Refactor cmake

[pcchen]

I create this issue to document information needed to refactor cmake related material.

Minimum requirements to build cytnx. Here is the incomplete list.

For C++

• boost
• blas/lapack. Effectively, we only support MKL and OpenBLAS. Maybe we only officially support these two.

For python wrapper

• python
• pybind11

For GPU support

• CUDA
• cuTENSOR
• cuQuantum

For OpenMP

Some references

• https://cliutils.gitlab.io/modern-cmake/README.html

[pcchen]

CMP0167 https://cmake.org/cmake/help/latest/policy/CMP0167.html

The FindBoost module is removed in cmake 3.30.

find_package(Boost CONFIG) finds the upstream package directly, without the find module.

Examples

• https://cliutils.gitlab.io/modern-cmake/chapters/packages/Boost.html

[pcchen]

CMP0146 https://cmake.org/cmake/help/latest/policy/CMP0146.html

The FindCUDA module is removed in cmake 3.27.

Examples

• https://cliutils.gitlab.io/modern-cmake/chapters/packages/CUDA.html

[pcchen]

OpenMP examples

• https://cliutils.gitlab.io/modern-cmake/chapters/packages/OpenMP.html

[pcchen]

CUDA Quantum and CMake¶

[pcchen]

OpenBLAS

If libopenblas is installed in /opt/OpenBLAS/

On Linux, after setting

export C_INCLUDE_PATH=/opt/OpenBLAS/include
export CPLUS_INCLUDE_PATH=/opt/OpenBLAS/include
export LD_LIBRARY_PATH=/opt/OpenBLAS/lib

cmake can find OpenBLAS and one can compile cytnx.

On MacOS (Apple silicon), still cmake cannot find OpenBLAS.

On Linux and MacOS, one can use CMAKE_PREFIX_PATH.

CMAKE_PREFIX_PATH=/opt/OpenBLAS cmake ..

then cmake can find OpenBLAS. But it will fail to compile because it cannot find lapacke.h.
One still have to set the include_path as above. Then it will compile OK.

==> Next thing is to try to use submodules to handle OpenBLAS.

[IvanaGyro]

Status consolidation (checked against current master, a969a66c)

I agree that #473 should no longer drive a new monolithic “refactor CMake” PR: its requirements have been decomposed into focused issues and PRs. However, it would be inaccurate to call every item complete. The all-platform Pixi workspace is still an open PR, and a few CMake discovery/export tasks remain open.

CMake foundation and policy work

• Presets / modern package configuration: PR Replace Install.sh with CMakePresets.json #579 replaced Install.sh with CMakePresets.json; PR Modernize CytnxConfig.cmake.in and re-find PUBLIC imported-target deps #809 modernized CytnxConfig.cmake and re-finds public imported dependencies; PR build(cmake): support MSVC toolchains #1110 added MSVC CMake portability; and PR build: stop tuning HPTT for the build host, and add Windows CUDA presets #1149 added the usable Windows CUDA presets.
• CMP0146 / FindCUDA: resolved in the actual configuration path: current master uses enable_language(CUDA) plus find_package(CUDAToolkit), not FindCUDA. [build] CUDA configure output shows blank "CUDA Version"/"CUDA Toolkit Root" — uses legacy FindCUDA variables #949 remains only to clean up stale legacy variables in the configure-time status output; PR build: require GCC>=13, CUDA>=12 (12 & 13), cuTENSOR>=2.0 with early actionable errors; harden CUDA finders #950 is the broader open CUDA/CMake hardening umbrella.
• CMP0167 / Boost: not fully resolved. Installed-package discovery is config-mode after PR Modernize CytnxConfig.cmake.in and re-find PUBLIC imported-target deps #809, and PR fix(ci): unblock the CUDA wheel build — CMAKE_PREFIX_PATH separator and Boost pin #1123 fixed the CUDA-wheel CMAKE_PREFIX_PATH handling plus its Boost pin, but the non-Windows root build still uses bare find_package(Boost REQUIRED). That direct follow-up is FindBoost is deprecated #1122.

