org.apache.sysml.runtime.matrix.data

Class LibMatrixCUDA

java.lang.Object

org.apache.sysml.runtime.matrix.data.LibMatrixCUDA

public class LibMatrixCUDA
extends Object

Field Summary

Fields

Modifier and Type	Field and Description
static jcuda.jcublas.cublasHandle	cublasHandle
static jcuda.jcudnn.cudnnHandle	cudnnHandle
static jcuda.jcusparse.cusparseHandle	cusparseHandle
static JCudaKernels	kernels

Constructor Summary

Constructors

Constructor and Description
LibMatrixCUDA()

Method Summary

Modifier and Type	Method and Description
static void	axpy(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in2, String outputName, double constant) Performs daxpy operation
static void	batchNormalizationBackward(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout, org.apache.sysml.runtime.controlprogram.caching.MatrixObject scale, org.apache.sysml.runtime.controlprogram.caching.MatrixObject ret, org.apache.sysml.runtime.controlprogram.caching.MatrixObject retScale, org.apache.sysml.runtime.controlprogram.caching.MatrixObject retBias, double epsilon) This method computes the backpropagation errors for image, scale and bias of batch normalization layer
static void	batchNormalizationForwardInference(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject scale, org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias, org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningMean, org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningVar, org.apache.sysml.runtime.controlprogram.caching.MatrixObject ret, double epsilon) Performs the forward BatchNormalization layer computation for inference
static void	batchNormalizationForwardTraining(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject scale, org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias, org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningMean, org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningVar, org.apache.sysml.runtime.controlprogram.caching.MatrixObject ret, org.apache.sysml.runtime.controlprogram.caching.MatrixObject retRunningMean, org.apache.sysml.runtime.controlprogram.caching.MatrixObject retRunningVar, double epsilon, double exponentialAverageFactor) Performs the forward BatchNormalization layer computation for training
static void	biasAdd(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject input, org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock) Performs the operation corresponding to the DML script: ones = matrix(1, rows=1, cols=Hout*Wout) output = input + matrix(bias %*% ones, rows=1, cols=F*Hout*Wout) This operation is often followed by conv2d and hence we have introduced bias_add(input, bias) built-in function
static void	biasMultiply(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject input, org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock) Performs the operation corresponding to the DML script: ones = matrix(1, rows=1, cols=Hout*Wout) output = input * matrix(bias %*% ones, rows=1, cols=F*Hout*Wout) This operation is often followed by conv2d and hence we have introduced bias_add(input, bias) built-in function
protected static void	bothSparseMatmult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left, org.apache.sysml.runtime.controlprogram.caching.MatrixObject right, boolean isLeftTransposed, boolean isRightTransposed) Sparse C = Sparse op(A) * Sparse op(B) Reroutes call to sparse matrix-vector mult if needed
static void	conv2d(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject filter, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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)
static void	conv2dBackwardData(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject filter, org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, 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) This method computes the backpropogation errors for previous layer of convolution operation
static void	conv2dBackwardFilter(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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) This method computes the backpropogation errors for filter of convolution operation
static void	conv2dBiasAdd(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias, org.apache.sysml.runtime.controlprogram.caching.MatrixObject filter, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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)
protected static void	denseDenseMatmult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left1, org.apache.sysml.runtime.controlprogram.caching.MatrixObject right1, boolean isLeftTransposed1, boolean isRightTransposed1) Dense dense matrix multiply C = op(A) * op(B), A and B are dense matrices
static void	denseDenseMatmult(String instName, jcuda.Pointer output, int leftRows1, int leftCols1, int rightRows1, int rightCols1, boolean isLeftTransposed1, boolean isRightTransposed1, jcuda.Pointer leftPtr, jcuda.Pointer rightPtr) Dense-dense matrix multiply C = op(A) * op(B), A and B are dense matrices On the host, the matrices are in row-major format; cuBLAS expects them in column-major format.
