From 99e2c1e6666263a8538bff6205be92a7a0c08451 Mon Sep 17 00:00:00 2001
From: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Date: Mon, 12 Dec 2016 09:08:53 +0000
Subject: [PATCH] Kernels options

---
 lib/qcd/action/fermion/StaggeredKernels.cc | 28 ++++++++++++++++++----
 lib/qcd/action/fermion/StaggeredKernels.h  |  6 +++++
 2 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/lib/qcd/action/fermion/StaggeredKernels.cc b/lib/qcd/action/fermion/StaggeredKernels.cc
index b437199c..6608f8de 100644
--- a/lib/qcd/action/fermion/StaggeredKernels.cc
+++ b/lib/qcd/action/fermion/StaggeredKernels.cc
@@ -30,6 +30,8 @@ directory
 namespace Grid {
 namespace QCD {
 
+int StaggeredKernelsStatic::Opt;
+
 template <class Impl>
 StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
 
@@ -189,8 +191,17 @@ void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, Dou
   SiteSpinor naive;
   int oneLink  =0;
   int threeLink=1;
-  DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
-  DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
+  switch(Opt) {
+  case OptHandUnroll:
+    DhopSiteDepthHand(st,lo,U,buf,sF,sU,in,naive,oneLink);
+    DhopSiteDepthHand(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
+    break;
+  case OptGeneric:
+  default:
+    DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
+    DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
+    break;
+  }
   out._odata[sF] =-naive-naik;
 };
 template <class Impl>
@@ -201,8 +212,17 @@ void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, Double
   SiteSpinor naive;
   int oneLink  =0;
   int threeLink=1;
-  DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
-  DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
+  switch(Opt) {
+  case OptHandUnroll:
+    DhopSiteDepthHand(st,lo,U,buf,sF,sU,in,naive,oneLink);
+    DhopSiteDepthHand(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
+    break;
+  case OptGeneric:
+  default:
+    DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
+    DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
+    break;
+  }
   out._odata[sF] =naive+naik;
 };
 
diff --git a/lib/qcd/action/fermion/StaggeredKernels.h b/lib/qcd/action/fermion/StaggeredKernels.h
index f51a4b37..5a6cb45c 100644
--- a/lib/qcd/action/fermion/StaggeredKernels.h
+++ b/lib/qcd/action/fermion/StaggeredKernels.h
@@ -37,6 +37,9 @@ namespace QCD {
   ////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 class StaggeredKernelsStatic { 
  public:
+  enum { OptGeneric, OptHandUnroll, OptInlineAsm };
+  // S-direction is INNERMOST and takes no part in the parity.
+  static int Opt;  // these are a temporary hack
 };
  
 template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , public StaggeredKernelsStatic { 
@@ -52,6 +55,9 @@ public:
 
    void DhopSiteDepth(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteSpinor * buf,
 		     int sF, int sU, const FermionField &in, SiteSpinor &out,int threeLink);
+
+   void DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteSpinor * buf,
+		     int sF, int sU, const FermionField &in, SiteSpinor &out,int threeLink);
       
    void DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU, SiteSpinor * buf,
 		int sF, int sU, const FermionField &in, FermionField &out);