UserKernelGroup

Functions
void	OpenKernelGroup (char *GroupName)
	Open a user kernel group. More...
void	CloseKernelGroup ()
	Close currently open user kernel group. More...
void	CloseKernelGroupNoMerge ()
	Close currently open user kernel group setting KER_OPT_KEEP_GROUP for nodes in the group. More...
CKernelCall_T *	UserKernelCall (char *CallName, KernelCallLocationT CallLocation, ArgBindingDescr_T **BindingList)
	Call a user kernel in a user kernel group. More...
Object_T *	KerGroupArg (char *KerArgName, Object_Type_T ObjType, int ArgSize, int ItemSize, char *CArgName)
ArgBindingDescr_T *	KG_ArgOper (char *ArgName, char Oper, int Value)
	Binds argument to a user kernel group argument and combine it with Value using Oper. More...
StackedTensors_T *	AT_StackedTensors (char *OutTensorName, int Count,...)
	Creates a stacked tensor, resulting object is OutTensorName, Count in tensors passed as names are stacked according to list order. More...
void	AddStackedTensors (char *OutTensorName, int Count,...)
	Add a stacked tensor to the open graph. More...
StackedTensors_T *	AT_StackedTensorsList (int Count,...)
	Creates a list of stacked tensors. More...
CKernel_Arg_T *	StackedTensorsLookup (StackedTensors_T *List, NameT *Name)
	Given input tensor Name returns corresponding Out Stacked Tensor if it exists. More...
void	UserKernelGroup (char *GroupName, CKernel_Arg_T **CArg, CKernelCall_T **CCalls)
	Declare a user kernel group. More...
KernelGroup_T *	UserKernelGroupK (char *GroupName, unsigned int IterCount, CKernel_Arg_T **CArg, StackedTensors_T *StackedTensors, CKernelCall_T **CCalls, Object_T **KerArg)

Description

Functions in this group should be used to create groups of user kernels

A user kernel group starts with an OpenKernelGroup() followed by multiple UserKernel() and CloseKernelGroup()

Then the group header and how user kernels are called in the group are captured in UserKernelGroup()

Function Documentation

void AddStackedTensors	(	char *	OutTensorName,
		int	Count,
			...
	)

Adds a stacked tensor to the graph. The stacked tensor creates a series of aliases for the tensor OutTensorName which should be declared as an output or local. The aliases can be used as outputs of kernels and will be assembled in the order declared into the output tensor.

StackedTensors_T* AT_StackedTensors	(	char *	OutTensorName,
		int	Count,
			...
	)

Creates a list of stacked tensors, resulting object is OutTensorName, Count in tensors passed as names are stacked according to list order

StackedTensors_T* AT_StackedTensorsList	(	int	Count,
			...
	)

Creates a list of stacked tensors

void CloseKernelGroup

(

)

Close currently open user kernel group.

void CloseKernelGroupNoMerge

(

)

Close currently open user kernel group setting KER_OPT_KEEP_GROUP for nodes in the group

Object_T* KerGroupArg	(	char *	KerArgName,
		Object_Type_T	ObjType,
		int	ArgSize,
		int	ItemSize,
		char *	CArgName
	)

Parameters

KerArgName	User kernel group argument name
ObjType	Argument type, O_IN and/or O_OUT can be used only
ArgSize	Number of elements of this argument
ItemSize	Size in byte of the elementary data type of this argument
CArgName	Corresponding name in the User Kernel Group definition

ArgBindingDescr_T* KG_ArgOper	(	char *	ArgName,
		char	Oper,
		int	Value
	)

Binds argument to a user kernel group argument and combine it with Value using Oper

See K_Arg() for ArgName and ArgSelect. Supported operation are defined in ArgBindingOper. Value is a signed immediate constant.

Parameters

ArgName	A tiled user kernel argument name
Oper	Operation, see ArgBindingOper. Valid: + - * / %
Value	A signed immediate value

void OpenKernelGroup

(

char *

GroupName

)

Open a user kernel group

Parameters

GroupName

C name to be used for the group

CKernel_Arg_T* StackedTensorsLookup	(	StackedTensors_T *	List,
		NameT *	Name
	)

Given input tensor Name returns corresponding Out Stacked Tensor if it exists

Parameters

List	A list of stacked tensor
Name	Input tensor name to search

CKernelCall_T* UserKernelCall	(	char *	CallName,
		KernelCallLocationT	CallLocation,
		ArgBindingDescr_T **	BindingList
	)

Call a user kernel in a user kernel group

Parameters

CallName	A user kernel name
CallLocation	Here only LOC_IN_GROUP is valid
BindingList	See Bindings(). restricted to C_Arg(), C_ArgIndex(), C_ArgIndirect() and Imm()

void UserKernelGroup	(	char *	GroupName,
		CKernel_Arg_T **	CArg,
		CKernelCall_T **	CCalls
	)

Declare a user kernel group

Parameters

GroupName	Group's name as in OpenKernelGroup()
CArg	See Carg() and TCArg()
CCalls	See Calls() and UserKernelCall()

KernelGroup_T* UserKernelGroupK	(	char *	GroupName,
		unsigned int	IterCount,
		CKernel_Arg_T **	CArg,
		StackedTensors_T *	StackedTensors,
		CKernelCall_T **	CCalls,
		Object_T **	KerArg
	)

Parameters

GroupName	Group's name as in OpenKernelGroup()
IterCount	Number of time this group should be iterated, usually 1 but used for grouped convolution
CArg	See Carg() and TCArg()
StackedTensors	List of stacked tensors
CCalls	See Calls() and UserKernelCall()
KerArg	Kernel Group arguments. Restricted to KerGroupArg() as argument of KerArgs()