org.apache.sysds.runtime.matrix.data

Class LibMatrixCuDNN

java.lang.Object

org.apache.sysds.runtime.matrix.data.LibMatrixCUDA

org.apache.sysds.runtime.matrix.data.LibMatrixCuDNN

public class LibMatrixCuDNN
extends LibMatrixCUDA

This class contains method that invoke CuDNN operations.

Field Summary

Fields inherited from class org.apache.sysds.runtime.matrix.data.LibMatrixCUDA

cudaSupportFunctions, customKernelSuffix, sizeOfDataType

Constructor Summary

Constructors

Constructor and Description
LibMatrixCuDNN()

Method Summary

Modifier and Type	Method and Description
static void	batchNormalizationBackward(GPUContext gCtx, String instName, MatrixObject image, MatrixObject dout, MatrixObject scale, MatrixObject dX, MatrixObject dScale, MatrixObject dBias, double epsilon, MatrixObject resultSaveMean, MatrixObject resultSaveInvVariance) This method computes the backpropagation errors for image, scale and bias of batch normalization layer
static void	batchNormalizationForwardInference(GPUContext gCtx, String instName, MatrixObject image, MatrixObject scale, MatrixObject bias, MatrixObject runningMean, MatrixObject runningVar, MatrixObject ret, double epsilon) Performs the forward BatchNormalization layer computation for inference
static void	batchNormalizationForwardTraining(GPUContext gCtx, String instName, MatrixObject image, MatrixObject scale, MatrixObject bias, MatrixObject runningMean, MatrixObject runningVar, MatrixObject ret, MatrixObject retRunningMean, MatrixObject retRunningVar, double epsilon, double exponentialAverageFactor, MatrixObject resultSaveMean, MatrixObject resultSaveInvVariance) Performs the forward BatchNormalization layer computation for training
static void	conv2d(GPUContext gCtx, String instName, MatrixObject image, MatrixObject filter, 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, double intermediateMemoryBudget) Performs a 2D convolution
static void	conv2dBackwardData(GPUContext gCtx, String instName, MatrixObject filter, MatrixObject dout, 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, double intermediateMemoryBudget) This method computes the backpropogation errors for previous layer of convolution operation
static void	conv2dBackwardFilter(GPUContext gCtx, String instName, MatrixObject image, MatrixObject dout, 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, double intermediateMemoryBudget) This method computes the backpropogation errors for filter of convolution operation
static void	conv2dBiasAdd(GPUContext gCtx, String instName, MatrixObject image, MatrixObject bias, MatrixObject filter, 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, double intermediateMemoryBudget) Does a 2D convolution followed by a bias_add
static jcuda.Pointer	getDensePointerForCuDNN(GPUContext gCtx, MatrixObject image, String instName, int numRows, int numCols) Convenience method to get jcudaDenseMatrixPtr.
static void	lstm(ExecutionContext ec, GPUContext gCtx, String instName, jcuda.Pointer X, jcuda.Pointer wPointer, jcuda.Pointer out0, jcuda.Pointer c0, boolean return_sequences, String outputName, String cyName, int N, int M, int D, int T) Computes the forward pass for an LSTM layer with M neurons.
static void	lstmBackward(ExecutionContext ec, GPUContext gCtx, String instName, jcuda.Pointer x, jcuda.Pointer hx, jcuda.Pointer cx, jcuda.Pointer wPointer, String doutName, String dcyName, String dxName, String dwName, String dbName, String dhxName, String dcxName, boolean return_sequences, int N, int M, int D, int T)
static void	pooling(GPUContext gCtx, String instName, MatrixObject image, 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, LibMatrixDNN.PoolingType poolingType, double intermediateMemoryBudget) performs maxpooling on GPU by exploiting cudnnPoolingForward(...)
static void	poolingBackward(GPUContext gCtx, String instName, MatrixObject image, MatrixObject dout, MatrixObject maxpoolOutput, 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, LibMatrixDNN.PoolingType poolingType, double intermediateMemoryBudget) Performs maxpoolingBackward on GPU by exploiting cudnnPoolingBackward(...) This method computes the backpropogation errors for previous layer of maxpooling operation
static void	relu(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject in, String outputName) Performs the relu operation on the GPU.
static void	softmax(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject in1, String outputName) Performs an "softmax" operation on a matrix on the GPU

