template<typename AReal = Float_t>
class TMVA::DNN::TCuda< AReal >

The TCuda architecture class.

Low-level interface class for CUDA computing architectures. Contains as public types the declaration of the scalar, matrix and buffer types for this architecture as well as the remaining functions in the low-level interface in the form of static members.

Definition at line 64 of file Cuda.h.

Public Types
using	ActivationDescriptor_t = CudaActivationDescriptor

using	AFloat = AReal

using	AlgorithmBackward_t = CudaConvolutionBwdDataAlgo

using	AlgorithmDataType_t = DummyCudaDataType

using	AlgorithmForward_t = CudaConvolutionFwdAlgo

using	AlgorithmHelper_t = CudaConvolutionBwdFilterAlgo

using	BNormDescriptors_t = TDNNGenDescriptors< BNormLayer_t >

using	BNormLayer_t = TBatchNormLayer< TCuda< AReal > >

using	ConvDescriptors_t = CNN::TCNNDescriptors< ConvLayer_t >

using	ConvLayer_t = CNN::TConvLayer< TCuda< AReal > >

using	ConvolutionDescriptor_t = CudaConvolutionDescriptor

using	ConvWorkspace_t = CNN::TCNNWorkspace< ConvLayer_t >

using	DeviceBuffer_t = TCudaDeviceBuffer< AFloat >

using	DropoutDescriptor_t = CudaDropoutDescriptor

using	EmptyDescriptor_t = CudaEmptyDescriptor

using	FilterDescriptor_t = CudaFilterDescriptor

using	GenLayer_t = VGeneralLayer< TCuda< AReal > >

using	HostBuffer_t = TCudaHostBuffer< AFloat >

using	Matrix_t = TCudaMatrix< AFloat >

using	PoolingDescriptor_t = CudaPoolingDescriptor

using	PoolingDescriptors_t = CNN::TCNNDescriptors< PoolingLayer_t >

using	PoolingLayer_t = CNN::TMaxPoolLayer< TCuda< AReal > >

using	PoolingWorkspace_t = CNN::TCNNWorkspace< PoolingLayer_t >

using	RecurrentDescriptor_t = DummyCudaDataType

using	ReduceTensorDescriptor_t = DummyCudaDataType

using	RNNDescriptors_t = RNN::TRNNDescriptors< TCuda< AReal > >

using	RNNWorkspace_t = RNN::TRNNWorkspace< TCuda< AReal > >

using	Scalar_t = AFloat

using	Tensor_t = TCudaTensor< AFloat >

using	TensorDescriptor_t = DummyCudaDataType

Public Member Functions
template<typename AMatrix_t >
void	CopyDiffArch (std::vector< TCudaMatrix< AFloat > > &B, const std::vector< AMatrix_t > &A)

template<typename AMatrix_t >
void	CopyDiffArch (TCudaMatrix< AFloat > &B, const AMatrix_t &A)

void	Multiply (TCudaMatrix< double > &C, const TCudaMatrix< double > &A, const TCudaMatrix< double > &B)

void	Multiply (TCudaMatrix< float > &C, const TCudaMatrix< float > &A, const TCudaMatrix< float > &B)

void	MultiplyTranspose (TCudaMatrix< double > &output, const TCudaMatrix< double > &input, const TCudaMatrix< double > &Weights)

void	MultiplyTranspose (TCudaMatrix< float > &output, const TCudaMatrix< float > &input, const TCudaMatrix< float > &Weights)

void	ScaleAdd (TCudaMatrix< double > &B, const TCudaMatrix< double > &A, double alpha)

void	ScaleAdd (TCudaMatrix< float > &B, const TCudaMatrix< float > &A, float alpha)

void	SumColumns (TCudaMatrix< double > &B, const TCudaMatrix< double > &A, double alpha, double beta)

void	SumColumns (TCudaMatrix< float > &B, const TCudaMatrix< float > &A, float alpha, float beta)

void	SumRows (TCudaMatrix< double > &B, const TCudaMatrix< double > &A)

void	SumRows (TCudaMatrix< float > &B, const TCudaMatrix< float > &A)

void	TransposeMultiply (TCudaMatrix< double > &C, const TCudaMatrix< double > &A, const TCudaMatrix< double > &B, double alpha, double beta)

void	TransposeMultiply (TCudaMatrix< float > &C, const TCudaMatrix< float > &A, const TCudaMatrix< float > &B, float alpha, float beta)

Static Public Member Functions
static void	ConvLayerForward (Tensor_t &output, Tensor_t &inputActivationFunc, const Tensor_t &input, const Matrix_t &weights, const Matrix_t &biases, const DNN::CNN::TConvParams &params, EActivationFunction activFunc, Tensor_t &, const ConvDescriptors_t &, ConvWorkspace_t &)
	Forward propagation in the Convolutional layer.

static Tensor_t	CreateTensor (DeviceBuffer_t buffer, size_t b, size_t t, size_t w)

static Tensor_t	CreateTensor (DeviceBuffer_t buffer, size_t n, size_t c, size_t h, size_t w)

static Tensor_t	CreateTensor (size_t b, size_t t, size_t w)

static Tensor_t	CreateTensor (size_t n, size_t c, size_t h, size_t w)

static void	CreateWeightTensors (std::vector< Matrix_t > &newWeights, const std::vector< Matrix_t > &weights)

static void	FreeConvWorkspace (TWorkspace *&)
	Only used for certain cudnn on-device memory.

static void	FreePoolDropoutWorkspace (TWorkspace *&)

static void	FreeRNNWorkspace (TWorkspace *&)

static TMVA::Experimental::MemoryLayout	GetTensorLayout ()

static void	InitializeActivationDescriptor (ActivationDescriptor_t &, EActivationFunction, double=0.0)

static void	InitializeBNormDescriptors (TDescriptors &, BNormLayer_t )
	Initialize CNN data/operator descriptors.

static void	InitializeConvDescriptors (TDescriptors &, ConvLayer_t )

static void	InitializeConvWorkspace (TWorkspace &, TDescriptors &, const DNN::CNN::TConvParams &, ConvLayer_t *)

static void	InitializeGRUDescriptors (TDescriptors &, GenLayer_t )

static void	InitializeGRUTensors (GenLayer_t *)

static void	InitializeGRUWorkspace (TWorkspace &, TDescriptors &, GenLayer_t *)

static void	InitializeLSTMDescriptors (TDescriptors &, GenLayer_t )

static void	InitializeLSTMTensors (GenLayer_t *)

static void	InitializeLSTMWorkspace (TWorkspace &, TDescriptors &, GenLayer_t *)

static void	InitializePoolDescriptors (TDescriptors &, PoolingLayer_t )

static void	InitializePoolDropoutWorkspace (TWorkspace &, TDescriptors &, const DNN::CNN::TConvParams &, PoolingLayer_t *)

static void	InitializeRNNDescriptors (TDescriptors &, GenLayer_t )

static void	InitializeRNNTensors (GenLayer_t *)

static void	InitializeRNNWorkspace (TWorkspace &, TDescriptors &, GenLayer_t *)

static bool	IsCudnn ()

static void	PrepareInternals (Tensor_t &)
	Dummy placeholder - preparation is currently only required for the CUDA architecture.

static void	ReleaseBNormDescriptors (TDescriptors *&)

static void	ReleaseConvDescriptors (TDescriptors *&)
	Release CNN data/operator descriptors.

static void	ReleaseDescriptor (ActivationDescriptor_t &)

static void	ReleasePoolDescriptors (TDescriptors *&)

static void	ReleaseRNNDescriptors (TDescriptors *&)

Forward Propagation
Low-level functions required for the forward propagation of activations through the network.
static void	MultiplyTranspose (Matrix_t &output, const Matrix_t &input, const Matrix_t &weights)
	Matrix-multiply `input` with the transpose of `weights` and write the results into `output`.

static void	MultiplyTranspose (Tensor_t &output, const Tensor_t &input, const Matrix_t &weights)

static void	AddRowWise (Matrix_t &output, const Matrix_t &biases)
	Add the vectors biases row-wise to the matrix output.

static void	AddRowWise (Tensor_t &output, const Matrix_t &biases)

Backward Propagation (Dense Layers)
Low-level functions required for the forward propagation of activations through the network.
static void	Backward (Tensor_t &activationGradientsBackward, Matrix_t &weightGradients, Matrix_t &biasGradients, const Tensor_t &df, const Tensor_t &activationGradients, const Matrix_t &weights, const Tensor_t &activationBackward)
	Perform the complete backward propagation step.

static void	ScaleAdd (Matrix_t &A, const Matrix_t &B, Scalar_t beta=1.0)
	Adds a the elements in matrix B scaled by c to the elements in the matrix A.

static void	Copy (Matrix_t &B, const Matrix_t &A)

template<typename AMatrix_t >
static void	CopyDiffArch (Matrix_t &B, const AMatrix_t &A)

static void	ScaleAdd (Tensor_t &A, const Tensor_t &B, Scalar_t beta=1.0)
	Above functions extended to vectors.

