tilelang.profiler.bench
=======================

.. py:module:: tilelang.profiler.bench

.. autoapi-nested-parse::

   Profiler and benchmarking utilities for PyTorch functions.



Attributes
----------

.. autoapisummary::

   tilelang.profiler.bench.IS_CUDA
   tilelang.profiler.bench.device
   tilelang.profiler.bench.Event


Classes
-------

.. autoapisummary::

   tilelang.profiler.bench.suppress_stdout_stderr


Functions
---------

.. autoapisummary::

   tilelang.profiler.bench.do_bench


Module Contents
---------------

.. py:class:: suppress_stdout_stderr

   Context manager to suppress stdout and stderr output.

   Source: https://github.com/deepseek-ai/DeepGEMM/blob/main/deep_gemm/testing/bench.py


   .. py:method:: __enter__()


   .. py:method:: __exit__(*_)


.. py:data:: IS_CUDA

.. py:data:: device
   :value: 'cuda:0'


.. py:data:: Event

.. py:function:: do_bench(fn, warmup = 25, rep = 100, _n_warmup = 0, _n_repeat = 0, quantiles = None, fast_flush = True, backend = 'event', return_mode = 'mean')

   Benchmark the runtime of a PyTorch function with L2 cache management.

   This function provides accurate GPU kernel timing by:
   - Clearing L2 cache between runs for consistent measurements
   - Auto-calculating warmup and repeat counts based on kernel runtime
   - Supporting multiple profiling backends (CUDA events or CUPTI)
   - Offering flexible result aggregation (mean/median/min/max/quantiles)

   :param fn: Function to benchmark
   :param warmup: Target warmup time in milliseconds (default: 25)
   :param rep: Target total benchmark time in milliseconds (default: 100)
   :param _n_warmup: Manual override for warmup iterations (default: 0 = auto)
   :param _n_repeat: Manual override for benchmark iterations (default: 0 = auto)
   :param quantiles: Performance percentiles to compute (e.g., [0.5, 0.95])
   :param fast_flush: Use faster L2 cache flush with int32 vs int8 (default: True)
   :param backend: Profiler backend - "event" (CUDA events) or "cupti" (default: "event")
   :param return_mode: Result aggregation method - "mean", "median", "min", or "max"

   :returns: Runtime in milliseconds (float) or list of quantile values if quantiles specified