tvm_tirx.testing.bench.CudaProfiler

class tvm_tirx.testing.bench.CudaProfiler(*args, **kwargs)

Bases: object

A lightweight wrapper around T.timer_* CUDA intrinsics.

Stores repeated arguments used by timer_init/start/end/finalize so users can call concise methods in kernels. Intended to mirror Pipeline/TileScheduler helpers.

When profiler_enabled is False (or a false-y Expr), calls to init/start/end/finalize become no-ops. This allows constructing a profiler unconditionally and eliminating external if PROFILER_ON: guards.

init(group_id)

Parameters:

group_id (Expr)

start(event_type, leader=None)

Parameters:

• event_type (Enum)

• leader (Expr | bool | None)

end(event_type, leader=None)

Parameters:

• event_type (Enum)

• leader (Expr | bool | None)

finalize(leader=None)

Parameters:

leader (Expr | bool | None)