Add Windows OpenMP 2.0 compatibility

[tony-davis]

Summary

This PR adds compatibility for OpenMP 2.0 implementations (e.g., MSVC, older GCC) that lack functions introduced in OpenMP 3.0. BLIS currently uses omp_get_active_level() and omp_get_max_active_levels() which are not available in OpenMP 2.0, preventing compilation on Windows with MSVC and other OpenMP 2.0 environments.

Changes

frame/thread/bli_thread.c

Added OpenMP 2.0 compatibility shims that detect OpenMP version using the standard _OPENMP macro:

// OpenMP 2.0 compatibility: Provide fallbacks for OpenMP 3.0 functions
// Some compilers (e.g., MSVC, older GCC) only support OpenMP 2.0 which lacks:
//   - omp_get_active_level() (OpenMP 3.0)
//   - omp_get_max_active_levels() (OpenMP 3.0)
// The _OPENMP macro is set to version-specific values by OpenMP-compliant compilers:
//   200203 = OpenMP 2.0, 200811 = OpenMP 3.0, 201107 = OpenMP 3.1, etc.
#if defined(BLIS_ENABLE_OPENMP) && defined(_OPENMP) && _OPENMP < 200811
static inline int omp_get_active_level(void) {
    return 0;  // Always assume top-level (no nested parallelism support)
}
static inline int omp_get_max_active_levels(void) {
    return 1;  // OpenMP 2.0 doesn't support nested parallelism
}
#endif

Why _OPENMP < 200811 instead of _MSC_VER?

• ✅ Portable: Works with any OpenMP 2.x implementation, not just MSVC
• ✅ Future-proof: Won't break when MSVC adds OpenMP 3.0 support
• ✅ Accurate: Detects actual OpenMP version rather than compiler vendor
• ✅ Standard: _OPENMP is defined by all OpenMP-compliant compilers

(Credit: GitHub Copilot suggested this improvement over the original _MSC_VER check)

Testing

Environment:

• Windows 11
• Visual Studio 2022 (MSVC 19.50.35721.0)
• OpenMP 2.0 (via MSVC's /openmp flag)

Results:

• ✅ BLIS builds successfully with OpenMP threading enabled
• ✅ ILP64 support functional
• ✅ Zen4-optimized kernels compile and link correctly
• ✅ No performance overhead (shims only active for OpenMP 2.0)

Compatibility

Backward Compatibility: ✅ Fully maintained

• OpenMP 3.0+ systems: No changes (native functions used)
• OpenMP 2.0 systems: Shims provide safe fallbacks
• No OpenMP: No changes (shims not compiled)

Impact:

• Nested parallelism disabled on OpenMP 2.0 systems (rare in BLAS workloads)
• Safe defaults: single-threaded assumptions for compatibility

Checklist

• Code follows BLIS coding style
• Changes are minimal and focused
• Backward compatibility maintained
• Tested on target platform (Windows/MSVC)
• No performance regression on existing platforms
• Documentation inline via comments

[Copilot]

Pull request overview

This PR enables BLIS to build on Windows using GNU-style Clang compilers (e.g., LLVM/Clang built from source) and adds compatibility for OpenMP 2.0 implementations. The changes support AMD's ROCm testing infrastructure while maintaining full backward compatibility with existing MSVC/Clang-CL toolchains.

Changes:

• Added compiler frontend detection in CMake to distinguish GNU-style Clang from MSVC/Clang-CL on Windows
• Added OpenMP 2.0 compatibility shims for functions introduced in OpenMP 3.0

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
CMakeLists.txt	Adds detection of Clang with GNU-style frontend on Windows to use -fopenmp-simd instead of /openmp:experimental
frame/thread/bli_thread.c	Provides fallback implementations of omp_get_active_level() and omp_get_max_active_levels() for OpenMP 2.0 environments

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.