From 19f9867f96fd8b7a3aaf5a1a2d3fe02511730c43 Mon Sep 17 00:00:00 2001
From: Tobias Ribizel <ribizel@kit.edu>
Date: Fri, 5 Feb 2021 13:26:12 +0100
Subject: [PATCH] add stopping criterion norm2 call

---
 core/solver/bicg.cpp     | 5 +++--
 core/solver/bicgstab.cpp | 5 +++--
 core/solver/cg.cpp       | 3 ++-
 core/solver/cgs.cpp      | 5 +++--
 core/solver/fcg.cpp      | 3 ++-
 5 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/core/solver/bicg.cpp b/core/solver/bicg.cpp
index 264ac0b3b75..52f866c8efe 100644
--- a/core/solver/bicg.cpp
+++ b/core/solver/bicg.cpp
@@ -186,14 +186,15 @@ void Bicg<ValueType>::apply_impl(const LinOp *b, LinOp *x) const
 
     /* Memory movement summary:
      * Per iteration:
-     * 13.5n * values + matrix/preconditioner storage
+     * 14n * values + matrix/preconditioner storage
      * Two iterations:
-     * 27n * values + matrix/preconditioner storage + conj storage
+     * 28n * values + matrix/preconditioner storage + conj storage
      * 2x SpMV:                4n * values + storage + conj storage
      * 2x Preconditioner:      4n * values + storage + conj storage
      * 2x dot                  4n
      * 1x step 1 (axpys)       6n
      * 1x step 2 (axpys)       9n
+     * 1x norm2 residual        n
      */
     while (true) {
         get_preconditioner()->apply(r.get(), z.get());
diff --git a/core/solver/bicgstab.cpp b/core/solver/bicgstab.cpp
index 1d3071912fe..114415f46ed 100644
--- a/core/solver/bicgstab.cpp
+++ b/core/solver/bicgstab.cpp
@@ -139,9 +139,9 @@ void Bicgstab<ValueType>::apply_impl(const LinOp *b, LinOp *x) const
 
     /* Memory movement summary:
      * Per iteration:
-     * 14.5n * values + matrix/preconditioner storage
+     * 15.5n * values + matrix/preconditioner storage
      * Two iterations:
-     * 29n * values + 2 * matrix/preconditioner storage
+     * 31n * values + 2 * matrix/preconditioner storage
      * 2x SpMV:                4n * values + 2 * storage
      * 2x Preconditioner:      4n * values + 2 * storage
      * 3x dot                  6n
@@ -149,6 +149,7 @@ void Bicgstab<ValueType>::apply_impl(const LinOp *b, LinOp *x) const
      * 1x step 1 (fused axpys) 4n
      * 1x step 2 (axpy)        3n
      * 1x step 3 (fused axpys) 7n
+     * 2x norm2 residual       2n
      */
     while (true) {
         ++iter;
diff --git a/core/solver/cg.cpp b/core/solver/cg.cpp
index 7ee32b687c9..4340b6023b9 100644
--- a/core/solver/cg.cpp
+++ b/core/solver/cg.cpp
@@ -129,12 +129,13 @@ void Cg<ValueType>::apply_impl(const LinOp *b, LinOp *x) const
 
     int iter = -1;
     /* Memory movement summary:
-     * 17n * values + matrix/preconditioner storage
+     * 18n * values + matrix/preconditioner storage
      * 1x SpMV:           2n * values + storage
      * 1x Preconditioner: 2n * values + storage
      * 2x dot             4n
      * 1x step 1 (axpy)   3n
      * 1x step 2 (axpys)  6n
+     * 1x norm2 residual   n
      */
     while (true) {
         get_preconditioner()->apply(r.get(), z.get());
diff --git a/core/solver/cgs.cpp b/core/solver/cgs.cpp
index ef9e5029253..38e62e5f60c 100644
--- a/core/solver/cgs.cpp
+++ b/core/solver/cgs.cpp
@@ -140,15 +140,16 @@ void Cgs<ValueType>::apply_impl(const LinOp *b, LinOp *x) const
     int iter = 0;
     /* Memory movement summary:
      * Per iteration:
-     * 13.5n * values + matrix/preconditioner storage
+     * 14n * values + matrix/preconditioner storage
      * Two iterations:
-     * 27n * values + 2 * matrix/preconditioner storage
+     * 28n * values + 2 * matrix/preconditioner storage
      * 2x SpMV:                4n * values + 2 * storage
      * 2x Preconditioner:      4n * values + 2 * storage
      * 2x dot                  4n
      * 1x step 1 (fused axpys) 5n
      * 1x step 2 (fused axpys) 4n
      * 1x step 3 (axpys)       6n
+     * 1x norm2 residual        n
      */
     while (true) {
         r->compute_dot(r_tld.get(), rho.get());
diff --git a/core/solver/fcg.cpp b/core/solver/fcg.cpp
index 7104368e711..33e8a683a49 100644
--- a/core/solver/fcg.cpp
+++ b/core/solver/fcg.cpp
@@ -131,12 +131,13 @@ void Fcg<ValueType>::apply_impl(const LinOp *b, LinOp *x) const
 
     int iter = -1;
     /* Memory movement summary:
-     * 20n * values + matrix/preconditioner storage
+     * 21n * values + matrix/preconditioner storage
      * 1x SpMV:                2n * values + storage
      * 1x Preconditioner:      2n * values + storage
      * 3x dot                  6n
      * 1x step 1 (axpy)        3n
      * 1x step 2 (fused axpys) 7n
+     * 1x norm2 residual        n
      */
     while (true) {
         get_preconditioner()->apply(r.get(), z.get());