protected static void	denseSparseMatmult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, org.apache.sysml.runtime.controlprogram.caching.MatrixObject right, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left, boolean isLeftTransposed, boolean isRightTransposed, int transA, int transB, int m, int n, int k) C = op(A) * op(B) where A is dense and B is sparse If B is ultrasparse, A is converted to a sparse matrix and sparseSparseMatmult(MatrixObject, int, int, int, int, int, CSRPointer, CSRPointer) is invoked otherwise B is converted to a dense matrix and denseDenseMatmult(Pointer, int, int, int, int, boolean, boolean, Pointer, Pointer) is invoked.
protected static void	eitherSparseMatmult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left, org.apache.sysml.runtime.controlprogram.caching.MatrixObject right, boolean isLeftTransposed, boolean isRightTransposed) One of the matrices is sparse, the other dense C = op(A) x op(B)
static void	exp(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1, String outputName) Performs an "exp" operation on a matrix on the GPU
static boolean	isInSparseFormat(org.apache.sysml.runtime.controlprogram.caching.MatrixObject mo)
static org.apache.sysml.runtime.controlprogram.caching.MatrixObject	matmult(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left1, org.apache.sysml.runtime.controlprogram.caching.MatrixObject right1, String outputName, boolean isLeftTransposed1, boolean isRightTransposed1) Matrix multiply on GPU Examines sparsity and shapes and routes call to appropriate method from cuBLAS or cuSparse C = op(A) x op(B)
static void	matmultTSMM(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left, String outputName, boolean isLeftTransposed) Performs tsmm, A %*% A' or A' %*% A, on GPU by exploiting cublasDsyrk(...)
static void	matrixScalarArithmetic(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in2, String outputName, boolean isLeftTransposed, boolean isRightTransposed, org.apache.sysml.runtime.matrix.operators.BinaryOperator op) Performs elementwise operation specified by op of two input matrices in1 and in2
static void	matrixScalarArithmetic(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in, String outputName, boolean isInputTransposed, org.apache.sysml.runtime.matrix.operators.ScalarOperator op) Entry point to perform elementwise matrix-scalar operation specified by op
static void	maxpooling(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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) performs maxpooling on GPU by exploiting cudnnPoolingForward(...)
static void	maxpoolingBackward(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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) Performs maxpoolingBackward on GPU by exploiting cudnnPoolingBackward(...) This method computes the backpropogation errors for previous layer of maxpooling operation
static void	performMaxpooling(String instName, jcuda.Pointer x, jcuda.jcudnn.cudnnTensorDescriptor xDesc, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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)
static jcuda.Pointer	pointerTo(double value)
static void	relu(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in, String outputName) Performs the relu operation on the GPU.
static void	reluBackward(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject input, org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock) This method computes the backpropagation errors for previous layer of relu operation
static void	reluMaxpooling(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject image, org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock, 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) performs relu followed by maxpooling on GPU by exploiting cudnnPoolingForward(...)
protected static void	sparseDenseMatmult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, org.apache.sysml.runtime.controlprogram.caching.MatrixObject left, org.apache.sysml.runtime.controlprogram.caching.MatrixObject right, boolean isLeftTransposed, boolean isRightTransposed, int transA, int transB, int m, int n, int k) * C = op(A) * op(B) where A is sparse and B is dense If A is ultrasparse, B is converted to a sparse matrix and sparseSparseMatmult(MatrixObject, int, int, int, int, int, CSRPointer, CSRPointer) is invoked otherwise A is converted to a dense matrix and denseDenseMatmult(Pointer, int, int, int, int, boolean, boolean, Pointer, Pointer) is invoked.
protected static void	sparseMatrixDenseVectorMult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, JCudaObject.CSRPointer A, jcuda.Pointer B_dense, int transA, int m, int k) C = op(A) x B A is a sparse matrix, B is a dense vector
protected static void	sparseMatrixVectorMult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, int transA, int m, int n, int k, JCudaObject.CSRPointer A, JCudaObject.CSRPointer B) Does a sparse matrix-vector multiply.
protected static void	sparseSparseMatmult(String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject output, int transA, int transB, int m, int n, int k, JCudaObject.CSRPointer A, JCudaObject.CSRPointer B) Does a sparse-sparse Matrix multiply C = op(A) x op(B), A, B are sparse matrices
static void	transpose(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in, String outputName) Transposes the input matrix using cublasDgeam
static void	unaryAggregate(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec, String instName, org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1, String output, org.apache.sysml.runtime.matrix.operators.AggregateUnaryOperator op) Entry point to perform Unary aggregate operations on the GPU.

