jama.gpu

Class GPU

java.lang.Object

jama.gpu.GPU

public class GPU
extends Object

This is the main class to perform computation on the GPU. Create a GPU via the factory methods create(..)

Example

 GPU gpu = GPU.create(); // Creates a gpu with the JavaCL.createBestContext()
 FloatMatrix C = gpu.multiplyLocal(A, B);
 

Note that this is currently in experimental stadium. You cannot influence
the parameters of the openCL kernels. So blocksize, workgroupsize and
localworkgroupsize are hardcoded. This may change in future releases.

Version:

2.0.0-SNAPSHOT

Author:

Nepomuk Seiler

Method Summary

Methods

Modifier and Type	Method and Description
static GPU	create()
static GPU	create(com.nativelibs4java.opencl.CLContext context)
protected static org.bridj.Pointer<Float>	matrixToPointer(FloatMatrix matrix)
protected static org.bridj.Pointer<Double>	matrixToPointer(Matrix matrix)
FloatMatrix	multiply(FloatMatrix A, FloatMatrix B) Use the naive openCL kernel, which computes more exact results than the local variant.
FloatMatrix	multiplyLocal(FloatMatrix A, FloatMatrix B) Use the local optimized openCL kernel, which may yield worse results than the non local variant.
protected static FloatMatrix	pointerToFloatMatrix(org.bridj.Pointer<Float> pointer, int rows, int cols)
protected static Matrix	pointerToMatrix(org.bridj.Pointer<Double> pointer, int rows, int cols)
protected static FloatMatrix	removeZeroPadding(FloatMatrix result, FloatMatrix A, FloatMatrix B) Removes the zero padding from the result matrix.
protected static int	workgroupSize(int size, int blocksize)
protected static FloatMatrix	zeroPadding(FloatMatrix matrix, int workgroupSize) Returns a padded matrix with width/height of a correct factor determined by the workgroupSize.

Methods inherited from class java.lang.Object

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

Method Detail

create

public static GPU create()

create

public static GPU create(com.nativelibs4java.opencl.CLContext context)

multiply

public FloatMatrix multiply(FloatMatrix A,
                   FloatMatrix B)
                     throws IOException

Use the naive openCL kernel, which computes more exact results than the local variant.

Parameters:

A - - Left input matrix

B - - Right input matrix

Returns:

A * B

Throws:

IOException

multiplyLocal

public FloatMatrix multiplyLocal(FloatMatrix A,
                        FloatMatrix B)
                          throws IOException

Use the local optimized openCL kernel, which may yield worse results than the non local variant.

Parameters:

A - - left input matrix

B - - right input matrix

Returns:

A * B

Throws:

IOException

matrixToPointer

protected static org.bridj.Pointer<Double> matrixToPointer(Matrix matrix)

matrixToPointer

protected static org.bridj.Pointer<Float> matrixToPointer(FloatMatrix matrix)

pointerToMatrix

protected static Matrix pointerToMatrix(org.bridj.Pointer<Double> pointer,
                     int rows,
                     int cols)

pointerToFloatMatrix

protected static FloatMatrix pointerToFloatMatrix(org.bridj.Pointer<Float> pointer,
                               int rows,
                               int cols)

workgroupSize

protected static int workgroupSize(int size,
                int blocksize)

Parameters:

size -

Returns:

zeroPadding

protected static FloatMatrix zeroPadding(FloatMatrix matrix,
                      int workgroupSize)

Returns a padded matrix with width/height of a correct factor determined by the workgroupSize. If the dimensions are already correct the identical matrix is returned.

Parameters:

matrix -

workgroupSize -

Returns:

removeZeroPadding

protected static FloatMatrix removeZeroPadding(FloatMatrix result,
                            FloatMatrix A,
                            FloatMatrix B)

Removes the zero padding from the result matrix.

Parameters:

result - - result matrix created with zero padded matrices

A - - original non-padded matrix

B - - original non-padded matrix

Returns:

non padded result matrix