tc::CudaCache Class Reference

#include <compilation_cache.h>

Inheritance diagram for tc::CudaCache:

Collaboration diagram for tc::CudaCache:

Classes
struct	CachedEntry
struct	RetrievalResult

Public Member Functions
	CudaCache ()=default
	CudaCache (const CudaCacheProto &buf)
CudaCacheProto	toProtobuf () const
void	cacheKernel (const std::string &id, const MappingOptions &options, const std::vector< const DLTensor * > &inputs, const std::vector< const DLTensor * > &outputs, const std::string &kernelSpecializedName, const std::vector< int > &kernelParameters, const std::string &cudaSource, const Grid &grid, const Block &block)
std::unique_ptr< RetrievalResult >	retrieveKernel (const std::string &id, const MappingOptions &options, const std::vector< const DLTensor * > &inputs, const std::vector< const DLTensor * > &outputs) const
void	removeEntriesNotInOptionsCache (const OptionsCache &oc)
template<typename C , typename InputTy >
auto	searchKernelImpl (C &c, const std::string &id, const MappingOptions &options, const std::vector< InputTy > &inputs, const std::vector< InputTy > &outputs) -> decltype(c.searchKernel(id, options, inputs, outputs))
Public Member Functions inherited from tc::Cache< CudaCache >
size_t	size () const
void	clear ()

Private Types
using	Protobuf = CudaCacheProto

Private Member Functions
CachedEntry *	searchKernel (const std::string &id, const MappingOptions &options, const std::vector< detail::TensorInfo > &inputs, const std::vector< detail::TensorInfo > &outputs)
CachedEntry *	searchKernel (const std::string &id, const MappingOptions &options, const std::vector< const DLTensor * > &inputs, const std::vector< const DLTensor * > &outputs)
const CachedEntry *	searchKernel (const std::string &id, const MappingOptions &options, const std::vector< const DLTensor * > &inputs, const std::vector< const DLTensor * > &outputs) const

Static Private Member Functions
static std::shared_ptr < CudaCache > &	getGlobalSharedCache ()
template<typename C , typename TensorTy >
static auto	searchKernelImpl (C &c, const std::string &id, const MappingOptions &options, const std::vector< TensorTy > &inputs, const std::vector< TensorTy > &outputs) -> decltype(c.searchKernel(id, options, inputs, outputs))

Private Attributes
std::vector< CachedEntry >	entries_

Friends
class	Cache< CudaCache >

Additional Inherited Members
Static Public Member Functions inherited from tc::Cache< CudaCache >
static void	enableCache ()
static void	disableCache ()
static void	dumpCacheToProtobuf (const std::string &filename)
static void	loadCacheFromProtobuf (const std::string &filename)
static void	loadCacheFromProtobuf (const Protobuf &buf)
static std::shared_ptr< CudaCache >	getCache ()
static bool	cacheEnabled ()
Public Attributes inherited from tc::Cache< CudaCache >
int	numberAttemptedRetrievals
int	numberSuccessfulRetrievals
int	numberCacheAttemps
Protected Attributes inherited from tc::Cache< CudaCache >
std::mutex	mtx_

Detailed Description

CudaCache stores the Cuda source of optimized kernels

Member Typedef Documentation

using tc::CudaCache::Protobuf = CudaCacheProto

private

Constructor & Destructor Documentation

tc::CudaCache::CudaCache ( )

default

tc::CudaCache::CudaCache

(

const CudaCacheProto &

buf

)

Member Function Documentation

void tc::CudaCache::cacheKernel	(	const std::string &	id,
		const MappingOptions &	options,
		const std::vector< const DLTensor * > &	inputs,
		const std::vector< const DLTensor * > &	outputs,
		const std::string &	kernelSpecializedName,
		const std::vector< int > &	kernelParameters,
		const std::string &	cudaSource,
		const Grid &	grid,
		const Block &	block
	)

If op was previously cached and the inputs' shape, isl options, and the target device are the same then this is a noop Else (cudaSource, grid, block) is stored in the cache

static std::shared_ptr<CudaCache>& tc::CudaCache::getGlobalSharedCache ( )

staticprivate

void tc::CudaCache::removeEntriesNotInOptionsCache

(

const OptionsCache &

oc

)

std::unique_ptr<RetrievalResult> tc::CudaCache::retrieveKernel	(	const std::string &	id,
		const MappingOptions &	options,
		const std::vector< const DLTensor * > &	inputs,
		const std::vector< const DLTensor * > &	outputs
	)		const

Returns the cache entry that matches op (id, isl options, target device) and inputs' shapes.

CachedEntry* tc::CudaCache::searchKernel ( const std::string &  id, const MappingOptions &  options, const std::vector< detail::TensorInfo > &  inputs, const std::vector< detail::TensorInfo > &  outputs  )

private

SearchKernel (through SearchKernelImpl) searches op in the cache if a cached entry that corresponds to the op's configuration (MappingOptions and TargetDevice) and the shape of inputs matches it is returned

CachedEntry* tc::CudaCache::searchKernel ( const std::string &  id, const MappingOptions &  options, const std::vector< const DLTensor * > &  inputs, const std::vector< const DLTensor * > &  outputs  )

private

const CachedEntry* tc::CudaCache::searchKernel ( const std::string &  id, const MappingOptions &  options, const std::vector< const DLTensor * > &  inputs, const std::vector< const DLTensor * > &  outputs  ) const

private

template<typename C , typename InputTy >

auto tc::CudaCache::searchKernelImpl	(	C &	c,
		const std::string &	id,
		const MappingOptions &	options,
		const std::vector< InputTy > &	inputs,
		const std::vector< InputTy > &	outputs
	)		-> decltype(c.searchKernel(id, options, inputs, outputs))

template<typename C , typename TensorTy >

static auto tc::CudaCache::searchKernelImpl ( C &  c, const std::string &  id, const MappingOptions &  options, const std::vector< TensorTy > &  inputs, const std::vector< TensorTy > &  outputs  ) -> decltype(c.searchKernel(id, options, inputs, outputs))

staticprivate

CudaCacheProto tc::CudaCache::toProtobuf

(

)

const

Friends And Related Function Documentation

friend class Cache< CudaCache >

friend

Member Data Documentation

std::vector<CachedEntry> tc::CudaCache::entries_

private

---

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

• include/tc/core/compilation_cache.h
• include/tc/core/compilation_cache-inl.h