Methods inherited from class java.lang.Object

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

Field Detail

cudnnHandle

public static jcuda.jcudnn.cudnnHandle cudnnHandle

cublasHandle

public static jcuda.jcublas.cublasHandle cublasHandle

cusparseHandle

public static jcuda.jcusparse.cusparseHandle cusparseHandle

kernels

public static JCudaKernels kernels

Constructor Detail

LibMatrixCUDA

public LibMatrixCUDA()

Method Detail

isInSparseFormat

public static boolean isInSparseFormat(org.apache.sysml.runtime.controlprogram.caching.MatrixObject mo)

conv2dBiasAdd

public static void conv2dBiasAdd(String instName,
                                 org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                 org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias,
                                 org.apache.sysml.runtime.controlprogram.caching.MatrixObject filter,
                                 org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                                 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)
                          throws DMLRuntimeException

Throws:

DMLRuntimeException

conv2d

public static void conv2d(String instName,
                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject filter,
                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                          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)
                   throws DMLRuntimeException

Throws:

DMLRuntimeException

pointerTo

public static jcuda.Pointer pointerTo(double value)

reluBackward

public static void reluBackward(String instName,
                                org.apache.sysml.runtime.controlprogram.caching.MatrixObject input,
                                org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout,
                                org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock)
                         throws DMLRuntimeException

This method computes the backpropagation errors for previous layer of relu operation

Parameters:

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

input - input image

dout - next layer error propogation

outputBlock - output

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

biasMultiply

public static void biasMultiply(String instName,
                                org.apache.sysml.runtime.controlprogram.caching.MatrixObject input,
                                org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias,
                                org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock)
                         throws DMLRuntimeException

Performs the operation corresponding to the DML script: ones = matrix(1, rows=1, cols=Hout*Wout) output = input * matrix(bias %*% ones, rows=1, cols=F*Hout*Wout) This operation is often followed by conv2d and hence we have introduced bias_add(input, bias) built-in function

Parameters:

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

input - input image

bias - bias

outputBlock - output

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

biasAdd

public static void biasAdd(String instName,
                           org.apache.sysml.runtime.controlprogram.caching.MatrixObject input,
                           org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias,
                           org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock)
                    throws DMLRuntimeException

Performs the operation corresponding to the DML script: ones = matrix(1, rows=1, cols=Hout*Wout) output = input + matrix(bias %*% ones, rows=1, cols=F*Hout*Wout) This operation is often followed by conv2d and hence we have introduced bias_add(input, bias) built-in function

Parameters:

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

input - input image

bias - bias

outputBlock - output

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

batchNormalizationForwardInference

public static void batchNormalizationForwardInference(String instName,
                                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject scale,
                                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias,
                                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningMean,
                                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningVar,
                                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject ret,
                                                      double epsilon)
                                               throws DMLRuntimeException

Performs the forward BatchNormalization layer computation for inference

Parameters:

instName - name of the instruction

image - input image

scale - scale (as per CuDNN) and gamma as per original paper: shape [1, C, 1, 1]

bias - bias (as per CuDNN) and beta as per original paper: shape [1, C, 1, 1]

runningMean - running mean accumulated during training phase: shape [1, C, 1, 1]

runningVar - running variance accumulated during training phase: shape [1, C, 1, 1]

ret - normalized input

epsilon - epsilon value used in the batch normalization formula

Throws:

DMLRuntimeException - if error occurs

batchNormalizationForwardTraining

public static void batchNormalizationForwardTraining(String instName,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject scale,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject bias,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningMean,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject runningVar,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject ret,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject retRunningMean,
                                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject retRunningVar,
                                                     double epsilon,
                                                     double exponentialAverageFactor)
                                              throws DMLRuntimeException

