Refactor abstract layer for `matmul_veclib()` and enable blas on Windows by ThreeMonth03 · Pull Request #835 · solvcon/modmesh

ThreeMonth03 · 2026-05-26T14:37:44Z

Installs openblas on the Windows platform, includes cblas when implementing matmul_veclib(), and refactors testcases.

Implement abstract layer math/blas_compat.hpp. Move the existing abstraction layer linalg/lapack_compat.hpp to linalg/math_compat.hpp.

For issue #830

ThreeMonth03

@yungyuc Please take a look.

ThreeMonth03 · 2026-05-27T14:29:18Z

+      # Reference: https://www.openmathlib.org/OpenBLAS/docs/install/
+      - name: install OpenBLAS
+        run: |
+          vcpkg install openblas:x64-windows
+          $openblas_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$openblas_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+


This library doesn't includes lapack, so I think we need another pull request to write the script about source build on Windows.

Openblas includes LAPACK. Why this openblas you install using vcpkg does not?

ThreeMonth03 · 2026-05-27T14:32:58Z

Remove the check of cblas because we assume that every machine has installed cblas now.

ThreeMonth03 · 2026-05-27T14:34:07Z

+    def assert_matmul(self, lhs, rhs, expected):
+        result = lhs.matmul(rhs)
+
+        self.assertEqual(list(result.shape), list(expected.shape))
+        np.testing.assert_array_almost_equal(result.ndarray, expected)
+        return result
+
+    def assert_matmul_fast(self, lhs, rhs, expected, matmul_result):
+        fast_result = lhs.matmul_fast(rhs)
+
+        self.assertEqual(list(fast_result.shape), list(expected.shape))
+        np.testing.assert_array_almost_equal(fast_result.ndarray, expected)
+        np.testing.assert_array_almost_equal(fast_result.ndarray,
+                                             matmul_result.ndarray)
+        return fast_result

    def assert_matmul_veclib(self, lhs, rhs, expected, matmul_result):
-        veclib_unavailable = (
-            "SimpleArray::matmul_veclib(): "
-            "Accelerate/CBLAS matmul is only "
-            "available on Apple Silicon"
-        )
-
-        try:
-            veclib_result = lhs.matmul_veclib(rhs)
-        except RuntimeError as exc:
-            # FIXME: Split veclib backend coverage into dedicated tests and
-            # mark unsupported platforms as expected failures once a follow-up
-            # issue is filed.
-            self.assertEqual(str(exc), veclib_unavailable)
-            return
+        veclib_result = lhs.matmul_veclib(rhs)


Add helper function for checking same patterns.

ThreeMonth03 · 2026-05-27T14:34:56Z

I'm not sure the abstract layer should be put in folder /math.

It's OK to keep lapack_compact.hpp here under math/ for now.

ThreeMonth03 · 2026-05-27T14:35:21Z

Abstract layer of blas.

ThreeMonth03 · 2026-05-27T14:38:22Z

+namespace modmesh
+{
+
+namespace blas


I think these api should be wrapped in something like namespace blas, but I barely see another namespace in namespace modmesh, except namespace detail;.

No, we do not use an additional namespace for them. Please remove the namespace blas.

ThreeMonth03 · 2026-05-27T14:39:49Z

+          Complex<double> const * matrix,
+          Complex<double> const * vector,
+          Complex<double> * result,
+          MatrixOperation operation = MatrixOperation::none);


These functions are used in matmul_vec_vec_veclib();, matmul_vec_mat_veclib(); and return matmul_mat_vec_veclib();.

ThreeMonth03 · 2026-05-27T14:40:17Z

Remove abstract layer here.

yungyuc

Clarify: why vcpkg openblas does not have lapack.
Remove the namespace blas.
Name the wrapper to vendor code as *_blas().

yungyuc · 2026-05-28T00:02:59Z

+      # Reference: https://www.openmathlib.org/OpenBLAS/docs/install/
+      - name: install OpenBLAS
+        run: |
+          vcpkg install openblas:x64-windows
+          $openblas_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$openblas_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+


Openblas includes LAPACK. Why this openblas you install using vcpkg does not?

yungyuc · 2026-05-28T00:04:48Z

It's OK to keep lapack_compact.hpp here under math/ for now.

yungyuc · 2026-05-28T00:05:21Z

+#pragma once
+
+/*
+ * Copyright (c) 2026, modmesh contributors.


yungyuc · 2026-05-28T00:06:20Z

+namespace modmesh
+{
+
+namespace blas


No, we do not use an additional namespace for them. Please remove the namespace blas.

