Class ReDeMatMultDispatcher

java.lang.Object

org.flag4j.linalg.ops.dispatch.BiTensorOpDispatcher<Matrix,Matrix,Matrix>

org.flag4j.linalg.ops.dispatch.ReDeMatMultDispatcher

public final class ReDeMatMultDispatcher
extends BiTensorOpDispatcher<Matrix,Matrix,Matrix>

A dispatcher that selects the most suitable matrix multiplication kernel for two real dense matrices.

This class implements a threshold- and shape-based decision tree to optimize performance for various matrix multiplication scenarios (e.g., small matrices, square matrices, wide/tall matrices, matrix-vector products, etc.). It maintains a cache of recently used shape/kernels to further improve performance for repeated multiplication patterns.

Usage:

• Use dispatch(Matrix, Matrix) to multiply two matrices, dynamically choosing the best kernel.
• This class is a singleton; call getInstance() to retrieve the instance if you need direct access to the underlying dispatcher.

Configuration:

The dispatcher reads various thresholds (e.g., ASPECT_THRESH, SML_THRESH, etc.) from ReDeMatMultDispatchConfigs, allowing external tuning without modifying code.

Thread Safety:

• The dispatcher itself does not maintain mutable state beyond a cached kernel lookup, which is also thread-safe.
• dispatch(Matrix, Matrix) is safe to call from multiple threads.

See Also:

• BiTensorOpDispatcher
• ReDeMatMultDispatchConfigs

Field Summary

Fields inherited from class org.flag4j.linalg.ops.dispatch.BiTensorOpDispatcher

DEFAULT_CACHE_SIZE

Method Summary

Modifier and Type	Method	Description
static Matrix	dispatch(Matrix a, Matrix b)	Dispatches the multiplication of two matrices to an appropriate implementation based on the shapes of the two matrices.
protected BiFunction<Matrix,Matrix,Matrix>	getFunc(Shape aShape, Shape bShape, int data1Length, int data2Length)	Computes the appropriate function to use when computing the matrix multiplication between two matrices.
static ReDeMatMultDispatcher	getInstance()	Gets the singleton instance of this class.
protected void	validateShapes(Shape aShape, Shape bShape)	Validates the shapes are valid for the operation.

Methods inherited from class org.flag4j.linalg.ops.dispatch.BiTensorOpDispatcher

dispatch_

Methods inherited from class java.lang.Object

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

Method Details

validateShapes

protected void validateShapes(Shape aShape,
 Shape bShape)

Validates the shapes are valid for the operation.

Specified by:

validateShapes in class BiTensorOpDispatcher<Matrix,Matrix,Matrix>

Parameters:

aShape - Shape of first tensor in the operation.

bShape - Shape of second tensor in the operation.

getInstance

public static ReDeMatMultDispatcher getInstance()

Gets the singleton instance of this class. If this class has not been instanced, a new instance will be created.

Returns:

The singleton instance of this class.

dispatch

public static Matrix dispatch(Matrix a,
 Matrix b)

Dispatches the multiplication of two matrices to an appropriate implementation based on the shapes of the two matrices.

Parameters:

a - Left matrix in the matrix multiplication.

b - Right matrix in the matrix multiplication.

Returns:

The matrix product of a and b.

getFunc

protected BiFunction<Matrix,Matrix,Matrix> getFunc(Shape aShape,
 Shape bShape,
 int data1Length,
 int data2Length)

Computes the appropriate function to use when computing the matrix multiplication between two matrices.

Specified by:

getFunc in class BiTensorOpDispatcher<Matrix,Matrix,Matrix>

Parameters:

aShape - Shape of the first matrix in the matrix multiplication problem.

bShape - Shape of the second matrix in the matrix multiplication problem.

data1Length - Full length of the data array within the first matrix.

data2Length - Full length of the data array within the second matrix.

Returns:

The appropriate function to use when computing the matrix multiplication between two matrices.