From 04ae7929a39795cfd623c3b47e855da5c4f53fed Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Thu, 22 Dec 2016 17:53:22 +0000
Subject: [PATCH] BGQ or KNL assembler now

---
 lib/qcd/action/fermion/WilsonKernelsAsm.cc | 132 +--------------------
 1 file changed, 2 insertions(+), 130 deletions(-)

diff --git a/lib/qcd/action/fermion/WilsonKernelsAsm.cc b/lib/qcd/action/fermion/WilsonKernelsAsm.cc
index d7a9edd3..ab805f4f 100644
--- a/lib/qcd/action/fermion/WilsonKernelsAsm.cc
+++ b/lib/qcd/action/fermion/WilsonKernelsAsm.cc
@@ -53,136 +53,8 @@ WilsonKernels<Impl >::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,
   assert(0);
 }
 
-#if defined(AVX512) 
-#include <simd/Intel512wilson.h>
-
-    ///////////////////////////////////////////////////////////
-    // If we are AVX512 specialise the single precision routine
-    ///////////////////////////////////////////////////////////
-
-#include <simd/Intel512single.h>
-    
-static Vector<vComplexF> signsF;
-
-  template<typename vtype>    
-  int setupSigns(Vector<vtype>& signs ){
-    Vector<vtype> bother(2);
-    signs = bother;
-    vrsign(signs[0]);
-    visign(signs[1]);
-    return 1;
-  }
-
-  static int signInitF = setupSigns(signsF);
-  
-#define label(A)  ilabel(A)
-#define ilabel(A) ".globl\n"  #A ":\n" 
-  
-#define MAYBEPERM(A,perm) if (perm) { A ; }
-#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
-#define FX(A) WILSONASM_ ##A
-#define COMPLEX_TYPE vComplexF
-#define signs signsF
-  
-#undef KERNEL_DAG
-template<> void 
-WilsonKernels<WilsonImplF>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-      
-#define KERNEL_DAG
-template<> void 
-WilsonKernels<WilsonImplF>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
-						   int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-				    
-#undef VMOVIDUP
-#undef VMOVRDUP
-#undef MAYBEPERM
-#undef MULT_2SPIN
-#undef FX 
-#define FX(A) DWFASM_ ## A
-#define MAYBEPERM(A,B) 
-//#define VMOVIDUP(A,B,C)                                  VBCASTIDUPf(A,B,C)
-//#define VMOVRDUP(A,B,C)                                  VBCASTRDUPf(A,B,C)
-#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
-				    
-#undef KERNEL_DAG
-template<> void 
-WilsonKernels<DomainWallVec5dImplF>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-				    
-#define KERNEL_DAG
-template<> void 
-WilsonKernels<DomainWallVec5dImplF>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-#undef COMPLEX_TYPE
-#undef signs
-#undef VMOVRDUP
-#undef MAYBEPERM
-#undef MULT_2SPIN
-#undef FX 
-	
-///////////////////////////////////////////////////////////
-// If we are AVX512 specialise the double precision routine
-///////////////////////////////////////////////////////////
-
-#include <simd/Intel512double.h>
-    
-static Vector<vComplexD> signsD;
-#define signs signsD
-static int signInitD = setupSigns(signsD);
-    
-#define MAYBEPERM(A,perm) if (perm) { A ; }
-#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
-#define FX(A) WILSONASM_ ##A
-#define COMPLEX_TYPE vComplexD
-  
-#undef KERNEL_DAG
-template<> void 
-WilsonKernels<WilsonImplD>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-      
-#define KERNEL_DAG
-template<> void 
-WilsonKernels<WilsonImplD>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
-						   int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-				    
-#undef VMOVIDUP
-#undef VMOVRDUP
-#undef MAYBEPERM
-#undef MULT_2SPIN
-#undef FX 
-#define FX(A) DWFASM_ ## A
-#define MAYBEPERM(A,B) 
-//#define VMOVIDUP(A,B,C)                                  VBCASTIDUPd(A,B,C)
-//#define VMOVRDUP(A,B,C)                                  VBCASTRDUPd(A,B,C)
-#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
-				    
-#undef KERNEL_DAG
-template<> void 
-WilsonKernels<DomainWallVec5dImplD>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-				    
-#define KERNEL_DAG
-template<> void 
-WilsonKernels<DomainWallVec5dImplD>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
-#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
-	
-#undef COMPLEX_TYPE
-#undef signs
-#undef VMOVRDUP
-#undef MAYBEPERM
-#undef MULT_2SPIN
-#undef FX 
-
-#endif //AVX512
+#include <qcd/action/fermion/WilsonKernelsAsmAvx512.h>
+#include <qcd/action/fermion/WilsonKernelsAsmQPX.h>
 
 #define INSTANTIATE_ASM(A)\
 template void WilsonKernels<A>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,\