From 5050833b42196629438ed674fee22b02e35d61fa Mon Sep 17 00:00:00 2001
From: nmeyer-ur <nils.meyer@ur.de>
Date: Tue, 2 Jun 2020 13:08:57 +0200
Subject: [PATCH] revert changes due to performance penalty in Wilson using MPI

---
 Grid/simd/Grid_a64fx-fixedsize.h | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/Grid/simd/Grid_a64fx-fixedsize.h b/Grid/simd/Grid_a64fx-fixedsize.h
index 6f27a4ec..602d56f6 100644
--- a/Grid/simd/Grid_a64fx-fixedsize.h
+++ b/Grid/simd/Grid_a64fx-fixedsize.h
@@ -419,7 +419,6 @@ struct Conj{
   }
 };
 
-/*
 struct TimesMinusI{
   // Complex float
   inline vecf operator()(vecf a, vecf b){
@@ -442,11 +441,14 @@ struct TimesMinusI{
     return svneg_m(a_v, pg_odd, a_v);
   }
 };
-*/
 
 // alternative implementation using fcadd
 // this is not optimal because we have  op1 = op2 + TimesMinusI(op3)  etc
 // ideally we have                      AddTimesMinusI(op1,op2,op3)
+//
+// makes performance worse in Benchmark_wilson using MPI
+// increases halogtime and gathertime
+/*
 struct TimesMinusI{
   // Complex float
   inline vecf operator()(vecf a, vecf b){
@@ -463,8 +465,8 @@ struct TimesMinusI{
     return svcadd_x(pg1, z_v, a, 270);
   }
 };
+*/
 
-/*
 struct TimesI{
   // Complex float
   inline vecf operator()(vecf a, vecf b){
@@ -487,11 +489,15 @@ struct TimesI{
     return svneg_m(a_v, pg_even, a_v);
   }
 };
-*/
+
 
 // alternative implementation using fcadd
 // this is not optimal because we have  op1 = op2 + TimesI(op3)  etc
 // ideally we have                      AddTimesI(op1,op2,op3)
+//
+// makes performance worse in Benchmark_wilson using MPI
+// increases halogtime and gathertime
+/*
 struct TimesI{
   // Complex float
   inline vecf operator()(vecf a, vecf b){
@@ -508,7 +514,7 @@ struct TimesI{
     return svcadd_x(pg1, z_v, a, 90);
   }
 };
-
+*/
 
 
 struct PrecisionChange {