6.1.2. Datatype Constructors

Contiguous. The simplest datatype constructor is MPI_TYPE_CONTIGUOUS, which allows replication of a datatype into contiguous locations.

newtype is the datatype obtained by concatenating count copies of oldtype. Concatenation is defined using extent as the size of the concatenated copies.

Example Let oldtype have type map { ( double, 0), ( char, 8) } , with extent 16, and let count = 3. The type map of the datatype returned by newtype is

{ ( double, 0), ( char, 8), ( double, 16), ( char, 24), ( double, 32), ( char, 40) };

i.e., alternating double and char elements, with displacements 0, 8, 16, 24, 32, 40.

In general, assume that the type map of oldtype is

{ ( type₀, disp₀), ... , ( type_n-1, disp_n-1) } ,

with extent ex. Then newtype has a type map with count · n entries defined by:

Vector. The procedure MPI_TYPE_VECTOR is a more general constructor that allows replication of a datatype into locations that consist of equally spaced blocks. Each block is obtained by concatenating the same number of copies of the old datatype. The spacing between blocks is a multiple of the extent of the old datatype.

Example Assume, again, that oldtype has type map { ( double, 0), ( char, 8) } , with extent 16. A call to MPI_TYPE_VECTOR (2, 3, 4, oldtype, newtype) will create the datatype with type map,

{ ( double, 0), ( char, 8), ( double, 16), ( char, 24), ( double, 32), ( char, 40),

( double, 64), ( char, 72), ( double, 80), ( char, 88), ( double, 96), ( char, 104) } .

That is, two blocks with three copies each of the old type, with a stride of 4 elements (4 · 16 bytes) between the start of each block.

Example A call to MPI_TYPE_VECTOR (3, 1, -2, oldtype, newtype) will create the datatype,

{ ( double, 0), ( char, 8), ( double, -32), ( char, -24), ( double, -64), ( char, -56) } .

In general, assume that oldtype has type map,

{ ( type₀, disp₀), ..., ( type_n-1, disp_n-1) } ,

with extent ex. Let bl be the blocklength. The newly created datatype has a type map with count · bl · n

entries:

{ ( type₀, disp₀), ... , ( type_n-1 , disp_n-1),

( type₀ , disp₀ + ex) , ... , ( type_n-1 , disp_n-1 + ex ), ...,

( type₀ , disp₀ + ( bl -1) · ex ) , ... , ( type_n-1 , disp_n-1 + ( bl -1) · ex ) ,

( type₀ , disp₀ + stride · ex ) , ... , ( type_n-1 , disp_n-1 + stride · ex ), ... ,

( type₀ , disp₀ + ( stride + bl -1) · ex ) , ... , ( type_n-1, disp_n-1 + ( stride + bl -1) · ex ) , ... ,

( type₀ , disp₀ + stride · ( count-1) · ex ) , ... ,

( type_n-1 , disp_n-1 + stride · ( count -1) · ex ) , ... ,

( type₀ , disp₀ + ( stride · ( count -1) + bl -1) · ex ) , ... ,

( type_n-1, disp_n-1 + ( stride · ( count -1) + bl -1) · ex ) } .

A call to MPI_TYPE_CONTIGUOUS (count, oldtype, newtype) is equivalent to a call to MPI_TYPE_VECTOR (count, 1, 1, oldtype, newtype), or to a call to MPI_TYPE_VECTOR (1, count, n, oldtype, newtype), where n is an arbitrary integer value.

Hvector. The procedure MPI_TYPE_CREATE_HVECTOR is identical to MPI_TYPE_VECTOR, except that stride is given in bytes, rather than in elements. The use for both types of vector constructors is illustrated in Section Examples. ( H stands for ``heterogeneous'').

Assume that oldtype has type map,

{ ( type₀, disp₀), ..., ( type_n-1, disp_n-1) } ,

with extent ex. Let bl be the blocklength. The newly created datatype has a type map with count · bl · n

entries:

{ ( type₀, disp₀), ... , ( type_n-1 , disp_n-1),

( type₀ , disp₀ + ex) , ... , ( type_n-1 , disp_n-1 + ex ), ...,

( type₀ , disp₀ + ( bl -1) · ex ) , ... , ( type_n-1 , disp_n-1 + ( bl -1) · ex ) ,

( type₀ , disp₀ + stride ) , ... , ( type_n-1 , disp_n-1 + stride ) , ... ,

( type₀ , disp₀ + stride + ( bl -1) · ex ) , ... ,

( type_n-1, disp_n-1 + stride + ( bl -1) · ex ) , ... ,

( type₀ , disp₀ + stride · ( count-1) ) , ... , ( type_n-1 , disp_n-1 + stride · ( count -1) ) , ... ,

( type₀ , disp₀ + stride · ( count -1) + ( bl -1) · ex ) , ... ,

( type_n-1, disp_n-1 + stride · ( count -1) + ( bl -1) · ex ) } .

Indexed. The procedure MPI_TYPE_INDEXED allows replication of an old datatype into a sequence of blocks (each block is a concatenation of the old datatype), where each block can contain a different number of copies and have a different displacement. All block displacements are multiples of the old type extent.

Example Let oldtype have type map { ( double, 0), ( char, 8) } , with extent 16. Let B = (3, 1) and let D = (4, 0). A call to MPI_TYPE_INDEXED (2, B, D, oldtype, newtype) returns a datatype with type map,

{ ( double, 64), ( char, 72), ( double, 80), ( char, 88), ( double, 96), ( char, 104),

( double, 0), ( char, 8) } .

