python_report_interface

Python Report Interface

This folder contains various sample scripts to illustrate the use of NVIDIA Nsight Compute's Python Report Interface.

The interface is provided as a python module in the Nsight Compute installation. It allows you to load the data from Nsight Compute's profile reports in python for analysis and post-processing in your own workflows.

For an introduction to the Python Report Interface, please have a look at our online documentation. You may also be interested in the full API documentation.

Contents

The collection of sample scripts currently contains the following Jupyter Notebooks:

• Breakdown_metrics.ipynb: Find and iterate over breakdown metrics
• Kernel_name_based_filtering.ipynb: Filter IAction objects w.r.t. their name base
• Metric_attributes.ipynb: Query various properties of IMetric objects
• NVTX_support.ipynb: Filter kernels based on NVTX ranges and retrieve NVTX event attributes
• Opcode_instanced_metrics.ipynb: Traverse opcode-instanced metrics along with their SASS instruction types
• Source_correlated_metrics.ipynb: Find and analyze metrics that are correlated with SASS/CUDA-C code

Below scripts cover more advanced content by extending the topics in the previous notebooks:

• Aggregate_instruction_statistics.ipynb: Combines and extends Opcode_instanced_metrics and Source_correlated_metrics

Importing ncu_report

When executing the sample notebooks, make sure you can import the Python module ncu_report. It can usually be found in the extras/python subfolder of an Nsight Compute installation. You can either add its path to your PYTHONPATH environment variable or use the site library to add the path at runtime:

import site

# Use this with the path containing the `ncu_report` module
site.addsitedir("/path/to/Nsight/Compute/extras/python")