LucasBoTang · Oct 27, 2026 · Oct 27, 2026
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
@@ -16,40 +16,50 @@ concurrency:
 
 jobs:
   build:
-    name: Linux / CUDA ${{ matrix.cuda }}
-    runs-on: ubuntu-latest
-    container: nvidia/cuda:${{ matrix.cuda }}-devel-ubuntu22.04
+    name: ${{ matrix.os }} / CUDA ${{ matrix.cuda }}
+    runs-on: ${{ matrix.os }}
+    env:
+      CMAKE_BUILD_PARALLEL_LEVEL: "2"
     strategy:
       fail-fast: false
       matrix:
-        cuda: ["12.4.0", "13.3.0"]
+        os: [ubuntu-22.04, windows-2022]
+        cuda: ["12.4.0", "12.5.0", "12.6.0", "12.8.0", "12.9.0", "13.0.0", "13.1.0"]
     defaults:
       run:
         shell: bash
 
     steps:
+      - uses: actions/checkout@v5
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - uses: ilammy/msvc-dev-cmd@v1
+        if: runner.os == 'Windows'
+        with:
+          arch: x64
+
+      - uses: Jimver/cuda-toolkit@v0.2.30
+        with:
+          cuda: ${{ matrix.cuda }}
+          linux-local-args: '["--toolkit"]'
+          use-github-cache: 'false'
+          log-file-suffix: '${{ matrix.os }}-${{ matrix.cuda }}'
+
       - name: Install system dependencies
+        if: runner.os == 'Linux'
         run: |
-          # CUDA is preinstalled in the image; these dependencies only need Ubuntu sources.
-          # Ignore extra repositories so NVIDIA index sync failures do not block this step.
-          test -s /etc/apt/sources.list
-          apt_options=(
-            -o Dir::Etc::sourcelist=/etc/apt/sources.list
-            -o Dir::Etc::sourceparts=-
-          )
-          apt-get "${apt_options[@]}" update
-          apt-get "${apt_options[@]}" install -y --no-install-recommends \
-            build-essential git ca-certificates python3-dev python3-venv zlib1g-dev
-
-      - uses: actions/checkout@v5
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends zlib1g-dev
 
       - name: Install build tools
         run: |
-          python3 -m venv "$RUNNER_TEMP/build-env"
-          "$RUNNER_TEMP/build-env/bin/python" -m pip install --upgrade pip "cmake>=3.24" ninja
-          echo "$RUNNER_TEMP/build-env/bin" >> "$GITHUB_PATH"
+          python -m pip install --upgrade pip "cmake>=3.24" ninja
 
       - name: Configure
+        # Check every toolkit's build; run the test suites once per OS.
         run: |
           cmake -S . -B build -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
@@ -58,18 +68,31 @@ jobs:
             -DHPDLP_BUILD_SHARED_LIB=ON \
             -DHPDLP_BUILD_CLI=ON \
             -DHPDLP_BUILD_PYTHON=OFF \
-            -DHPDLP_BUILD_TESTS=ON
+            -DHPDLP_BUILD_TESTS=${{ matrix.cuda == '13.1.0' && 'ON' || 'OFF' }}
 
       - name: Build
-        run: cmake --build build --parallel 2
+        run: cmake --build build
+
+      - name: Add PSLP to the Windows DLL search path
+        if: runner.os == 'Windows' && matrix.cuda == '13.1.0'
+        shell: pwsh
+        run: |
+          "$env:GITHUB_WORKSPACE\build\_deps\pslp-build" >> $env:GITHUB_PATH
 
       - name: Run CPU tests
+        if: matrix.cuda == '13.1.0'
+        env:
+          LD_LIBRARY_PATH: ${{ env.CUDA_PATH }}/lib64
         run: ctest --test-dir build -L cpu --output-on-failure
 
       - name: Install Python package
+        if: matrix.cuda == '13.1.0'
         run: |
           python -m pip install ".[test]" \
             --config-settings=cmake.define.CMAKE_CUDA_ARCHITECTURES=80
 
       - name: Run Python tests
+        if: matrix.cuda == '13.1.0'
+        env:
+          LD_LIBRARY_PATH: ${{ env.CUDA_PATH }}/lib64
         run: python -m pytest -m "not gpu"
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -16,7 +16,7 @@ project(hpdlp VERSION 0.0.1 LANGUAGES C CXX CUDA)
 add_compile_definitions(HPDLP_VERSION="${PROJECT_VERSION}")
 
 # Language standards
-set(CMAKE_C_STANDARD 99)
+set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
@@ -31,15 +31,6 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
                                                      "MinSizeRel" "RelWithDebInfo")
 endif()
 
-# Global compile flags
-add_compile_definitions(CUSPARSE_ENABLE_EXPERIMENTAL_API)
-add_compile_options(-Wall -Wextra)
-
-# Windows compatibility
-if(WIN32)
-  add_compile_definitions(strtok_r=strtok_s)
-endif()
-
 # -----------------------------------------------------------------------------
 # CONTROL OPTIONS
 # -----------------------------------------------------------------------------
@@ -51,8 +42,9 @@ option(HPDLP_BUILD_SHARED_LIB "Build the H-PDLP shared library" ON)
 cmake_dependent_option(HPDLP_BUILD_PYTHON "Build the H-PDLP Python bindings" OFF
                        "HPDLP_BUILD_STATIC_LIB" OFF)
 