That is, three copies of the old type starting at displacement 64, and one copy starting at displacement 0.

In general, assume that oldtype has type map,

{ ( type₀, disp₀), ..., ( type_n-1, disp_n-1) } ,

with extent ex. Let B be the array_of_blocklengths argument and D be the array_of_displacements argument. The newly created datatype has n · ∑_i=0 ^count-1 B[i] entries:

{ ( type₀, disp₀ + D[0] · ex ) , ... , ( type_n-1 , disp_n-1 + D[0] · ex ) , ... ,

( type₀ , disp₀ + ( D[0] + B[0] -1) · ex) ,...,

( type_n-1 , disp_n-1 + ( D[0] + B[0] -1) · ex ) , ...,

( type₀, disp₀ + D[count-1] · ex ) , ... , ( type_n-1 , disp_n-1 + D[count-1] · ex ) , ... ,

( type₀ , disp₀ + ( D[count-1] + B[count-1] -1) · ex) ,...,

( type_n-1 , disp_n-1 + ( D[count-1] + B[count-1] -1) · ex ) } .

A call to MPI_TYPE_VECTOR (count, blocklength, stride, oldtype, newtype) is equivalent to a call to MPI_TYPE_INDEXED (count, B, D, oldtype, newtype) where

and

Hindexed. The procedure MPI_TYPE_CREATE_HINDEXED is identical to MPI_TYPE_INDEXED, except that block displacements in array_of_displacements are specified in bytes, rather than in multiples of the oldtype extent.

Assume that oldtype has type map,

{ ( type₀, disp₀), ..., ( type_n-1, disp_n-1) } ,

with extent ex. Let B be the array_of_blocklengths argument and D be the array_of_displacements argument. The newly created datatype has a type map with n · ∑_i=0 ^count-1 B[i] entries:

{ ( type₀, disp₀ + D[0] ) , ... , ( type_n-1 , disp_n-1 + D[0] ) , ... ,

( type₀ , disp₀ + D[0] +( B[0] -1) · ex) ,...,

( type_n-1 , disp_n-1 + D[0] +( B[0] -1) · ex ) , ...,

( type₀, disp₀ + D[count-1] ) , ... , ( type_n-1 , disp_n-1 + D[count-1] ) , ... ,

( type₀ , disp₀ + D[count-1] +( B[count-1] -1) · ex) ,...,

( type_n-1 , disp_n-1 + D[count-1] +( B[count-1] -1) · ex ) } .

Indexed_block. This procedure is the same as MPI_TYPE_INDEXED except that the blocklength is the same for all blocks. There are many codes using indirect addressing arising from unstructured grids where the blocksize is always 1 (gather/scatter). The following convenience procedure allows for constant blocksize and arbitrary displacements.

Hindexed_block. The procedure MPI_TYPE_CREATE_HINDEXED_BLOCK is identical to MPI_TYPE_CREATE_INDEXED_BLOCK, except that block displacements in array_of_displacements are specified in bytes, rather than in multiples of the oldtype extent.

Struct. MPI_TYPE_CREATE_STRUCT is the most general type constructor. It further generalizes MPI_TYPE_CREATE_HINDEXED in that it allows each block to consist of replications of different datatypes.

Example Let type1 have type map,

{ ( double, 0), ( char, 8) } ,

with extent 16. Let B = (2, 1, 3), D = (0, 16, 26), and T = (MPI_FLOAT, type1, MPI_CHAR). Then a call to MPI_TYPE_CREATE_STRUCT(3, B, D, T, newtype) returns a datatype with type map,

{ ( float, 0), ( float, 4), ( double, 16), ( char, 24), ( char, 26), ( char, 27), ( char, 28) } .

That is, two copies of MPI_FLOAT starting at 0, followed by one copy of type1 starting at 16, followed by three copies of MPI_CHAR, starting at 26. In this example, we assume that a float occupies four bytes.

In general, let T be the array_of_types argument, where T[i] is a handle to,

typemap_i = { ( type₀ⁱ , disp₀ⁱ ) , ... , ( type_ni-1ⁱ , disp_ni-1ⁱ ) } ,

with extent ex_i. Let B be the array_of_blocklength argument and D be the array_of_displacements argument.Let c be the count argument. Then the newly created datatype has a type map with ∑_i=0 ^c-1 B[i] · n_i

entries:

{ ( type₀⁰ , disp₀⁰ + D[0]) , ... , ( type_n0⁰ , disp_n0⁰ + D[0] ) , ... ,

( type₀⁰ , disp₀⁰ + D[0] + ( B[0]-1) · ex₀ ) , ... , ( type_n0⁰ , disp_n0⁰ + D[0] + ( B[0]-1) · ex₀ ) , ... ,

( type₀ ^c-1 , disp₀ ^c-1 + D[c-1]) , ... , ( type_{n c-1-1} ^c-1 , disp_{n c-1-1} ^c-1 + D[c-1] ) , ... ,

( type₀ ^c-1 , disp₀ ^c-1 + D[c-1] + ( B[c-1]-1) · ex _c-1 ) , ... ,

( type_{n c-1-1} ^c-1 , disp_{n c-1-1} ^c-1 + D[c-1] + ( B[c-1]-1) · ex _c-1 ) } .

A call to MPI_TYPE_CREATE_HINDEXED (count, B, D, oldtype, newtype) is equivalent to a call to MPI_TYPE_CREATE_STRUCT (count, B, D, T, newtype), where each entry of T is equal to oldtype.