opennn::cudnn_frontend Namespace Reference

Classes
struct	GraphSlot

Typedefs
using	VariantPack = unordered_map<shared_ptr<graph::Tensor_attributes>, void*>

Functions
int	device_sm_version ()
bool	frontend_enabled ()
bool	frontend_verbose ()
bool	bn_frontend_enabled ()
bool	deterministic_algorithms_enabled ()
bool	conv_energy_autotune_enabled ()
float	conv_energy_tolerance ()
int64_t	conv_energy_window_ms ()
bool	conv_energy_verbose ()
bool	conv_energy_stage_active ()
bool	graph_timing_enabled ()
profiler::Stats &	graph_timing_stats ()
template<typename TensorMap>
void	execute_graph (graph::Graph &graph, TensorMap &tensors, void *workspace, const string &what, const string &timing_label)
void *	shared_workspace (int64_t bytes)
template<typename GraphCache, typename Body>
bool	run_frontend (GraphCache &cache, const char *label, Body &&body)
void	throw_frontend_unavailable (const string &what)
int64_t	element_bytes (DataType_t dtype)
DataType_t	to_dtype (Type t)
vector< int64_t >	nhwc_strides (int64_t c, int64_t h, int64_t w)
shared_ptr< graph::Graph >	new_graph (Type dtype=Type::FP32)
shared_ptr< graph::Tensor_attributes >	nhwc_tensor (graph::Graph &graph, const char *name, int64_t n, int64_t c, int64_t h, int64_t w)
void	set_nhwc_output (shared_ptr< graph::Tensor_attributes > &tensor, int64_t n, int64_t c, int64_t h, int64_t w)
vector< int64_t >	bhsd_strides (int64_t h, int64_t s, int64_t d, bool interleaved)
shared_ptr< graph::Tensor_attributes >	bhsd_tensor (graph::Graph &graph, const char *name, int64_t b, int64_t h, int64_t s, int64_t d, bool interleaved)
void	set_bhsd_output (shared_ptr< graph::Tensor_attributes > &tensor, int64_t b, int64_t h, int64_t s, int64_t d, bool interleaved)
shared_ptr< graph::Tensor_attributes >	per_channel_tensor (graph::Graph &graph, const char *name, int64_t channels)
void	set_per_channel_output (shared_ptr< graph::Tensor_attributes > &tensor, int64_t channels)
template<typename... Args>
string	timing_label (format_string< Args... > fmt, Args &&... args)
shared_ptr< graph::Tensor_attributes >	scalar_tensor (graph::Graph &graph, const char *name, DataType_t dtype, bool pass_by_value=false, int64_t batch=1)
int64_t	candidate_limit_from (long long value)
int64_t	autotune_candidate_limit ()
int64_t	autotune_candidate_limit (const string &tag)
vector< HeurMode_t >	heuristic_modes ()
const vector< NumericalNote_t > &	conv_engine_notes ()
bool	build_top_candidates (graph::Graph &graph, int64_t limit)
bool	plan_cache_enabled ()
bool	sdpa_autotune_enabled ()
int64_t	sdpa_workspace_cap_override_bytes ()
const std::filesystem::path &	plan_cache_directory ()
std::filesystem::path	plan_cache_file (const graph::Graph &graph)
bool	load_cached_plan (graph::Graph &graph, const cudnnHandle_t handle, int64_t &workspace_bytes)
void	store_cached_plan (graph::Graph &graph)
int64_t	autotune_workspace_bytes (const graph::Graph &graph)
int64_t	sdpa_workspace_cap_bytes (int64_t batch, int64_t heads, int64_t q_seq, int64_t src_seq, int64_t head_dim)
bool	finalize_attention (graph::Graph &graph, const string &tag, int64_t &workspace_bytes, bool allow_autotune=false, int64_t workspace_cap=0)
shared_ptr< graph::Tensor_attributes >	seq_len_scalar (graph::Graph &graph, const char *name, int64_t batch=1)
bool	finalize (graph::Graph &graph, int64_t &workspace_bytes, const string &tag, const bool fp32_graph=false)
void	report_autotune_skipped (const char *tag, const char *reason)
bool	is_convolution_tag (const char *tag)
template<typename TensorMap>
void	select_plan_by_energy (graph::Graph &graph, TensorMap &tensors, void *workspace, const char *tag)
template<typename TensorMap>
void	autotune_now (bool &pending, graph::Graph &graph, TensorMap &tensors, int64_t &workspace_bytes, const char *tag=nullptr)
template<typename TensorMap>
void	autotune_with_scratch (bool &pending, graph::Graph &graph, const TensorMap &tensors, int64_t &workspace_bytes, const char *tag=nullptr)
template<typename TensorMap>
void	run_slot (GraphSlot &slot, TensorMap &tensors, const char *what, const string &timing_label, bool scratch_autotune)

