Mila::Dnn::Compute::CudaTimer Class Reference

GPU-accurate interval timer using a CUDA event pair. More...

Public Member Functions
	CudaTimer ()
	Construct a CudaTimer and allocate its CUDA event pair.
	CudaTimer (const CudaTimer &)=delete
	CudaTimer (CudaTimer &&)=delete
	~CudaTimer ()
	Destructor.
float	elapsedMilliseconds () const
	Returns the elapsed GPU time in milliseconds between start and stop.
CudaTimer &	operator= (const CudaTimer &)=delete
CudaTimer &	operator= (CudaTimer &&)=delete
void	start (cudaStream_t stream)
	Record the start event into the given CUDA stream.
void	stop (cudaStream_t stream)
	Record the stop event into the given CUDA stream.

Private Attributes
cudaEvent_t	start_event_ { nullptr }
cudaEvent_t	stop_event_ { nullptr }

Detailed Description

GPU-accurate interval timer using a CUDA event pair.

Records start and stop events on a given CUDA stream. Reading the elapsed time via elapsedMilliseconds() synchronizes on the stop event, ensuring the GPU has reached that point before the measurement is read back.

This approach avoids inserting a full cudaStreamSynchronize() at every measurement boundary, so it is suitable for fine-grained per-operation profiling during the beta.1 pass.

Example usage:

CudaTimer timer;
timer.start( stream );
// ... dispatch CUDA kernels ...
timer.stop( stream );
float elapsed_ms = timer.elapsedMilliseconds();

Thread safety: A CudaTimer instance must not be used concurrently from multiple threads. Each CUDA stream should have its own CudaTimer.

Constructor & Destructor Documentation

CudaTimer() [1/3]

Mila::Dnn::Compute::CudaTimer::CudaTimer ( )

inline

Construct a CudaTimer and allocate its CUDA event pair.

Exceptions

std::runtime_error

if either CUDA event cannot be created.

Here is the caller graph for this function:

~CudaTimer()

Mila::Dnn::Compute::CudaTimer::~CudaTimer ( )

inline

Destructor.

Releases both CUDA events.

CudaTimer() [2/3]

Mila::Dnn::Compute::CudaTimer::CudaTimer ( const CudaTimer & )

delete

Here is the call graph for this function:

CudaTimer() [3/3]

Mila::Dnn::Compute::CudaTimer::CudaTimer ( CudaTimer && )

delete

Here is the call graph for this function:

Member Function Documentation

elapsedMilliseconds()

float Mila::Dnn::Compute::CudaTimer::elapsedMilliseconds ( ) const

inlinenodiscard

Returns the elapsed GPU time in milliseconds between start and stop.

Synchronizes on the stop event (cudaEventSynchronize) before reading back the elapsed time. The caller's thread blocks until the GPU has reached the stop event.

Returns

Elapsed time in milliseconds as reported by the CUDA event subsystem.

Exceptions

std::runtime_error

if synchronization or elapsed time query fails.

operator=() [1/2]

CudaTimer & Mila::Dnn::Compute::CudaTimer::operator= ( const CudaTimer & )

delete

Here is the call graph for this function:

operator=() [2/2]

CudaTimer & Mila::Dnn::Compute::CudaTimer::operator= ( CudaTimer && )

delete

Here is the call graph for this function:

start()

void Mila::Dnn::Compute::CudaTimer::start ( cudaStream_t stream )

inline

Record the start event into the given CUDA stream.

The GPU timestamp is recorded at the point the stream reaches this event, which may be later than the host-side call.

Parameters

stream

CUDA stream on which to record the start event.

Exceptions

std::runtime_error

if event recording fails.

stop()

void Mila::Dnn::Compute::CudaTimer::stop ( cudaStream_t stream )

inline

Record the stop event into the given CUDA stream.

The GPU timestamp is recorded at the point the stream reaches this event. Call elapsedMilliseconds() afterwards to read the measured interval.

Parameters

stream

CUDA stream on which to record the stop event.

Exceptions

std::runtime_error

if event recording fails.

Member Data Documentation

start_event_

cudaEvent_t Mila::Dnn::Compute::CudaTimer::start_event_ { nullptr }

private

stop_event_

cudaEvent_t Mila::Dnn::Compute::CudaTimer::stop_event_ { nullptr }

private

---

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

• /__w/Mila/Mila/Mila/Src/Dnn/Compute/Devices/Cuda/Profiling/CudaTimer.ixx