.TH Constants 3 "11/12/2019" " " "MPI"

.SH NAME

Constants \-  Meaning of MPI's defined constants 

.SH DATA TYPES

Note that the Fortran types should only be used in Fortran programs,

and the C types should only be used in C programs.  For example,

it is in error to use 

.I MPI_INT

for a Fortran INTEGER.

Datatypes are of type 

.I MPI_Datatype

in C, type 

.I INTEGER

in Fortran,

and 

.I Type(MPI_Datatype)

in Fortran08

.SH C DATATYPES

.PD 0

.TP

.B MPI_CHAR 

- char

.PD 1

.PD 0

.TP

.B MPI_SIGNED_CHAR 

- signed char

.PD 1

.PD 0

.TP

.B MPI_UNSIGNED_CHAR 

- unsigned char

.PD 1

.PD 0

.TP

.B MPI_BYTE 

- See standard; like unsigned char

.PD 1

.PD 0

.TP

.B MPI_WCHAR 

- wide character (wchar_t)

.PD 1

.PD 0

.TP

.B MPI_SHORT 

- short

.PD 1

.PD 0

.TP

.B MPI_UNSIGNED_SHORT 

- unsigned short

.PD 1

.PD 0

.TP

.B MPI_INT   

- int

.PD 1

.PD 0

.TP

.B MPI_UNSIGNED 

- unsigned int

.PD 1

.PD 0

.TP

.B MPI_LONG 

- long

.PD 1

.PD 0

.TP

.B MPI_UNSIGNED_LONG 

- unsigned long

.PD 1

.PD 0

.TP

.B MPI_LONG_LONG_INT 

- long long

.PD 1

.PD 0

.TP

.B MPI_LONG_LONG 

- synonyn for 

.I MPI_LONG_LONG_INT

.PD 1

.PD 0

.TP

.B MPI_UNSIGNED_LONG_LONG 

- unsigned long long

.PD 1

.PD 0

.TP

.B MPI_FLOAT 

- float

.PD 1

.PD 0

.TP

.B MPI_DOUBLE 

- double

.PD 1

.PD 0

.TP

.B MPI_LONG_DOUBLE  

- long double (some systems may not implement this)

.PD 1

.PD 0

.TP

.B MPI_INT8_T  

- int8_t

.PD 1

.PD 0

.TP

.B MPI_INT16_T 

- int16_t

.PD 1

.PD 0

.TP

.B MPI_INT32_T 

- int32_t

.PD 1

.PD 0

.TP

.B MPI_INT64_T 

- int64_t

.PD 1

.PD 0

.TP

.B MPI_UINT8_T  

- uint8_t

.PD 1

.PD 0

.TP

.B MPI_UINT16_T 

- uint16_t

.PD 1

.PD 0

.TP

.B MPI_UINT32_T 

- uint32_t

.PD 1

.PD 0

.TP

.B MPI_UINT64_T 

- uint64_t

.PD 1

.PD 0

.TP

.B MPI_C_BOOL 

- _Bool

.PD 1

.PD 0

.TP

.B MPI_C_FLOAT_COMPLEX 

- float _Complex

.PD 1

.PD 0

.TP

.B MPI_C_COMPLEX 

- float _Complex

.PD 1

.PD 0

.TP

.B MPI_C_DOUBLE_COMPLEX 

- double _Complex

.PD 1

.PD 0

.TP

.B MPI_C_LONG_DOUBLE_COMPLEX 

- long double _Complex

.PD 1

The following are datatypes for the MPI functions 

.I MPI_MAXLOC

and

.I MPI_MINLOC

\&.