Performs the forward BatchNormalization layer computation for training

Parameters:

instName - name of the instruction

image - input image

scale - scale (as per CuDNN) and gamma as per original paper: shape [1, C, 1, 1]

bias - bias (as per CuDNN) and beta as per original paper: shape [1, C, 1, 1]

runningMean - running mean accumulated during training phase: shape [1, C, 1, 1]

runningVar - running variance accumulated during training phase: shape [1, C, 1, 1]

ret - (output) normalized input

retRunningMean - (output) running mean accumulated during training phase: shape [1, C, 1, 1]

retRunningVar - (output) running variance accumulated during training phase: shape [1, C, 1, 1]

epsilon - epsilon value used in the batch normalization formula

exponentialAverageFactor - factor used in the moving average computation

Throws:

DMLRuntimeException - if error occurs

batchNormalizationBackward

public static void batchNormalizationBackward(String instName,
                                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout,
                                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject scale,
                                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject ret,
                                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject retScale,
                                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject retBias,
                                              double epsilon)
                                       throws DMLRuntimeException

This method computes the backpropagation errors for image, scale and bias of batch normalization layer

Parameters:

instName - name of the instruction

image - input image

dout - input errors of shape C, H, W

scale - scale (as per CuDNN) and gamma as per original paper: shape [1, C, 1, 1]

ret - (output) backpropagation errors for previous layer

retScale - backpropagation error for scale

retBias - backpropagation error for bias

epsilon - epsilon value used in the batch normalization formula

Throws:

DMLRuntimeException - if error occurs

conv2dBackwardFilter

public static void conv2dBackwardFilter(String instName,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                                        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)
                                 throws DMLRuntimeException

This method computes the backpropogation errors for filter of convolution operation

Parameters:

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

image - input image

dout - errors from next layer

outputBlock - output errors

N - number of images

C - number of channels

H - height

W - width

K - number of filters

R - filter height

S - filter width

pad_h - pad height

pad_w - pad width

stride_h - stride height

stride_w - stride width

P - output activation height

Q - output activation width

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

conv2dBackwardData

public static void conv2dBackwardData(String instName,
                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject filter,
                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout,
                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                      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)
                               throws DMLRuntimeException

This method computes the backpropogation errors for previous layer of convolution operation

Parameters:

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

filter - filter used in conv2d

dout - errors from next layer

output - output errors

N - number of images

C - number of channels

H - height

W - width

K - number of filters

R - filter height

S - filter width

pad_h - pad height

pad_w - pad width

stride_h - stride height

stride_w - stride width

P - output activation height

Q - output activation width

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

maxpooling

public static void maxpooling(String instName,
                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                              org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                              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)
                       throws DMLRuntimeException

performs maxpooling on GPU by exploiting cudnnPoolingForward(...)

Parameters:

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

image - image as matrix object

outputBlock - output matrix

N - batch size

C - number of channels

H - height of image

W - width of image

K - number of filters

R - height of filter

S - width of filter

pad_h - vertical padding

pad_w - horizontal padding

stride_h - horizontal stride

stride_w - vertical stride

P - (H - R + 1 + 2*pad_h)/stride_h

Q - (W - S + 1 + 2*pad_w)/stride_w

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

performMaxpooling

public static void performMaxpooling(String instName,
                                     jcuda.Pointer x,
                                     jcuda.jcudnn.cudnnTensorDescriptor xDesc,
                                     org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                                     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)
                              throws DMLRuntimeException

Throws:

DMLRuntimeException

reluMaxpooling

public static void reluMaxpooling(String instName,
                                  org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                  org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                                  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)
                           throws DMLRuntimeException

performs relu followed by maxpooling on GPU by exploiting cudnnPoolingForward(...)

Parameters:

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

image - image as matrix object

outputBlock - output matrix

N - batch size

C - number of channels

H - height of image

W - width of image

K - number of filters

R - height of filter

S - width of filter

pad_h - vertical padding

pad_w - horizontal padding

stride_h - horizontal stride

stride_w - vertical stride

P - (H - R + 1 + 2*pad_h)/stride_h

Q - (W - S + 1 + 2*pad_w)/stride_w

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

maxpoolingBackward

public static void maxpoolingBackward(String instName,
                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject image,
                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject dout,
                                      org.apache.sysml.runtime.controlprogram.caching.MatrixObject outputBlock,
                                      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)
                               throws DMLRuntimeException

Performs maxpoolingBackward on GPU by exploiting cudnnPoolingBackward(...) This method computes the backpropogation errors for previous layer of maxpooling operation

Parameters:

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

image - image as matrix object

dout - delta matrix, output of previous layer

outputBlock - output matrix

N - batch size

C - number of channels

H - height of image

W - width of image

K - number of filters

R - height of filter

S - width of filter

pad_h - vertical padding

pad_w - horizontal padding

stride_h - horizontal stride

stride_w - vertical stride

P - (H - R + 1 + 2*pad_h)/stride_h

Q - (W - S + 1 + 2*pad_w)/stride_w

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

relu

public static void relu(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                        String instName,
                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject in,
                        String outputName)
                 throws DMLRuntimeException

Performs the relu operation on the GPU.

