scope_profiler.gpu_timing

Optional CUDA-event timing backends for asynchronous GPU work.

Classes

Name	Description
CuPyCUDATimingBackend	CUDA-event timing through CuPy's current CUDA stream.
GPUTimingBackend	Minimal backend contract used by GPU-timed regions.
TorchCUDATimingBackend	CUDA-event timing through PyTorch's current CUDA stream.

Functions

Name	Description
resolve_gpu_timing_backend	Resolve a CUDA timing backend name or validate a backend-like object.

CuPyCUDATimingBackendclass#

View source

class CuPyCUDATimingBackend()

CUDA-event timing through CuPy’s current CUDA stream.

nameattributeclass attributeinstance attribute#

View source

name = 'cupy'

elapsed_time_nsmethod#

View source

def elapsed_time_ns(start_event, end_event) -> int

record_eventmethod#

View source

def record_event()

synchronizemethod#

View source

def synchronize() -> None

Wait for all work on the current CUDA device.

GPUTimingBackendclass#

View source

class GPUTimingBackend(Protocol)

Bases: Protocol

Minimal backend contract used by GPU-timed regions.

nameattributeinstance attribute#

View source

name: str

elapsed_time_nsmethod#

View source

def elapsed_time_ns(start_event, end_event) -> int

Return elapsed device time between two events in nanoseconds.

record_eventmethod#

View source

def record_event()

Record and return an event on the backend’s current CUDA stream.

TorchCUDATimingBackendclass#

View source

class TorchCUDATimingBackend()

CUDA-event timing through PyTorch’s current CUDA stream.

nameattributeclass attributeinstance attribute#

View source

name = 'torch'

elapsed_time_nsmethod#

View source

def elapsed_time_ns(start_event, end_event) -> int

record_eventmethod#

View source

def record_event()

synchronizemethod#

View source

def synchronize() -> None

Wait for all work on the current CUDA device.

resolve_gpu_timing_backendfunction#

View source

def resolve_gpu_timing_backend(backend='auto') -> GPUTimingBackend

Resolve a CUDA timing backend name or validate a backend-like object.