static void	Copy (Tensor_t &A, const Tensor_t &B)

template<typename ATensor_t >
static void	CopyDiffArch (Tensor_t &A, const ATensor_t &B)

template<typename AMatrix_t >
static void	CopyDiffArch (std::vector< Matrix_t > &A, const std::vector< AMatrix_t > &B)

Activation Functions
For each activation function, the low-level interface contains two routines. One that applies the activation function to a matrix and one that evaluate the derivatives of the activation function at the elements of a given matrix and writes the results into the result matrix.
static void	ActivationFunctionForward (Tensor_t &X, EActivationFunction activFunct, const ActivationDescriptor_t activationDescr, const double coef=0.0, const AFloat alpha=1, const AFloat beta=0)

static void	ActivationFunctionBackward (Tensor_t &dX, const Tensor_t &Y, const Tensor_t &dY, const Tensor_t &X, EActivationFunction activFunct, const ActivationDescriptor_t activationDescr, const AFloat alpha=1, const AFloat beta=0)
	Computes the gradient of the activation function.

static void	IdentityDerivative (Tensor_t &B, const Tensor_t &A)

static void	Relu (Tensor_t &B)

static void	ReluDerivative (Tensor_t &B, const Tensor_t &A)

static void	Sigmoid (Tensor_t &B)

static void	SigmoidDerivative (Tensor_t &B, const Tensor_t &A)

static void	Tanh (Tensor_t &B)

static void	TanhDerivative (Tensor_t &B, const Tensor_t &A)

static void	FastTanh (Tensor_t &B)

static void	FastTanhDerivative (Tensor_t &B, const Tensor_t &A)

static void	SymmetricRelu (Tensor_t &B)

static void	SymmetricReluDerivative (Tensor_t &B, const Tensor_t &A)

static void	SoftSign (Tensor_t &B)

static void	SoftSignDerivative (Tensor_t &B, const Tensor_t &A)

static void	Gauss (Tensor_t &B)

static void	GaussDerivative (Tensor_t &B, const Tensor_t &A)

Loss Functions
Loss functions compute a scalar value given the `output` of the network for a given training input and the expected network prediction `Y` that quantifies the quality of the prediction. For each function also a routing that computes the gradients (suffixed by Gradients) must be provided for the starting of the backpropagation algorithm.
static Scalar_t	MeanSquaredError (const Matrix_t &Y, const Matrix_t &output, const Matrix_t &weights)

static void	MeanSquaredErrorGradients (Matrix_t &dY, const Matrix_t &Y, const Matrix_t &output, const Matrix_t &weights)

static Scalar_t	CrossEntropy (const Matrix_t &Y, const Matrix_t &output, const Matrix_t &weights)
	Sigmoid transformation is implicitly applied, thus `output` should hold the linear activations of the last layer in the net.

static void	CrossEntropyGradients (Matrix_t &dY, const Matrix_t &Y, const Matrix_t &output, const Matrix_t &weights)

static Scalar_t	SoftmaxCrossEntropy (const Matrix_t &Y, const Matrix_t &output, const Matrix_t &weights)
	Softmax transformation is implicitly applied, thus `output` should hold the linear activations of the last layer in the net.

static void	SoftmaxCrossEntropyGradients (Matrix_t &dY, const Matrix_t &Y, const Matrix_t &output, const Matrix_t &weights)

Output Functions
Output functions transform the activations `output` of the output layer in the network to a valid prediction `YHat` for the desired usage of the network, e.g. the identity function for regression or the sigmoid transformation for two-class classification.
static void	Sigmoid (Matrix_t &YHat, const Matrix_t &)

static void	Softmax (Matrix_t &YHat, const Matrix_t &)

Regularization
For each regularization type two functions are required, one named `<Type>Regularization` that evaluates the corresponding regularization functional for a given weight matrix and the `Add<Type>RegularizationGradients`, that adds the regularization component in the gradients to the provided matrix.
static Scalar_t	L1Regularization (const Matrix_t &W)

static void	AddL1RegularizationGradients (Matrix_t &A, const Matrix_t &W, Scalar_t weightDecay)

static Scalar_t	L2Regularization (const Matrix_t &W)

static void	AddL2RegularizationGradients (Matrix_t &A, const Matrix_t &W, Scalar_t weightDecay)

Initialization
For each initialization method, one function in the low-level interface is provided. The naming scheme is Initialize<Type> for a given initialization method Type.
static void	InitializeGauss (Matrix_t &A)

static void	InitializeUniform (Matrix_t &A)

static void	InitializeIdentity (Matrix_t &A)

static void	InitializeZero (Matrix_t &A)

static void	InitializeZero (Tensor_t &A)

static void	InitializeGlorotNormal (Matrix_t &A)
	Truncated normal initialization (Glorot, called also Xavier normal) The values are sample with a normal distribution with stddev = sqrt(2/N_input + N_output) and values larger than 2 * stddev are discarded See Glorot & Bengio, AISTATS 2010 - http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf.

static void	InitializeGlorotUniform (Matrix_t &A)
	Sample from a uniform distribution in range [ -lim,+lim] where lim = sqrt(6/N_in+N_out).

static TRandom &	GetRandomGenerator ()

static void	SetRandomSeed (size_t seed)

Dropout
static void	DropoutForward (Tensor_t &A, TDescriptors descriptors, TWorkspace workspace, Scalar_t p)
	Apply dropout with activation probability `p` to the given tensor `A` and scale the result by reciprocal of `p`.

static void	DropoutForward (Matrix_t &A, Scalar_t p)

static void	DropoutBackward (Tensor_t &, TDescriptors , TWorkspace )

Batch Normalization Layer Propagation
static void	BatchNormLayerForwardTraining (int axis, const Tensor_t &x, Tensor_t &y, Matrix_t &gamma, Matrix_t &beta, Matrix_t &mean, Matrix_t &, Matrix_t &iVariance, Matrix_t &runningMeans, Matrix_t &runningVars, Scalar_t nTrainedBatches, Scalar_t momentum, Scalar_t epsilon, const TensorDescriptor_t &bnParDescriptor)
	The input from each batch are normalized during training to have zero mean and unit variance and they are then scaled by two parameter, different for each input variable:

static void	BatchNormLayerForwardInference (int axis, const Tensor_t &x, Matrix_t &gamma, Matrix_t &beta, Tensor_t &y, const Matrix_t &runningMeans, const Matrix_t &runningVars, Scalar_t epsilon, const TensorDescriptor_t &)
	During inference the inputs are not normalized using the batch mean but the previously computed at running mean and variance.

static void	BatchNormLayerBackward (int axis, const Tensor_t &x, const Tensor_t &dy, Tensor_t &dx, Matrix_t &gamma, Matrix_t &dgamma, Matrix_t &dbeta, const Matrix_t &mean, const Matrix_t &variance, const Matrix_t &iVariance, Scalar_t epsilon, const TensorDescriptor_t &)

Forward Propagation in Convolutional Layer
static size_t	calculateDimension (size_t imgDim, size_t fltDim, size_t padding, size_t stride)
	Calculate how many neurons "fit" in the output layer, given the input as well as the layer's hyperparameters.

static void	Im2col (Matrix_t &A, const Matrix_t &B, size_t imgHeight, size_t imgWidth, size_t fltHeight, size_t fltWidth, size_t strideRows, size_t strideCols, size_t zeroPaddingHeight, size_t zeroPaddingWidth)
	Transform the matrix B in local view format, suitable for convolution, and store it in matrix A.

static void	Im2colIndices (std::vector< int > &V, const Matrix_t &B, size_t nLocalViews, size_t imgHeight, size_t imgWidth, size_t fltHeight, size_t fltWidth, size_t strideRows, size_t strideCols, size_t zeroPaddingHeight, size_t zeroPaddingWidth)

static void	Im2colFast (Matrix_t &A, const Matrix_t &B, const std::vector< int > &V)

static void	RotateWeights (Matrix_t &A, const Matrix_t &B, size_t filterDepth, size_t filterHeight, size_t filterWidth, size_t numFilters)
	Rotates the matrix `B`, which is representing a weights, and stores them in the matrix `A`.

static void	AddConvBiases (Matrix_t &output, const Matrix_t &biases)
	Add the biases in the Convolutional Layer.

Backward Propagation in Convolutional Layer
static void	ConvLayerBackward (Tensor_t &activationGradientsBackward, Matrix_t &weightGradients, Matrix_t &biasGradients, Tensor_t &df, Tensor_t &activationGradients, const Matrix_t &weights, const Tensor_t &activationBackward, const Tensor_t &outputTensor, EActivationFunction activFunc, const ConvDescriptors_t &, ConvWorkspace_t &, size_t batchSize, size_t inputHeight, size_t inputWidth, size_t depth, size_t height, size_t width, size_t filterDepth, size_t filterHeight, size_t filterWidth, size_t nLocalViews)
	Perform the complete backward propagation step in a Convolutional Layer.

static void	CalculateConvActivationGradients (Tensor_t &activationGradientsBackward, const Tensor_t &df, const Matrix_t &weights, size_t batchSize, size_t inputHeight, size_t inputWidth, size_t depth, size_t height, size_t width, size_t filterDepth, size_t filterHeight, size_t filterWidth)
	Utility function for calculating the activation gradients of the layer before the convolutional layer.

static void	CalculateConvWeightGradients (Matrix_t &weightGradients, const Tensor_t &df, const Tensor_t &activations_backward, size_t batchSize, size_t inputHeight, size_t inputWidth, size_t depth, size_t height, size_t width, size_t filterDepth, size_t filterHeight, size_t filterWidth, size_t nLocalViews)
	Utility function for calculating the weight gradients of the convolutional layer.

static void	CalculateConvBiasGradients (Matrix_t &biasGradients, const Tensor_t &df, size_t batchSize, size_t depth, size_t nLocalViews)
	Utility function for calculating the bias gradients of the convolutional layer.

