Upgrade to Kokkos v4.5.01

.gitmodules

@@ -14,7 +14,7 @@
 	url = https://github.com/kokkos/kokkos.git
 [submodule "vendor/kokkos-kernels"]
 	path = vendor/kokkos-kernels
-	url = https://github.com/yasahi-hpc/kokkos-kernels.git
+	url = https://github.com/kokkos/kokkos-kernels.git
 [submodule "vendor/doxygen-awesome-css"]
 	path = vendor/doxygen-awesome-css
 	url = https://github.com/jothepro/doxygen-awesome-css.git

CMakeLists.txt

@@ -48,7 +48,7 @@ set(DDC_Kokkos_DEPENDENCY_POLICY
 set_property(CACHE DDC_Kokkos_DEPENDENCY_POLICY PROPERTY STRINGS "${DDC_DEPENDENCY_POLICIES}")
 if("${DDC_Kokkos_DEPENDENCY_POLICY}" STREQUAL "AUTO")
     if(NOT TARGET Kokkos::kokkos)
-        find_package(Kokkos 4.4...4.5 QUIET)
+        find_package(Kokkos 4.4...<5 QUIET)
         if(NOT Kokkos_FOUND)
             ddc_configure_kokkos()
         else()
@@ -58,7 +58,7 @@ if("${DDC_Kokkos_DEPENDENCY_POLICY}" STREQUAL "AUTO")
 elseif("${DDC_Kokkos_DEPENDENCY_POLICY}" STREQUAL "EMBEDDED")
     ddc_configure_kokkos()
 elseif("${DDC_Kokkos_DEPENDENCY_POLICY}" STREQUAL "INSTALLED")
-    find_package(Kokkos 4.4...4.5 REQUIRED)
+    find_package(Kokkos 4.4...<5 REQUIRED)
     ddc_check_required_kokkos_options()
 endif()
 
@@ -208,19 +208,15 @@ if("${DDC_BUILD_KERNELS_SPLINES}")
     )
     if("${DDC_KokkosKernels_DEPENDENCY_POLICY}" STREQUAL "AUTO")
         if(NOT TARGET Kokkos::kokkoskernels)
-            # fork https://github.com/yasahi-hpc/kokkos-kernels
-            # on branch develop-spline-kernels-v2
-            find_package(KokkosKernels QUIET)
+            find_package(KokkosKernels 4.5.1...<5 QUIET)
             if(NOT KokkosKernels_FOUND)
                 ddc_configure_kokkos_kernels()
             endif()
         endif()
     elseif("${DDC_KokkosKernels_DEPENDENCY_POLICY}" STREQUAL "EMBEDDED")
         ddc_configure_kokkos_kernels()
     elseif("${DDC_KokkosKernels_DEPENDENCY_POLICY}" STREQUAL "INSTALLED")
-        # fork https://github.com/yasahi-hpc/kokkos-kernels
-        # on branch develop-spline-kernels-v2
-        find_package(KokkosKernels REQUIRED)
+        find_package(KokkosKernels 4.5.1...<5 REQUIRED)
     endif()
 
     add_library(ddc_splines INTERFACE)

README.md

@@ -35,7 +35,7 @@ To use DDC core, one needs the following dependencies:
 
 * a C++17-compliant compiler
 * CMake 3.22...<4
-* Kokkos 4.4...4.5
+* Kokkos 4.4...<5
 * (optional, micro benchmarking) Benchmark 1.8...<2
 * (optional, documentation) Doxygen 1.8.13...<2
 * (optional, unit-testing) GoogleTest 1.14...<2
@@ -48,8 +48,7 @@ To use DDC components, one needs the following dependencies:
   * PDI 1.6...<2
 * (optional, spline interpolation) DDC::splines
   * Ginkgo 1.8.0
-  * Kokkos Kernels fork <https://github.com/yasahi-hpc/kokkos-kernels> on branch develop-spline-kernels-v2
-
+  * Kokkos Kernels 4.5.1...<5
 
 ## Getting the code and basic configuration
 
@@ -84,3 +83,7 @@ One can find the formatting style in the file `.clang-format`.
 
