Skip to content

[FEAT] Support the profiler client for cuTile kernels #524

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
pd-hung wants to merge 1 commit into from
Draft

[FEAT] Support the profiler client for cuTile kernels #524

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 8 additions & 3 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,7 +123,7 @@ uv sync --extra test # tests but no NKI support
### Testing
* To run core TileLens tests, run `pytest tests/`.
* (if NKI installed) To run NKI-specific tests, run `pytest tests/ -m nki`.
* To run cuTile-specific tests, run `pytest tests/unit/test_cutile.py tests/end_to_end/test_cutile.py tests/cutile/test_cutile_tracer.py`.
* To run cuTile-specific tests, run `pytest tests/unit/test_cutile.py tests/end_to_end/test_cutile.py tests/cutile/`.
* To run all tests (Triton + NKI), run `pytest tests/ -m ""`.
* To run visualizer web UI tests, run `npm run test:frontend`.

Expand Down Expand Up @@ -187,7 +187,9 @@ simulation runtimes live under `tilelens/core/simulation/`.

### cuTile tracing

cuTile kernels run on the CPU with NumPy arrays and support tracing and visualization:
cuTile kernels run on the CPU with NumPy arrays or Torch tensors and support tracing
and visualization. Like Triton's interpreter, device tensors are copied to the host
for the run and copied back afterwards:

```sh
uv sync --extra test
Expand All @@ -205,7 +207,10 @@ kernel[grid](lhs, rhs, result)
```

Use contiguous NumPy arrays for visualization. cuTile analysis currently supports
only the tracer client; numerical behavior may differ from GPU execution.
the tracer and profiler clients; numerical behavior may differ from GPU execution.
The profiler reports load/store bytes, out-of-bounds tile elements, and for-loop
lengths, and `tile-profiler` runs `ct.launch` calls through the interpreter.
It skips the buffer-load check, which only applies to Triton on AMD GPUs.

## Analysis Clients

Expand Down
Loading
Loading