From: Martin Kronbichler <martin.kronbichler@it.uu.se>
Date: Wed, 16 Feb 2022 16:27:01 +0000 (+0100)
Subject: Improve instruction-level parallelism for precondition_SSOR
X-Git-Tag: v9.4.0-rc1~479^2
X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=refs%2Fpull%2F13399%2Fhead;p=dealii.git

Improve instruction-level parallelism for precondition_SSOR
---

diff --git a/include/deal.II/lac/sparse_matrix.templates.h b/include/deal.II/lac/sparse_matrix.templates.h
index 74fba0265f..cd512ff7b8 100644
--- a/include/deal.II/lac/sparse_matrix.templates.h
+++ b/include/deal.II/lac/sparse_matrix.templates.h
@@ -1501,7 +1501,11 @@ SparseMatrix<number>::precondition_SSOR(
           const size_type first_right_of_diagonal_index =
             pos_right_of_diagonal[row];
           number s = 0;
-          for (size_type j = first_right_of_diagonal_index; j < end_row; ++j)
+          // go through the column from the end towards the diagonal in order
+          // to delay the use of the newly computed "dst" values on
+          // out-of-order-execution hardware
+          for (size_type j = end_row - 1; j >= first_right_of_diagonal_index;
+               --j)
             s += val[j] * number(dst(cols->colnums[j]));
 
           *dst_ptr -= s * omega;