.PD 0

.TP

.B MPI_FLOAT_INT 

- 

.I struct { float, int }

.PD 1

.PD 0

.TP

.B MPI_LONG_INT  

- 

.I struct { long, int }

.PD 1

.PD 0

.TP

.B MPI_DOUBLE_INT 

- 

.I struct { double, int }

.PD 1

.PD 0

.TP

.B MPI_SHORT_INT  

- 

.I struct { short, int }

.PD 1

.PD 0

.TP

.B MPI_2INT       

- 

.I struct { int, int }

.PD 1

.PD 0

.TP

.B MPI_LONG_DOUBLE_INT 

- 

.I struct { long double, int }

; this

is an 

.B optional

type, and may be set to 

.I MPI_DATATYPE_NULL

.PD 1

Special datatypes for C and Fortran

.PD 0

.TP

.B MPI_PACKED 

- For 

.I MPI_Pack

and 

.I MPI_Unpack

.PD 1

.PD 0

.TP

.B MPI_UB 

- For 

.I MPI_Type_struct

; an upper-bound indicator.  Removed in MPI 3

.PD 1

.PD 0

.TP

.B MPI_LB 

- For 

.I MPI_Type_struct

; a lower-bound indicator. Removed in MPI 3

.PD 1

.SH FORTRAN DATATYPES

.PD 0

.TP

.B MPI_REAL 

- 

.I REAL

.PD 1

.PD 0

.TP

.B MPI_INTEGER 

- 

.I INTEGER

.PD 1

.PD 0

.TP

.B MPI_LOGICAL 

- 

.I LOGICAL

.PD 1

.PD 0

.TP

.B MPI_DOUBLE_PRECISION 

- 

.I DOUBLE PRECISION

.PD 1

.PD 0

.TP

.B MPI_COMPLEX 

- 

.I COMPLEX

.PD 1

.PD 0

.TP

.B MPI_DOUBLE_COMPLEX 

- 

.I complex*16

(or 

.I complex*32

) where supported.

.PD 1

The following datatypes are optional

.PD 0

.TP

.B MPI_INTEGER1  

- 

.I integer*1

if supported

.PD 1

.PD 0

.TP

.B MPI_INTEGER2  

- 

.I integer*2

if supported

.PD 1

.PD 0

.TP

.B MPI_INTEGER4  

- 

.I integer*4

if supported

.PD 1

.PD 0

.TP

.B MPI_INTEGER8  

- 

.I integer*8

if supported

.PD 1

.PD 0

.TP

.B MPI_INTEGER16 

- 

.I integer*16

if supported

.PD 1

.PD 0

.TP

.B MPI_REAL4     

- 

.I real*4

if supported

.PD 1

.PD 0

.TP

.B MPI_REAL8     

- 

.I real*8

if supported

.PD 1

.PD 0

.TP

.B MPI_REAL16    

- 

.I real*16

if supported

.PD 1

.PD 0

.TP

.B MPI_COMPLEX8  

- 

.I complex*8

if supported

.PD 1

.PD 0

.TP

.B MPI_COMPLEX16 

- 

.I complex*16

if supported

.PD 1

.PD 0

.TP

.B MPI_COMPLEX32 

- 

.I complex*32

if supported

.PD 1

The following are datatypes for the MPI functions 

.I MPI_MAXLOC

and

.I MPI_MINLOC

\&.

In Fortran, these datatype always consist of

two elements of the same Fortran type.