yungyuc · 2026-05-28T00:07:19Z

 */

 #include <modmesh/buffer/small_vector.hpp>
-#include <modmesh/math/Complex.hpp>


Good catch.

yungyuc · 2026-05-28T00:08:04Z

@@ -0,0 +1,402 @@
+/*
+ * Copyright (c) 2026, modmesh contributors.


yungyuc · 2026-05-28T00:09:59Z

    void check_inner(size_t lhs_idx, size_t rhs_idx) const;
    void check_tiles() const;
    A matmul_vec_vec();
+    A matmul_vec_vec_veclib();


Name the wrapper to vendor code as *_blas(). veclib is for Apple accelerate/veclib, not all vendor library.

veclib was used elsewhere for vendor lib, because initially the only vendor lib was only accelerate/veclib. A separate issue should be created to refactor.

ThreeMonth03

Clarify: why vcpkg openblas does not have lapack.
Remove the namespace blas.
Name the wrapper to vendor code as *_blas().

I think there are some existing problem in CI now. @yungyuc, please review this pull request after CI works as usual.

ThreeMonth03 · 2026-05-28T14:12:02Z

+      # vcpkg's openblas:x64-windows provides CBLAS but not LAPACK,
+      # since the vcpkg port explicitly builds openblas without LAPACK.
+      # References:
+      # - https://github.com/microsoft/vcpkg/blob/9733af874ef1d4c357ae0d7b4c0fc669e5a5e012/ports/openblas/portfile.cmake#L48-L56
+      - name: install OpenBLAS
+        run: |
+          vcpkg install openblas:x64-windows
+          $openblas_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$openblas_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+


In the reference, vcpkg port builds openblas without turning on the flag -DBUILD_WITHOUT_LAPACK=ON.
It means that vcpkg install openblas:x64-windows doesn't install lapack.

Please also clarify why vcpkg openblas turns off lapack.

modmesh uses both BLAS and LAPACK. The Linux installation has both. If Windows installation of openblas does not include both, please provide a way to supplement LAPACK.

We can discuss more.

vcpkg also affords lapack port for windows, so lapack could be installed by the following command:

vcpkg install lapack:x64-windows

vcpkg's developers mentioned that they prefer dividing LAPACK and OpenBLAS to two ports, so vcpkg openblas turns off lapack. However, they don't explain the motivation furthermore.

In that case, we should install both blas and lapack so that the dependency is aligned across macos, linux, and windows.

ThreeMonth03 · 2026-05-28T14:24:16Z

+float dot_blas(size_t size, float const * lhs, float const * rhs);
+double dot_blas(size_t size, double const * lhs, double const * rhs);
+Complex<float> dot_blas(size_t size,
+                        Complex<float> const * lhs,
+                        Complex<float> const * rhs);
+Complex<double> dot_blas(size_t size,
+                         Complex<double> const * lhs,
+                         Complex<double> const * rhs);


Remove namespace blas; and add suffix _blas for function names.

yungyuc

Remaining points:

Clarify: why vcpkg openblas does not have lapack.
*_veclib() should also be renamed to *_blas when they are on the same call stack of the wrappers already renamed.

@yungyuc, please review this pull request after CI works as usual.

It's OK. The passing CI jobs provided sufficient confidence.

yungyuc · 2026-05-28T23:56:49Z

+      # vcpkg's openblas:x64-windows provides CBLAS but not LAPACK,
+      # since the vcpkg port explicitly builds openblas without LAPACK.
+      # References:
+      # - https://github.com/microsoft/vcpkg/blob/9733af874ef1d4c357ae0d7b4c0fc669e5a5e012/ports/openblas/portfile.cmake#L48-L56
+      - name: install OpenBLAS
+        run: |
+          vcpkg install openblas:x64-windows
+          $openblas_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$openblas_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+


Please also clarify why vcpkg openblas turns off lapack.

modmesh uses both BLAS and LAPACK. The Linux installation has both. If Windows installation of openblas does not include both, please provide a way to supplement LAPACK.

We can discuss more.

yungyuc · 2026-05-29T00:00:58Z

+ * Perform matrix multiplication using vendor BLAS when available.
 */
 template <typename A, typename T>
 A SimpleArrayMixinCalculators<A, T>::matmul_veclib(A const & other) const


SimpleArrayMixinClaculator::matmul_veclib() should also be renamed to matmul_blas like that in SimpleArrayMatmulHelper, because they are on the same call stack.