Forward Propagation in Max Pooling Layer
static void	Downsample (Tensor_t &A, Tensor_t &B, const Tensor_t &C, const PoolingDescriptors_t &, PoolingWorkspace_t &, size_t imgHeight, size_t imgWidth, size_t fltHeight, size_t fltWidth, size_t strideRows, size_t strideCols)
	Downsample the matrix `C` to the matrix `A`, using max operation, such that the winning indices are stored in matrix `B`.

Backward Propagation in Max Pooling Layer
static void	MaxPoolLayerBackward (Tensor_t &activationGradientsBackward, const Tensor_t &activationGradients, const Tensor_t &indexMatrix, const Tensor_t &, const Tensor_t &, const PoolingDescriptors_t &, PoolingWorkspace_t &, size_t imgHeight, size_t imgWidth, size_t fltHeight, size_t fltWidth, size_t strideRows, size_t strideCols, size_t nLocalViews)
	Perform the complete backward propagation step in a Pooling Layer.

Forward and Backward Propagation in Reshape Layer
static void	Reshape (Matrix_t &A, const Matrix_t &B)
	Transform the matrix `B` to a matrix with different dimensions `A`.

static void	Flatten (Tensor_t &A, const Tensor_t &B)
	Flattens the tensor `B`, such that each matrix, is stretched in one row, resulting with a matrix `A`.

static void	Deflatten (Tensor_t &A, const Tensor_t &B)
	Transforms each row of `B` to a matrix and stores it in the tensor `B`.

static void	Rearrange (Tensor_t &out, const Tensor_t &in)
	Rearrage data according to time fill B x T x D out with T x B x D matrix in.

static Matrix_t &	RecurrentLayerBackward (Matrix_t &state_gradients_backward, Matrix_t &input_weight_gradients, Matrix_t &state_weight_gradients, Matrix_t &bias_gradients, Matrix_t &df, const Matrix_t &state, const Matrix_t &weights_input, const Matrix_t &weights_state, const Matrix_t &input, Matrix_t &input_gradient)
	Backward pass for Recurrent Networks.

static void	RNNForward (const Tensor_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, Tensor_t &, Matrix_t &, Matrix_t &, const RNNDescriptors_t &, RNNWorkspace_t &, bool)

static void	RNNBackward (const Tensor_t &, const Matrix_t &, const Matrix_t &, const Tensor_t &, const Tensor_t &, const Matrix_t &, const Matrix_t &, const Tensor_t &, Tensor_t &, Matrix_t &, Matrix_t &, Tensor_t &, const RNNDescriptors_t &, RNNWorkspace_t &)

static Matrix_t &	LSTMLayerBackward (Matrix_t &state_gradients_backward, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &)

static Matrix_t &	GRULayerBackward (Matrix_t &state_gradients_backward, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, const Matrix_t &, Matrix_t &, bool)
	Backward pass for GRU Network.

Additional Arithmetic Functions
Additional arithmetic on CUDA matrices used to implement the low-level interface.
static void	Multiply (Matrix_t &C, const Matrix_t &A, const Matrix_t &B)
	Standard multiplication of two matrices `A` and `B` with the result being written into C.

static void	TransposeMultiply (Matrix_t &output, const Matrix_t &input, const Matrix_t &Weights, Scalar_t alpha=1.0, Scalar_t beta=0.)
	Matrix multiplication of two matrices `A` and `B^T` (transposed) with the result being written into C.

static void	Hadamard (Tensor_t &A, const Tensor_t &B)
	In-place Hadamard (element-wise) product of matrices `A` and `B` with the result being written into `A`.

static void	Hadamard (Matrix_t &A, const Matrix_t &B)

static void	SumColumns (Matrix_t &B, const Matrix_t &A, Scalar_t alpha=1.0, Scalar_t beta=0.)
	Sum columns of (m x n) matrix `A` and write the results into the first m elements in `A`.

static Scalar_t	Sum (const Matrix_t &A)
	Compute the sum of all elements in `A`.

static bool	AlmostEquals (const Matrix_t &A, const Matrix_t &B, double epsilon=0.1)
	Check two matrices for equality, taking floating point arithmetic errors into account.

static void	ConstAdd (Matrix_t &A, Scalar_t beta)
	Add the constant `beta` to all the elements of matrix `A` and write the result into `A`.

static void	ConstMult (Matrix_t &A, Scalar_t beta)
	Multiply the constant `beta` to all the elements of matrix `A` and write the result into `A`.

static void	ReciprocalElementWise (Matrix_t &A)
	Reciprocal each element of the matrix `A` and write the result into `A`.

static void	SquareElementWise (Matrix_t &A)
	Square each element of the matrix `A` and write the result into `A`.

static void	SqrtElementWise (Matrix_t &A)
	Square root each element of the matrix `A` and write the result into `A`.

static void	AdamUpdate (Matrix_t &A, const Matrix_t &M, const Matrix_t &V, Scalar_t alpha, Scalar_t eps)
	Adam updates.

static void	AdamUpdateFirstMom (Matrix_t &A, const Matrix_t &B, Scalar_t beta)

static void	AdamUpdateSecondMom (Matrix_t &A, const Matrix_t &B, Scalar_t beta)

static void	PrintTensor (const Tensor_t &A, const std::string name="Cuda-tensor", bool=false)

static void	SumRows (Matrix_t &B, const Matrix_t &A)
	extra functions defined only for CPU architecture !!!

Static Private Attributes
static TRandom *	fgRandomGen = nullptr

#include <TMVA/DNN/Architectures/Cuda.h>

Member Typedef Documentation

◆ ActivationDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::ActivationDescriptor_t = CudaActivationDescriptor

Definition at line 78 of file Cuda.h.

◆ AFloat

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::AFloat = AReal

Definition at line 70 of file Cuda.h.

◆ AlgorithmBackward_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::AlgorithmBackward_t = CudaConvolutionBwdDataAlgo

Definition at line 87 of file Cuda.h.

◆ AlgorithmDataType_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::AlgorithmDataType_t = DummyCudaDataType

Definition at line 89 of file Cuda.h.

◆ AlgorithmForward_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::AlgorithmForward_t = CudaConvolutionFwdAlgo

Definition at line 86 of file Cuda.h.

◆ AlgorithmHelper_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::AlgorithmHelper_t = CudaConvolutionBwdFilterAlgo

Definition at line 88 of file Cuda.h.

◆ BNormDescriptors_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::BNormDescriptors_t = TDNNGenDescriptors<BNormLayer_t>

Definition at line 97 of file Cuda.h.

◆ BNormLayer_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::BNormLayer_t = TBatchNormLayer<TCuda<AReal> >

Definition at line 96 of file Cuda.h.

◆ ConvDescriptors_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::ConvDescriptors_t = CNN::TCNNDescriptors<ConvLayer_t>

Definition at line 100 of file Cuda.h.

◆ ConvLayer_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::ConvLayer_t = CNN::TConvLayer<TCuda<AReal> >

Definition at line 99 of file Cuda.h.

◆ ConvolutionDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::ConvolutionDescriptor_t = CudaConvolutionDescriptor

Definition at line 79 of file Cuda.h.

◆ ConvWorkspace_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::ConvWorkspace_t = CNN::TCNNWorkspace<ConvLayer_t>

Definition at line 101 of file Cuda.h.

◆ DeviceBuffer_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::DeviceBuffer_t = TCudaDeviceBuffer<AFloat>

Definition at line 75 of file Cuda.h.

◆ DropoutDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::DropoutDescriptor_t = CudaDropoutDescriptor

Definition at line 81 of file Cuda.h.

◆ EmptyDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::EmptyDescriptor_t = CudaEmptyDescriptor

Definition at line 93 of file Cuda.h.

◆ FilterDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::FilterDescriptor_t = CudaFilterDescriptor

Definition at line 80 of file Cuda.h.

◆ GenLayer_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::GenLayer_t = VGeneralLayer<TCuda<AReal> >

Definition at line 95 of file Cuda.h.

◆ HostBuffer_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::HostBuffer_t = TCudaHostBuffer<AFloat>

Definition at line 76 of file Cuda.h.

◆ Matrix_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::Matrix_t = TCudaMatrix<AFloat>

Definition at line 73 of file Cuda.h.

◆ PoolingDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::PoolingDescriptor_t = CudaPoolingDescriptor

Definition at line 83 of file Cuda.h.

◆ PoolingDescriptors_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::PoolingDescriptors_t = CNN::TCNNDescriptors<PoolingLayer_t>

Definition at line 103 of file Cuda.h.

