Refactoring

src/BLR/BLRBatch.cpp

@@ -36,9 +36,6 @@
 #include "dense/KBLASWrapper.hpp"
 #endif
 
-// for gpu::round_up
-#include "sparse/fronts/FrontalMatrixGPUKernels.hpp"
-
 namespace strumpack {
   namespace BLR {
 

src/BLR/BLRExtendAdd.hpp

@@ -35,8 +35,8 @@
 namespace strumpack {
 
   // forward declarations
-  template<typename scalar_t,typename integer_t> class FrontalMatrix;
-  template<typename scalar_t,typename integer_t> class FrontalMatrixMPI;
+  template<typename scalar_t,typename integer_t> class Front;
+  template<typename scalar_t,typename integer_t> class FrontMPI;
   template<typename scalar_t,typename integer_t> class FrontBLRMPI;
 
   namespace BLR {
@@ -46,8 +46,8 @@ namespace strumpack {
       using DistM_t = DistributedMatrix<scalar_t>;
       using BLR_t = BLRMatrix<scalar_t>;
       using BLRMPI_t = BLRMatrixMPI<scalar_t>;
-      using F_t = FrontalMatrix<scalar_t,integer_t>;
-      using FMPI_t = FrontalMatrixMPI<scalar_t,integer_t>;
+      using F_t = Front<scalar_t,integer_t>;
+      using FMPI_t = FrontMPI<scalar_t,integer_t>;
       using FBLRMPI_t = FrontBLRMPI<scalar_t,integer_t>;
       using VI_t = std::vector<std::size_t>;
       using VVS_t = std::vector<std::vector<scalar_t>>;

src/BLR/BLRMatrix.GPU.cpp

@@ -38,7 +38,7 @@
 #include "BLRBatch.hpp"
 
 #include "dense/GPUWrapper.hpp"
-#include "sparse/fronts/FrontalMatrixGPUKernels.hpp"
+#include "sparse/fronts/FrontGPUKernels.hpp" // for replace_pivots
 
 namespace strumpack {
   namespace BLR {

src/BLR/BLRMatrixMPI.GPU.cpp

@@ -36,10 +36,7 @@
 #include "BLRTileBLAS.hpp"
 #include "BLRBatch.hpp"
 #include "dense/GPUWrapper.hpp"
-
-// TODO why include this?
-#include "sparse/fronts/FrontalMatrixGPUKernels.hpp"
-
+#include "sparse/fronts/FrontGPUKernels.hpp" // for replace_pivots
 
 namespace strumpack {
   namespace BLR {

src/dense/GPUWrapper.hpp

@@ -31,11 +31,31 @@
 
 #include <memory>
 #include "DenseMatrix.hpp"
+#include "sparse/fronts/FrontGPUStructs.hpp"
 
 namespace strumpack {
   namespace gpu {
 #if defined(STRUMPACK_USE_GPU)
 
+    // constexpr
+    inline int align_max_struct() {
+      auto m = sizeof(std::complex<double>);
+      m = std::max(m, sizeof(gpu::FrontData<std::complex<double>>));
+      m = std::max(m, sizeof(gpu::AssembleData<std::complex<double>>));
+      m = std::max(m, sizeof(Triplet<std::complex<double>>));
+      int k = 16;
+      while (k < int(m)) k *= 2;
+      return k;
+    }
+
+    inline std::size_t round_up(std::size_t n) {
+      int k = align_max_struct();
+      return std::size_t((n + k - 1) / k) * k;
+    }
+    template<typename T> T* aligned_ptr(void* p) {
+      return (T*)(round_up(uintptr_t(p)));
+    }
+
     class Stream {
     public:
       Stream();

src/sparse/EliminationTree.cpp

@@ -31,7 +31,7 @@
 
 #include "EliminationTree.hpp"
 #include "fronts/FrontFactory.hpp"
-#include "fronts/FrontalMatrix.hpp"
+#include "fronts/Front.hpp"
 #include "SeparatorTree.hpp"
 
 namespace strumpack {
@@ -51,7 +51,7 @@ namespace strumpack {
   EliminationTree<scalar_t,integer_t>::~EliminationTree() {}
 
   template<typename scalar_t,typename integer_t>
-  FrontalMatrix<scalar_t,integer_t>*
+  Front<scalar_t,integer_t>*
   EliminationTree<scalar_t,integer_t>::root() const {
     return root_.get();
   }
@@ -123,7 +123,7 @@ namespace strumpack {
   }
 
   template<typename scalar_t,typename integer_t>
-  std::unique_ptr<FrontalMatrix<scalar_t,integer_t>>
+  std::unique_ptr<Front<scalar_t,integer_t>>
   EliminationTree<scalar_t,integer_t>::setup_tree
   (const SPOptions<scalar_t>& opts, const SpMat_t& A,
    SeparatorTree<integer_t>& sep_tree,

src/sparse/EliminationTree.hpp

@@ -39,15 +39,15 @@
 
 namespace strumpack {
 
-  template<typename scalar_t,typename integer_t> class FrontalMatrix;
+  template<typename scalar_t,typename integer_t> class Front;
   template<typename integer_t> class SeparatorTree;
 