ThreeMonth03

Clarify: why vcpkg openblas does not have lapack.
*_veclib() should also be renamed to *_blas() when they are on the same call stack of the wrappers already renamed.
I missed to modify the public api name, and I fix it now. @yungyuc , please see the discussion.

ThreeMonth03 · 2026-05-29T04:15:30Z

+      # vcpkg's openblas:x64-windows provides CBLAS but not LAPACK,
+      # since the vcpkg port explicitly builds openblas without LAPACK.
+      # References:
+      # - https://github.com/microsoft/vcpkg/blob/9733af874ef1d4c357ae0d7b4c0fc669e5a5e012/ports/openblas/portfile.cmake#L48-L56
+      - name: install OpenBLAS
+        run: |
+          vcpkg install openblas:x64-windows
+          $openblas_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$openblas_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+


vcpkg also affords lapack port for windows, so lapack could be installed by the following command:

vcpkg install lapack:x64-windows

vcpkg's developers mentioned that they prefer dividing LAPACK and OpenBLAS to two ports, so vcpkg openblas turns off lapack. However, they don't explain the motivation furthermore.

ThreeMonth03 · 2026-05-29T04:17:03Z

@@ -434,8 +434,8 @@ class MODMESH_PYTHON_WRAPPER_VISIBILITY WrapSimpleArray
                { self.idiv(scalar); })
            .def("imatmul", [](wrapped_type & self, wrapped_type const & other)
                 { self.imatmul(other); })
-            .def("imatmul_veclib", [](wrapped_type & self, wrapped_type const & other)
-                 { self.imatmul_veclib(other); })
+            .def("imatmul_blas", [](wrapped_type & self, wrapped_type const & other)
+                 { self.imatmul_blas(other); })


Rename suffix *_veclib() to *_blas().

yungyuc

Install both blas and lapack on windows.

yungyuc · 2026-05-29T04:27:44Z

+      # vcpkg's openblas:x64-windows provides CBLAS but not LAPACK,
+      # since the vcpkg port explicitly builds openblas without LAPACK.
+      # References:
+      # - https://github.com/microsoft/vcpkg/blob/9733af874ef1d4c357ae0d7b4c0fc669e5a5e012/ports/openblas/portfile.cmake#L48-L56
+      - name: install OpenBLAS
+        run: |
+          vcpkg install openblas:x64-windows
+          $openblas_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$openblas_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+


In that case, we should install both blas and lapack so that the dependency is aligned across macos, linux, and windows.

yungyuc · 2026-05-29T04:29:40Z

@@ -434,8 +434,8 @@ class MODMESH_PYTHON_WRAPPER_VISIBILITY WrapSimpleArray
                { self.idiv(scalar); })
            .def("imatmul", [](wrapped_type & self, wrapped_type const & other)
                 { self.imatmul(other); })
-            .def("imatmul_veclib", [](wrapped_type & self, wrapped_type const & other)
-                 { self.imatmul_veclib(other); })
+            .def("imatmul_blas", [](wrapped_type & self, wrapped_type const & other)
+                 { self.imatmul_blas(other); })


ThreeMonth03

Install both blas and lapack on windows.

@yungyuc I install LAPACK on windows runner. Please review it.

ThreeMonth03 · 2026-05-29T07:54:21Z

+      - name: install BLAS and LAPACK
+        run: |
+          vcpkg install openblas:x64-windows lapack:x64-windows
+          $vcpkg_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$vcpkg_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+
      - name: dependency by pip


Install LAPACK and BLAS in windows runner.

ThreeMonth03 · 2026-05-29T07:57:12Z

+# Let CMake find vcpkg-provided OpenBLAS/LAPACK headers, import libraries, and
+# package metadata during configure and link on Windows.
+ifeq ($(OS),Windows_NT)
+CMAKE_TOOLCHAIN_FILE ?= C:/vcpkg/scripts/buildsystems/vcpkg.cmake
+CMAKE_ARGS += -DCMAKE_TOOLCHAIN_FILE=$(CMAKE_TOOLCHAIN_FILE)
+CMAKE_ARGS += -DVCPKG_TARGET_TRIPLET=x64-windows
+endif
+


To ensure makefile works as usual on Windows, additional flags are added in makefile.

