From: Martin Kronbichler <martin.kronbichler@uni-a.de>
Date: Thu, 8 Sep 2022 06:48:13 +0000 (+0200)
Subject: Merge pull request #14245 from peterrum/relaxation_pre_post_a_p
X-Git-Tag: v9.5.0-rc1~985
X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=1504c8c0b9cac4867dee3ac2828b186a273f9199;p=dealii.git

Merge pull request #14245 from peterrum/relaxation_pre_post_a_p

PreconditionRelaxation: add specialization
---

1504c8c0b9cac4867dee3ac2828b186a273f9199
diff --cc include/deal.II/lac/precondition.h
index b5702d9959,04779e88cc..7f060aca6f
--- a/include/deal.II/lac/precondition.h
+++ b/include/deal.II/lac/precondition.h
@@@ -1173,22 -1175,115 +1176,115 @@@ namespace interna
          }
      }
  
-     // 2) specialized implementation for inverse-diagonal preconditioner
+     // 2) specialized implementation with a preconditioner and a matrix that
+     // accepts ranges
+     template <
+       typename MatrixType,
+       typename PreconditionerType,
+       typename VectorType,
+       std::enable_if_t<
+         has_vmult_with_std_functions_for_precondition<PreconditionerType,
+                                                       VectorType> &&
+           has_vmult_with_std_functions_for_precondition<MatrixType, VectorType>,
+         int> * = nullptr>
+     void
+     step_operations(const MatrixType &        A,
+                     const PreconditionerType &preconditioner,
+                     VectorType &              dst,
+                     const VectorType &        src,
+                     const double              relaxation,
+                     VectorType &              tmp,
+                     VectorType &,
+                     const unsigned int i,
+                     const bool         transposed)
+     {
+       using Number = typename VectorType::value_type;
+ 
+       if (i == 0)
+         {
+           Number *      dst_ptr = dst.begin();
+           const Number *src_ptr = src.begin();
+ 
+           preconditioner.vmult(
+             dst,
+             src,
+             [&](const unsigned int start_range, const unsigned int end_range) {
+               // zero 'dst' before running the vmult operation
+               if (end_range > start_range)
+                 std::memset(dst.begin() + start_range,
+                             0,
+                             sizeof(Number) * (end_range - start_range));
+             },
+             [&](const unsigned int start_range, const unsigned int end_range) {
+               if (relaxation == 1.0)
+                 return; // nothing to do
+ 
+               const auto src_ptr = src.begin();
+               const auto dst_ptr = dst.begin();
+ 
+               DEAL_II_OPENMP_SIMD_PRAGMA
+               for (std::size_t i = start_range; i < end_range; ++i)
+                 dst_ptr[i] *= relaxation;
+             });
+         }
+       else
+         {
+           tmp.reinit(src, true);
+ 
+           Assert(transposed == false, ExcNotImplemented());
+ 
+           A.vmult(
+             tmp,
+             src,
+             [&](const unsigned int start_range, const unsigned int end_range) {
+               // zero 'tmp' before running the vmult
+               // operation
+               if (end_range > start_range)
+                 std::memset(tmp.begin() + start_range,
+                             0,
+                             sizeof(Number) * (end_range - start_range));
+             },
+             [&](const unsigned int start_range, const unsigned int end_range) {
+               const auto src_ptr = src.begin();
+               const auto tmp_ptr = tmp.begin();
+ 
+               if (relaxation == 1.0)
+                 {
+                   DEAL_II_OPENMP_SIMD_PRAGMA
+                   for (std::size_t i = start_range; i < end_range; ++i)
+                     tmp_ptr[i] = src_ptr[i] - tmp_ptr[i];
+                 }
+               else
+                 {
+                   // note: we scale the residual here to be able to add into
+                   // the dst vector, which contains the solution from the last
+                   // iteration
+                   DEAL_II_OPENMP_SIMD_PRAGMA
+                   for (std::size_t i = start_range; i < end_range; ++i)
+                     tmp_ptr[i] = relaxation * (src_ptr[i] - tmp_ptr[i]);
+                 }
+             });
+ 
+           preconditioner.vmult(dst, tmp);
+         }
+     }
+ 
+     // 3) specialized implementation for inverse-diagonal preconditioner
      template <typename MatrixType,
                typename VectorType,
 -              std::enable_if_t<
 -                !IsBlockVector<VectorType>::value &&
 -                  !has_vmult_with_std_functions<MatrixType,
 -                                                VectorType,
 -                                                DiagonalMatrix<VectorType>>,
 -                VectorType> * = nullptr>
 +              std::enable_if_t<!IsBlockVector<VectorType>::value &&
 +                                 !has_vmult_with_std_functions<
 +                                   MatrixType,
 +                                   VectorType,
 +                                   dealii::DiagonalMatrix<VectorType>>,
 +                               VectorType> * = nullptr>
      void
 -    step_operations(const MatrixType &                A,
 -                    const DiagonalMatrix<VectorType> &preconditioner,
 -                    VectorType &                      dst,
 -                    const VectorType &                src,
 -                    const double                      relaxation,
 -                    VectorType &                      tmp,
 +    step_operations(const MatrixType &                        A,
 +                    const dealii::DiagonalMatrix<VectorType> &preconditioner,
 +                    VectorType &                              dst,
 +                    const VectorType &                        src,
 +                    const double                              relaxation,
 +                    VectorType &                              tmp,
                      VectorType &,
                      const unsigned int i,
                      const bool         transposed)