bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam > Class Template Reference

Inheritance diagram for bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >:

Collaboration diagram for bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >:

Classes
struct	FusedMultiHeadAttentionKernelInfo

Public Types
using	KernelMeta = TKernelMeta
using	KernelParam = TKernelParam

Public Member Functions
uint64_t	hashID (unsigned int s, unsigned int d) const
virtual uint64_t	hashID (const KernelMeta &kernelMeta) const
	TFusedMultiHeadAttentionXMMAKernel (const TKernelMeta *pMetaStart, unsigned int nMetaCount, Data_type type, unsigned int sm)
void	loadXMMAKernels ()
bool	isValid (int s) const
virtual void	run (TKernelParam &params, cudaStream_t ss) const
virtual	~TFusedMultiHeadAttentionXMMAKernel ()=default

Protected Attributes
nvinfer1::CUDADriverWrapper	mDriver
Data_type	mDataType
const TKernelMeta *	mKernelMeta
unsigned int	mKernelMetaCount
unsigned int	mSM
std::unordered_map< const unsigned char *, CUmodule >	mModules
std::unordered_map< uint64_t, FusedMultiHeadAttentionKernelInfo >	mFunctions
std::set< int >	mValidSequences

Member Typedef Documentation

KernelMeta

template<typename TKernelMeta , typename TKernelParam >

using bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::KernelMeta = TKernelMeta

KernelParam

template<typename TKernelMeta , typename TKernelParam >

using bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::KernelParam = TKernelParam

Constructor & Destructor Documentation

TFusedMultiHeadAttentionXMMAKernel()

template<typename TKernelMeta , typename TKernelParam >

bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::TFusedMultiHeadAttentionXMMAKernel ( const TKernelMeta *  pMetaStart, unsigned int  nMetaCount, Data_type  type, unsigned int  sm  )

inline

~TFusedMultiHeadAttentionXMMAKernel()

template<typename TKernelMeta , typename TKernelParam >

virtual bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::~TFusedMultiHeadAttentionXMMAKernel ( )

virtualdefault

Member Function Documentation

hashID() [1/2]

template<typename TKernelMeta , typename TKernelParam >

uint64_t bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::hashID ( unsigned int  s, unsigned int  d  ) const

inline

Here is the caller graph for this function:

hashID() [2/2]

template<typename TKernelMeta , typename TKernelParam >

virtual uint64_t bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::hashID ( const KernelMeta &  kernelMeta ) const

inlinevirtual

Reimplemented in bert::FusedMultiHeadAttentionXMMAKernelV2.

loadXMMAKernels()

template<typename TKernelMeta , typename TKernelParam >

void bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::loadXMMAKernels ( )

inline

isValid()

template<typename TKernelMeta , typename TKernelParam >

bool bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::isValid ( int  s ) const

inline

run()

template<typename TKernelMeta , typename TKernelParam >

virtual void bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::run ( TKernelParam &  params, cudaStream_t  ss  ) const

inlinevirtual

Reimplemented in bert::FusedMultiHeadAttentionXMMAKernelV2.

Here is the caller graph for this function:

Member Data Documentation

mDriver

template<typename TKernelMeta , typename TKernelParam >

nvinfer1::CUDADriverWrapper bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mDriver

protected

mDataType

template<typename TKernelMeta , typename TKernelParam >

Data_type bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mDataType

protected

mKernelMeta

template<typename TKernelMeta , typename TKernelParam >

const TKernelMeta* bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mKernelMeta

protected

mKernelMetaCount

template<typename TKernelMeta , typename TKernelParam >

unsigned int bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mKernelMetaCount

protected

mSM

template<typename TKernelMeta , typename TKernelParam >

unsigned int bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mSM

protected

mModules

template<typename TKernelMeta , typename TKernelParam >

std::unordered_map<const unsigned char*, CUmodule> bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mModules

protected

mFunctions

template<typename TKernelMeta , typename TKernelParam >

std::unordered_map<uint64_t, FusedMultiHeadAttentionKernelInfo> bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mFunctions

protected

mValidSequences

template<typename TKernelMeta , typename TKernelParam >

std::set<int> bert::TFusedMultiHeadAttentionXMMAKernel< TKernelMeta, TKernelParam >::mValidSequences

protected

---

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

• fused_multihead_attention.h