MPLAPACK Binary128 Setup

Uni20 can optionally use MPLAPACK 3.0.0 or newer as its binary128 BLAS/LAPACK backend. This enables uni20::float128, its complex counterpart, and the scalar-generic tensor, BlockTensor, Krylov, and tensor-network paths backed by those provider operations.

The default dependency mode is AUTO: CMake prefers a compatible installed MPLAPACK package with the binary128 component and otherwise fetches the MPLAPACK 3.0.0 release. Configure the ordinary fetched path with:

cmake -S . -B ./build_codex/build_gcc13_debug_mplapack \
  -DCMAKE_BUILD_TYPE=Debug \
  -DUNI20_ENABLE_FLOAT128=ON -DUNI20_BACKEND_MPLAPACK=ON

The fetched build selects reference components for the enabled scalar families: binary128, native binary80 when available (UNI20_ENABLE_FLOAT80=AUTO), and the MPFR/MPC backend when UNI20_ENABLE_MPREAL=ON. MPLAPACK's other scalar backends, optimized duplicate library, examples, tests, benchmarks, CUDA, and OpenCL targets remain disabled.

With UNI20_BACKEND_MPLAPACK=OFF, float128 and complex256 still work: the binary128 package supplies scalar headers and math, but Uni20 does not register its BLAS/LAPACK kernels. See scalar configuration.

Dependency Selection

Use UNI20_USE_SYSTEM_MPLAPACK to select the normal Uni20 dependency modes:

value	behavior
AUTO	Prefer an installed MPLAPACK 3.0.0 or newer package with binary128; otherwise fetch v3.0.0.
ON	Require a compatible installed package and fail if it is unavailable.
OFF	Always fetch the pinned v3.0.0 source.

MPLAPACK 3.0.0 exposes one self-contained library per precision. Uni20 consumes:

mplapack::mplapack_binary128

The pre-release split between separate mpblas_binary128 and mplapack_binary128 libraries is not part of the supported package contract.

Optional System Build

To provide a system package explicitly, build MPLAPACK 3.0.0 or newer as GNU++23 with only the required backend:

cmake -S ../mplapack -B ./build_codex/mplapack_binary128 \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX="$PWD/build_codex/mplapack_binary128_install" \
  -DMPLAPACK_CXX_STANDARD=23 \
  -DMPLAPACK_CXX_EXTENSIONS=ON \
  -DMPLAPACK_ENABLE_DOUBLE=OFF \
  -DMPLAPACK_ENABLE_MPFR=OFF \
  -DMPLAPACK_ENABLE_GMP=OFF \
  -DMPLAPACK_ENABLE_QD=OFF \
  -DMPLAPACK_ENABLE_DD=OFF \
  -DMPLAPACK_ENABLE_BINARY80=OFF \
  -DMPLAPACK_ENABLE_BINARY128=ON \
  -DMPLAPACK_ENABLE_OPT=OFF \
  -DMPLAPACK_ENABLE_CUDA=OFF \
  -DMPLAPACK_ENABLE_OPENCL=OFF \
  -DMPLAPACK_BUILD_EXAMPLES=OFF \
  -DMPLAPACK_BUILD_TESTS=OFF \
  -DMPLAPACK_BUILD_BENCHMARKS=OFF
 
cmake --build ./build_codex/mplapack_binary128 --parallel
cmake --install ./build_codex/mplapack_binary128

The configure output should report a real binary128 backend, for example:

binary128: _Float128 + strfromf128

Then require that installed package when configuring Uni20:

cmake -S . -B ./build_codex/build_gcc13_debug_mplapack \
  -DCMAKE_BUILD_TYPE=Debug \
  -DUNI20_ENABLE_FLOAT128=ON -DUNI20_BACKEND_MPLAPACK=ON \
  -DUNI20_USE_SYSTEM_MPLAPACK=ON \
  -DCMAKE_PREFIX_PATH="$PWD/build_codex/mplapack_binary128_install"

Pointing directly at the MPLAPACK build tree also works while developing MPLAPACK itself:

cmake -S . -B ./build_codex/build_gcc13_debug_mplapack \
  -DCMAKE_BUILD_TYPE=Debug \
  -DUNI20_ENABLE_FLOAT128=ON -DUNI20_BACKEND_MPLAPACK=ON \
  -DUNI20_USE_SYSTEM_MPLAPACK=ON \
  -Dmplapack_DIR="$PWD/build_codex/mplapack_binary128"

Validate

Build and run the MPLAPACK-gated Uni20 probes:

cmake --build ./build_codex/build_gcc13_debug_mplapack --parallel 36 \
  --target \
    uni20_mplapack_binary128_tests \
    uni20_linalg_mplapack_binary128_tests \
    uni20_krylov_mplapack_binary128_tests \
    spin_half_heisenberg_dmrg_example
 
ctest --test-dir ./build_codex/build_gcc13_debug_mplapack \
  --output-on-failure \
  --parallel 36 \
  -R "MplapackBinary128|SpinHalfHeisenbergDmrg.*Float128"

Supported Upstream State

Uni20 requires the public MPLAPACK 3.0.0 package surface. The fetched dependency is pinned to tag v3.0.0; compatible installed packages must report version 3.0.0 or newer and provide the binary128 component.

Do not build a system MPLAPACK package for Uni20 in GNU++17 mode. The package may build successfully in GNU++17 mode, but the installed headers can record GNU++17 feature-test results. When those headers are later consumed from Uni20's C++23 build, fallback overload decisions may not match the consuming translation unit. Building MPLAPACK itself as GNU++23 keeps the installed configuration aligned with Uni20. The fetched configuration sets this automatically.

Troubleshooting

If UNI20_USE_SYSTEM_MPLAPACK=ON cannot find MPLAPACK, check the package path:

find ./build_codex/mplapack_binary128_install -name mplapackConfig.cmake

Then pass either the install prefix through CMAKE_PREFIX_PATH or the directory containing mplapackConfig.cmake through mplapack_DIR.

If Uni20 configures but binary128 wrappers are disabled or tests are skipped, inspect the generated MPLAPACK configuration:

rg "MPLAPACK_BINARY128_MODE" \
  ./build_codex/mplapack_binary128/include/mplapack_config.h \
  ./build_codex/mplapack_binary128_install/include/mplapack/mplapack_config.h

MPLAPACK_BINARY128_MODE_LDBL is not accepted as Uni20's binary128 backend; use a real binary128 mode such as _Float128 or quadmath.