scanner::BatchedKernel Class Reference

Interface for a unit of computation in a pipeline. More...

#include <kernel.h>

Inheritance diagram for scanner::BatchedKernel:

Public Member Functions
	BatchedKernel (const KernelConfig &config)
virtual void	execute_kernel (const StenciledBatchedElements &input_columns, BatchedElements &output_columns)
	For internal use.
Public Member Functions inherited from scanner::BaseKernel
	BaseKernel (const KernelConfig &config)
virtual void	validate (proto::Result *result)
	Checks if kernel arguments are valid. More...
virtual void	fetch_resources (proto::Result *result)
	Downloads any resources necessary for the kernel to run. More...
virtual void	setup_with_resources (proto::Result *result)
	Runs any setup code that relies on fetched resources. More...
virtual void	new_stream (const std::vector< u8 > &args)
	Called when the Kernel is about to process a new stream. More...
virtual void	reset ()
	Requests that kernel resets its logical state. More...
virtual void	set_profiler (Profiler *profiler)
	For internal use.

Protected Member Functions
virtual void	execute (const BatchedElements &input_columns, BatchedElements &output_columns)=0
	Runs the op on input elements and produces equal number of output elements. More...

Additional Inherited Members
Public Attributes inherited from scanner::BaseKernel
Profiler *	profiler_ = nullptr
Static Public Attributes inherited from scanner::BaseKernel
static const i32	UnlimitedDevices = 0

Detailed Description

Interface for a unit of computation in a pipeline.

Kernels form the core of Scanner's interface. They are essentially functions that take elements of inputs and produce an equal number elements of outputs. Kernels are stateful operators that get reset when provided non-contiguous batches of input. See KernelFactory for how an op defines what hardware it can use for its computation.

Member Function Documentation

virtual void scanner::BatchedKernel::execute ( const BatchedElements &  input_columns, BatchedElements &  output_columns  )

protectedpure virtual

Runs the op on input elements and produces equal number of output elements.

Parameters

input_columns	vector of columns, where each column is a vector of inputs and each input is a byte array
output_columns	op output, each column must have same length as the number of input elements

Evaluate gets run on batches of inputs. At the beginning of a pipeline this is raw RGB images from the input images/videos, and after that the input becomes whatever was returned by the previous op.

Number of output columns must be non-zero.

Implemented in scanner::ImageEncoderKernel, scanner::SampleKernel, scanner::SampleKernel, scanner::SliceKernel, scanner::SpaceKernel, and scanner::UnsliceKernel.

---

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

• scanner/api/kernel.h
• scanner/api/kernel.cpp