Methods inherited from class org.apache.sysds.runtime.matrix.data.LibMatrixCUDA

abs, acos, asin, atan, axpy, biasAdd, biasMultiply, cbind, ceil, channelSums, computeNNZ, cos, cosh, cumulativeScan, cumulativeSumProduct, denseTranspose, deviceCopy, double2float, exp, float2double, floor, getCudaKernels, getDenseMatrixOutputForGPUInstruction, getDensePointer, getNnz, isInSparseFormat, log, matmultTSMM, matrixMatrixArithmetic, matrixMatrixRelational, matrixScalarArithmetic, matrixScalarOp, matrixScalarRelational, one, rbind, reluBackward, resetFloatingPointPrecision, round, sigmoid, sign, sin, sinh, sliceOperations, solve, sqrt, tan, tanh, toInt, transpose, unaryAggregate, zero

Methods inherited from class java.lang.Object

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

Constructor Detail

LibMatrixCuDNN

public LibMatrixCuDNN()

Method Detail

conv2dBiasAdd

public static void conv2dBiasAdd(GPUContext gCtx,
                                 String instName,
                                 MatrixObject image,
                                 MatrixObject bias,
                                 MatrixObject filter,
                                 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,
                                 double intermediateMemoryBudget)

Does a 2D convolution followed by a bias_add

Parameters:

gCtx - a valid GPUContext

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

image - input image matrix object

bias - bias matrix object

filter - filter matrix object

output - output matrix object

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

intermediateMemoryBudget - intermediate memory budget

conv2d

public static void conv2d(GPUContext gCtx,
                          String instName,
                          MatrixObject image,
                          MatrixObject filter,
                          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,
                          double intermediateMemoryBudget)

Performs a 2D convolution

Parameters:

gCtx - a valid GPUContext

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

image - input matrix object

filter - filter matrix object

outputBlock - output matrix object

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

intermediateMemoryBudget - intermediate memory budget

softmax

public static void softmax(ExecutionContext ec,
                           GPUContext gCtx,
                           String instName,
                           MatrixObject in1,
                           String outputName)

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

Parameters:

ec - execution context

gCtx - a valid GPUContext

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

in1 - input matrix

outputName - output matrix name

conv2dBackwardFilter

public static void conv2dBackwardFilter(GPUContext gCtx,
                                        String instName,
                                        MatrixObject image,
                                        MatrixObject dout,
                                        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,
                                        double intermediateMemoryBudget)

This method computes the backpropogation errors for filter of convolution operation

Parameters:

gCtx - a valid GPUContext

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

intermediateMemoryBudget - intermediate memory budget

conv2dBackwardData

public static void conv2dBackwardData(GPUContext gCtx,
                                      String instName,
                                      MatrixObject filter,
                                      MatrixObject dout,
                                      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,
                                      double intermediateMemoryBudget)

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

Parameters:

gCtx - a valid GPUContext

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

intermediateMemoryBudget - intermediate memory budget

pooling

public static void pooling(GPUContext gCtx,
                           String instName,
                           MatrixObject image,
                           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,
                           LibMatrixDNN.PoolingType poolingType,
                           double intermediateMemoryBudget)

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

Parameters:

gCtx - a valid GPUContext

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

poolingType - type of pooling

intermediateMemoryBudget - intermediate memory budget

poolingBackward

public static void poolingBackward(GPUContext gCtx,
                                   String instName,
                                   MatrixObject image,
                                   MatrixObject dout,
                                   MatrixObject maxpoolOutput,
                                   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,
                                   LibMatrixDNN.PoolingType poolingType,
                                   double intermediateMemoryBudget)

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

Parameters:

gCtx - a valid GPUContext

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

image - image as matrix object

dout - delta matrix, output of previous layer

maxpoolOutput - (optional and can be null) output of maxpool forward function

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

poolingType - type of pooling

intermediateMemoryBudget - intermediate memory budget

relu

public static void relu(ExecutionContext ec,
                        GPUContext gCtx,
                        String instName,
                        MatrixObject in,
                        String outputName)

Performs the relu operation on the GPU.

