This directory contains examples demonstrating how to use Nsight Python for profiling and visualizing CUDA kernel performance.

• Python 3.10+
• CUDA-capable GPU
• NVIDIA Nsight Compute (for profiling)

The examples require additional packages beyond the base nsight package:

Most examples use PyTorch for GPU operations:

# Install PyTorch with CUDA support matching your system (e.g., CUDA 12.6, 12.9, 13.0)
# Replace cuXXX with your CUDA version (e.g., cu126, cu129, cu130)
pip install torch --index-url https://download.pytorch.org/whl/cuXXX

Visit pytorch.org for installation commands matching your specific CUDA version.

For the cuTile example (13_cutile.py):

Refer Installing cuTile Python.

Refer cuTile Python Prerequisites section.

For the Triton example (07_triton_minimal.py):

pip install triton

Start with 00_minimal.py:

cd examples
python 00_minimal.py

This will profile a simple matrix multiplication and generate a plot showing the performance.

• 00_minimal.py - Simplest possible benchmark

  • Basic @nsight.analyze.kernel usage
  • Single parameter configuration sweep
  • Default time-based profiling

• 01_compare_throughput.py - Comparing implementations

  • Multiple annotated regions (different matmul implementations)
  • Using NSight Compute metrics (DRAM throughput)
  • Using @nsight.annotate as a function decorator
  • Printing collected data with print_data=True

• 02_parameter_sweep.py - Sweeping parameters

  • Multiple configuration values
  • Visualizing performance across problem sizes

• 03_custom_metrics.py - Computing TFLOPs

  • Using derive_metric to compute custom metric
  • Understanding the metric function signature
  • Transforming time measurements into performance metrics

• 04_multi_parameter.py - Multiple parameters

  • Using itertools.product() for parameter combinations
  • Flexible metric functions with *conf pattern
  • Handling multiple configuration dimensions

• 05_subplots.py - Creating subplot grids

  • Using row_panels and col_panels
  • Organizing multi-dimensional data visually
  • Creating publication-ready plots

• 06_plot_customization.py - Advanced plotting

  • Customizing plot appearance
  • Using plot_callback for advanced control
  • Line plots vs bar charts
  • Annotating data points

• 07_triton_minimal.py - Profiling Triton kernels

  • Integrating Triton GPU kernels
  • Using variant_fields and variant_annotations
  • Comparing against PyTorch baselines with normalize_against
  • Showing speedup metrics

• 08_multiple_metrics.py - Collecting multiple metrics

  • Collecting multiple metrics with using a sequence of metric names
  • Merged results with "Metric" column in DataFrame
  • @plot decorator incompatible with multiple metrics

• 09_advanced_metric_custom.py - Computing advanced custom metric

  • Using derive_metric to compute custom metric from multiple metrics

• 10_multiple_kernels_combine.py - Combining metrics from multiple kernels

  • Using combine_kernel_metrics to aggregate metrics from multiple kernels
  • Summing metrics from consecutive kernel executions

• 11_output_csv.py - Outputting to CSV

  • Using output_csv parameter to enable/disable CSV file generation
  • Using output_prefix to specify output file location and naming
  • Reading and displaying generated CSV files

• 12_cutile.py - Profiling a cuTile kernel

  • Comparing against PyTorch kernel for different problem sizes

• 13_custom_info_collectors.py - Custom information collectors

  • Using @nsight.experimental.collect(scope=...)
  • Collecting information once, per configuration, per run, or per annotation
  • Understanding collection scopes and aggregation behavior
  • Tracking execution order and system state during profiling