 > [!WARNING]
 > The formatting might not give the same result with different versions of a tool.
+
+## Known issues
+
+* Kokkos 4.5.0 embeds a version of mdspan that is not compatible with DDC, see <https://github.com/kokkos/mdspan/pull/368>. This issue has been fixed in Kokkos 4.5.1.

cmake/DDCConfig.cmake.in

@@ -14,7 +14,7 @@ endmacro()
 
 set(DDC_BUILD_DOUBLE_PRECISION @DDC_BUILD_DOUBLE_PRECISION@)
 
-ddc_find_dependency(Kokkos 4.4...4.5)
+ddc_find_dependency(Kokkos 4.4...<5)
 
 if(@DDC_BUILD_KERNELS_FFT@)
    ddc_find_dependency(KokkosFFT 0.2.1...<1)
@@ -28,9 +28,7 @@ if(@DDC_BUILD_KERNELS_SPLINES@)
    list(PREPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_LIST_DIR})
    ddc_find_dependency(LAPACKE)
    list(POP_FRONT CMAKE_MODULE_PATH)
-   # fork https://github.com/yasahi-hpc/kokkos-kernels
-   # on branch develop-spline-kernels-v2
-   ddc_find_dependency(KokkosKernels)
+   ddc_find_dependency(KokkosKernels 4.5.1...<5)
 endif()
 
 if(@DDC_BUILD_PDI_WRAPPER@)

docs/CMakeLists.txt

@@ -21,7 +21,7 @@ set(DOXYGEN_EXAMPLE_PATH "${DDC_SOURCE_DIR}/examples")
 set(DOXYGEN_EXPAND_ONLY_PREDEF YES)
 set(DOXYGEN_EXTRACT_ALL YES CACHE STRING "")
 set(DOXYGEN_EXCLUDE_SYMBOLS "detail")
-set(DOXYGEN_EXCLUDE_PATTERNS "*/experimental/*;*/detail/*")
+set(DOXYGEN_EXCLUDE_PATTERNS "*/detail/*;*/experimental/*;*/kokkos-kernels-ext/*")
 set(DOXYGEN_EXTRACT_LOCAL_CLASSES YES)
 set(DOXYGEN_FULL_PATH_NAMES NO)
 set(DOXYGEN_HTML_OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/html")

include/ddc/kernels/splines/kokkos-kernels-ext/KokkosBatched_Gbtrs.hpp

@@ -0,0 +1,59 @@
+// Copyright (C) The DDC development team, see COPYRIGHT.md file
+//
+// SPDX-License-Identifier: MIT
+
+// clang-format off
+// NOLINTBEGIN(*)
+
+#ifndef KOKKOSBATCHED_GBTRS_HPP_
+#define KOKKOSBATCHED_GBTRS_HPP_
+
+#include <KokkosBatched_Util.hpp>
+
+/// \author Yuuichi Asahi (yuuichi.asahi@cea.fr)
+
+namespace KokkosBatched {
+
+/// \brief Serial Batched Gbtrs:
+///
+/// Solve A_l x_l = b_l for all l = 0, ..., N
+///   with a general band matrix A using the LU factorization computed
+///   by gbtrf.
+///
+/// \tparam AViewType: Input type for the matrix, needs to be a 2D view
+/// \tparam BViewType: Input type for the right-hand side and the solution,
+/// needs to be a 1D view
+/// \tparam PivViewType: Integer type for pivot indices, needs to be a 1D view
+///
+/// \param A [in]: A is a ldab by n banded matrix.
+/// Details of the LU factorization of the band matrix A, as computed by
+/// gbtrf. U is stored as an upper triangular band matrix with KL+KU
+/// superdiagonals in rows 1 to KL+KU+1, and the multipliers used during
+/// the factorization are stored in rows KL+KU+2 to 2*KL+KU+1.
+/// \param b [inout]: right-hand side and the solution
+/// \param piv [in]: The pivot indices; for 1 <= i <= N, row i of the matrix
+/// was interchanged with row piv(i).
+/// \param kl [in]: kl specifies the number of subdiagonals within the band
+/// of A. kl >= 0
+/// \param ku [in]: ku specifies the number of superdiagonals within the band
+/// of A. ku >= 0
+///
+/// No nested parallel_for is used inside of the function.
+///
+
+template <typename ArgTrans, typename ArgAlgo>
+struct SerialGbtrs {
+  template <typename AViewType, typename BViewType, typename PivViewType>
+  KOKKOS_INLINE_FUNCTION static int invoke(const AViewType &A,
+                                           const BViewType &b,
+                                           const PivViewType &piv, const int kl,
+                                           const int ku);
+};
+}  // namespace KokkosBatched
+
+#include "KokkosBatched_Gbtrs_Serial_Impl.hpp"
+
+#endif  // KOKKOSBATCHED_GBTRS_HPP_
+
+// NOLINTEND(*)
+// clang-format on

include/ddc/kernels/splines/kokkos-kernels-ext/KokkosBatched_Gbtrs_Serial_Impl.hpp