◆ PoolingLayer_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::PoolingLayer_t = CNN::TMaxPoolLayer<TCuda<AReal> >

Definition at line 102 of file Cuda.h.

◆ PoolingWorkspace_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::PoolingWorkspace_t = CNN::TCNNWorkspace<PoolingLayer_t>

Definition at line 104 of file Cuda.h.

◆ RecurrentDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::RecurrentDescriptor_t = DummyCudaDataType

Definition at line 91 of file Cuda.h.

◆ ReduceTensorDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::ReduceTensorDescriptor_t = DummyCudaDataType

Definition at line 90 of file Cuda.h.

◆ RNNDescriptors_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::RNNDescriptors_t = RNN::TRNNDescriptors<TCuda<AReal> >

Definition at line 106 of file Cuda.h.

◆ RNNWorkspace_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::RNNWorkspace_t = RNN::TRNNWorkspace<TCuda<AReal> >

Definition at line 107 of file Cuda.h.

◆ Scalar_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::Scalar_t = AFloat

Definition at line 71 of file Cuda.h.

◆ Tensor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::Tensor_t = TCudaTensor<AFloat>

Definition at line 74 of file Cuda.h.

◆ TensorDescriptor_t

template<typename AReal = Float_t>

using TMVA::DNN::TCuda< AReal >::TensorDescriptor_t = DummyCudaDataType

Definition at line 84 of file Cuda.h.

Member Function Documentation

◆ ActivationFunctionBackward()

template<typename AFloat >

void TMVA::DNN::TCuda< AFloat >::ActivationFunctionBackward	(	Tensor_t &	dX,
		const Tensor_t &	Y,
		const Tensor_t &	dY,
		const Tensor_t &	X,
		EActivationFunction	activFunct,
		const ActivationDescriptor_t	activationDescr,
		const AFloat	alpha = `1`,
		const AFloat	beta = `0`
	)

static

Computes the gradient of the activation function.

Definition at line 37 of file ActivationFunctions.cu.

◆ ActivationFunctionForward()

template<typename AFloat >

void TMVA::DNN::TCuda< AFloat >::ActivationFunctionForward	(	Tensor_t &	X,
		EActivationFunction	activFunct,
		const ActivationDescriptor_t	activationDescr,
		const double	coef = `0.0`,
		const AFloat	alpha = `1`,
		const AFloat	beta = `0`
	)

static

Definition at line 28 of file ActivationFunctions.cu.

◆ AdamUpdate()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::AdamUpdate	(	Matrix_t &	A,
		const Matrix_t &	M,
		const Matrix_t &	V,
		Scalar_t	alpha,
		Scalar_t	eps
	)

static

Adam updates.

Definition at line 425 of file Arithmetic.cu.

◆ AdamUpdateFirstMom()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::AdamUpdateFirstMom	(	Matrix_t &	A,
		const Matrix_t &	B,
		Scalar_t	beta
	)

static

Definition at line 441 of file Arithmetic.cu.

◆ AdamUpdateSecondMom()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::AdamUpdateSecondMom	(	Matrix_t &	A,
		const Matrix_t &	B,
		Scalar_t	beta
	)

static

Definition at line 455 of file Arithmetic.cu.

◆ AddConvBiases()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::AddConvBiases	(	Matrix_t &	output,
		const Matrix_t &	biases
	)

static

Add the biases in the Convolutional Layer.

Definition at line 432 of file Propagation.cu.

◆ AddL1RegularizationGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::AddL1RegularizationGradients	(	Matrix_t &	A,
		const Matrix_t &	W,
		Scalar_t	weightDecay
	)

static

Definition at line 43 of file Regularization.cu.

◆ AddL2RegularizationGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::AddL2RegularizationGradients	(	Matrix_t &	A,
		const Matrix_t &	W,
		Scalar_t	weightDecay
	)

static

Definition at line 76 of file Regularization.cu.

◆ AddRowWise() [1/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::AddRowWise	(	Matrix_t &	output,
		const Matrix_t &	biases
	)

static

Add the vectors biases row-wise to the matrix output.

◆ AddRowWise() [2/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::AddRowWise	(	Tensor_t &	output,
		const Matrix_t &	biases
	)

inlinestatic

Definition at line 217 of file Cuda.h.

◆ AlmostEquals()

template<typename AReal = Float_t>

bool TMVA::DNN::TCuda< AFloat >::AlmostEquals	(	const Matrix_t &	A,
		const Matrix_t &	B,
		double	epsilon = `0.1`
	)

static

Check two matrices for equality, taking floating point arithmetic errors into account.

Checks two matrices for element-wise equality.

Template Parameters

AFloat An architecture-specific floating point number type.

Parameters

A	The first matrix.
B	The second matrix.
epsilon	Equality tolerance, needed to address floating point arithmetic.

Returns: Whether the two matrices can be considered equal element-wise

Definition at line 291 of file Arithmetic.cu.

◆ Backward()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Backward	(	Tensor_t &	activationGradientsBackward,
		Matrix_t &	weightGradients,
		Matrix_t &	biasGradients,
		const Tensor_t &	df,
		const Tensor_t &	activationGradients,
		const Matrix_t &	weights,
		const Tensor_t &	activationBackward
	)

static

Perform the complete backward propagation step.

If the provided activationGradientsBackward matrix is not empty, compute the gradients of the objective function with respect to the activations of the previous layer (backward direction). Also compute the weight and the bias gradients. Modifies the values in df and thus produces only a valid result, if it is applied the first time after the corresponding forward propagation has been per- formed.

Definition at line 91 of file Propagation.cu.

◆ BatchNormLayerBackward()

template<typename AFloat >

void TMVA::DNN::TCuda< AFloat >::BatchNormLayerBackward	(	int	axis,
		const Tensor_t &	x,
		const Tensor_t &	dy,
		Tensor_t &	dx,
		Matrix_t &	gamma,
		Matrix_t &	dgamma,
		Matrix_t &	dbeta,
		const Matrix_t &	mean,
		const Matrix_t &	variance,
		const Matrix_t &	iVariance,
		Scalar_t	epsilon,
		const TensorDescriptor_t &
	)

static

Definition at line 754 of file Propagation.cu.

◆ BatchNormLayerForwardInference()

template<typename AFloat >

void TMVA::DNN::TCuda< AFloat >::BatchNormLayerForwardInference	(	int	axis,
		const Tensor_t &	x,
		Matrix_t &	gamma,
		Matrix_t &	beta,
		Tensor_t &	y,
		const Matrix_t &	runningMeans,
		const Matrix_t &	runningVars,
		Scalar_t	epsilon,
		const TensorDescriptor_t &
	)

static

During inference the inputs are not normalized using the batch mean but the previously computed at running mean and variance.

Definition at line 743 of file Propagation.cu.

◆ BatchNormLayerForwardTraining()

template<typename AFloat >

void TMVA::DNN::TCuda< AFloat >::BatchNormLayerForwardTraining	(	int	axis,
		const Tensor_t &	x,
		Tensor_t &	y,
		Matrix_t &	gamma,
		Matrix_t &	beta,
		Matrix_t &	mean,
		Matrix_t &	,
		Matrix_t &	iVariance,
		Matrix_t &	runningMeans,
		Matrix_t &	runningVars,
		Scalar_t	nTrainedBatches,
		Scalar_t	momentum,
		Scalar_t	epsilon,
		const TensorDescriptor_t &	bnParDescriptor
	)

static

The input from each batch are normalized during training to have zero mean and unit variance and they are then scaled by two parameter, different for each input variable:

a scale factor \gamma gamma
an offset \beta beta

Definition at line 729 of file Propagation.cu.

◆ CalculateConvActivationGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::CalculateConvActivationGradients	(	Tensor_t &	activationGradientsBackward,
		const Tensor_t &	df,
		const Matrix_t &	weights,
		size_t	batchSize,
		size_t	inputHeight,
		size_t	inputWidth,
		size_t	depth,
		size_t	height,
		size_t	width,
		size_t	filterDepth,
		size_t	filterHeight,
		size_t	filterWidth
	)

static

Utility function for calculating the activation gradients of the layer before the convolutional layer.

Definition at line 324 of file Propagation.cu.

◆ CalculateConvBiasGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::CalculateConvBiasGradients	(	Matrix_t &	biasGradients,
		const Tensor_t &	df,
		size_t	batchSize,
		size_t	depth,
		size_t	nLocalViews
	)

static

Utility function for calculating the bias gradients of the convolutional layer.

Definition at line 416 of file Propagation.cu.

◆ CalculateConvWeightGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::CalculateConvWeightGradients	(	Matrix_t &	weightGradients,
		const Tensor_t &	df,
		const Tensor_t &	activations_backward,
		size_t	batchSize,
		size_t	inputHeight,
		size_t	inputWidth,
		size_t	depth,
		size_t	height,
		size_t	width,
		size_t	filterDepth,
		size_t	filterHeight,
		size_t	filterWidth,
		size_t	nLocalViews
	)

static

Utility function for calculating the weight gradients of the convolutional layer.

Definition at line 369 of file Propagation.cu.