.PD 0

.TP

.B MPI_2INTEGER 

- 

.I INTEGER,INTEGER

.PD 1

.PD 0

.TP

.B MPI_2REAL    

- 

.I REAL, REAL

.PD 1

.PD 0

.TP

.B MPI_2DOUBLE_PRECISION 

- 

.I DOUBLE PRECISION, DOUBLE PRECISION

.PD 1

MPI Datatypes for MPI Types

.PD 0

.TP

.B MPI_AINT 

- Datatype for an 

.I MPI_Aint

.PD 1

.PD 0

.TP

.B MPI_OFFSET 

- Datatype for an 

.I MPI_Offset

.PD 1

.PD 0

.TP

.B MPI_COUNT 

- Datatype for an 

.I MPI_Count

.PD 1

.SH MPI DATATYPE COMBINER NAMES

.PD 0

.TP

.B MPI_COMBINER_NAMED            

- a named predefined datatype

.PD 1

.PD 0

.TP

.B MPI_COMBINER_DUP              

- MPI_TYPE_DUP

.PD 1

.PD 0

.TP

.B MPI_COMBINER_CONTIGUOUS       

- MPI_TYPE_CONTIGUOUS

.PD 1

.PD 0

.TP

.B MPI_COMBINER_VECTOR           

- MPI_TYPE_VECTOR

.PD 1

.PD 0

.TP

.B MPI_COMBINER_HVECTOR_INTEGER  

- Removed in MPI-3

.PD 1

.PD 0

.TP

.B MPI_COMBINER_HVECTOR          

- MPI_TYPE_CREATE_HVECTOR

.PD 1

.PD 0

.TP

.B MPI_COMBINER_INDEXED          

- MPI_TYPE_INDEXED

.PD 1

.PD 0

.TP

.B MPI_COMBINER_HINDEXED_INTEGER 

- Removed in MPI-3

.PD 1

.PD 0

.TP

.B MPI_COMBINER_HINDEXED         

- MPI_TYPE_CREATE_HINDEXED

.PD 1

.PD 0

.TP

.B MPI_COMBINER_INDEXED_BLOCK    

- MPI_TYPE_CREATE_INDEXED_BLOCK

.PD 1

.PD 0

.TP

.B MPI_COMBINER_STRUCT_INTEGER   

- Removed in MPI-3

.PD 1

.PD 0

.TP

.B MPI_COMBINER_STRUCT           

- MPI_TYPE_CREATE_STRUCT

.PD 1

.PD 0

.TP

.B MPI_COMBINER_SUBARRAY         

- MPI_TYPE_CREATE_SUBARRAY

.PD 1

.PD 0

.TP

.B MPI_COMBINER_DARRAY           

- MPI_TYPE_CREATE_DARRAY

.PD 1

.PD 0

.TP

.B MPI_COMBINER_F90_REAL         

- MPI_TYPE_CREATE_F90_REAL

.PD 1

.PD 0

.TP

.B MPI_COMBINER_F90_COMPLEX      

- MPI_TYPE_CREATE_F90_COMPLEX

.PD 1

.PD 0

.TP

.B MPI_COMBINER_F90_INTEGER      

- MPI_TYPE_CREATE_F90_INTEGER

.PD 1

.PD 0

.TP

.B MPI_COMBINER_RESIZED          

- MPI_TYPE_CREATE_RESIZED

.PD 1

.PD 0

.TP

.B MPI_COMBINER_HINDEXED_BLOCK   

- MPI_TYPE_CREATE_HINDEXED_BLOCK

.PD 1

.SH MPI DATATYPE TYPE CLASSES

MPI Type classes used with routines to return Fortran types with defined

precision and range

.PD 0

.TP

.B MPI_TYPECLASS_REAL    

- 

.I REAL

.PD 1

.PD 0

.TP

.B MPI_TYPECLASS_INTEGER 

- 

.I INTEGER

.PD 1

.PD 0

.TP

.B MPI_TYPECLASS_COMPLEX 

- 

.I COMPLEX

.PD 1

.SH MPI DARRAY AND SUBARRAY VALUES

These values are used to create a datatype with the 

.I DARRAY

and 

.I SUBARRAY

constructors.

.PD 0

.TP

.B MPI_ORDER_C              

- Row-major order (as used by C)

.PD 1

.PD 0

.TP

.B MPI_ORDER_FORTRAN        

- Column-major order (as used by Fortran)

.PD 1

.PD 0

.TP

.B MPI_DISTRIBUTE_BLOCK     

- Block distribution

.PD 1

.PD 0

.TP

.B MPI_DISTRIBUTE_CYCLIC    

- Cyclic distribution

.PD 1

.PD 0

.TP

.B MPI_DISTRIBUTE_NONE      

- This dimension is not distributed

.PD 1

.PD 0

.TP

.B MPI_DISTRIBUTE_DFLT_DARG 

- Use the default distribution

.PD 1

.SH COMMUNICATORS

Communicators are of type 

.I MPI_Comm

in C, 

.I INTEGER

in Fortran, and

.I Type(MPI_Comm)

in Fortran08

.PD 0

.TP

.B MPI_COMM_WORLD 

- Contains all of the processes

.PD 1

.PD 0

.TP

.B MPI_COMM_SELF 

- Contains only the calling process

.PD 1

.SH KIND OF COMMUNICATOR FOR 'MPI_COMM_SPLIT_TYPE'

.PD 0

.TP

.B MPI_COMM_TYPE_SHARED 

- All processes that can share memory are grouped into

the same communicator.

.PD 1

.SH GROUPS

Groups are of type 

.I MPI_Group

in C, 

.I INTEGER

in Fortran,

and 

.I Type(MPI_Group)

in Fortran08

.PD 0

.TP

.B MPI_GROUP_EMPTY 

- A group containing no members.

.PD 1

.SH RESULTS OF THE COMPARE OPERATIONS ON GROUPS AND COMMUNICATORS

.PD 0

.TP

.B MPI_IDENT 

- Identical

.PD 1

.PD 0

.TP

.B MPI_CONGRUENT  

- (only for 

.I MPI_COMM_COMPARE

) The groups are identical

.PD 1

.PD 0

.TP

.B MPI_SIMILAR 

- Same members, but in a different order

.PD 1

.PD 0

.TP

.B MPI_UNEQUAL 

- Different

.PD 1

.SH COLLECTIVE OPERATIONS

The collective combination operations (e.g., 

.I MPI_REDUCE

, 

.I MPI_ALLREDUCE

,

.I MPI_REDUCE_SCATTER

, and 

.I MPI_SCAN

) take a combination operation.