   // TODO rename this to SuperNodalTree?
   template<typename scalar_t,typename integer_t>
   class EliminationTree {
     using SpMat_t = CompressedSparseMatrix<scalar_t,integer_t>;
     using DenseM_t = DenseMatrix<scalar_t>;
-    using F_t = FrontalMatrix<scalar_t,integer_t>;
+    using F_t = Front<scalar_t,integer_t>;
 
   public:
     EliminationTree() {}

src/sparse/EliminationTreeMPI.cpp

@@ -35,7 +35,7 @@
 #include "ordering/MatrixReorderingMPI.hpp"
 #include "dense/DistributedMatrix.hpp"
 #include "fronts/FrontFactory.hpp"
-#include "fronts/FrontalMatrixMPI.hpp"
+#include "fronts/FrontMPI.hpp"
 
 namespace strumpack {
 

src/sparse/EliminationTreeMPI.hpp

@@ -48,7 +48,7 @@ namespace strumpack {
     using DistM_t = DistributedMatrix<scalar_t>;
     using DenseM_t = DenseMatrix<scalar_t>;
     using DenseMW_t = DenseMatrixWrapper<scalar_t>;
-    using F_t = FrontalMatrix<scalar_t,integer_t>;
+    using F_t = Front<scalar_t,integer_t>;
     using SepRange = std::pair<integer_t,integer_t>;
 
   public:

src/sparse/EliminationTreeMPIDist.cpp

@@ -34,8 +34,8 @@
 #include "PropMapSparseMatrix.hpp"
 #include "ordering/MatrixReorderingMPI.hpp"
 #include "fronts/FrontFactory.hpp"
-#include "fronts/FrontalMatrix.hpp"
-#include "fronts/FrontalMatrixMPI.hpp"
+#include "fronts/Front.hpp"
+#include "fronts/FrontMPI.hpp"
 
 namespace strumpack {
 
@@ -628,7 +628,7 @@ namespace strumpack {
 
 
   template<typename scalar_t,typename integer_t>
-  std::unique_ptr<FrontalMatrix<scalar_t,integer_t>>
+  std::unique_ptr<Front<scalar_t,integer_t>>
   EliminationTreeMPIDist<scalar_t,integer_t>::prop_map
   (const Opts_t& opts, std::vector<std::vector<integer_t>>& local_upd,
    std::vector<float>& local_subtree_work,
@@ -699,7 +699,7 @@ namespace strumpack {
    * [P0_sib,P0_sib+P_sib)
    */
   template<typename scalar_t,typename integer_t>
-  std::unique_ptr<FrontalMatrix<scalar_t,integer_t>>
+  std::unique_ptr<Front<scalar_t,integer_t>>
   EliminationTreeMPIDist<scalar_t,integer_t>::prop_map_sub_graphs
   (const Opts_t& opts, const RedistSubTree<integer_t>& tree,
    int P0, int P, int P0_sib, int P_sib, const MPIComm& fcomm, int level) {

src/sparse/EliminationTreeMPIDist.hpp

@@ -39,7 +39,7 @@ namespace strumpack {
 
   // forward declarations
   template<typename scalar_t,typename integer_t> class MatrixReorderingMPI;
-  template<typename scalar_t,typename integer_t> class FrontalMatrixMPI;
+  template<typename scalar_t,typename integer_t> class FrontMPI;
   template<typename scalar_t,typename integer_t> class CSRMatrixMPI;
   template<typename integer_t> class RedistSubTree;
 
@@ -48,8 +48,8 @@ namespace strumpack {
     public EliminationTreeMPI<scalar_t,integer_t> {
     using DenseM_t = DenseMatrix<scalar_t>;
     using DenseMW_t = DenseMatrixWrapper<scalar_t>;
-    using F_t = FrontalMatrix<scalar_t,integer_t>;
-    using FMPI_t = FrontalMatrixMPI<scalar_t,integer_t>;
+    using F_t = Front<scalar_t,integer_t>;
+    using FMPI_t = FrontMPI<scalar_t,integer_t>;
     using DistM_t = DistributedMatrix<scalar_t>;
     using Opts_t = SPOptions<scalar_t>;
     using SepRange = std::pair<integer_t,integer_t>;