◆ calculateDimension()

template<typename AFloat >

size_t TMVA::DNN::TCuda< AFloat >::calculateDimension	(	size_t	imgDim,
		size_t	fltDim,
		size_t	padding,
		size_t	stride
	)

static

Calculate how many neurons "fit" in the output layer, given the input as well as the layer's hyperparameters.

Definition at line 151 of file Propagation.cu.

◆ ConstAdd()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::ConstAdd	(	Matrix_t &	A,
		Scalar_t	beta
	)

static

Add the constant beta to all the elements of matrix A and write the result into A.

Definition at line 357 of file Arithmetic.cu.

◆ ConstMult()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::ConstMult	(	Matrix_t &	A,
		Scalar_t	beta
	)

static

Multiply the constant beta to all the elements of matrix A and write the result into A.

Definition at line 371 of file Arithmetic.cu.

◆ ConvLayerBackward()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::ConvLayerBackward	(	Tensor_t &	activationGradientsBackward,
		Matrix_t &	weightGradients,
		Matrix_t &	biasGradients,
		Tensor_t &	df,
		Tensor_t &	activationGradients,
		const Matrix_t &	weights,
		const Tensor_t &	activationBackward,
		const Tensor_t &	outputTensor,
		EActivationFunction	activFunc,
		const ConvDescriptors_t &	,
		ConvWorkspace_t &	,
		size_t	batchSize,
		size_t	inputHeight,
		size_t	inputWidth,
		size_t	depth,
		size_t	height,
		size_t	width,
		size_t	filterDepth,
		size_t	filterHeight,
		size_t	filterWidth,
		size_t	nLocalViews
	)

static

Perform the complete backward propagation step in a Convolutional Layer.

If the provided activationGradientsBackward matrix is not empty, compute the gradients of the objective function with respect to the activations of the previous layer (backward direction). Also compute the weight and the bias gradients. Modifies the values in df and thus produces only a valid result, if it is applied the first time after the corresponding forward propagation has been per- formed.

Definition at line 276 of file Propagation.cu.

◆ ConvLayerForward()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::ConvLayerForward	(	Tensor_t &	output,
		Tensor_t &	inputActivationFunc,
		const Tensor_t &	input,
		const Matrix_t &	weights,
		const Matrix_t &	biases,
		const DNN::CNN::TConvParams &	params,
		EActivationFunction	activFunc,
		Tensor_t &	,
		const ConvDescriptors_t &	,
		ConvWorkspace_t &
	)

static

Forward propagation in the Convolutional layer.

Definition at line 236 of file Propagation.cu.

◆ Copy() [1/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Copy	(	Matrix_t &	B,
		const Matrix_t &	A
	)

static

◆ Copy() [2/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Copy	(	Tensor_t &	A,
		const Tensor_t &	B
	)

static

◆ CopyDiffArch() [1/5]

template<typename AReal = Float_t>

template<typename AMatrix_t >

static void TMVA::DNN::TCuda< AReal >::CopyDiffArch	(	Matrix_t &	B,
		const AMatrix_t &	A
	)

static

◆ CopyDiffArch() [2/5]

template<typename AReal = Float_t>

template<typename AMatrix_t >

static void TMVA::DNN::TCuda< AReal >::CopyDiffArch	(	std::vector< Matrix_t > &	A,
		const std::vector< AMatrix_t > &	B
	)

static

◆ CopyDiffArch() [3/5]

template<typename AReal = Float_t>

template<typename AMatrix_t >

void TMVA::DNN::TCuda< AReal >::CopyDiffArch	(	std::vector< TCudaMatrix< AFloat > > &	B,
		const std::vector< AMatrix_t > &	A
	)

Definition at line 851 of file Cuda.h.

◆ CopyDiffArch() [4/5]

template<typename AReal = Float_t>

template<typename AMatrix_t >

void TMVA::DNN::TCuda< AReal >::CopyDiffArch	(	TCudaMatrix< AFloat > &	B,
		const AMatrix_t &	A
	)

Definition at line 839 of file Cuda.h.

◆ CopyDiffArch() [5/5]

template<typename AReal = Float_t>

template<typename ATensor_t >

static void TMVA::DNN::TCuda< AReal >::CopyDiffArch	(	Tensor_t &	A,
		const ATensor_t &	B
	)

static

◆ CreateTensor() [1/4]

template<typename AReal = Float_t>

static Tensor_t TMVA::DNN::TCuda< AReal >::CreateTensor	(	DeviceBuffer_t	buffer,
		size_t	b,
		size_t	t,
		size_t	w
	)

inlinestatic

Definition at line 121 of file Cuda.h.

◆ CreateTensor() [2/4]

template<typename AReal = Float_t>

static Tensor_t TMVA::DNN::TCuda< AReal >::CreateTensor	(	DeviceBuffer_t	buffer,
		size_t	n,
		size_t	c,
		size_t	h,
		size_t	w
	)

inlinestatic

Definition at line 118 of file Cuda.h.

◆ CreateTensor() [3/4]

template<typename AReal = Float_t>

static Tensor_t TMVA::DNN::TCuda< AReal >::CreateTensor	(	size_t	b,
		size_t	t,
		size_t	w
	)

inlinestatic

Definition at line 114 of file Cuda.h.

◆ CreateTensor() [4/4]

template<typename AReal = Float_t>

static Tensor_t TMVA::DNN::TCuda< AReal >::CreateTensor	(	size_t	n,
		size_t	c,
		size_t	h,
		size_t	w
	)

inlinestatic

Definition at line 111 of file Cuda.h.

◆ CreateWeightTensors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::CreateWeightTensors	(	std::vector< Matrix_t > &	newWeights,
		const std::vector< Matrix_t > &	weights
	)

inlinestatic

Definition at line 132 of file Cuda.h.

◆ CrossEntropy()

template<typename AReal = Float_t>

AFloat TMVA::DNN::TCuda< AFloat >::CrossEntropy	(	const Matrix_t &	Y,
		const Matrix_t &	output,
		const Matrix_t &	weights
	)

static

Sigmoid transformation is implicitly applied, thus output should hold the linear activations of the last layer in the net.

Definition at line 68 of file LossFunctions.cu.

◆ CrossEntropyGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::CrossEntropyGradients	(	Matrix_t &	dY,
		const Matrix_t &	Y,
		const Matrix_t &	output,
		const Matrix_t &	weights
	)

static

Definition at line 88 of file LossFunctions.cu.

◆ Deflatten()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Deflatten	(	Tensor_t &	A,
		const Tensor_t &	B
	)

static

Transforms each row of B to a matrix and stores it in the tensor B.

Deflatten a matrix into a vector of matrices.

Parameters

[out]	A	Output matrices. Each element will be a part of the input.
[in]	B	Input flat matrix.
[in]	size	Number of matrices in the output vector.
[in]	nRows	Number of rows in each matrix of the output vector.
[in]	nCols	Number of columns on each matrix of the output vector.

Each row in the input matrix is the concatenation of the same row in each of the output matrices. Passing an std::vector to a CUDA kernel is a non trivial task that requires manually allocating and copying to device memory - details in comments within the function's body. Launching one thread per input element.

Definition at line 670 of file Propagation.cu.

◆ Downsample()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Downsample	(	Tensor_t &	A,
		Tensor_t &	B,
		const Tensor_t &	C,
		const PoolingDescriptors_t &	,
		PoolingWorkspace_t &	,
		size_t	imgHeight,
		size_t	imgWidth,
		size_t	fltHeight,
		size_t	fltWidth,
		size_t	strideRows,
		size_t	strideCols
	)

static

Downsample the matrix C to the matrix A, using max operation, such that the winning indices are stored in matrix B.

Downsampling function used as the forward propagation step of a Max-Pooling layer.

Parameters

[out]	A	The output matrix. Each row corresponds to a slice and each element is the max within a receptive field.
[out]	B	The winning indices matrix. Each element is the index of the max element.
[in]	C	The input matrix. Each row is a slice.
[in]	imgHeight	The heigh of the input.
[in]	imgWidth	The output of the input.
[in]	fltHeight	Height of the kernel.
[in]	fltWidth	Width of the kernel.
[in]	strideRows	stride size in the horizontal dimension.
[in]	strideCols	stride size in the vertical dimension.

Each output element is the maximum of the receptive field. We also save the winning indices to facilitate back-propagation - we need to know which input element influenced the output and only apply the derivative correction to this particular element. The slicing process is the same as in a convolutional layer, however padding is set to 0.

Definition at line 468 of file Propagation.cu.

◆ DropoutBackward()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::DropoutBackward	(	Tensor_t &	,
		TDescriptors *	,
		TWorkspace *
	)

inlinestatic

Definition at line 469 of file Cuda.h.

◆ DropoutForward() [1/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::DropoutForward	(	Matrix_t &	A,
		Scalar_t	p
	)

inlinestatic

Definition at line 464 of file Cuda.h.

◆ DropoutForward() [2/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::DropoutForward	(	Tensor_t &	A,
		TDescriptors *	descriptors,
		TWorkspace *	workspace,
		Scalar_t	p
	)

static

Apply dropout with activation probability p to the given tensor A and scale the result by reciprocal of p.

◆ FastTanh()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::FastTanh ( Tensor_t & B )

inlinestatic

Definition at line 325 of file Cuda.h.

