Skip to content
My Code, My RulesMy Code, My Rules

PY-NUMB0007 · device_wide_numba_synchronization_in_stream_scope

Count device-wide Numba synchronization in modules that create streams.

This is a deterministic rule for python. Read its implementation.

Definition

Section titled “Definition”

Report cuda.synchronize() in a module that creates a Numba CUDA stream. The call waits for all work on the device, including independent work on other streams. Synchronize the stream or an event that represents the actual dependency instead.

Evidence

Section titled “Evidence”

Each finding identifies the device-wide wait. The value is the number of global Numba waits in modules that otherwise use streams.

Exceptions

Section titled “Exceptions”

An explicit benchmark boundary or process shutdown may need a device-wide wait. Retain such a call only with a project waiver that states the measured boundary.

Examples

Section titled “Examples”

Bad

Section titled “Bad”
stream = cuda.stream()
cuda.synchronize()

Good

Section titled “Good”
stream = cuda.stream()
stream.synchronize()

References

Section titled “References”
  • Cites “Numba CUDA documentation”, CUDA Kernel API and synchronization. Open reference
  • Cites “CUDA C++ Programming Guide”, explicit synchronization. Open reference