Dune Core Modules (2.5.0)
Collective communication interface and sequential default implementation. More...
#include <dune/common/parallel/collectivecommunication.hh>
Public Member Functions | |
CollectiveCommunication () | |
Construct default object. | |
CollectiveCommunication (const Communicator &) | |
Constructor with a given communicator. More... | |
int | rank () const |
Return rank, is between 0 and size()-1. | |
int | size () const |
Number of processes in set, is greater than 0. | |
template<typename T > | |
T | sum (T &in) const |
Compute the sum of the argument over all processes and return the result in every process. Assumes that T has an operator+. | |
template<typename T > | |
int | sum (T *inout, int len) const |
Compute the sum over all processes for each component of an array and return the result in every process. Assumes that T has an operator+. More... | |
template<typename T > | |
T | prod (T &in) const |
Compute the product of the argument over all processes and return the result in every process. Assumes that T has an operator*. | |
template<typename T > | |
int | prod (T *inout, int len) const |
Compute the product over all processes for each component of an array and return the result in every process. Assumes that T has an operator*. More... | |
template<typename T > | |
T | min (T &in) const |
Compute the minimum of the argument over all processes and return the result in every process. Assumes that T has an operator<. | |
template<typename T > | |
int | min (T *inout, int len) const |
Compute the minimum over all processes for each component of an array and return the result in every process. Assumes that T has an operator<. More... | |
template<typename T > | |
T | max (T &in) const |
Compute the maximum of the argument over all processes and return the result in every process. Assumes that T has an operator<. | |
template<typename T > | |
int | max (T *inout, int len) const |
Compute the maximum over all processes for each component of an array and return the result in every process. Assumes that T has an operator<. More... | |
int | barrier () const |
Wait until all processes have arrived at this point in the program. More... | |
template<typename T > | |
int | broadcast (T *inout, int len, int root) const |
Distribute an array from the process with rank root to all other processes. More... | |
template<typename T > | |
int | gather (T *in, T *out, int len, int root) const |
Gather arrays on root task. More... | |
template<typename T > | |
int | gatherv (T *in, int sendlen, T *out, int *recvlen, int *displ, int root) const |
Gather arrays of variable size on root task. More... | |
template<typename T > | |
int | scatter (T *send, T *recv, int len, int root) const |
Scatter array from a root to all other task. More... | |
template<typename T > | |
int | scatterv (T *send, int *sendlen, int *displ, T *recv, int recvlen, int root) const |
Scatter arrays of variable length from a root to all other tasks. More... | |
template<typename T > | |
int | allgather (T *sbuf, int count, T *rbuf) const |
Gathers data from all tasks and distribute it to all. More... | |
template<typename T > | |
int | allgatherv (T *in, int sendlen, T *out, int *recvlen, int *displ) const |
Gathers data of variable length from all tasks and distribute it to all. More... | |
template<typename BinaryFunction , typename Type > | |
int | allreduce (Type *inout, int len) const |
Compute something over all processes for each component of an array and return the result in every process. More... | |
template<typename BinaryFunction , typename Type > | |
void | allreduce (Type *in, Type *out, int len) const |
Compute something over all processes for each component of an array and return the result in every process. More... | |
Detailed Description
class Dune::CollectiveCommunication< Communicator >
Collective communication interface and sequential default implementation.
CollectiveCommunication offers an abstraction to the basic methods of parallel communication, following the message-passing paradigm. It allows one to switch parallel features on and off, without changing the code. Currently only MPI and sequential code are supported.
A CollectiveCommunication object is returned by all grids (also the sequential ones) in order to allow code to be written in a transparent way for sequential and parallel grids.
This class provides a default implementation for sequential grids. The number of processes involved is 1, any sum, maximum, etc. returns just its input argument and so on.
In specializations one can implement the real thing using appropriate communication functions, e.g. there exists an implementation using the Message Passing Interface (MPI), see Dune::CollectiveCommunication<MPI_Comm>.
Moreover, the communication subsystem used by an implementation is not visible in the interface, i.e. Dune grid implementations are not restricted to MPI.
- Template Parameters
-
Communicator The communicator type used by your message-passing implementation. For MPI this will be MPI_Comm. For sequential codes there is the dummy communicator No_Comm. It is assumed that if you want to specialize the CollectiveCommunication class for a message-passing system other than MPI, that message-passing system will have something equivalent to MPI communicators.
Constructor & Destructor Documentation
◆ CollectiveCommunication()
|
inline |
Constructor with a given communicator.
As this is implementation for the sequential setting, the communicator is a dummy and simply discarded.
Member Function Documentation
◆ allgather()
|
inline |
Gathers data from all tasks and distribute it to all.