Variables
constexpr size_t	graph_cache_capacity = 8
const auto	check_status

Typedef Documentation

VariantPack

using opennn::cudnn_frontend::VariantPack = unordered_map<shared_ptr<graph::Tensor_attributes>, void*>

Function Documentation

autotune_candidate_limit() [1/2]

int64_t opennn::cudnn_frontend::autotune_candidate_limit ( )

inline

autotune_candidate_limit() [2/2]

int64_t opennn::cudnn_frontend::autotune_candidate_limit ( const string & tag )

inline

autotune_now()

template<typename TensorMap>

void opennn::cudnn_frontend::autotune_now ( bool & pending, graph::Graph & graph, TensorMap & tensors, int64_t & workspace_bytes, const char * tag = nullptr )

inline

autotune_with_scratch()

template<typename TensorMap>

void opennn::cudnn_frontend::autotune_with_scratch ( bool & pending, graph::Graph & graph, const TensorMap & tensors, int64_t & workspace_bytes, const char * tag = nullptr )

inline

autotune_workspace_bytes()

int64_t opennn::cudnn_frontend::autotune_workspace_bytes ( const graph::Graph & graph )

inline

bhsd_strides()

vector< int64_t > opennn::cudnn_frontend::bhsd_strides ( int64_t h, int64_t s, int64_t d, bool interleaved )

inline

bhsd_tensor()

shared_ptr< graph::Tensor_attributes > opennn::cudnn_frontend::bhsd_tensor ( graph::Graph & graph, const char * name, int64_t b, int64_t h, int64_t s, int64_t d, bool interleaved )

inline

bn_frontend_enabled()

bool opennn::cudnn_frontend::bn_frontend_enabled ( )

inline

build_top_candidates()

bool opennn::cudnn_frontend::build_top_candidates ( graph::Graph & graph, int64_t limit )

inline

candidate_limit_from()

int64_t opennn::cudnn_frontend::candidate_limit_from ( long long value )

inline

conv_energy_autotune_enabled()

bool opennn::cudnn_frontend::conv_energy_autotune_enabled ( )

inline

conv_energy_stage_active()

bool opennn::cudnn_frontend::conv_energy_stage_active ( )

inline

conv_energy_tolerance()

float opennn::cudnn_frontend::conv_energy_tolerance ( )

inline

conv_energy_verbose()

bool opennn::cudnn_frontend::conv_energy_verbose ( )

inline

conv_energy_window_ms()

int64_t opennn::cudnn_frontend::conv_energy_window_ms ( )

inline

conv_engine_notes()

const vector< NumericalNote_t > & opennn::cudnn_frontend::conv_engine_notes ( )

inline

deterministic_algorithms_enabled()

bool opennn::cudnn_frontend::deterministic_algorithms_enabled ( )

inline

device_sm_version()

int opennn::cudnn_frontend::device_sm_version ( )

inline

element_bytes()

int64_t opennn::cudnn_frontend::element_bytes ( DataType_t dtype )

inline

execute_graph()

template<typename TensorMap>

void opennn::cudnn_frontend::execute_graph ( graph::Graph & graph, TensorMap & tensors, void * workspace, const string & what, const string & timing_label )

inline

finalize()

bool opennn::cudnn_frontend::finalize ( graph::Graph & graph, int64_t & workspace_bytes, const string & tag, const bool fp32_graph = false )

inline

finalize_attention()

bool opennn::cudnn_frontend::finalize_attention ( graph::Graph & graph, const string & tag, int64_t & workspace_bytes, bool allow_autotune = false, int64_t workspace_cap = 0 )

inline

frontend_enabled()

bool opennn::cudnn_frontend::frontend_enabled ( )

inline

frontend_verbose()

bool opennn::cudnn_frontend::frontend_verbose ( )

inline

graph_timing_enabled()

bool opennn::cudnn_frontend::graph_timing_enabled ( )

inline

graph_timing_stats()

profiler::Stats & opennn::cudnn_frontend::graph_timing_stats ( )

inline

heuristic_modes()

vector< HeurMode_t > opennn::cudnn_frontend::heuristic_modes ( )

inline

is_convolution_tag()

bool opennn::cudnn_frontend::is_convolution_tag ( const char * tag )

inline

load_cached_plan()