Parameters:

ec - currently active ExecutionContext

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

in - input matrix

outputName - name of the output matrix

Throws:

DMLRuntimeException - if an error occurs

matmultTSMM

public static void matmultTSMM(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                               String instName,
                               org.apache.sysml.runtime.controlprogram.caching.MatrixObject left,
                               String outputName,
                               boolean isLeftTransposed)
                        throws DMLRuntimeException

Performs tsmm, A %*% A' or A' %*% A, on GPU by exploiting cublasDsyrk(...)

Parameters:

ec - execution context

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

left - input matrix, as in a tsmm expression like A %*% A' or A' %*% A, we just need to check whether the left one is transposed or not, I named it 'left'

outputName - output matrix name

isLeftTransposed - if true, left transposed

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

matmult

public static org.apache.sysml.runtime.controlprogram.caching.MatrixObject matmult(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                                                                                   String instName,
                                                                                   org.apache.sysml.runtime.controlprogram.caching.MatrixObject left1,
                                                                                   org.apache.sysml.runtime.controlprogram.caching.MatrixObject right1,
                                                                                   String outputName,
                                                                                   boolean isLeftTransposed1,
                                                                                   boolean isRightTransposed1)
                                                                            throws DMLRuntimeException

Matrix multiply on GPU Examines sparsity and shapes and routes call to appropriate method from cuBLAS or cuSparse C = op(A) x op(B)

Parameters:

ec - Current ExecutionContext instance

instName - name of the invoking instruction to recordStatistics.

left1 - Matrix A

right1 - Matrix B

outputName - Name of the output matrix C (in code generated after LOP layer)

isLeftTransposed1 - op for A, transposed or not

isRightTransposed1 - op for B, tranposed or not

Returns:

output of matrix multiply

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

eitherSparseMatmult

protected static void eitherSparseMatmult(String instName,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject left,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject right,
                                          boolean isLeftTransposed,
                                          boolean isRightTransposed)
                                   throws DMLRuntimeException

One of the matrices is sparse, the other dense C = op(A) x op(B)

Parameters:

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

output - allocated output object for C on host to which GPU output will be attached

left - Matrix A on host

right - Matrix B on host

isLeftTransposed - op for A, tranposed or not

isRightTransposed - op for B, transposed or not

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

denseSparseMatmult

protected static void denseSparseMatmult(String instName,
                                         org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                         org.apache.sysml.runtime.controlprogram.caching.MatrixObject right,
                                         org.apache.sysml.runtime.controlprogram.caching.MatrixObject left,
                                         boolean isLeftTransposed,
                                         boolean isRightTransposed,
                                         int transA,
                                         int transB,
                                         int m,
                                         int n,
                                         int k)
                                  throws DMLRuntimeException

C = op(A) * op(B) where A is dense and B is sparse If B is ultrasparse, A is converted to a sparse matrix and sparseSparseMatmult(MatrixObject, int, int, int, int, int, CSRPointer, CSRPointer) is invoked otherwise B is converted to a dense matrix and denseDenseMatmult(Pointer, int, int, int, int, boolean, boolean, Pointer, Pointer) is invoked.

Parameters:

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

output - ?

right - ?

left - ?

isLeftTransposed - ?

isRightTransposed - ?

transA - ?

transB - ?

m - ?

n - ?

k - ?

