Skip to content

Releases

PrimalSolver 0.1.7

Choose a tag to compare

@github-actions github-actions released this 15 Nov 13:04

Command-line interface and language bindings for PrimalSolver 0.1.7.

Binaries

platform file
Linux x86_64 (static) primal-0.1.7-linux-x86_64
Linux aarch64 (static) primal-0.1.7-linux-aarch64
macOS arm64 primal-0.1.7-macos-arm64
Windows x86_64 primal-0.1.7-windows-x86_64.exe
Windows i686 primal-0.1.7-windows-i686.exe

SHA256SUMS carries the checksums. The two Linux binaries are built and
attached by .github/workflows/ci.yml; the macOS and Windows ones are
cross-compiled from the same tag tree (MinGW-w64 for Windows) because the
hosted macOS runners are not available to this repository, so CI does not
execute them. The 32-bit Windows build also emits three
-Wmisleading-indentation warnings from sdp.c under that GCC; the other
targets are warning-free.

What changed since v0.1.5

  • Julia package (julia/) — the shared library now comes only from
    PrimalSolver_jll: no deps/build.jl, no PRIMAL_LIB / PRIMAL_C_SRC, no
    lookups into a source checkout. A development build is selected with the
    JLLWrappers preference libprimal_path (see julia/README.md).
    Pkg.test() now also runs the 171 programs in examples/julia, the only
    cover of the low-level API in src/Model.jl.
  • CI — .github/workflows/ is published (it used to be gitignored, which
    is why the repository had no CI); the new julia.yml runs Pkg.test() with
    coverage on Julia 1.10 and 1.x, and ci.yml builds the C library and the
    CLI on Linux.
  • Conic solver — the sparse KKT work: cone members absorbed into their
    block, degree ordering of the pattern, fixed members solved outside the
    factorization, fewer polish rounds, and the objective-accuracy fix for pure
    EXP models. Benchmarks in BENCHMARKS.md.

Install

  • Julia: using Pkg; Pkg.add(["JuMP", "PrimalSolver"]) (registration of
    this version in the General registry follows this release).
  • Python: pip install primalsolver.
  • CLI: download the asset for your platform above and run
    primal model.mps --brief.

Known issues

primal --version in these binaries prints 0.1.4: the library carried two
version numbers — PRIMAL_getversion reported 0.1.4 while
PRIMAL_checkversion accepted only 0.1.0 — and the suite pinned both sides.
They are unified in 693534b (the version now lives once, in primal.h);
the binaries here are built from the tag tree and are left untouched.

PrimalSolver 0.1.0 — primal command line

Choose a tag to compare

@c-vision c-vision released this 29 Sep 23:21

PrimalSolver 0.1.0 — primal command line

A convex optimization solver written from scratch in portable C99 — LP, MILP,
QP, QCQP, SOCP, SDP, exponential and power cones, and mixed-integer conic
problems — with no dependencies beyond libc and libm.

This release ships the command-line front end primal: it reads a model
file, solves it and prints the solution.

Downloads

platform asset
macOS arm64 primal-0.1.0-macos-arm64
Linux x86_64 primal-0.1.0-linux-x86_64
Linux aarch64 primal-0.1.0-linux-aarch64
Windows x86_64 primal-0.1.0-windows-x86_64.exe
Windows i686 primal-0.1.0-windows-i686.exe

The Linux and Windows builds are statically linked: the Linux ones carry no
glibc dependency, the Windows ones need only the UCRT (Windows 10+). macOS ships
the Apple-silicon build; on Intel macOS build from source (below).

All built from 787cd46 with -std=c99 -Wall -Wextra -pedantic -O2, zero
warnings. macOS may quarantine the downloaded file:

xattr -d com.apple.quarantine ./primal-0.1.0-macos-arm64

Usage

./primal-0.1.0-<platform> model.lp

The input file is the interface. Formats, detected automatically: CPLEX LP
(.lp), MPS (.mps), OPF (.opf), CBF (.cbf). The algorithm is chosen from
the model (LP, QP, SOCP, SDP, exponential/power, MIP), so one binary covers
every class. Options include --brief, --full (duals and slacks),
--solution-file FILE, --write FILE, --max-iter N,
--tol-pfeas/--tol-dfeas/--tol-gap V and --param NAME=VALUE; --help.

A worked input file with its expected output is in the README, and one
ready-to-run input per problem class is under examples/.

Build from source

make            # builds out/primal
make test       # 4910 checks, 0 failures
make run-samples
make run-examples

POSIX C99 with gcc or clang on Linux, macOS and the BSDs; on Windows use
MinGW-w64 (x86_64-w64-mingw32-gcc or i686-w64-mingw32-gcc) and link with
-pthread -static.

License

Apache-2.0.