@@ -0,0 +1,170 @@
+// Copyright (C) The DDC development team, see COPYRIGHT.md file
+//
+// SPDX-License-Identifier: MIT
+
+// clang-format off
+// NOLINTBEGIN(*)
+
+#ifndef KOKKOSBATCHED_GBTRS_SERIAL_IMPL_HPP_
+#define KOKKOSBATCHED_GBTRS_SERIAL_IMPL_HPP_
+
+#include <Kokkos_Swap.hpp>
+#include <KokkosBatched_Util.hpp>
+#include <KokkosBlas2_gemv.hpp>
+#include <KokkosBatched_Tbsv.hpp>
+
+namespace KokkosBatched {
+
+template <typename AViewType, typename BViewType>
+KOKKOS_INLINE_FUNCTION static int checkGbtrsInput(
+    [[maybe_unused]] const AViewType &A, [[maybe_unused]] const BViewType &b,
+    [[maybe_unused]] const int kl, [[maybe_unused]] const int ku) {
+  static_assert(Kokkos::is_view_v<AViewType>,
+                "KokkosBatched::gbtrs: AViewType is not a Kokkos::View.");
+  static_assert(Kokkos::is_view_v<BViewType>,
+                "KokkosBatched::gbtrs: BViewType is not a Kokkos::View.");
+  static_assert(AViewType::rank == 2,
+                "KokkosBatched::gbtrs: AViewType must have rank 2.");
+  static_assert(BViewType::rank == 1,
+                "KokkosBatched::gbtrs: BViewType must have rank 1.");
+#if (KOKKOSKERNELS_DEBUG_LEVEL > 0)
+  if (kl < 0) {
+    Kokkos::printf(
+        "KokkosBatched::gbtrs: input parameter kl must not be less than 0: kl "
+        "= "
+        "%d\n",
+        kl);
+    return 1;
+  }
+
+  if (ku < 0) {
+    Kokkos::printf(
+        "KokkosBatched::gbtrs: input parameter ku must not be less than 0: ku "
+        "= "
+        "%d\n",
+        ku);
+    return 1;
+  }
+
+  const int lda = A.extent(0), n = A.extent(1);
+  if (lda < (2 * kl + ku + 1)) {
+    Kokkos::printf(
+        "KokkosBatched::gbtrs: leading dimension of A must be smaller than 2 * "
+        "kl + ku + 1: "
+        "lda = %d, kl = %d, ku = %d\n",
+        lda, kl, ku);
+    return 1;
+  }
+
+  const int ldb = b.extent(0);
+  if (ldb < Kokkos::max(1, n)) {
+    Kokkos::printf(
+        "KokkosBatched::gbtrs: leading dimension of b must be smaller than "
+        "max(1, n): "
+        "ldb = %d, n = %d\n",
+        ldb, n);
+    return 1;
+  }
+
+#endif
+  return 0;
+}
+
+//// Non-transpose ////
+template <>
+struct SerialGbtrs<Trans::NoTranspose, Algo::Level3::Unblocked> {
+  template <typename AViewType, typename BViewType, typename PivViewType>
+  KOKKOS_INLINE_FUNCTION static int invoke(const AViewType &A,
+                                           const BViewType &b,
+                                           const PivViewType &piv, const int kl,
+                                           const int ku) {
+    // Quick return if possible
+    const int n = A.extent(1);
+    if (n == 0) return 0;
+
+    auto info = checkGbtrsInput(A, b, kl, ku);
+    if (info) return info;
+
+    bool lonti = kl > 0;
+    const int kd = ku + kl + 1;
+    if (lonti) {
+      for (int j = 0; j < n - 1; ++j) {
+        const int lm = Kokkos::min(kl, n - j - 1);
+        auto l       = piv(j);
+        // If pivot index is not j, swap rows l and j in b
+        if (l != j) {
+          Kokkos::kokkos_swap(b(l), b(j));
+        }
+
+        // Perform a rank-1 update of the remaining part of the current column
+        // (ger)
+        for (int i = 0; i < lm; ++i) {
+          b(j + 1 + i) = b(j + 1 + i) - A(kd + i, j) * b(j);
+        }
+      }
+    }
+
+    // Solve U*X = b for each right hand side, overwriting B with X.
+    [[maybe_unused]] auto info_tbsv =
+        KokkosBatched::SerialTbsv<Uplo::Upper, Trans::NoTranspose,
+                                  Diag::NonUnit,
+                                  Algo::Trsv::Unblocked>::invoke(A, b, kl + ku);
+
+    return 0;
+  }
+};
+
+//// Transpose ////
+template <>
+struct SerialGbtrs<Trans::Transpose, Algo::Level3::Unblocked> {
+  template <typename AViewType, typename BViewType, typename PivViewType>
+  KOKKOS_INLINE_FUNCTION static int invoke(const AViewType &A,
+                                           const BViewType &b,
+                                           const PivViewType &piv, const int kl,
+                                           const int ku) {
+    // Quick return if possible
+    const int n = A.extent(1);
+    if (n == 0) return 0;
+
+    auto info = checkGbtrsInput(A, b, kl, ku);
+    if (info) return info;
+
+    bool lonti   = kl > 0;
+    const int kd = ku + kl + 1;
+
+    // Solve U*X = b for each right hand side, overwriting B with X.
+    [[maybe_unused]] auto info_tbsv =
+        KokkosBatched::SerialTbsv<Uplo::Upper, Trans::Transpose, Diag::NonUnit,
+                                  Algo::Tbsv::Unblocked>::invoke(A, b, kl + ku);
+
+    if (lonti) {
+      for (int j = n - 2; j >= 0; --j) {
+        const int lm = Kokkos::min(kl, n - j - 1);
+
+        // Gemv transposed
+        auto a = Kokkos::subview(b, Kokkos::pair(j + 1, j + 1 + lm));
+        auto x = Kokkos::subview(A, Kokkos::pair(kd, kd + lm), j);
+        auto y = Kokkos::subview(b, Kokkos::pair(j, j + lm));
+
+        [[maybe_unused]] auto info_gemv =
+            KokkosBlas::Impl::SerialGemvInternal<Algo::Gemv::Unblocked>::invoke(
+                1, a.extent(0), -1.0, a.data(), a.stride_0(), a.stride_0(),
+                x.data(), x.stride_0(), 1.0, y.data(), y.stride_0());
+
+        // If pivot index is not j, swap rows l and j in b
+        auto l = piv(j);
+        if (l != j) {
+          Kokkos::kokkos_swap(b(l), b(j));
+        }
+      }
+    }
+
+    return 0;
+  }
+};
+}  // namespace KokkosBatched
+
+#endif  // KOKKOSBATCHED_GBTRS_SERIAL_IMPL_HPP_
+
+// NOLINTEND(*)
+// clang-format on

