ParallelComputeLab targets C++20 on Linux x86-64, with AArch64 as a supported secondary platform. It builds with CMake ≥ 3.20 and either GCC or Clang.
| Dependency | Version | Required | How it is obtained |
|---|---|---|---|
| A C++20 compiler | GCC ≥ 11, Clang ≥ 15 | yes | system |
| CMake | ≥ 3.20 | yes | system |
| GoogleTest | 1.17.0 (pinned) | for tests | system package, else FetchContent |
| Google Benchmark | 1.9.4 (pinned) | for benchmarks | system package, else FetchContent |
| OpenMP | 4.5+ | no | system; the library builds without it |
| A CBLAS implementation | any | no | system; enables the GEMM baseline |
cmake/Dependencies.cmake prefers an installed system package and falls back to
downloading a pinned release with FetchContent. Network access is needed
only at configure time in that case; afterwards the test suite and the
benchmark harness run entirely offline.
To force the system-package path (this is what the portability CI job checks):
sudo apt-get install cmake ninja-build libgtest-dev libbenchmark-dev libopenblas-dev
cmake -S . -B build/sysdeps -DPCL_FETCH_DEPENDENCIES=OFFTo pin different upstream versions:
cmake -S . -B build -DPCL_GTEST_VERSION=1.18.0 -DPCL_BENCHMARK_VERSION=1.9.5cmake --preset gcc && cmake --build --preset gcc && ctest --preset gcc| Preset | What it is |
|---|---|
gcc |
GCC, Release (-O3 -DNDEBUG), OpenMP, portable ISA baseline |
clang |
the same sources with Clang, for compiler comparison |
gcc-avx2-native |
-march=native: not portable, only for inspecting generated code |
debug |
Debug with assertions and sanitizers off |
asan |
RelWithDebInfo + AddressSanitizer + UndefinedBehaviorSanitizer |
tsan |
RelWithDebInfo + ThreadSanitizer (must be its own build) |
portable-noopenmp |
no OpenMP, no explicit SIMD: the minimal baseline |
| Option | Default | Meaning |
|---|---|---|
PCL_BUILD_TESTS |
ON |
build the five test binaries |
PCL_BUILD_BENCHMARKS |
ON |
build pcl_benchmarks |
PCL_ENABLE_OPENMP |
ON |
compile the OpenMP kernels; falls back with a warning |
PCL_ENABLE_SIMD |
ON |
compile the explicit SIMD kernels |
PCL_ENABLE_BLAS |
ON |
link a CBLAS if one is found |
PCL_FETCH_DEPENDENCIES |
ON |
allow FetchContent when a package is missing |
PCL_SIMD_LEVEL |
portable |
portable, avx2 or avx512 (see below) |
PCL_NATIVE_ARCH |
OFF |
-march=native; the result is not portable |
PCL_WERROR |
OFF |
warnings are errors |
PCL_ENABLE_ASAN / UBSAN / TSAN |
OFF |
sanitizers (mutually exclusive configurations) |
The library never requires a wide instruction set to be correct:
- the portable baseline compiles with the compiler default, which on x86-64 is SSE2 and on AArch64 is FP/ASIMD;
- the AVX2 kernels are compiled with a function-level target attribute
(
__attribute__((target("avx2,fma")))) and are only entered after a runtime CPUID check, so a binary built on an AVX2 machine still runs on an SSE2-only one; PCL_DISABLE_SIMD(set byPCL_ENABLE_SIMD=OFF) forces every explicit kernel down to portable scalar code.
PCL_SIMD_LEVEL / PCL_NATIVE_ARCH change the flags the translation unit is
compiled with, which is a different thing from the runtime dispatch. Any build
that uses them must have its flags recorded with its results — the harness does
this automatically in <label>.meta.json under build.cxx_flags.
Sanitizers need separate build directories; ASan and TSan in particular cannot share a configuration (the build fails fast if you try).
cmake --preset asan && cmake --build --preset asan && ctest --preset asan
cmake --preset tsan && cmake --build --preset tsan && ctest --preset tsancmake -S . -B build/minimal \
-DPCL_ENABLE_OPENMP=OFF -DPCL_ENABLE_SIMD=OFF -DPCL_ENABLE_BLAS=OFF \
-DPCL_BUILD_TESTS=OFF -DPCL_BUILD_BENCHMARKS=OFF
cmake --build build/minimal -jThis is the configuration that proves the library needs nothing beyond a C++20 compiler and pthreads.
cmake --install build/gcc --prefix /usr/localExports the CMake target ParallelComputeLab::parallel_compute and installs the
public headers, including the generated build_config.hpp that records the
compiler settings.