tenzr::profiler¶

Classes¶

	Name
struct	tenzr::profiler::BottleneckReport Heuristic analysis of the dominant performance bottleneck.
class	tenzr::profiler::Profiler Process-wide collector of profile events with Chrome Trace export.
struct	tenzr::profiler::ProfileSummary Aggregated profile statistics for a recorded trace.
class	tenzr::profiler::ScopedTrace RAII helper that records a duration for its enclosing scope.
class	tenzr::profiler::StepMarker RAII helper that marks the start and end of a training step.
struct	tenzr::profiler::TopOp Aggregated timing statistics for one operation in a profile.
struct	tenzr::profiler::TraceEvent One Chrome Trace Event entry recorded by the profiler.
class	tenzr::profiler::TraceSession RAII helper that starts profiling on construction and exports on destruction.

Functions¶

	Name
bool	IsEnabled()
void	Start() Begin recording on the global profiler.
void	Stop() Stop recording on the global profiler.
void	Reset() Clear all events on the global profiler.
void	RecordCounter(std::string name, int64_t value, std::string cat ="memory")
void	RecordInstant(std::string name, std::string cat ="general", char scope ='t')
bool	ExportChromeTrace(const std::string & path)
ProfileSummary	AnalyzeTrace()
void	RecordStreamKernel(std::string name, int64_t stream_id, int64_t dur_us, std::string cat ="gpu_kernel", std::unordered_map< std::string, std::string > args ={})
void	RecordMemoryTransfer(std::string name, size_t bytes, bool h2d, int64_t dur_us)
void	RecordCollectiveComm(std::string name, size_t bytes, int64_t dur_us)
template <typename Fn ,typename... Args> auto	Trace(const std::string & path, Fn && fn, Args &&... args)

Functions Documentation¶

function IsEnabled¶

inline bool IsEnabled()

Return: true while the global profiler is recording.

function Start¶

inline void Start()

Begin recording on the global profiler.

function Stop¶

inline void Stop()

Stop recording on the global profiler.

function Reset¶

inline void Reset()

Clear all events on the global profiler.

function RecordCounter¶

inline void RecordCounter(
    std::string name,
    int64_t value,
    std::string cat ="memory"
)

Parameters:

name Counter name.
value Counter value.
cat Event category.

Record a counter on the global profiler.

function RecordInstant¶

inline void RecordInstant(
    std::string name,
    std::string cat ="general",
    char scope ='t'
)

Parameters:

name Event name.
cat Event category.
scope Instant scope ('g', 'p', or 't').

Record an instant event on the global profiler.

function ExportChromeTrace¶

inline bool ExportChromeTrace(
    const std::string & path
)

Parameters:

path Output file path.

Return: true on success.

Export the global profiler's trace.

function AnalyzeTrace¶

inline ProfileSummary AnalyzeTrace()

Return: Aggregated statistics from the global profiler.

function RecordStreamKernel¶

inline void RecordStreamKernel(
    std::string name,
    int64_t stream_id,
    int64_t dur_us,
    std::string cat ="gpu_kernel",
    std::unordered_map< std::string, std::string > args ={}
)

Parameters:

name Kernel name.
stream_id Stream identifier.
dur_us Duration in microseconds.
cat Event category.
args Additional event arguments.

Record a device kernel on the global profiler.

function RecordMemoryTransfer¶

inline void RecordMemoryTransfer(
    std::string name,
    size_t bytes,
    bool h2d,
    int64_t dur_us
)

Parameters:

name Transfer name.
bytes Number of bytes transferred.
h2d true for host-to-device, false for device-to-host.
dur_us Duration in microseconds.

Record a memory transfer on the global profiler.

function RecordCollectiveComm¶

inline void RecordCollectiveComm(
    std::string name,
    size_t bytes,
    int64_t dur_us
)

Parameters:

name Operation name.
bytes Number of bytes communicated.
dur_us Duration in microseconds.

Record a collective operation on the global profiler.

function Trace¶

template <typename Fn ,
typename... Args>
auto Trace(
    const std::string & path,
    Fn && fn,
    Args &&... args
)

Parameters:

path Output path for the trace file.
fn Callable to invoke.
args Arguments forwarded to fn.

Return: fn's result.

Functional profiling wrapper analogous to jax.profiler.trace: records a trace over the span of one fn invocation.