Parameters:

ec - currently active ExecutionContext

gCtx - a valid GPUContext

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

in - input matrix

outputName - name of the output matrix

lstm

public static void lstm(ExecutionContext ec,
                        GPUContext gCtx,
                        String instName,
                        jcuda.Pointer X,
                        jcuda.Pointer wPointer,
                        jcuda.Pointer out0,
                        jcuda.Pointer c0,
                        boolean return_sequences,
                        String outputName,
                        String cyName,
                        int N,
                        int M,
                        int D,
                        int T)
                 throws DMLRuntimeException

Computes the forward pass for an LSTM layer with M neurons. The input data has N sequences of T examples, each with D features.

Parameters:

ec - execution context

gCtx - gpu context

instName - name of the instruction

X - input matrix pointer

wPointer - weight matrix pointer

out0 - Outputs from previous timestep

c0 - Initial cell state

return_sequences - Whether to return `out` at all timesteps, or just for the final timestep.

outputName - name of the out variable. If `return_sequences` is True, outputs for all timesteps.

cyName - name of the output cell state. Cell state for final timestep.

N - minibatch size

M - hidden size

D - number of features

T - sequence length

Throws:

DMLRuntimeException - if error

lstmBackward

public static void lstmBackward(ExecutionContext ec,
                                GPUContext gCtx,
                                String instName,
                                jcuda.Pointer x,
                                jcuda.Pointer hx,
                                jcuda.Pointer cx,
                                jcuda.Pointer wPointer,
                                String doutName,
                                String dcyName,
                                String dxName,
                                String dwName,
                                String dbName,
                                String dhxName,
                                String dcxName,
                                boolean return_sequences,
                                int N,
                                int M,
                                int D,
                                int T)
                         throws DMLRuntimeException

Throws:

DMLRuntimeException

batchNormalizationForwardTraining

public static void batchNormalizationForwardTraining(GPUContext gCtx,
                                                     String instName,
                                                     MatrixObject image,
                                                     MatrixObject scale,
                                                     MatrixObject bias,
                                                     MatrixObject runningMean,
                                                     MatrixObject runningVar,
                                                     MatrixObject ret,
                                                     MatrixObject retRunningMean,
                                                     MatrixObject retRunningVar,
                                                     double epsilon,
                                                     double exponentialAverageFactor,
                                                     MatrixObject resultSaveMean,
                                                     MatrixObject resultSaveInvVariance)
                                              throws DMLRuntimeException

Performs the forward BatchNormalization layer computation for training

Parameters:

gCtx - a valid GPUContext

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

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

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

Throws:

DMLRuntimeException - if error occurs

batchNormalizationForwardInference

public static void batchNormalizationForwardInference(GPUContext gCtx,
                                                      String instName,
                                                      MatrixObject image,
                                                      MatrixObject scale,
                                                      MatrixObject bias,
                                                      MatrixObject runningMean,
                                                      MatrixObject runningVar,
                                                      MatrixObject ret,
                                                      double epsilon)
                                               throws DMLRuntimeException

Performs the forward BatchNormalization layer computation for inference

Parameters:

gCtx - a valid GPUContext

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

batchNormalizationBackward

public static void batchNormalizationBackward(GPUContext gCtx,
                                              String instName,
                                              MatrixObject image,
                                              MatrixObject dout,
                                              MatrixObject scale,
                                              MatrixObject dX,
                                              MatrixObject dScale,
                                              MatrixObject dBias,
                                              double epsilon,
                                              MatrixObject resultSaveMean,
                                              MatrixObject resultSaveInvVariance)
                                       throws DMLRuntimeException

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

Parameters:

gCtx - a valid GPUContext

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]

dX - (output) backpropagation errors for previous layer

dScale - backpropagation error for scale

dBias - backpropagation error for bias

epsilon - epsilon value used in the batch normalization formula

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

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

Throws:

DMLRuntimeException - if error occurs

getDensePointerForCuDNN

public static jcuda.Pointer getDensePointerForCuDNN(GPUContext gCtx,
                                                    MatrixObject image,
                                                    String instName,
                                                    int numRows,
                                                    int numCols)
                                             throws DMLRuntimeException

Convenience method to get jcudaDenseMatrixPtr. This method explicitly converts sparse to dense format, so use it judiciously.

Parameters:

gCtx - a valid GPUContext

image - input matrix object

instName - name of the instruction

numRows - expected number of rows

numCols - expected number of columns

Returns:

jcuda pointer

Throws:

DMLRuntimeException - if error occurs while sparse to dense conversion