Distributions

A distribution, also called a domain map, specifies the implementation of the domains and arrays that are mapped using it. That is, it defines how domain indices and array elements are mapped to locales, how they are stored in memory, and how operations such as accesses, iteration, and slicing are performed. Each domain and array is mapped using some distribution.

In general, a distribution describes domains and arrays that are partitioned across multiple locales. A layout is a distribution that describes domains and arrays that exist on a single locale.

A distribution is represented in the program with an instance of a distribution record. Chapel provides a set of standard distributions. Users can create distributions as well.

Distributions are presented as follows:

Distributions for Domain Types

Each domain type has a distribution associated with it. This distribution is used to map all domain values of this type (see Distributions for Domain Values).

If a domain type does not have a distribution specified for it explicitly as described below, a default distribution is provided by the Chapel implementation. Such a distribution will typically be a layout that maps the entire domain to the locale on which the current task is executed or the domain or array variable is declared.

A distribution can be specified explicitly using a dmapped clause:

mapped-domain-type:
  domain-type 'dmapped' distribution-record

distribution-record:
  expression

where the distribution-record expression produces an instance of a distribution record.

Warning

The dmapped keyword and the dmapped clause are currently unstable and may change in the future. Factory functions provided by the desired distribution, when available, should be used instead.

Example.

The code

use BlockDist;
var MyBlockDist = new blockDist({1..5,1..6});

creates an instance of the Block distribution record for two-dimensional domains and arrays with a bounding box of {1..5, 1..6} over all of the locales.

Example.

The code

use BlockDist;
var MyBlockDist = new blockDist({1..5,1..6});
type MyBlockedDom = domain(2) dmapped MyBlockDist;

defines the type of two-dimensional rectangular domains that are mapped using a Block distribution.

The following syntactic sugar is provided within the dmapped clause. If a dmapped clause starts with the name of a distribution record, it is considered to be an initialization expression as if preceded by new.

Example.

The code

use BlockDist;
type BlockDom = domain(2) dmapped blockDist({1..5,1..6});

is equivalent to

use BlockDist;
type BlockDom = domain(2) dmapped new blockDist({1..5,1..6});

Distributions for Domain Values

A domain value is always mapped using the distribution of that value’s type. The type inferred for a domain literal (see Rectangular Domain Values) has a default distribution.

Example.

In the following code

use BlockDist;
var MyDomLiteral = {1..2,1..3};
var MyBlockedDom: domain(2) dmapped blockDist({1..5,1..6}) = MyDomLiteral;

MyDomLiteral is a domain literal and so will be mapped using a default distribution. MyBlockedDom is given a type explicitly, therefore it will be mapped using a Block distribution.

A domain value’s distribution can be specified explicitly with a dmapped clause, in the same way as a domain type’s distribution.

mapped-domain-expression:
  domain-expression 'dmapped' distribution-record

Example.

In the following code

use BlockDist;
var MyBlockedDomLiteral1 = {1..2,1..3} dmapped new blockDist({1..5,1..6});
var MyBlockedDomLiteral2 = {1..2,1..3} dmapped blockDist({1..5,1..6});

both MyBlockedDomLiteral1 and MyBlockedDomLiteral2 will be mapped using a Block distribution.

Distributions for Arrays

Each array is mapped using the distribution of the domain over which the array was declared.

Example.

In the code

use BlockDist;
var Dom: domain(2) dmapped blockDist({1..5,1..6}) = {1..5,1..6};
var MyArray: [Dom] real;

the distribution used for MyArray is the Block distribution from the type of Dom.

Distributions Are Not Retained upon Domain Assignment

Domain assignment transfers only the index set of the right-hand side expression. The implementation of the left-hand side domain expression, including its distribution, is determined by its type and so does not change upon a domain assignment.

Example.

In the code

use BlockDist;
var Dom1: domain(2) dmapped blockDist({1..5,1..6}) = {1..5,1..6};
var Dom2: domain(2) = Dom1;

Dom2 is mapped using a default distribution, despite Dom1 having a Block distribution.

Example.

In the code

use BlockDist;
var Dom1: domain(2) dmapped blockDist({1..5,1..6}) = {1..5,1..6};
var Dom2 = Dom1;

Dom2 is mapped using the same distribution as Dom1. This is because the declaration of Dom2 lacks an explicit type specifier and so its type is defined to be the type of its initialization expression, Dom1. So in this situation the effect is that the distribution does transfer upon initialization.