From 392e0645137a988dea02f74e27ba2a8ad8182328 Mon Sep 17 00:00:00 2001
From: Antonin Portelli <antonin.portelli@me.com>
Date: Mon, 24 Oct 2016 19:24:21 +0100
Subject: [PATCH] fast local peek-poke

---
 lib/lattice/Lattice_peekpoke.h | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)

diff --git a/lib/lattice/Lattice_peekpoke.h b/lib/lattice/Lattice_peekpoke.h
index 9bece943..19d349c4 100644
--- a/lib/lattice/Lattice_peekpoke.h
+++ b/lib/lattice/Lattice_peekpoke.h
@@ -154,7 +154,7 @@ PARALLEL_FOR_LOOP
     template<class vobj,class sobj>
     void peekLocalSite(sobj &s,const Lattice<vobj> &l,std::vector<int> &site){
         
-      GridBase *grid=l._grid;
+      GridBase *grid = l._grid;
 
       typedef typename vobj::scalar_type scalar_type;
       typedef typename vobj::vector_type vector_type;
@@ -164,16 +164,18 @@ PARALLEL_FOR_LOOP
       assert( l.checkerboard== l._grid->CheckerBoard(site));
       assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
+      static const int words=sizeof(vobj)/sizeof(vector_type);
       int odx,idx;
       idx= grid->iIndex(site);
       odx= grid->oIndex(site);
 
-      std::vector<sobj> buf(Nsimd);
-
-      extract(l._odata[odx],buf);
+      scalar_type * vp = (scalar_type *)&l._odata[odx];
+      scalar_type * pt = (scalar_type *)&s;
+      
+      for(int w=0;w<words;w++){
+        pt[w] = vp[idx+w*Nsimd];
+      }
       
-      s = buf[idx];
-
       return;
     };
 
@@ -190,18 +192,17 @@ PARALLEL_FOR_LOOP
       assert( l.checkerboard== l._grid->CheckerBoard(site));
       assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
+      static const int words=sizeof(vobj)/sizeof(vector_type);
       int odx,idx;
       idx= grid->iIndex(site);
       odx= grid->oIndex(site);
 
-      std::vector<sobj> buf(Nsimd);
-
-      // extract-modify-merge cycle is easiest way and this is not perf critical
-      extract(l._odata[odx],buf);
+      scalar_type * vp = (scalar_type *)&l._odata[odx];
+      scalar_type * pt = (scalar_type *)&s;
       
-      buf[idx] = s;
-
-      merge(l._odata[odx],buf);
+      for(int w=0;w<words;w++){
+        vp[idx+w*Nsimd] = pt[w];
+      }
 
       return;
     };