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

sparseDenseMatmult

protected static void sparseDenseMatmult(String instName,
                                         org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                         org.apache.sysml.runtime.controlprogram.caching.MatrixObject left,
                                         org.apache.sysml.runtime.controlprogram.caching.MatrixObject right,
                                         boolean isLeftTransposed,
                                         boolean isRightTransposed,
                                         int transA,
                                         int transB,
                                         int m,
                                         int n,
                                         int k)
                                  throws DMLRuntimeException

* C = op(A) * op(B) where A is sparse and B is dense If A is ultrasparse, B is converted to a sparse matrix and sparseSparseMatmult(MatrixObject, int, int, int, int, int, CSRPointer, CSRPointer) is invoked otherwise A is converted to a dense matrix and denseDenseMatmult(Pointer, int, int, int, int, boolean, boolean, Pointer, Pointer) is invoked.

Parameters:

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

output - ?

left - ?

right - ?

isLeftTransposed - ?

isRightTransposed - ?

transA - ?

transB - ?

m - ?

n - ?

k - ?

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

sparseMatrixDenseVectorMult

protected static void sparseMatrixDenseVectorMult(String instName,
                                                  org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                                  JCudaObject.CSRPointer A,
                                                  jcuda.Pointer B_dense,
                                                  int transA,
                                                  int m,
                                                  int k)
                                           throws DMLRuntimeException

C = op(A) x B A is a sparse matrix, B is a dense vector

Parameters:

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

output - allocated output on the host, to which the GPU output C will be attached

A - sparse matrix A on the GPU

B_dense - dense matrix/vector B on the GPU

transA - op for A, tranposed or not

m - number of rows in A (not op(A))

k - number of cols in A or number of rows in B (not op(A) or op(B))

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

bothSparseMatmult

protected static void bothSparseMatmult(String instName,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject left,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject right,
                                        boolean isLeftTransposed,
                                        boolean isRightTransposed)
                                 throws DMLRuntimeException

Sparse C = Sparse op(A) * Sparse op(B) Reroutes call to sparse matrix-vector mult if needed

Parameters:

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

output - ?

instName - name of the invoking instruction to recordStatistics.

left - ?

right - ?

isLeftTransposed - ?

isRightTransposed - ?

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

sparseMatrixVectorMult

protected static void sparseMatrixVectorMult(String instName,
                                             org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                             int transA,
                                             int m,
                                             int n,
                                             int k,
                                             JCudaObject.CSRPointer A,
                                             JCudaObject.CSRPointer B)
                                      throws DMLRuntimeException

Does a sparse matrix-vector multiply. C = op(A) x B, A is a sparse matrix, B is a sparse vector with numCols = 1.

Parameters:

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

output - allocated output object C to which the GPU output matrix will be attached

transA - if A is to be transposed or not (the op in op(A))

m - number of rows in A (not op(A))

n - number of cols in A (not op(A))

k - number of rows in B, (cols in B is assumed to be 1)

A - left sparse matrix on GPU

B - right sparse vector on GPU

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

sparseSparseMatmult

protected static void sparseSparseMatmult(String instName,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                          int transA,
                                          int transB,
                                          int m,
                                          int n,
                                          int k,
                                          JCudaObject.CSRPointer A,
                                          JCudaObject.CSRPointer B)
                                   throws DMLRuntimeException

Does a sparse-sparse Matrix multiply C = op(A) x op(B), A, B are sparse matrices

Parameters:

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

output - allocated output object on host to which the GPU output matrix will be attached

transA - op for A - to be transposed or not

transB - op for B

m - number of rows in op(A)

n - number of cols in op(B)

k - number of cols in op(A) or rows in op(B)

A - left sparse matrix on GPU

B - right sparse matrix on GPU

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

denseDenseMatmult

protected static void denseDenseMatmult(String instName,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject output,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject left1,
                                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject right1,
                                        boolean isLeftTransposed1,
                                        boolean isRightTransposed1)
                                 throws DMLRuntimeException

Dense dense matrix multiply C = op(A) * op(B), A and B are dense matrices

Parameters:

instName - name of the invoking instruction to recordStatistics.

output - output object C on host with GPU data allocated

left1 - left matrix A on host (in row-major order)

right1 - right matrix B on host (in row-major order)

isLeftTransposed1 - op for A, transposed or not

isRightTransposed1 - op for B, transposed or not

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

denseDenseMatmult

public static void denseDenseMatmult(String instName,
                                     jcuda.Pointer output,
                                     int leftRows1,
                                     int leftCols1,
                                     int rightRows1,
                                     int rightCols1,
                                     boolean isLeftTransposed1,
                                     boolean isRightTransposed1,
                                     jcuda.Pointer leftPtr,
                                     jcuda.Pointer rightPtr)
                              throws DMLRuntimeException