+# The CLI uses getopt.h, libgen.h and unistd.h, which MSVC does not provide.
 cmake_dependent_option(HPDLP_BUILD_CLI "Build the H-PDLP command-line executable" ON
-                       "HPDLP_BUILD_STATIC_LIB" OFF)
+                       "HPDLP_BUILD_STATIC_LIB;NOT MSVC" OFF)
 
 cmake_dependent_option(HPDLP_BUILD_TESTS "Build the H-PDLP test suite" OFF
                        "HPDLP_BUILD_STATIC_LIB" OFF)
@@ -112,6 +104,25 @@ FetchContent_Declare(
 
 FetchContent_MakeAvailable(pslp)
 
+# Apply project flags after dependencies so they do not inherit our options.
+add_compile_definitions(CUSPARSE_ENABLE_EXPERIMENTAL_API)
+if(MSVC)
+  add_compile_options(
+    $<$<COMPILE_LANGUAGE:C,CXX>:/W4>
+    $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/W4>)
+  add_compile_definitions(_CRT_SECURE_NO_WARNINGS strtok_r=strtok_s)
+  if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA"
+     AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13)
+    # CUDA 13 CCCL headers require the conforming MSVC preprocessor.
+    # Keep CUDA 12 in its default mode.
+    add_compile_options($<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/Zc:preprocessor>)
+  endif()
+else()
+  add_compile_options(
+    "$<$<COMPILE_LANGUAGE:C,CXX>:-Wall;-Wextra>"
+    $<$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>:-Xcompiler=-Wall,-Wextra>)
+endif()
+
 # -----------------------------------------------------------------------------
 # SOURCE DISCOVERY & TARGET DEFINITION
 # -----------------------------------------------------------------------------
@@ -187,6 +198,7 @@ if(HPDLP_BUILD_SHARED_LIB)
   set_target_properties(
     hpdlp_shared
     PROPERTIES OUTPUT_NAME "hpdlp"
+               WINDOWS_EXPORT_ALL_SYMBOLS ON
                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}"
                CUDA_SEPARABLE_COMPILATION ON
                CUDA_RESOLVE_DEVICE_SYMBOLS ON)
@@ -245,6 +257,7 @@ if(NOT HPDLP_BUILD_PYTHON)
     install(
       TARGETS hpdlp_shared
       LIBRARY DESTINATION lib
+      ARCHIVE DESTINATION lib
       RUNTIME DESTINATION bin) # 'bin' for DLLs on Windows, 'lib' for .so on Linux
   endif()
 

diff --git a/README.md b/README.md
@@ -26,7 +26,7 @@ Maximization is supported through MPS `OBJSENSE` and the C/Python APIs. For MPS
 
 ## Installation
 
-Linux and WSL2 are the primary build environments. Native Windows builds have not been validated.
+Linux and WSL2 support all build targets. Native Windows builds with MSVC support the libraries and Python bindings; the CLI requires Linux or WSL2.
 
 - **GPU:** an NVIDIA GPU supported by the selected CUDA toolkit.
 - **CUDA Toolkit:** 12.4+ with NVCC, cuBLAS, cuSPARSE, and cuRAND.
@@ -116,7 +116,7 @@ cmake --build build -j
 ctest --test-dir build --output-on-failure
 ```
 
-See the [test guide](test/README.md) for CPU/GPU subsets and [Python tests](python/README.md#tests) for pytest instructions. CI builds everything and runs the CPU-only C tests and the non-GPU Python tests.
+See the [test guide](test/README.md) for CPU/GPU subsets and [Python tests](python/README.md#tests) for pytest instructions. CI builds on Linux and Windows across the supported CUDA versions. On CUDA 13.1, both platforms also run the CPU-only C tests and the non-GPU Python tests.
 
 ## Acknowledgements
 

diff --git a/test/unit/cpu/test_error_handling.c b/test/unit/cpu/test_error_handling.c
@@ -27,7 +27,11 @@ static lp_problem_t *create_small_problem(void)
 
 int main(void)
 {
+#ifdef _WIN32
+    _putenv_s("CUDA_VISIBLE_DEVICES", "-1");
+#else
     setenv("CUDA_VISIBLE_DEVICES", "", 1);
+#endif
 
     lp_problem_t *problem = create_small_problem();
     check("problem is created without a GPU", problem != NULL);

diff --git a/test/unit/gpu/test_active_set_masks.cu b/test/unit/gpu/test_active_set_masks.cu
@@ -6,11 +6,11 @@
 #include <stdio.h>
 
 __global__ void update_binding_constraint_mask_kernel(
-    const double *, const double *, const double *, const double *,
-    int *, bool *, active_set_mask_changes_t *, double, int, int, int);
+    const double *__restrict__, const double *, const double *__restrict__, const double *__restrict__,
+    int *__restrict__, bool *__restrict__, active_set_mask_changes_t *, double, int, int, int);
 __global__ void update_free_variable_mask_kernel(
-    const double *, const double *, const double *, const double *,
-    const double *, int *, bool *, active_set_mask_changes_t *, double,
+    const double *__restrict__, const double *__restrict__, const double *, const double *__restrict__,
+    const double *__restrict__, int *__restrict__, bool *__restrict__, active_set_mask_changes_t *, double,
     int, int, int);
 
 enum { NUM_ROWS = 10, NUM_COLUMNS = 12, NUM_ZERO_SCALE_COLUMNS = 4, NUM_GAUGES = 2 };