Perform pointwise operation with both boradcasting and reduction. More...

#include <include/dlprim/core/pointwise.hpp>

Public Member Functions
virtual size_t	workspace ()=0
	Get size of workspace in bytes needed.

virtual void	enqueue (std::vector< Tensor > xs, std::vector< Tensor > ys, Tensor &workspace, std::vector< double > parameters, std::vector< double > alpha, std::vector< double > beta, ExecutionContext const &e)=0
	Perform coputations. More...

Static Public Member Functions
static std::unique_ptr< PointwiseOperationBroadcastReduce >	create (Context &ctx, std::vector< TensorSpecs > xs, std::vector< TensorSpecs > ys, int weights_count, DataType weights_type, std::string const &compute_code, std::string const &reduce_init, std::string const &reduce)
	Create objects: More...

Detailed Description

Perform pointwise operation with both boradcasting and reduction.

Calculation is performed over a shape that xs and ys tensors are boradcaasted to.

For example xs have shapes: (64,10,5) and (64,10,1) and ys has shape (10,1) they all broadcast to 64,10,5 and reduction is performed over dimentsions 0 and 2

All ys tensors need to have same shape and be boradcastable to total shape

Optional parameters can be provided that avalible in code as w0... wN, Final ys are computed as ys[i] = alpha[i] * reduced_result + beta[i] * ys[i]

Member Function Documentation

static std::unique_ptr<PointwiseOperationBroadcastReduce> dlprim::core::PointwiseOperationBroadcastReduce::create	(	Context &	ctx,
		std::vector< TensorSpecs >	xs,
		std::vector< TensorSpecs >	ys,
		int	weights_count,
		DataType	weights_type,
		std::string const &	compute_code,
		std::string const &	reduce_init,
		std::string const &	reduce
	)

static

Create objects:

Parameters

xs	- vector of input tensor specs - such tensors are expected to be given to enqueue
ys	- vector of output tenorr specs - such tensors are expectred to be give to enqueue
weights_type	- type of weights parameters as provided
compute_code	- OpenCL code to compute values. You can use x0, x1, ... xN as input values for each x for xs y0,.., yN for each output and w0,...,wN for each weight. For example "y0 = x0 + w0 * x1;"
reduce_init	- initalization of reduction variables `reduce_yN` for example "reduce_y0 = 0;" or "reduce_y0=-FLT_MAX;"
reduce	- code for sum reduction "reduce_y0 += y0" or max reduction "reduce_y0 = max(reduce_y0,y0)"

virtual void dlprim::core::PointwiseOperationBroadcastReduce::enqueue	(	std::vector< Tensor >	xs,
		std::vector< Tensor >	ys,
		Tensor &	workspace,
		std::vector< double >	parameters,
		std::vector< double >	alpha,
		std::vector< double >	beta,
		ExecutionContext const &	e
	)

pure virtual

Perform coputations.

Parameters

xs	- vector of input tensor
ys	- vector of output tenors
alpha	- scale for ys, must match size of ys
beta	- scale for summation of previous ys, must match size of ys

The documentation for this class was generated from the following file:

include/dlprim/core/pointwise.hpp

Public Member Functions

Static Public Member Functions

Detailed Description

Member Function Documentation