Package org.apache.sysds.runtime.matrix.data

Class LibMatrixCuDNNConvolutionAlgorithm

java.lang.Object

org.apache.sysds.runtime.matrix.data.LibMatrixCuDNNConvolutionAlgorithm

All Implemented Interfaces:

AutoCloseable

public class LibMatrixCuDNNConvolutionAlgorithm
extends Object
implements AutoCloseable

This class is a wrapper that contain necessary data structures to invoke a cudnn convolution* functions (such as cudnnConvolutionForward, etc) It implements autocloseable to simplify the LibMatrixCuDNN code and also avoids potential memory leaks. The caller has to use the factory methods: cudnnGetConvolutionForwardAlgorithm, cudnnGetConvolutionBackwardFilterAlgorithm and cudnnGetConvolutionBackwardDataAlgorithm to get the LibMatrixCuDNNConvolutionAlgorithm object. The naming of this methods is consistent with that of CuDNN library.

Field Summary

Fields

Modifier and Type	Field	Description
int	algo
long	sizeInBytes
jcuda.Pointer	workSpace

Method Summary

Modifier and Type	Method	Description
void	close()	Deallocates the tensor and filter descriptors as well as allocated workspace
static LibMatrixCuDNNConvolutionAlgorithm	cudnnGetConvolutionBackwardDataAlgorithm(GPUContext gCtx, String instName, int N, int C, int H, int W, int K, int R, int S, int pad_h, int pad_w, int stride_h, int stride_w, int P, int Q, long workspaceLimit)	Factory method to get the algorithm wrapper for convolution backward data
static LibMatrixCuDNNConvolutionAlgorithm	cudnnGetConvolutionBackwardFilterAlgorithm(GPUContext gCtx, String instName, int N, int C, int H, int W, int K, int R, int S, int pad_h, int pad_w, int stride_h, int stride_w, int P, int Q, long workspaceLimit)	Factory method to get the algorithm wrapper for convolution backward filter
static LibMatrixCuDNNConvolutionAlgorithm	cudnnGetConvolutionForwardAlgorithm(GPUContext gCtx, String instName, int N, int C, int H, int W, int K, int R, int S, int pad_h, int pad_w, int stride_h, int stride_w, int P, int Q, long workspaceLimit)	Factory method to get the algorithm wrapper for convolution forward

Methods inherited from class java.lang.Object

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

Field Details

algo

public int algo

workSpace

public jcuda.Pointer workSpace

sizeInBytes

public long sizeInBytes

Method Details

close

public void close()

Deallocates the tensor and filter descriptors as well as allocated workspace

Specified by:

close in interface AutoCloseable

cudnnGetConvolutionForwardAlgorithm

public static LibMatrixCuDNNConvolutionAlgorithm cudnnGetConvolutionForwardAlgorithm(GPUContext gCtx,
 String instName,
 int N,
 int C,
 int H,
 int W,
 int K,
 int R,
 int S,
 int pad_h,
 int pad_w,
 int stride_h,
 int stride_w,
 int P,
 int Q,
 long workspaceLimit)

Factory method to get the algorithm wrapper for convolution forward

Parameters:

gCtx - a valid GPUContext

instName - the invoking instruction's name for record Statistics.

N - number of input images

C - number of channels

H - height of each image

W - width of each image

K - number of output "channels"

R - height of filter

S - width of filter

pad_h - padding height

pad_w - padding width

stride_h - stride height

stride_w - string width

P - output height

Q - output width

workspaceLimit - maximum intermediate memory to use

Returns:

algorithm wrapper

cudnnGetConvolutionBackwardFilterAlgorithm

public static LibMatrixCuDNNConvolutionAlgorithm cudnnGetConvolutionBackwardFilterAlgorithm(GPUContext gCtx,
 String instName,
 int N,
 int C,
 int H,
 int W,
 int K,
 int R,
 int S,
 int pad_h,
 int pad_w,
 int stride_h,
 int stride_w,
 int P,
 int Q,
 long workspaceLimit)

Factory method to get the algorithm wrapper for convolution backward filter

Parameters:

gCtx - a valid GPUContext

instName - the invoking instruction's name for record Statistics.

N - number of input images

C - number of channels

H - height of each image

W - width of each image

K - number of output "channels"

R - height of filter

S - width of filter

pad_h - padding height

pad_w - padding width

stride_h - stride height

stride_w - string width

P - output height

Q - output width

workspaceLimit - maximum intermediate memory to use

Returns:

algorithm wrapper

cudnnGetConvolutionBackwardDataAlgorithm

public static LibMatrixCuDNNConvolutionAlgorithm cudnnGetConvolutionBackwardDataAlgorithm(GPUContext gCtx,
 String instName,
 int N,
 int C,
 int H,
 int W,
 int K,
 int R,
 int S,
 int pad_h,
 int pad_w,
 int stride_h,
 int stride_w,
 int P,
 int Q,
 long workspaceLimit)

Factory method to get the algorithm wrapper for convolution backward data

Parameters:

gCtx - a valid GPUContext

instName - the invoking instruction's name for record Statistics.

N - number of input images

C - number of channels

H - height of each image

W - width of each image

K - number of output "channels"

R - height of filter

S - width of filter

pad_h - padding height

pad_w - padding width

stride_h - stride height

stride_w - string width

P - output height

Q - output width

workspaceLimit - maximum intermediate memory to use

Returns:

algorithm wrapper