Building optional components

Omnistat includes two optional components that can be built and installed to provide additional data collector capabilities.

Hardware counters
Kernel tracing

Both rely on C++ compilations via cmake and additional instructions for each optional component are outlined below.

Hardware Counters

The ROCprofiler extension provides access to low-level GPU hardware counters for in-depth performance analysis. There are different ways to build and install this extension depending on how Omnistat is installed.

Install with setuptools

This method builds the extension in-place, without installing an Omnistat package.

# Install build dependencies
pip install cmake-build-extension nanobind

# Build and install extension in place
BUILD_ROCPROFILER_SDK_EXTENSION=1 python setup.py build_ext --inplace

Install with pip

This method builds the extension and installs Omnistat as a package.

BUILD_ROCPROFILER_SDK_EXTENSION=1 pip install .[query]

With a venv virtual environment:

python -m venv ~/venv/omnistat
BUILD_ROCPROFILER_SDK_EXTENSION=1 ~/venv/omnistat/bin/python -m pip install .[query]

Kernel Tracing

The kernel tracing extension is a standalone C++ shared library (libomnistat_trace.so) that intercepts GPU kernel dispatches at runtime to collect per-kernel timing and execution metrics. Unlike the ROCprofiler extension above, it does not require a Python build step.

Requirements

ROCm 6.4+
libcurl
C++20 compiler
CMake 3.15+

Note

If the compiler does not support std::format, the build automatically fetches the fmt library via CMake’s FetchContent. For offline builds, download the fmt source tree ahead of time and point CMake at it:

cmake -S rocprofiler-sdk/ -B build-trace/ -DBUILD_KERNEL_TRACE_LIB=ON \
  -DFETCHCONTENT_SOURCE_DIR_FMT=/path/to/fmt

Build

cmake -S rocprofiler-sdk/ -B build-trace/ -DBUILD_KERNEL_TRACE_LIB=ON
cmake --build build-trace/

The resulting library is located at build-trace/libomnistat_trace.so. See Kernel Tracing metrics for usage instructions.