◆ FastTanhDerivative()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::FastTanhDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

inlinestatic

Definition at line 326 of file Cuda.h.

◆ Flatten()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Flatten	(	Tensor_t &	A,
		const Tensor_t &	B
	)

static

Flattens the tensor B, such that each matrix, is stretched in one row, resulting with a matrix A.

Flatten a vector of matrices into a single matrix.

Parameters

[out]	A	Output matrix.
[in]	B	Input vector. Each element is a matrix to be concatenated.
[in]	size	Number of matrices in the input vector.
[in]	nRows	Number of rows in each matrix of the input vector.
[in]	nCols	Number of columns on each matrix of the input vector.

Each row in the output matrix is the concatenation of the same row in each of the input matrices. Passing an std::vector to a CUDA kernel is a non trivial task that requires manually allocating and copying to device memory - details in comments within the function's body. Launching one thread per output element.

Definition at line 592 of file Propagation.cu.

◆ FreeConvWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::FreeConvWorkspace ( TWorkspace *& )

inlinestatic

Only used for certain cudnn on-device memory.

Definition at line 174 of file Cuda.h.

◆ FreePoolDropoutWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::FreePoolDropoutWorkspace ( TWorkspace *& )

inlinestatic

Definition at line 175 of file Cuda.h.

◆ FreeRNNWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::FreeRNNWorkspace ( TWorkspace *& )

inlinestatic

Definition at line 190 of file Cuda.h.

◆ Gauss()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Gauss ( Tensor_t & B )

static

Definition at line 218 of file ActivationFunctions.cu.

◆ GaussDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::GaussDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 231 of file ActivationFunctions.cu.

◆ GetRandomGenerator()

template<typename AFloat >

TRandom & TMVA::DNN::TCuda< AFloat >::GetRandomGenerator

static

Definition at line 37 of file Initialization.cu.

◆ GetTensorLayout()

template<typename AReal = Float_t>

static TMVA::Experimental::MemoryLayout TMVA::DNN::TCuda< AReal >::GetTensorLayout ( )

inlinestatic

Definition at line 109 of file Cuda.h.

◆ GRULayerBackward()

template<typename AReal = Float_t>

static Matrix_t & TMVA::DNN::TCuda< AReal >::GRULayerBackward	(	Matrix_t &	state_gradients_backward,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		Matrix_t &	,
		bool
	)

inlinestatic

Backward pass for GRU Network.

Definition at line 724 of file Cuda.h.

◆ Hadamard() [1/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Hadamard	(	Matrix_t &	A,
		const Matrix_t &	B
	)

static

◆ Hadamard() [2/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Hadamard	(	Tensor_t &	A,
		const Tensor_t &	B
	)

static

In-place Hadamard (element-wise) product of matrices A and B with the result being written into A.

◆ IdentityDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::IdentityDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 53 of file ActivationFunctions.cu.

◆ Im2col()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Im2col	(	Matrix_t &	A,
		const Matrix_t &	B,
		size_t	imgHeight,
		size_t	imgWidth,
		size_t	fltHeight,
		size_t	fltWidth,
		size_t	strideRows,
		size_t	strideCols,
		size_t	zeroPaddingHeight,
		size_t	zeroPaddingWidth
	)

static

Transform the matrix B in local view format, suitable for convolution, and store it in matrix A.

A helper for image operations that rearranges image regions into column vectors.

Parameters

[out]	A	The output matrix. Each row corresponds to a receptive field.
[in]	B	The input matrix. Each row corresponds to a row in the image view.
[in]	imgHeight	The heigh of the input.
[in]	imgWidth	The output of the input.
[in]	fltHeight	Height of the kernel.
[in]	fltWidth	Width of the kernel.
[in]	strideRows	stride size in the horizontal dimension.
[in]	strideCols	stride size in the vertical dimension.
[in]	zeroPaddingHeight	The padding in the horizontal dimension.
[in]	zeroPaddingWidth	The padding in the vertical dimension.

This transformation allows us to express a 2D convolution as a matrix multiplication. We can therefore harness the finely tuned GEMM implementation of cuBLAS to achieve maximum performance. This function can greatly speed-up propagation in TConvLayer.

Definition at line 183 of file Propagation.cu.

◆ Im2colFast()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Im2colFast	(	Matrix_t &	A,
		const Matrix_t &	B,
		const std::vector< int > &	V
	)

static

◆ Im2colIndices()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Im2colIndices	(	std::vector< int > &	V,
		const Matrix_t &	B,
		size_t	nLocalViews,
		size_t	imgHeight,
		size_t	imgWidth,
		size_t	fltHeight,
		size_t	fltWidth,
		size_t	strideRows,
		size_t	strideCols,
		size_t	zeroPaddingHeight,
		size_t	zeroPaddingWidth
	)

static

◆ InitializeActivationDescriptor()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeActivationDescriptor	(	ActivationDescriptor_t &	,
		EActivationFunction	,
		double	= `0.0`
	)

inlinestatic

Definition at line 156 of file Cuda.h.

◆ InitializeBNormDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeBNormDescriptors	(	TDescriptors *&	,
		BNormLayer_t *
	)

inlinestatic

Initialize CNN data/operator descriptors.

Not used at the moment.

Definition at line 147 of file Cuda.h.

◆ InitializeConvDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeConvDescriptors	(	TDescriptors *&	,
		ConvLayer_t *
	)

inlinestatic

Definition at line 152 of file Cuda.h.

◆ InitializeConvWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeConvWorkspace	(	TWorkspace *&	,
		TDescriptors *&	,
		const DNN::CNN::TConvParams &	,
		ConvLayer_t *
	)

inlinestatic

Definition at line 163 of file Cuda.h.

◆ InitializeGauss()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::InitializeGauss ( Matrix_t & A )

static

Definition at line 44 of file Initialization.cu.

◆ InitializeGlorotNormal()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::InitializeGlorotNormal ( Matrix_t & A )

static

Truncated normal initialization (Glorot, called also Xavier normal) The values are sample with a normal distribution with stddev = sqrt(2/N_input + N_output) and values larger than 2 * stddev are discarded See Glorot & Bengio, AISTATS 2010 - http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf.

Definition at line 90 of file Initialization.cu.

◆ InitializeGlorotUniform()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::InitializeGlorotUniform ( Matrix_t & A )

static

Sample from a uniform distribution in range [ -lim,+lim] where lim = sqrt(6/N_in+N_out).

This initialization is also called Xavier uniform see Glorot & Bengio, AISTATS 2010 - http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf

Definition at line 119 of file Initialization.cu.

◆ InitializeGRUDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeGRUDescriptors	(	TDescriptors *&	,
		GenLayer_t *
	)

inlinestatic

Definition at line 179 of file Cuda.h.

◆ InitializeGRUTensors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeGRUTensors ( GenLayer_t * )

inlinestatic

Definition at line 187 of file Cuda.h.

◆ InitializeGRUWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeGRUWorkspace	(	TWorkspace *&	,
		TDescriptors *&	,
		GenLayer_t *
	)

inlinestatic

Definition at line 183 of file Cuda.h.

◆ InitializeIdentity()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::InitializeIdentity ( Matrix_t & A )

static

Definition at line 142 of file Initialization.cu.

◆ InitializeLSTMDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeLSTMDescriptors	(	TDescriptors *&	,
		GenLayer_t *
	)

inlinestatic

Definition at line 178 of file Cuda.h.

◆ InitializeLSTMTensors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeLSTMTensors ( GenLayer_t * )

inlinestatic

Definition at line 186 of file Cuda.h.

◆ InitializeLSTMWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeLSTMWorkspace	(	TWorkspace *&	,
		TDescriptors *&	,
		GenLayer_t *
	)

inlinestatic

Definition at line 182 of file Cuda.h.

◆ InitializePoolDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializePoolDescriptors	(	TDescriptors *&	,
		PoolingLayer_t *
	)

inlinestatic

Definition at line 154 of file Cuda.h.

◆ InitializePoolDropoutWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializePoolDropoutWorkspace	(	TWorkspace *&	,
		TDescriptors *&	,
		const DNN::CNN::TConvParams &	,
		PoolingLayer_t *
	)

inlinestatic

Definition at line 167 of file Cuda.h.

◆ InitializeRNNDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeRNNDescriptors	(	TDescriptors *&	,
		GenLayer_t *
	)

inlinestatic

Definition at line 177 of file Cuda.h.

◆ InitializeRNNTensors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeRNNTensors ( GenLayer_t * )

inlinestatic

Definition at line 185 of file Cuda.h.

◆ InitializeRNNWorkspace()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeRNNWorkspace	(	TWorkspace *&	,
		TDescriptors *&	,
		GenLayer_t *
	)

inlinestatic

Definition at line 181 of file Cuda.h.

◆ InitializeUniform()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::InitializeUniform ( Matrix_t & A )

static

Definition at line 65 of file Initialization.cu.

◆ InitializeZero() [1/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeZero ( Matrix_t & A )

static

◆ InitializeZero() [2/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::InitializeZero ( Tensor_t & A )

static

◆ IsCudnn()

template<typename AReal = Float_t>

static bool TMVA::DNN::TCuda< AReal >::IsCudnn ( )

inlinestatic

Definition at line 139 of file Cuda.h.