bool opennn::cudnn_frontend::load_cached_plan ( graph::Graph & graph, const cudnnHandle_t handle, int64_t & workspace_bytes )

inline

new_graph()

shared_ptr< graph::Graph > opennn::cudnn_frontend::new_graph ( Type dtype = Type::FP32 )

inline

nhwc_strides()

vector< int64_t > opennn::cudnn_frontend::nhwc_strides ( int64_t c, int64_t h, int64_t w )

inline

nhwc_tensor()

shared_ptr< graph::Tensor_attributes > opennn::cudnn_frontend::nhwc_tensor ( graph::Graph & graph, const char * name, int64_t n, int64_t c, int64_t h, int64_t w )

inline

per_channel_tensor()

shared_ptr< graph::Tensor_attributes > opennn::cudnn_frontend::per_channel_tensor ( graph::Graph & graph, const char * name, int64_t channels )

inline

plan_cache_directory()

const std::filesystem::path & opennn::cudnn_frontend::plan_cache_directory ( )

inline

plan_cache_enabled()

bool opennn::cudnn_frontend::plan_cache_enabled ( )

inline

plan_cache_file()

std::filesystem::path opennn::cudnn_frontend::plan_cache_file ( const graph::Graph & graph )

inline

report_autotune_skipped()

void opennn::cudnn_frontend::report_autotune_skipped ( const char * tag, const char * reason )

inline

run_frontend()

template<typename GraphCache, typename Body>

bool opennn::cudnn_frontend::run_frontend	(	GraphCache &	cache,
		const char *	label,
		Body &&	body )

run_slot()

template<typename TensorMap>

void opennn::cudnn_frontend::run_slot ( GraphSlot & slot, TensorMap & tensors, const char * what, const string & timing_label, bool scratch_autotune )

inline

scalar_tensor()

shared_ptr< graph::Tensor_attributes > opennn::cudnn_frontend::scalar_tensor ( graph::Graph & graph, const char * name, DataType_t dtype, bool pass_by_value = false, int64_t batch = 1 )

inline

sdpa_autotune_enabled()

bool opennn::cudnn_frontend::sdpa_autotune_enabled ( )

inline

sdpa_workspace_cap_bytes()

int64_t opennn::cudnn_frontend::sdpa_workspace_cap_bytes ( int64_t batch, int64_t heads, int64_t q_seq, int64_t src_seq, int64_t head_dim )

inline

sdpa_workspace_cap_override_bytes()

int64_t opennn::cudnn_frontend::sdpa_workspace_cap_override_bytes ( )

inline

select_plan_by_energy()

template<typename TensorMap>

void opennn::cudnn_frontend::select_plan_by_energy ( graph::Graph & graph, TensorMap & tensors, void * workspace, const char * tag )

inline

seq_len_scalar()

shared_ptr< graph::Tensor_attributes > opennn::cudnn_frontend::seq_len_scalar ( graph::Graph & graph, const char * name, int64_t batch = 1 )

inline

set_bhsd_output()

void opennn::cudnn_frontend::set_bhsd_output ( shared_ptr< graph::Tensor_attributes > & tensor, int64_t b, int64_t h, int64_t s, int64_t d, bool interleaved )

inline

set_nhwc_output()

void opennn::cudnn_frontend::set_nhwc_output ( shared_ptr< graph::Tensor_attributes > & tensor, int64_t n, int64_t c, int64_t h, int64_t w )

inline

set_per_channel_output()

void opennn::cudnn_frontend::set_per_channel_output ( shared_ptr< graph::Tensor_attributes > & tensor, int64_t channels )

inline

shared_workspace()

void * opennn::cudnn_frontend::shared_workspace ( int64_t bytes )

inline

store_cached_plan()

void opennn::cudnn_frontend::store_cached_plan ( graph::Graph & graph )

inline

throw_frontend_unavailable()

void opennn::cudnn_frontend::throw_frontend_unavailable ( const string & what )

inline

timing_label()

template<typename... Args>

string opennn::cudnn_frontend::timing_label ( format_string< Args... > fmt, Args &&... args )

inline

to_dtype()

DataType_t opennn::cudnn_frontend::to_dtype ( Type t )

inline

Variable Documentation

check_status

const auto opennn::cudnn_frontend::check_status

inline

Initial value:

= [](auto status, const string& what) {
    throw_if(status.is_bad(),
             "cudnn-frontend {}: {}", what, status.get_message());
}

graph_cache_capacity

size_t opennn::cudnn_frontend::graph_cache_capacity = 8

inlineconstexpr