Draft of Python API for NVBench

The prototype is based on pybind11 to minimize boiler-plate
code needed to deal with move-only semantics of many nvbench
classes.

Author: oleksandr-pavlyk

python/.gitignore

@@ -0,0 +1,4 @@
+build
+nvbench_build
+nvbench_install
+__pycache__

python/CMakeLists.txt

@@ -0,0 +1,29 @@
+cmake_minimum_required(VERSION 3.30...4.0)
+
+project(${SKBUILD_PROJECT_NAME} LANGUAGES CXX)
+
+set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+
+find_package(Python REQUIRED COMPONENTS Development.Module)
+find_package(CUDAToolkit REQUIRED)
+
+include(FetchContent)
+
+FetchContent_Declare(
+    pybind11
+    URL https://github.com/pybind/pybind11/archive/refs/tags/v2.13.6.tar.gz
+    URL_HASH SHA256=e08cb87f4773da97fa7b5f035de8763abc656d87d5773e62f6da0587d1f0ec20
+    FIND_PACKAGE_ARGS NAMES pybind11
+)
+FetchContent_MakeAvailable(pybind11)
+
+find_package(nvbench CONFIG REQUIRED)
+
+pybind11_add_module(_nvbench MODULE src/py_nvbench.cpp)
+target_link_libraries(_nvbench PUBLIC nvbench::nvbench)
+target_link_libraries(_nvbench PRIVATE CUDA::cudart_static)
+set_target_properties(_nvbench PROPERTIES INSTALL_RPATH "$ORIGIN")
+
+install(TARGETS _nvbench DESTINATION cuda/nvbench)
+install(IMPORTED_RUNTIME_ARTIFACTS nvbench::nvbench DESTINATION cuda/nvbench)

python/README.md

@@ -0,0 +1,21 @@
+# CUDA Kernel Benchmarking Package
+
+This package provides Python API to CUDA Kernel Benchmarking Library `NVBench`.
+
+## Building
+
+### Build `NVBench` project
+
+```
+cd nvbench/python
+cmake -B nvbench_build --preset nvbench-ci -S $(pwd)/.. -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc -DNVBench_ENABLE_EXAMPLES=OFF -DCMAKE_INSTALL_PREFIX=$(pwd)/nvbench_install
+cmake --build nvbench_build/ --config Release --target install
+
+nvbench_DIR=$(pwd)/nvbench_install/lib/cmake CUDACXX=/usr/local/cuda/bin/nvcc pip install -e .
+```
+
+### Verify that package works
+
+```
+python test/run_1.py
+```

python/cuda/nvbench/__init__.py

@@ -0,0 +1,18 @@
+import importlib.metadata
+
+from cuda.bindings.path_finder import (  # type: ignore[import-not-found]
+    _load_nvidia_dynamic_library,
+)
+
+try:
+    __version__ = importlib.metadata.version("pynvbench")
+except Exception:
+    __version__ = "0.0.0dev"
+
+for libname in ("cupti", "nvperf_target", "nvperf_host"):
+    _load_nvidia_dynamic_library(libname)
+
+from ._nvbench import *  # noqa: E402, F403
+from ._nvbench import register, run_all_benchmarks  # noqa: E402
+
+__all__ = ["register", "run_all_benchmarks"]

python/pyproject.toml

@@ -0,0 +1,58 @@
+[build-system]
+requires = ["scikit-build-core>=0.10", "setuptools_scm"]
+build-backend = "scikit_build_core.build"
+
+[project]
+name = "pynvbench"
+description = "CUDA Kernel Benchmarking Package"
+authors = [{ name = "NVIDIA Corporation" }]
+classifiers = [
+  "Programming Language :: Python :: 3 :: Only",
+  "Environment :: GPU :: NVIDIA CUDA",
+  "License :: OSI Approved :: Apache Software License",
+]
+requires-python = ">=3.9"
+dependencies = [
+  # pathfinder
+  "cuda-bindings",
+
+  # Library expects to find shared libraries
+  # libcupti, libnvperf_target, libnvperf_host
+  # pathfinder is used to find it in the Python layout
+  "nvidia-cuda-cupti-cu12",
+
+  # The shared library
+  # libnvidia-ml must be installed system-wide
+  # (Debian package provider: libnvidia-compute)
+]
+dynamic = ["version"]
+readme = { file = "README.md", content-type = "text/markdown" }
+
+[project.optional-dependencies]
+test = ["pytest", "cupy-cuda12x", "numba"]
+
+[project.urls]
+Homepage = "https://developer.nvidia.com/"
+
+[tool.scikit-build]
+minimum-version = "build-system.requires"
+build-dir = "build/{wheel_tag}"
+
+[tool.scikit-build.cmake]
+version = ">=3.30.4"
+args = []
+build-type = "Release"
+source-dir = "."
+
+[tool.scikit-build.ninja]
+version = ">=1.11"
+make-fallback = true
+
+[tool.scikit-build.metadata.version]
+provider = "scikit_build_core.metadata.setuptools_scm"
+
+[tool.setuptools_scm]
+root = ".."
+
+[tool.scikit-build.wheel.packages]
+"cuda/nvbench" = "cuda/nvbench"

python/src/README.md

@@ -0,0 +1,17 @@
+
+```
+g++ py_nvbench.cpp                                    \
+   -shared -fPIC                                      \
+   -I ${HOME}/repos/pybind11/include                  \
+   -I ${HOME}/repos/pynvbench/nvbench_dir/include     \
+   -I /usr/local/cuda/include                         \
+   $(python3-config --includes)                       \
+   $(python3-config --libs)                           \
+   -L ${HOME}/repos/pynvbench/nvbench_dir/lib/        \
+   -lnvbench                                          \
+   -Wl,-rpath,${HOME}/repos/pynvbench/nvbench_dir/lib \
+   -L /usr/local/cuda/lib64/                          \
+   -lcudart                                           \
+   -Wl,-rpath,/usr/local/cuda/lib64                   \
+   -o _nvbench$(python3-config --extension-suffix)
+```