Skip to content

input-casting-policies

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

input-casting-policies

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
benchmark
benchmark
 
 
docs
docs
 
 
examples
examples
 
 
include/stdlib/ndarray
include/stdlib/ndarray
 
 
lib
lib
 
 
test
test
 
 
README.md
README.md
 
 
manifest.json
manifest.json
 
 
package.json
package.json
 
 

README.md

Policies

List of input ndarray casting policies.

Usage

var policies = require( '@stdlib/ndarray/input-casting-policies' );

policies()

Returns a list of input ndarray casting policies.

var out = policies();
// e.g., returns [ 'none', 'promoted', ... ]

The output array contains the following policies:

  • none: no guidance on specific casting behavior. An input ndarray may or may not be cast depending on the needs of an implementation.
  • promoted: cast an input ndarray to a promoted data type.
  • accumulation: cast an input ndarray to a data type amenable to accumulation.
  • output: cast an input ndarray to the data type of the output ndarray.

Notes

  • The following is some general guidance for the casting policies listed above:

    • none: applies whenever an input ndarray casting behavior should be entirely left up to an implementation. For example, an implementation may choose to cast internally in order to take advantage of specialized algorithms operating on specific data types.
    • promoted: applies whenever an input ndarray should be cast to the data type resolved from applying the rules of type promotion to an implementation's input and output ndarrays. For example, suppose an implementation is computing the sum and the data type of the input ndarray is int32 and the data type of the output ndarray is float32. In this scenario, casting int32 to float32 is not desirable as not all int32 values can be safely represented in float32, thus potentially leading to significant accumulated numerical error. Instead, we can promote int32 to float64, compute the sum, and then downcast the result for more accurate summation.
    • accumulation: applies whenever an input ndarray should be cast to a data type amenable to accumulation, irrespective of the output ndarray or other input ndarrays. For example, suppose an implementation is computing the sum and determining whether the sum passes a threshold, and further suppose that the data type of the input ndarray is int8 and the data type of the output ndarray is bool. In this scenario, as int8 has a small range of values, computing the sum has a high risk of overflow, rendering the results potentially meaningless, and type promotion is not applicable. As such, an implementation may prefer to internally cast int8 to a data type more amenable to accumulation, such as int32.
    • output: applies whenever an input ndarray should always be cast to the data type of the output ndarray. This might apply when an implementation wraps a type homogeneous interface, such as those commonly found in BLAS/LAPACK routines.

  • Whether an implementation supports a casting policy depends on the implementation. Supporting casting policies is mainly envisioned for generalized utilities wrapping lower-level APIs and needing to accommodate varied use cases (e.g., @stdlib/ndarray/base/unary-reduce-strided1d-dispatch). Exposing casting policies as part of user-facing APIs is generally not a good idea.

Examples

var indexOf = require( '@stdlib/utils/index-of' );
var policies = require( '@stdlib/ndarray/input-casting-policies' );

var POLICIES = policies();

function isPolicy( str ) {
    if ( indexOf( POLICIES, str ) === -1 ) {
        return false;
    }
    return true;
}

var bool = isPolicy( 'none' );
// returns true

bool = isPolicy( 'output' );
// returns true

bool = isPolicy( 'promoted' );
// returns true

bool = isPolicy( 'beep' );
// returns false