◆ L1Regularization()

template<typename AReal = Float_t>

AFloat TMVA::DNN::TCuda< AFloat >::L1Regularization ( const Matrix_t & W )

static

Definition at line 27 of file Regularization.cu.

◆ L2Regularization()

template<typename AReal = Float_t>

AFloat TMVA::DNN::TCuda< AFloat >::L2Regularization ( const Matrix_t & W )

static

Definition at line 60 of file Regularization.cu.

◆ LSTMLayerBackward()

template<typename AReal = Float_t>

static Matrix_t & TMVA::DNN::TCuda< AReal >::LSTMLayerBackward	(	Matrix_t &	state_gradients_backward,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Matrix_t &
	)

inlinestatic

Definition at line 701 of file Cuda.h.

◆ MaxPoolLayerBackward()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::MaxPoolLayerBackward	(	Tensor_t &	activationGradientsBackward,
		const Tensor_t &	activationGradients,
		const Tensor_t &	indexMatrix,
		const Tensor_t &	,
		const Tensor_t &	,
		const PoolingDescriptors_t &	,
		PoolingWorkspace_t &	,
		size_t	imgHeight,
		size_t	imgWidth,
		size_t	fltHeight,
		size_t	fltWidth,
		size_t	strideRows,
		size_t	strideCols,
		size_t	nLocalViews
	)

static

Perform the complete backward propagation step in a Pooling Layer.

Based on the winning indices stored in the index matrix, it just forwards the activation gradients to the previous layer.

Definition at line 499 of file Propagation.cu.

◆ MeanSquaredError()

template<typename AReal = Float_t>

AFloat TMVA::DNN::TCuda< AFloat >::MeanSquaredError	(	const Matrix_t &	Y,
		const Matrix_t &	output,
		const Matrix_t &	weights
	)

static

Definition at line 28 of file LossFunctions.cu.

◆ MeanSquaredErrorGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::MeanSquaredErrorGradients	(	Matrix_t &	dY,
		const Matrix_t &	Y,
		const Matrix_t &	output,
		const Matrix_t &	weights
	)

static

Definition at line 48 of file LossFunctions.cu.

◆ Multiply() [1/3]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Multiply	(	Matrix_t &	C,
		const Matrix_t &	A,
		const Matrix_t &	B
	)

static

Standard multiplication of two matrices A and B with the result being written into C.

◆ Multiply() [2/3]

void TMVA::DNN::TCuda< double >::Multiply	(	TCudaMatrix< double > &	C,
		const TCudaMatrix< double > &	A,
		const TCudaMatrix< double > &	B
	)

Definition at line 55 of file Arithmetic.cu.

◆ Multiply() [3/3]

void TMVA::DNN::TCuda< float >::Multiply	(	TCudaMatrix< float > &	C,
		const TCudaMatrix< float > &	A,
		const TCudaMatrix< float > &	B
	)

Definition at line 28 of file Arithmetic.cu.

◆ MultiplyTranspose() [1/4]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::MultiplyTranspose	(	Matrix_t &	output,
		const Matrix_t &	input,
		const Matrix_t &	weights
	)

static

Matrix-multiply input with the transpose of weights and write the results into output.

◆ MultiplyTranspose() [2/4]

void TMVA::DNN::TCuda< double >::MultiplyTranspose	(	TCudaMatrix< double > &	output,
		const TCudaMatrix< double > &	input,
		const TCudaMatrix< double > &	Weights
	)

Definition at line 52 of file Propagation.cu.

◆ MultiplyTranspose() [3/4]

void TMVA::DNN::TCuda< float >::MultiplyTranspose	(	TCudaMatrix< float > &	output,
		const TCudaMatrix< float > &	input,
		const TCudaMatrix< float > &	Weights
	)

Definition at line 28 of file Propagation.cu.

◆ MultiplyTranspose() [4/4]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::MultiplyTranspose	(	Tensor_t &	output,
		const Tensor_t &	input,
		const Matrix_t &	weights
	)

inlinestatic

Definition at line 208 of file Cuda.h.

◆ PrepareInternals()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::PrepareInternals ( Tensor_t & )

inlinestatic

Dummy placeholder - preparation is currently only required for the CUDA architecture.

Definition at line 545 of file Cuda.h.

◆ PrintTensor()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::PrintTensor	(	const Tensor_t &	A,
		const std::string	name = `"Cuda-tensor"`,
		bool	= `false`
	)

static

Definition at line 860 of file Cuda.h.

◆ Rearrange()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AReal >::Rearrange	(	Tensor_t &	out,
		const Tensor_t &	in
	)

static

Rearrage data according to time fill B x T x D out with T x B x D matrix in.

Definition at line 548 of file Propagation.cu.

◆ ReciprocalElementWise()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::ReciprocalElementWise ( Matrix_t & A )

static

Reciprocal each element of the matrix A and write the result into A.

Definition at line 385 of file Arithmetic.cu.

◆ RecurrentLayerBackward()

template<typename AReal = Float_t>

TCudaMatrix< AFloat > & TMVA::DNN::TCuda< AFloat >::RecurrentLayerBackward	(	Matrix_t &	state_gradients_backward,
		Matrix_t &	input_weight_gradients,
		Matrix_t &	state_weight_gradients,
		Matrix_t &	bias_gradients,
		Matrix_t &	df,
		const Matrix_t &	state,
		const Matrix_t &	weights_input,
		const Matrix_t &	weights_state,
		const Matrix_t &	input,
		Matrix_t &	input_gradient
	)

static

Backward pass for Recurrent Networks.

Definition at line 29 of file RecurrentPropagation.cu.

◆ ReleaseBNormDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ReleaseBNormDescriptors ( TDescriptors *& )

inlinestatic

Definition at line 161 of file Cuda.h.

◆ ReleaseConvDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ReleaseConvDescriptors ( TDescriptors *& )

inlinestatic

Release CNN data/operator descriptors.

Not used at the moment.

Definition at line 159 of file Cuda.h.

◆ ReleaseDescriptor()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ReleaseDescriptor ( ActivationDescriptor_t & )

inlinestatic

Definition at line 172 of file Cuda.h.

◆ ReleasePoolDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ReleasePoolDescriptors ( TDescriptors *& )

inlinestatic

Definition at line 160 of file Cuda.h.

◆ ReleaseRNNDescriptors()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ReleaseRNNDescriptors ( TDescriptors *& )

inlinestatic

Definition at line 189 of file Cuda.h.

◆ Relu()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Relu ( Tensor_t & B )

static

Definition at line 68 of file ActivationFunctions.cu.

◆ ReluDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::ReluDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 81 of file ActivationFunctions.cu.

◆ Reshape()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Reshape	(	Matrix_t &	A,
		const Matrix_t &	B
	)

static

Transform the matrix B to a matrix with different dimensions A.

Definition at line 535 of file Propagation.cu.

◆ RNNBackward()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::RNNBackward	(	const Tensor_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Tensor_t &	,
		const Tensor_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Tensor_t &	,
		Tensor_t &	,
		Matrix_t &	,
		Matrix_t &	,
		Tensor_t &	,
		const RNNDescriptors_t &	,
		RNNWorkspace_t &
	)

inlinestatic

Definition at line 693 of file Cuda.h.

◆ RNNForward()

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::RNNForward	(	const Tensor_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		const Matrix_t &	,
		Tensor_t &	,
		Matrix_t &	,
		Matrix_t &	,
		const RNNDescriptors_t &	,
		RNNWorkspace_t &	,
		bool
	)

inlinestatic

Definition at line 687 of file Cuda.h.

◆ RotateWeights()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::RotateWeights	(	Matrix_t &	A,
		const Matrix_t &	B,
		size_t	filterDepth,
		size_t	filterHeight,
		size_t	filterWidth,
		size_t	numFilters
	)

static

Rotates the matrix B, which is representing a weights, and stores them in the matrix A.

Definition at line 207 of file Propagation.cu.

◆ ScaleAdd() [1/4]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ScaleAdd	(	Matrix_t &	A,
		const Matrix_t &	B,
		Scalar_t	beta = `1.0`
	)

static

Adds a the elements in matrix B scaled by c to the elements in the matrix A.

This is required for the weight update in the gradient descent step.

◆ ScaleAdd() [2/4]

void TMVA::DNN::TCuda< double >::ScaleAdd	(	TCudaMatrix< double > &	B,
		const TCudaMatrix< double > &	A,
		double	alpha
	)

Definition at line 330 of file Arithmetic.cu.

◆ ScaleAdd() [3/4]

void TMVA::DNN::TCuda< float >::ScaleAdd	(	TCudaMatrix< float > &	B,
		const TCudaMatrix< float > &	A,
		float	alpha
	)

Definition at line 317 of file Arithmetic.cu.

◆ ScaleAdd() [4/4]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::ScaleAdd	(	Tensor_t &	A,
		const Tensor_t &	B,
		Scalar_t	beta = `1.0`
	)

static

Above functions extended to vectors.