include/ddc/kernels/splines/kokkos-kernels-ext/KokkosBatched_Getrs.hpp

@@ -0,0 +1,48 @@
+// Copyright (C) The DDC development team, see COPYRIGHT.md file
+//
+// SPDX-License-Identifier: MIT
+
+// clang-format off
+// NOLINTBEGIN(*)
+
+#ifndef KOKKOSBATCHED_GETRS_HPP_
+#define KOKKOSBATCHED_GETRS_HPP_
+
+#include <KokkosBatched_Util.hpp>
+
+/// \author Yuuichi Asahi (yuuichi.asahi@cea.fr)
+
+namespace KokkosBatched {
+
+/// \brief Serial Batched Getrs:
+/// Solve a system of linear equations
+///   A * x = b or A**T * x = b
+///   with a general N-by-N matrix A using LU factorization computed
+///   by Getrf.
+/// \tparam AViewType: Input type for the matrix, needs to be a 2D view
+/// \tparam PivViewType: Input type for the pivot indices, needs to be a 1D view
+/// \tparam BViewType: Input type for the right-hand side and the solution,
+/// needs to be a 1D view
+///
+/// \param A [inout]: A is a m by n general matrix, a rank 2 view
+/// \param piv [out]: On exit, the pivot indices, a rank 1 view
+/// \param B [inout]: right-hand side and the solution, a rank 1 view
+///
+/// No nested parallel_for is used inside of the function.
+///
+
+template <typename ArgTrans, typename ArgAlgo>
+struct SerialGetrs {
+  template <typename AViewType, typename PivViewType, typename BViewType>
+  KOKKOS_INLINE_FUNCTION static int invoke(const AViewType &A,
+                                           const PivViewType &piv,
+                                           const BViewType &b);
+};
+}  // namespace KokkosBatched
+
+#include "KokkosBatched_Getrs_Serial_Impl.hpp"
+
+#endif  // KOKKOSBATCHED_GETRS_HPP_
+
+// NOLINTEND(*)
+// clang-format on