Original dependency checklist

• Boost: PR build(wheels): source arpack/openblas/boost from conda-forge (Linux + macOS) #1057 moved the Linux/macOS wheel-native dependency source to conda-forge; PR fix(ci): unblock the CUDA wheel build — CMAKE_PREFIX_PATH separator and Boost pin #1123 fixes the CUDA wheel’s CMake prefix/Boost compatibility. The remaining CMake-policy work is FindBoost is deprecated #1122; reproducible contributor provisioning is in PR build: add Pixi development environments for every supported platform #1150.
• OpenBLAS / MKL / LAPACK / LAPACKE: PR Refactor MKL integration to use LP64 interface #605 established the MKL LP64 direction; PR Vendor FindLAPACKE via Inria morse_cmake submodule #800 adopted Morse’s maintained FindLAPACKE; PR build(wheels): source arpack/openblas/boost from conda-forge (Linux + macOS) #1057 provisions conda-forge OpenBLAS/LAPACKE for wheels. LAPACKE library configuration #516 remains the broader LAPACKE discussion. Relocatable downstream usage requirements are tracked by Fix exported CMake usage requirements and relocatable dependencies #1120 and its active implementation PR build(cmake): export dependency usage requirements as imported targets #1157.
• Python / pybind11: PR cmake: require pybind11 >= 3.0, fall back to FetchContent v3.0.0 #755 established the pybind11 3.x direction and PR build(cmake): support MSVC toolchains #1110 uses config-mode Python/pybind11 discovery on MSVC. Testing and documentation for new python build system: scikit-build-core #638 remains the test/documentation record; PR build: add Pixi development environments for every supported platform #1150 will pin/provision the development-side Python and pybind11 toolchain.
• CUDA / cuTENSOR / cuQuantum: PR cmake: portable CUDA arch default + require CMake 3.25 for device LTO #875 made CUDA configuration GPU-less-host friendly and raised the CMake floor; PR fix(cmake): make cuTENSOR/cuQuantum finders honest and version-aware #993 made the custom finders honest/version-aware; PR fix(cmake): search all cuTENSOR per-CUDA-major lib dirs, not just the host's #1131 fixed their per-CUDA-major library-layout search; PR ci(release): publish cytnx-cuda GPU wheels to PyPI (nightly + stable) #1023 established the CUDA PyPI wheel/runtime path; and PR build: stop tuning HPTT for the build host, and add Windows CUDA presets #1149 provides Windows CUDA presets. The remaining direct work is spread across CUDA related cmake options #462 and CUDA, CUTENSOR, CUQUANTUM #580 (policy/options), [build] Clarify compiler/CUDA/cuTENSOR requirements and fix FindCUTENSOR.cmake #945/[build] FindCUTENSOR/FindCUQUANTUM hard-code an outdated directory layout; tarball (2.x) and apt installs both break #946/[build] Installed CUDA extension is not rpath-pinned to the build toolkit; runtime loads system/apt CUDA runtime #948/[build] CUDA configure output shows blank "CUDA Version"/"CUDA Toolkit Root" — uses legacy FindCUDA variables #949/[build] Decide supported CUDA toolkit under cuTENSOR >= 2.0 + cuQuantum >= 24.0 (CUDA 11? CUDA 13?) #962 (requirements, finder/rpath/diagnostics), [build] FindCUTENSOR/FindCUQUANTUM: exported CUTENSOR_ROOT/CUQUANTUM_ROOT silently overrides an explicit -D #1129 and [build] Honor CMAKE_PREFIX_PATH for cuTENSOR/cuQuantum instead of requiring the bespoke CUTENSOR_ROOT/CUQUANTUM_ROOT flags #1136 (root-variable and CMAKE_PREFIX_PATH behavior), build: require GCC>=13, CUDA>=12 (12 & 13), cuTENSOR>=2.0 with early actionable errors; harden CUDA finders #950 (partly split umbrella), and build(cmake): export dependency usage requirements as imported targets #1157 (exported SDK usage requirements). Do not count PR docs(install): document cuTENSOR/cuQuantum as required for CUDA builds #1130 as a lasting documentation completion: it was reverted by PR Revert "docs(install): document cuTENSOR/cuQuantum as required for CUDA builds (#1130)" #1134.
• OpenMP: PR Fix HPTT build when using OpenBLAS #581 fixed propagation for the static HPTT dependency, and PR build(wheels): source arpack/openblas/boost from conda-forge (Linux + macOS) #1057 deliberately uses an OpenMP OpenBLAS build to avoid competing thread pools. Run GitHub action with all allowed configurations and cache the build result #583 remains the broader CI-configuration coverage tracker.

Pixi / reproducible development environments

• Use git submodule to handle external dependence #454 is closed with the decision to use managed packages for external build dependencies rather than make Boost, pybind11, or OpenBLAS external git submodules.
• Track Windows builds and PyPI releases #1114 is the Windows build/release umbrella. The closed Windows-only predecessor PR build(windows): add Pixi development environments #1117 was rebased into the current all-platform Pixi PR build: add Pixi development environments for every supported platform #1150.
• build: manage the Linux and macOS development environment with Pixi #1140 defines the Linux/macOS extension of that model: one locked, five-platform workspace covering Boost, OpenBLAS/MKL + LAPACKE, Python/pybind11, CUDA/cuTENSOR, contributor tools, and existing CMake presets.
• PR build: add Pixi development environments for every supported platform #1150 is the active implementation of that proposal. It is not merged yet, so pixi.toml/pixi.lock are not currently on master.
• No CI job exercises pixi.toml, so the development environment can rot silently #1158 tracks the required CI smoke coverage for the manifest, lockfile, and Pixi tasks after the workspace lands.

In short: #473 remains a useful historical index, while the dependency-provisioning work now belongs to #1140/PR #1150/#1158 and the remaining CMake correctness work belongs to #1122, #1120/PR #1157, and the CUDA finder/export follow-ups above. There is no need for a new umbrella implementation under #473, but it should not be closed as fully completed until those active items are resolved or explicitly declined.