This operation is of type 

.I MPI_Op

in C and of type 

.I INTEGER

in Fortran.

The predefined operations are

.PD 0

.TP

.B MPI_MAX 

- return the maximum

.PD 1

.PD 0

.TP

.B MPI_MIN 

- return the minumum

.PD 1

.PD 0

.TP

.B MPI_SUM 

- return the sum

.PD 1

.PD 0

.TP

.B MPI_PROD 

- return the product

.PD 1

.PD 0

.TP

.B MPI_LAND 

- return the logical and

.PD 1

.PD 0

.TP

.B MPI_BAND 

- return the bitwise and

.PD 1

.PD 0

.TP

.B MPI_LOR 

- return the logical or

.PD 1

.PD 0

.TP

.B MPI_BOR 

- return the bitwise of

.PD 1

.PD 0

.TP

.B MPI_LXOR 

- return the logical exclusive or

.PD 1

.PD 0

.TP

.B MPI_BXOR 

- return the bitwise exclusive or

.PD 1

.PD 0

.TP

.B MPI_MINLOC 

- return the minimum and the location (actually, the value of

the second element of the structure where the minimum of

the first is found)

.PD 1

.PD 0

.TP

.B MPI_MAXLOC 

- return the maximum and the location

.PD 1

.PD 0

.TP

.B MPI_REPLACE 

- replace b with a

.PD 1

.PD 0

.TP

.B MPI_NO_OP 

- perform no operation

.PD 1

.SH NOTES ON COLLECTIVE OPERATIONS

The reduction functions (

.I MPI_Op

) do not return an error value.  As a result,

if the functions detect an error, all they can do is either call 

.I MPI_Abort

or silently skip the problem.  Thus, if you change the error handler from

.I MPI_ERRORS_ARE_FATAL

to something else, for example, 

.I MPI_ERRORS_RETURN

,

then no error may be indicated.

The reason for this is the performance problems in ensuring that

all collective routines return the same error value.

Note that not all datatypes are valid for these functions.  For example,

.I MPI_COMPLEX

is not valid for 

.I MPI_MAX

and 

.I MPI_MIN

\&.

In addition, the MPI

1.1 standard did not include the C types 

.I MPI_CHAR

and 

.I MPI_UNSIGNED_CHAR

among the lists of arithmetic types for operations like 

.I MPI_SUM

\&.

However,

since the C type 

.I char

is an integer type (like 

.I short

), it should have been

included.  The MPI Forum will probably include 

.I char

and 

.I unsigned char

as a clarification to MPI 1.1; until then, users are advised that MPI

implementations may not accept 

.I MPI_CHAR

and 

.I MPI_UNSIGNED_CHAR

as valid

datatypes for 

.I MPI_SUM

, 

.I MPI_PROD

, etc.  MPICH does allow these datatypes.

.SH PERMANENT KEY VALUES

These are the same in C and Fortran

.PD 0

.TP

.B MPI_TAG_UB 

- Largest tag value

.PD 1

.PD 0

.TP

.B MPI_HOST 

- Rank of process that is host, if any

.PD 1

.PD 0

.TP

.B MPI_IO 

- Rank of process that can do I/O

.PD 1

.PD 0

.TP

.B MPI_WTIME_IS_GLOBAL 

- Has value 1 if 

.I MPI_WTIME

is globally synchronized.

.PD 1

.PD 0

.TP

.B MPI_UNIVERSE_SIZE 

- Number of available processes.  See the standard for

a description of limitations on this value

.PD 1

.PD 0

.TP

.B MPI_LASTUSEDCODE 

- Last used MPI error code (check - code or class?)

.PD 1

.PD 0

.TP

.B MPI_APPNUM 

- Application number, starting from 0.  See the standard for

.I MPI_COMM_SPAWN_MULTIPLE

and 

.I mpiexec

for details

.PD 1

.SH NULL OBJECTS

.PD 0

.TP

.B MPI_COMM_NULL          

- Null communicator

.PD 1

.PD 0

.TP

.B MPI_OP_NULL            

- Null operation

.PD 1

.PD 0

.TP

.B MPI_GROUP_NULL         

- Null group

.PD 1

.PD 0

.TP

.B MPI_DATATYPE_NULL      

- Null datatype

.PD 1

.PD 0

.TP

.B MPI_REQUEST_NULL       

- Null request

.PD 1

.PD 0

.TP

.B MPI_ERRHANDLER_NULL    

- Null error handler

.PD 1

.PD 0

.TP