_t stride)
void comm_gather_hostname(char *hostname_recv_buf)
Gather all hostnames.
Topology * comm_create_topology(int ndim, const int *dims, QudaCommsMap rank_from_coords, void *map_data)
void comm_allreduce_max(double *data)
const char * comm_dim_topology_string()
Return a string that defines the comm topology (for use as a tuneKey)
MsgHandle * comm_declare_receive_displaced(void *buffer, const int displacement[], size_t nbytes)
int comm_query(MsgHandle *mh)
void comm_allreduce_int(int *data)
void comm_wait(MsgHandle *mh)
void comm_broadcast(void *data, size_t nbytes)
void exit(int) __attribute__((noreturn))
void comm_free(MsgHandle *mh)
const char * comm_dim_partitioned_string()
Return a string that defines the comm partitioning (used as a tuneKey)
char * comm_hostname(void)
void comm_abort(int status)
void comm_allreduce_xor(uint64_t *data)
void comm_gather_gpuid(int *gpuid_recv_buf)
Gather all GPU ids.
unsigned long long uint64_t
void comm_init(int ndim, const int *dims, QudaCommsMap rank_from_coords, void *map_data)
int(* QudaCommsMap)(const int *coords, void *fdata)
static char partition_string[16]
char * strncpy(char *__dst, const char *__src, size_t __n)
void comm_set_default_topology(Topology *topo)
MsgHandle * comm_declare_strided_receive_displaced(void *buffer, const int displacement[], size_t blksize, int nblocks, size_t stride)
void comm_start(MsgHandle *mh)
MsgHandle * comm_declare_send_displaced(void *buffer, const int displacement[], size_t nbytes)
static char topology_string[16]
void comm_allreduce(double *data)
void comm_allreduce_array(double *data, size_t size)
MsgHandle * comm_declare_strided_send_displaced(void *buffer, const int displacement[], size_t blksize, int nblocks, size_t stride)