tvm_tirx.backend.cuda.iket#
NVIDIA IKET annotations and run-iket orchestration.
profile runs an explicit replayable command. run is intended for a
script’s __main__ block: the parent process asks run-iket to replay the
original script or python -m invocation, while the injected tracker and
capture processes execute the supplied callable and then exit.
Classes#
|
An error while validating or running an official IKET profile. |
|
Published artifacts from a successful official IKET profile. |
|
TIRx annotations compiled for NVIDIA's official IKET runtime. |
Functions#
Validate the supported official runtime before an instrumented CUBIN is loaded. |
|
|
Profile a replayable command with NVIDIA's official |
|
Replay the active script under |
Device annotations#
IketProfiler provides checkpoints and two range forms inside TVMScript:
Method |
Meaning |
|---|---|
|
Record a named checkpoint. |
|
Bracket nested ranges in stack order. |
|
Connect a start and end through the returned uint32 token. |
Capture and artifacts#
run(workload, ...) replays the active Python script or module under
the official IKET tracker and capture processes. Put GPU initialization,
compilation, and launches inside the callable, and invoke run from
the script’s __main__ block. For a separate driver, profile takes
a replayable command such as [sys.executable, "profile_workload.py"];
that command compiles and runs its instrumented kernels without starting
another capture.
output_dir selects the published artifact directory. The postprocess
option accepts "json", "perfetto", "html", "all", or
"none". Use clobber=True to replace an existing output directory.
The function signatures list timeout and capture-limit options.
IketProfileResult.trace_json, perfetto, and html return the
single artifact of each requested format. For captures producing multiple
artifacts, use json_traces, perfetto_traces, and html_reports
instead. The Profiling with IKET guide contains a
complete annotation and capture example.