Dense-dense matrix multiply C = op(A) * op(B), A and B are dense matrices On the host, the matrices are in row-major format; cuBLAS expects them in column-major format. What we have as input is t(A) and t(B), t(X) = transpose of X. We do t(B) %*% t(A) to get t(C); If we were to calculate t(t(C), we would get the resultant matrix C, but this would be in column-major format. What we really want is t(C). This we already have as the result of t(B) %*% t(A).

Parameters:

instName - name of the invoking instruction to recordStatistics.

output - output allocated on GPU in column major format

leftRows1 - number of rows in A

leftCols1 - number of cols in A

rightRows1 - number of rows in B

rightCols1 - number of cols in B

isLeftTransposed1 - op for A, transposed or not

isRightTransposed1 - op for B, transposed or not

leftPtr - A allocated on the GPU in row-major format

rightPtr - B allocated on the GPU in row-major format

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

unaryAggregate

public static void unaryAggregate(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                                  String instName,
                                  org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1,
                                  String output,
                                  org.apache.sysml.runtime.matrix.operators.AggregateUnaryOperator op)
                           throws DMLRuntimeException

Entry point to perform Unary aggregate operations on the GPU. The execution context object is used to allocate memory for the GPU.

Parameters:

ec - Instance of ExecutionContext, from which the output variable will be allocated

instName - name of the invoking instruction to recordStatistics.

in1 - input matrix

output - output matrix/scalar name

op - Instance of AggregateUnaryOperator which encapsulates the direction of reduction/aggregation and the reduction operation.

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

matrixScalarArithmetic

public static void matrixScalarArithmetic(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                                          String instName,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject in,
                                          String outputName,
                                          boolean isInputTransposed,
                                          org.apache.sysml.runtime.matrix.operators.ScalarOperator op)
                                   throws DMLRuntimeException

Entry point to perform elementwise matrix-scalar operation specified by op

Parameters:

ec - execution context

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

in - input matrix

outputName - output matrix name

isInputTransposed - true if input transposed

op - scalar operator

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

matrixScalarArithmetic

public static void matrixScalarArithmetic(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                                          String instName,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1,
                                          org.apache.sysml.runtime.controlprogram.caching.MatrixObject in2,
                                          String outputName,
                                          boolean isLeftTransposed,
                                          boolean isRightTransposed,
                                          org.apache.sysml.runtime.matrix.operators.BinaryOperator op)
                                   throws DMLRuntimeException

Performs elementwise operation specified by op of two input matrices in1 and in2

Parameters:

ec - execution context

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

in1 - input matrix 1

in2 - input matrix 2

outputName - output matrix name

isLeftTransposed - true if left-transposed

isRightTransposed - true if right-transposed

op - binary operator

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

transpose

public static void transpose(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                             String instName,
                             org.apache.sysml.runtime.controlprogram.caching.MatrixObject in,
                             String outputName)
                      throws DMLRuntimeException

Transposes the input matrix using cublasDgeam

Parameters:

ec - execution context

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

in - input matrix

outputName - output matrix name

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

exp

public static void exp(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                       String instName,
                       org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1,
                       String outputName)
                throws DMLRuntimeException

Performs an "exp" operation on a matrix on the GPU

Parameters:

ec - execution context

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

in1 - input matrix

outputName - output matrix name

Throws:

DMLRuntimeException - if DMLRuntimeException occurs

axpy

public static void axpy(org.apache.sysml.runtime.controlprogram.context.ExecutionContext ec,
                        String instName,
                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject in1,
                        org.apache.sysml.runtime.controlprogram.caching.MatrixObject in2,
                        String outputName,
                        double constant)
                 throws DMLRuntimeException

Performs daxpy operation

Parameters:

ec - execution context

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

in1 - input matrix 1

in2 - input matrix 2

outputName - output matrix name

constant - pointer constant

Throws:

DMLRuntimeException - if DMLRuntimeException occurs