The block of data sent from the jth process is received by every process and placed in the jth block of the buffer recvbuf.
- Parameters
-
[in] sbuf The buffer with the data to send. Has to be the same for each task. [in] count The number of elements to send by any process. [out] rbuf The receive buffer for the data. Has to be of size notasks*count, with notasks being the number of tasks in the communicator.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ allgatherv()
|
inline |
Gathers data of variable length from all tasks and distribute it to all.
The block of data sent from the jth process is received by every process and placed in the jth block of the buffer out.
- Parameters
-
[in] in The send buffer with the data to send. [in] sendlen The number of elements to send on each task. [out] out The buffer to store the received data in. [in] recvlen An array with size equal to the number of processes containing the number of elements to recieve from process i at position i, i.e. the number that is passed as sendlen argument to this function in process i. [in] displ An array with size equal to the number of processes. Data recieved from process i will be written starting at out+displ[i].
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ allreduce() [1/2]
|
inline |
Compute something over all processes for each component of an array and return the result in every process.
The template parameter BinaryFunction is the type of the binary function to use for the computation
- Parameters
-
in The array to compute on. out The array to store the results in. len The number of components in the array
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ allreduce() [2/2]
|
inline |
Compute something over all processes for each component of an array and return the result in every process.
The template parameter BinaryFunction is the type of the binary function to use for the computation
- Parameters
-
inout The array to compute on. len The number of components in the array
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ barrier()
|
inline |
Wait until all processes have arrived at this point in the program.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ broadcast()
|
inline |
Distribute an array from the process with rank root to all other processes.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ gather()
|
inline |
Gather arrays on root task.
Each process sends its in array of length len to the root process (including the root itself). In the root process these arrays are stored in rank order in the out array which must have size len * number of processes.
- Parameters
-
[in] in The send buffer with the data to send. [out] out The buffer to store the received data in. Might have length zero on non-root tasks. [in] len The number of elements to send on each task. [in] root The root task that gathers the data.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ gatherv()
|
inline |
Gather arrays of variable size on root task.
Each process sends its in array of length sendlen to the root process (including the root itself). In the root process these arrays are stored in rank order in the out array.
- Parameters
-
[in] in The send buffer with the data to be sent [in] sendlen The number of elements to send on each task [out] out The buffer to store the received data in. May have length zero on non-root tasks. [in] recvlen An array with size equal to the number of processes containing the number of elements to receive from process i at position i, i.e. the number that is passed as sendlen argument to this function in process i. May have length zero on non-root tasks. [out] displ An array with size equal to the number of processes. Data received from process i will be written starting at out+displ[i] on the root process. May have length zero on non-root tasks. [in] root The root task that gathers the data.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ max()
|
inline |
Compute the maximum over all processes for each component of an array and return the result in every process. Assumes that T has an operator<.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ min()
|
inline |
Compute the minimum over all processes for each component of an array and return the result in every process. Assumes that T has an operator<.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ prod()
|
inline |
Compute the product over all processes for each component of an array and return the result in every process. Assumes that T has an operator*.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ scatter()
|
inline |
Scatter array from a root to all other task.
The root process sends the elements with index from k*len to (k+1)*len-1 in its array to task k, which stores it at index 0 to len-1.
- Parameters
-
[in] send The array to scatter. Might have length zero on non-root tasks. [out] recv The buffer to store the received data in. Upon completion of the method each task will have same data stored there as the one in send buffer of the root task before. [in] len The number of elements in the recv buffer. [in] root The root task that gathers the data.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ scatterv()
|
inline |
Scatter arrays of variable length from a root to all other tasks.
The root process sends the elements with index from send+displ[k] to send+displ[k]-1 in its array to task k, which stores it at index 0 to recvlen-1.
- Parameters
-
[in] send The array to scatter. May have length zero on non-root tasks. [in] sendlen An array with size equal to the number of processes containing the number of elements to scatter to process i at position i, i.e. the number that is passed as recvlen argument to this function in process i. [in] displ An array with size equal to the number of processes. Data scattered to process i will be read starting at send+displ[i] on root the process. [out] recv The buffer to store the received data in. Upon completion of the method each task will have the same data stored there as the one in send buffer of the root task before. [in] recvlen The number of elements in the recv buffer. [in] root The root task that gathers the data.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
◆ sum()
|
inline |
Compute the sum over all processes for each component of an array and return the result in every process. Assumes that T has an operator+.
- Returns
- MPI_SUCCESS (==0) if successful, an MPI error code otherwise
The documentation for this class was generated from the following file:
- dune/common/parallel/collectivecommunication.hh