def simple_broadcast_v5(x, root, myid, numprocs, buffer): """ Same as simple_broadcast_v4 except that it uses Pypar's bypass mode. The use of bypass mode implies that the programmer has to define his own buffers on the receiving side (same as when 'use_buffer' is True) and that he is limited to send numpy arrays and nothing else ! Hence, this function works only with numpy arrays. Input parameters: - x: data to broadcast. - root: rank of the process that initiates the broadcast. - myid: rank of the process calling the function. - buffer: well-dimensioned user-defined buffer for receiving 'x'. Return value: the broadcasted data. """ list = range(0, root) + range(root + 1, numprocs) # The process of rank 'root' sends 'x' to all the other processes. if root == myid: for i in list: # Determine automatically the size of bsend's buffer. pypar.push_for_alloc(x, bypass=True) # Allocate and attach bsend's buffer. pypar.alloc_and_attach() # Send data to process of rank 'i'. pypar.bsend(x, i, bypass=True) # Detach and deallocate bsend's buffer. pypar.detach_and_dealloc() buffer = x # All processes with rank distinct from 'root' start receiving. else: buffer = pypar.receive(root, buffer, bypass=True) return buffer
def simple_broadcast_v3(x, root, myid, numprocs): """ Alternative implementation to 'simple_broadcast_v1()'. Make all bsend calls using a single buffer and a single call to 'mpi_alloc_and_attach()' and 'mpi_detach_and_dealloc()'. This implementation illustrates how to use 'push_for_alloc()' in order to allocate a single buffer, big enough, to handle all sends within one call to 'mpi_alloc_and_attach()' and 'mpi_detach_and_dealloc()'. Input parameters: - x: data to broadcast. - root: rank of the process that initiates the broadcast. - myid: rank of the process calling the function. Return value: the broadcasted data. """ list = range(0, root) + range(root + 1, numprocs) # The process of rank 'root' sends 'x' to all other processes. if root == myid: # Determine the size of bsend's buffer. for i in list: pypar.push_for_alloc(x) # Allocate and attach bsend's buffer. pypar.alloc_and_attach() # Make all sends. for i in list: pypar.bsend(x, i) # Deallocate and detach bsend's buffer. pypar.detach_and_dealloc() x_ = x # All processes with rank distinct from 'root' start receiving. else: x_ = pypar.receive(root) return x_
def simple_broadcast_v1(x, root, myid, numprocs): """ Broadcast implementation using bsend() w/o explicitly defining receive buffers. Input parameters: - x: data to broadcast. - root: rank of the process that initiates the broadcast. - myid: rank of the process calling the function. Return value: the broadcasted data. """ list = range(0, root) + range(root + 1, numprocs) # The process of rank 'root' sends 'x' to all the other processes. if root == myid: for i in list: # Determine automatically the size of bsend's buffer. pypar.push_for_alloc(x) # Allocate and attach bsend's buffer. pypar.alloc_and_attach() # Send data to process of rank 'i'. pypar.bsend(x, i) # Detach and deallocate bsend's buffer. pypar.detach_and_dealloc() x_ = x # All processes with rank distinct from 'root' start receiving. else: x_ = pypar.receive(root) return x_
def simple_broadcast_v2(x, root, myid, numprocs): """ Optimized version of 'simple_broadcast_v1()'. Since every bsend needed to achieve the broadcast requires a buffer of the same size, it is sufficient to allocate that buffer once and to use it for all bsend calls. This illustrates the use of functions alloc(), attach(), dealloc() detach() instead of mpi_alloc_and_attach() and mpi_detach_and_dealloc(). As it can be seen the former alternatives give a finer control. Note: a call to ... alloc_and_attach() ... is equivalent to: ... alloc() attach() ... Input parameters: - x: data to broadcast. - root: rank of the process that initiates the broadcast. - myid: rank of the process calling the function. Return value: the broadcasted data. """ list = range(0, root) + range(root + 1, numprocs) # The process of rank 'root' sends 'x' to all the other processes. if root == myid: # Determine automatically the size of bsend's buffer. pypar.push_for_alloc(x) # Allocate bsend's buffer. pypar.alloc() for i in list: # Attach bsend's buffer. pypar.attach() # Send data to process of rank 'i'. pypar.bsend(x, i) # Detach bsend's buffer. pypar.detach() x_ = x # Deallocate bsend's buffer. pypar.dealloc() # All processes with rank distinct from 'root' start receiving. else: x_ = pypar.receive(root) return x_
