Package com.yahoo.tensor.functions

Class ReduceJoin<NAMETYPE extends Name>

java.lang.Object

com.yahoo.tensor.functions.TensorFunction<NAMETYPE>

com.yahoo.tensor.functions.CompositeTensorFunction<NAMETYPE>

com.yahoo.tensor.functions.ReduceJoin<NAMETYPE>

public class ReduceJoin<NAMETYPE extends Name>
extends CompositeTensorFunction<NAMETYPE>

An optimization for tensor expressions where a join immediately follows a reduce. Evaluating this as one operation is significantly more efficient than evaluating each separately. This implementation optimizes the case where the reduce is done on the same dimensions as the join. A particularly efficient evaluation is done if there is one common dimension that is joined and reduced on, which is a common case as it covers vector and matrix like multiplications.

Author:

lesters

Constructor Summary

Constructors

Constructor	Description
ReduceJoin(Reduce<NAMETYPE> reduce, Join<NAMETYPE> join)
ReduceJoin(TensorFunction<NAMETYPE> argumentA, TensorFunction<NAMETYPE> argumentB, java.util.function.DoubleBinaryOperator combinator, Reduce.Aggregator aggregator, java.util.List<java.lang.String> dimensions)

Method Summary

Modifier and Type	Method	Description
java.util.List<TensorFunction<NAMETYPE>>	arguments()	Returns the function arguments of this node in the order they are applied
boolean	canOptimize(Tensor a, Tensor b)	Tests whether or not the reduce is over the join dimensions.
Tensor	evaluate(EvaluationContext<NAMETYPE> context)	Evaluates this by first converting it to a primitive function
PrimitiveTensorFunction<NAMETYPE>	toPrimitive()	Translate this function - and all of its arguments recursively - to a tree of primitive functions only.
java.lang.String	toString(ToStringContext context)	Return a string representation of this context.
TensorFunction<NAMETYPE>	withArguments(java.util.List<TensorFunction<NAMETYPE>> arguments)	Returns a copy of this tensor function with the arguments replaced by the given list of arguments.

Methods inherited from class com.yahoo.tensor.functions.CompositeTensorFunction

type

Methods inherited from class com.yahoo.tensor.functions.TensorFunction

evaluate, toString

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Constructor Detail

ReduceJoin

public ReduceJoin(Reduce<NAMETYPE> reduce,
                  Join<NAMETYPE> join)

ReduceJoin

public ReduceJoin(TensorFunction<NAMETYPE> argumentA,
                  TensorFunction<NAMETYPE> argumentB,
                  java.util.function.DoubleBinaryOperator combinator,
                  Reduce.Aggregator aggregator,
                  java.util.List<java.lang.String> dimensions)

Method Detail

arguments

public java.util.List<TensorFunction<NAMETYPE>> arguments()

Description copied from class: TensorFunction

Returns the function arguments of this node in the order they are applied

Specified by:

arguments in class TensorFunction<NAMETYPE extends Name>

withArguments

public TensorFunction<NAMETYPE> withArguments(java.util.List<TensorFunction<NAMETYPE>> arguments)

Description copied from class: TensorFunction

Returns a copy of this tensor function with the arguments replaced by the given list of arguments.

Specified by:

withArguments in class TensorFunction<NAMETYPE extends Name>

toPrimitive

public PrimitiveTensorFunction<NAMETYPE> toPrimitive()

Description copied from class: TensorFunction

Translate this function - and all of its arguments recursively - to a tree of primitive functions only.

Specified by:

toPrimitive in class TensorFunction<NAMETYPE extends Name>

Returns:

a tree of primitive functions implementing this

evaluate

public final Tensor evaluate(EvaluationContext<NAMETYPE> context)

Description copied from class: CompositeTensorFunction

Evaluates this by first converting it to a primitive function

Overrides:

evaluate in class CompositeTensorFunction<NAMETYPE extends Name>

Parameters:

context - a context which must be passed to all nested functions when evaluating

canOptimize

public boolean canOptimize(Tensor a,
                           Tensor b)

Tests whether or not the reduce is over the join dimensions. The remaining logic in this class assumes this to be true. If no dimensions are given, the join must be on all tensor dimensions.

Returns:

true if the implementation can optimize evaluation given the two tensors.

toString

public java.lang.String toString(ToStringContext context)

Description copied from class: TensorFunction

Return a string representation of this context.

Specified by:

toString in class TensorFunction<NAMETYPE extends Name>

Parameters:

context - a context which must be passed to all nested functions when requesting the string value