Good catch.

yungyuc

LGTM

yungyuc · 2026-05-29T11:12:57Z

+      - name: install BLAS and LAPACK
+        run: |
+          vcpkg install openblas:x64-windows lapack:x64-windows
+          $vcpkg_bin_path = @(
+            "C:\vcpkg\installed\x64-windows\bin",
+            "C:\vcpkg\installed\x64-windows\debug\bin"
+          ) -join ";"
+          echo "$vcpkg_bin_path" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+
      - name: dependency by pip


yungyuc · 2026-05-29T11:15:07Z

+# Let CMake find vcpkg-provided OpenBLAS/LAPACK headers, import libraries, and
+# package metadata during configure and link on Windows.
+ifeq ($(OS),Windows_NT)
+CMAKE_TOOLCHAIN_FILE ?= C:/vcpkg/scripts/buildsystems/vcpkg.cmake
+CMAKE_ARGS += -DCMAKE_TOOLCHAIN_FILE=$(CMAKE_TOOLCHAIN_FILE)
+CMAKE_ARGS += -DVCPKG_TARGET_TRIPLET=x64-windows
+endif
+


Good catch.

ThreeMonth03 marked this pull request as draft May 26, 2026 14:37

ThreeMonth03 force-pushed the optimize-matmul-fast branch 4 times, most recently from 998822d to b6febde Compare May 27, 2026 14:11

ThreeMonth03 marked this pull request as ready for review May 27, 2026 14:41

ThreeMonth03 commented May 27, 2026

View reviewed changes

yungyuc requested changes May 28, 2026

View reviewed changes

yungyuc changed the title ~~Support matmul_veclib() with cblas on Linux and Windows platform~~ Refactor abstract layer for matmul_veclib() and enable blas on Windows May 28, 2026

yungyuc assigned ThreeMonth03 May 28, 2026

yungyuc requested review from KHLee529, chestercheng, j8xixo12 and tigercosmos May 28, 2026 00:13

yungyuc added performance Profiling, runtime, and memory consumption array Multi-dimensional array implementation labels May 28, 2026

yungyuc added this to tensor operations May 28, 2026

github-project-automation Bot moved this to Todo in tensor operations May 28, 2026

yungyuc moved this from Todo to In Progress in tensor operations May 28, 2026

ThreeMonth03 force-pushed the optimize-matmul-fast branch 2 times, most recently from fcc2654 to 9fcfa1b Compare May 28, 2026 14:14

ThreeMonth03 commented May 28, 2026

View reviewed changes

ThreeMonth03 requested a review from yungyuc May 28, 2026 14:31

ThreeMonth03 force-pushed the optimize-matmul-fast branch from 9fcfa1b to dd3dcaf Compare May 28, 2026 16:33

yungyuc requested changes May 29, 2026

View reviewed changes

ThreeMonth03 force-pushed the optimize-matmul-fast branch from dd3dcaf to 30c3008 Compare May 29, 2026 03:26

ThreeMonth03 commented May 29, 2026

View reviewed changes

yungyuc requested changes May 29, 2026

View reviewed changes

ThreeMonth03 force-pushed the optimize-matmul-fast branch 2 times, most recently from 445efd0 to a43243c Compare May 29, 2026 07:06

Support matmul_veclib() with cblas on Linux and Windows platform

a4a77cb

ThreeMonth03 force-pushed the optimize-matmul-fast branch from a43243c to a4a77cb Compare May 29, 2026 08:17

ThreeMonth03 commented May 29, 2026

View reviewed changes

ThreeMonth03 requested a review from yungyuc May 29, 2026 08:26

yungyuc approved these changes May 29, 2026

View reviewed changes

yungyuc merged commit 3ae6690 into solvcon:master May 29, 2026
16 of 17 checks passed

github-project-automation Bot moved this from In Progress to Done in tensor operations May 29, 2026

yungyuc linked an issue May 29, 2026 that may be closed by this pull request

Enable SimpleArray::matmul_veclib() on Linux and Windows using openblas #830

Closed

2 tasks

yungyuc mentioned this pull request May 29, 2026

Enable SimpleArray::matmul_veclib() on Linux and Windows using openblas #830

Closed

2 tasks

		@@ -0,0 +1,402 @@
		/*
		* Copyright (c) 2026, modmesh contributors.

Conversation

ThreeMonth03 commented May 26, 2026 • edited by yungyuc Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ThreeMonth03 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yungyuc left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ThreeMonth03 left a comment

Choose a reason for hiding this comment

Uh oh!

ThreeMonth03 May 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yungyuc left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ThreeMonth03 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yungyuc left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

ThreeMonth03 commented May 26, 2026 •

edited by yungyuc

Loading

yungyuc left a comment •

edited

Loading

ThreeMonth03 May 28, 2026 •

edited

Loading

yungyuc left a comment •

edited

Loading

yungyuc left a comment •

edited

Loading