src/sparse/fronts/CMakeLists.txt

@@ -1,16 +1,16 @@
 target_sources(strumpack
   PRIVATE
-  ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixHIP.hip
+  ${CMAKE_CURRENT_LIST_DIR}/FrontHIP.hip
   ${CMAKE_CURRENT_LIST_DIR}/FrontFactory.cpp
-  ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrix.cpp
-  ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixDense.cpp
-  ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixDense.hpp
+  ${CMAKE_CURRENT_LIST_DIR}/Front.cpp
+  ${CMAKE_CURRENT_LIST_DIR}/FrontDense.cpp
+  ${CMAKE_CURRENT_LIST_DIR}/FrontDense.hpp
   ${CMAKE_CURRENT_LIST_DIR}/FrontHSS.cpp
   ${CMAKE_CURRENT_LIST_DIR}/FrontHSS.hpp
   ${CMAKE_CURRENT_LIST_DIR}/FrontBLR.cpp
   ${CMAKE_CURRENT_LIST_DIR}/FrontBLR.hpp
   ${CMAKE_CURRENT_LIST_DIR}/FrontFactory.hpp
-  ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrix.hpp)
+  ${CMAKE_CURRENT_LIST_DIR}/Front.hpp)
 
 install(FILES
   FrontFactory.hpp
@@ -19,10 +19,10 @@ install(FILES
 if(STRUMPACK_USE_MPI)
   target_sources(strumpack
     PRIVATE
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixMPI.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixMPI.hpp
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixDenseMPI.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixDenseMPI.hpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontMPI.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontMPI.hpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontDenseMPI.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontDenseMPI.hpp
     ${CMAKE_CURRENT_LIST_DIR}/FrontHSSMPI.cpp
     ${CMAKE_CURRENT_LIST_DIR}/FrontHSSMPI.hpp
     ${CMAKE_CURRENT_LIST_DIR}/FrontBLRMPI.cpp
@@ -34,10 +34,10 @@ endif()
 if(STRUMPACK_USE_CUDA OR STRUMPACK_USE_HIP OR STRUMPACK_USE_SYCL)
   target_sources(strumpack
     PRIVATE
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixGPU.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixGPU.hpp
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixGPUKernels.hpp)
-
+    ${CMAKE_CURRENT_LIST_DIR}/FrontGPU.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontGPU.hpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontGPUKernels.hpp
+    ${CMAKE_CURRENT_LIST_DIR}/FrontGPUStructs.hpp)
 endif()
 
 if(STRUMPACK_USE_MAGMA)
@@ -48,15 +48,15 @@ endif()
 
 if(STRUMPACK_USE_CUDA)
   target_sources(strumpack PRIVATE
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixCUDA.cu
+    ${CMAKE_CURRENT_LIST_DIR}/FrontCUDA.cu
     ${CMAKE_CURRENT_LIST_DIR}/FrontGPUSPD.cu
     ${CMAKE_CURRENT_LIST_DIR}/FrontGPUSPD.cpp
     ${CMAKE_CURRENT_LIST_DIR}/FrontGPUSPD.hpp)
 endif()
 
 if(STRUMPACK_USE_SYCL)
   target_sources(strumpack PRIVATE
-    ${CMAKE_CURRENT_LIST_DIR}/FrontalMatrixSYCL.cpp)
+    ${CMAKE_CURRENT_LIST_DIR}/FrontSYCL.cpp)
 endif()
 
 if(STRUMPACK_USE_ZFP OR STRUMPACK_USE_SZ3)

src/sparse/fronts/ExtendAdd.cpp

@@ -28,8 +28,8 @@
  */
 
 #include "ExtendAdd.hpp"
-#include "FrontalMatrix.hpp"
-#include "FrontalMatrixMPI.hpp"
+#include "Front.hpp"
+#include "FrontMPI.hpp"
 #include "BLR/BLRMatrixMPI.hpp"
 
 #include "sparse/CompressedSparseMatrix.hpp"

src/sparse/fronts/ExtendAdd.hpp

@@ -35,8 +35,8 @@
 namespace strumpack {
 
   // forward declarations
-  template<typename scalar_t,typename integer_t> class FrontalMatrix;
-  template<typename scalar_t,typename integer_t> class FrontalMatrixMPI;
+  template<typename scalar_t,typename integer_t> class Front;
+  template<typename scalar_t,typename integer_t> class FrontMPI;
   namespace BLR {
     template<typename scalar_t> class BLRMatrixMPI;
   }
@@ -45,8 +45,8 @@ namespace strumpack {
     using DenseM_t = DenseMatrix<scalar_t>;
     using DistM_t = DistributedMatrix<scalar_t>;
     using BLRMPI_t = BLR::BLRMatrixMPI<scalar_t>;
-    using F_t = FrontalMatrix<scalar_t,integer_t>;
-    using FMPI_t = FrontalMatrixMPI<scalar_t,integer_t>;
+    using F_t = Front<scalar_t,integer_t>;
+    using FMPI_t = FrontMPI<scalar_t,integer_t>;
     using VI_t = std::vector<std::size_t>;
     using VVS_t = std::vector<std::vector<scalar_t>>;