◆ SetRandomSeed()

template<typename AFloat >

void TMVA::DNN::TCuda< AFloat >::SetRandomSeed ( size_t seed )

static

Definition at line 31 of file Initialization.cu.

◆ Sigmoid() [1/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Sigmoid	(	Matrix_t &	YHat,
		const Matrix_t &
	)

static

◆ Sigmoid() [2/2]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::Sigmoid ( Tensor_t & B )

static

◆ SigmoidDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SigmoidDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 111 of file ActivationFunctions.cu.

◆ Softmax()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Softmax	(	Matrix_t &	YHat,
		const Matrix_t &
	)

static

Definition at line 42 of file OutputFunctions.cu.

◆ SoftmaxCrossEntropy()

template<typename AReal = Float_t>

AFloat TMVA::DNN::TCuda< AFloat >::SoftmaxCrossEntropy	(	const Matrix_t &	Y,
		const Matrix_t &	output,
		const Matrix_t &	weights
	)

static

Softmax transformation is implicitly applied, thus output should hold the linear activations of the last layer in the net.

Definition at line 108 of file LossFunctions.cu.

◆ SoftmaxCrossEntropyGradients()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SoftmaxCrossEntropyGradients	(	Matrix_t &	dY,
		const Matrix_t &	Y,
		const Matrix_t &	output,
		const Matrix_t &	weights
	)

static

Definition at line 128 of file LossFunctions.cu.

◆ SoftSign()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SoftSign ( Tensor_t & B )

static

Definition at line 188 of file ActivationFunctions.cu.

◆ SoftSignDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SoftSignDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 201 of file ActivationFunctions.cu.

◆ SqrtElementWise()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SqrtElementWise ( Matrix_t & A )

static

Square root each element of the matrix A and write the result into A.

Definition at line 411 of file Arithmetic.cu.

◆ SquareElementWise()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SquareElementWise ( Matrix_t & A )

static

Square each element of the matrix A and write the result into A.

Definition at line 398 of file Arithmetic.cu.

◆ Sum()

template<typename AReal = Float_t>

AFloat TMVA::DNN::TCuda< AFloat >::Sum ( const Matrix_t & A )

static

Compute the sum of all elements in A.

Definition at line 172 of file Arithmetic.cu.

◆ SumColumns() [1/3]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::SumColumns	(	Matrix_t &	B,
		const Matrix_t &	A,
		Scalar_t	alpha = `1.0`,
		Scalar_t	beta = `0.`
	)

static

Sum columns of (m x n) matrix A and write the results into the first m elements in A.

◆ SumColumns() [2/3]

void TMVA::DNN::TCuda< double >::SumColumns	(	TCudaMatrix< double > &	B,
		const TCudaMatrix< double > &	A,
		double	alpha,
		double	beta
	)

Definition at line 213 of file Arithmetic.cu.

◆ SumColumns() [3/3]

void TMVA::DNN::TCuda< float >::SumColumns	(	TCudaMatrix< float > &	B,
		const TCudaMatrix< float > &	A,
		float	alpha,
		float	beta
	)

Definition at line 189 of file Arithmetic.cu.

◆ SumRows() [1/3]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::SumRows	(	Matrix_t &	B,
		const Matrix_t &	A
	)

static

extra functions defined only for CPU architecture !!!

Sum rows of (m x n) matrix A and write the results into the first m elements in B.

◆ SumRows() [2/3]

void TMVA::DNN::TCuda< double >::SumRows	(	TCudaMatrix< double > &	B,
		const TCudaMatrix< double > &	A
	)

Definition at line 259 of file Arithmetic.cu.

◆ SumRows() [3/3]

void TMVA::DNN::TCuda< float >::SumRows	(	TCudaMatrix< float > &	B,
		const TCudaMatrix< float > &	A
	)

Definition at line 236 of file Arithmetic.cu.

◆ SymmetricRelu()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SymmetricRelu ( Tensor_t & B )

static

Definition at line 158 of file ActivationFunctions.cu.

◆ SymmetricReluDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::SymmetricReluDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 171 of file ActivationFunctions.cu.

◆ Tanh()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::Tanh ( Tensor_t & B )

static

Definition at line 128 of file ActivationFunctions.cu.

◆ TanhDerivative()

template<typename AReal = Float_t>

void TMVA::DNN::TCuda< AFloat >::TanhDerivative	(	Tensor_t &	B,
		const Tensor_t &	A
	)

static

Definition at line 141 of file ActivationFunctions.cu.

◆ TransposeMultiply() [1/3]

template<typename AReal = Float_t>

static void TMVA::DNN::TCuda< AReal >::TransposeMultiply	(	Matrix_t &	output,
		const Matrix_t &	input,
		const Matrix_t &	Weights,
		Scalar_t	alpha = `1.0`,
		Scalar_t	beta = `0.`
	)

static

Matrix multiplication of two matrices A and B^T (transposed) with the result being written into C.

◆ TransposeMultiply() [2/3]

void TMVA::DNN::TCuda< double >::TransposeMultiply	(	TCudaMatrix< double > &	C,
		const TCudaMatrix< double > &	A,
		const TCudaMatrix< double > &	B,
		double	alpha,
		double	beta
	)

Definition at line 109 of file Arithmetic.cu.

◆ TransposeMultiply() [3/3]

void TMVA::DNN::TCuda< float >::TransposeMultiply	(	TCudaMatrix< float > &	C,
		const TCudaMatrix< float > &	A,
		const TCudaMatrix< float > &	B,
		float	alpha,
		float	beta
	)

Definition at line 82 of file Arithmetic.cu.

Member Data Documentation

◆ fgRandomGen

template<typename AFloat >

TRandom * TMVA::DNN::TCuda< AFloat >::fgRandomGen = nullptr

staticprivate

Definition at line 67 of file Cuda.h.

tmva/tmva/inc/TMVA/DNN/Architectures/Cuda.h

tmva/tmva/src/DNN/Architectures/Cuda/ActivationFunctions.cu

tmva/tmva/src/DNN/Architectures/Cuda/Arithmetic.cu

tmva/tmva/src/DNN/Architectures/Cuda/Initialization.cu

tmva/tmva/src/DNN/Architectures/Cuda/LossFunctions.cu

tmva/tmva/src/DNN/Architectures/Cuda/OutputFunctions.cu

tmva/tmva/src/DNN/Architectures/Cuda/Propagation.cu

tmva/tmva/src/DNN/Architectures/Cuda/RecurrentPropagation.cu

tmva/tmva/src/DNN/Architectures/Cuda/Regularization.cu

Public Types

Public Member Functions

Static Public Member Functions

Static Private Attributes

Member Typedef Documentation

◆ ActivationDescriptor_t

◆ AFloat

◆ AlgorithmBackward_t

◆ AlgorithmDataType_t

◆ AlgorithmForward_t

◆ AlgorithmHelper_t

◆ BNormDescriptors_t

◆ BNormLayer_t

◆ ConvDescriptors_t

◆ ConvLayer_t

◆ ConvolutionDescriptor_t

◆ ConvWorkspace_t

◆ DeviceBuffer_t

◆ DropoutDescriptor_t

◆ EmptyDescriptor_t

◆ FilterDescriptor_t

◆ GenLayer_t

◆ HostBuffer_t

◆ Matrix_t

◆ PoolingDescriptor_t

◆ PoolingDescriptors_t

◆ PoolingLayer_t

◆ PoolingWorkspace_t

◆ RecurrentDescriptor_t

◆ ReduceTensorDescriptor_t

◆ RNNDescriptors_t

◆ RNNWorkspace_t

◆ Scalar_t

◆ Tensor_t

◆ TensorDescriptor_t

Member Function Documentation

◆ ActivationFunctionBackward()

◆ ActivationFunctionForward()

◆ AdamUpdate()

◆ AdamUpdateFirstMom()

◆ AdamUpdateSecondMom()

◆ AddConvBiases()

◆ AddL1RegularizationGradients()

◆ AddL2RegularizationGradients()

◆ AddRowWise() [1/2]

◆ AddRowWise() [2/2]

◆ AlmostEquals()

◆ Backward()

◆ BatchNormLayerBackward()

◆ BatchNormLayerForwardInference()

◆ BatchNormLayerForwardTraining()

◆ CalculateConvActivationGradients()

◆ CalculateConvBiasGradients()

◆ CalculateConvWeightGradients()

◆ calculateDimension()

◆ ConstAdd()

◆ ConstMult()

◆ ConvLayerBackward()

◆ ConvLayerForward()

◆ Copy() [1/2]

◆ Copy() [2/2]

◆ CopyDiffArch() [1/5]

◆ CopyDiffArch() [2/5]

◆ CopyDiffArch() [3/5]

◆ CopyDiffArch() [4/5]

◆ CopyDiffArch() [5/5]

◆ CreateTensor() [1/4]

◆ CreateTensor() [2/4]

◆ CreateTensor() [3/4]

◆ CreateTensor() [4/4]

◆ CreateWeightTensors()

◆ CrossEntropy()

◆ CrossEntropyGradients()

◆ Deflatten()

◆ Downsample()

◆ DropoutBackward()

◆ DropoutForward() [1/2]

◆ DropoutForward() [2/2]

◆ FastTanh()

◆ FastTanhDerivative()