Reintroducing the hand unrolled loops

2024-11-09 23:45:36 +00:00 · 2015-09-08 17:45:30 +01:00 · 2015-09-08 17:45:30 +01:00 · 2f38ebc446
commit 2f38ebc446
parent 638d6675ee
9 changed files with 134 additions and 210 deletions
--- a/18
+++ b/18
@ -19,11 +19,13 @@ Policies:
 * Support different boundary conditions (finite temp, chem. potential ... )
 * Support different fermion representations? 
  - contained entirely within the integrator presently
-  - Sign of force term.
-  - Rename "Ta" as too unclear
+
+- Sign of force term.

 - Reversibility test.

+- Rename "Ta" as too unclear
+
 - Lanczos

 - Rectangle gauge actions.
@ -31,11 +33,13 @@ Policies:
  Symanzik,
  ... etc...

- Prepare multigrid for HMC.
- Alternate setup schemes.
-* Support for ILDG --- ugly, not done
-* Flavour matrices?
-* FFTnD ?
+- Prepare multigrid for HMC. - Alternate setup schemes.
+
+- Support for ILDG --- ugly, not done
+
+- Flavour matrices?
+
+- FFTnD ?

 ================================================================
 * Hacks and bug fixes to clean up and Audits
--- a/106
+++ b/106
@ -6359,105 +6359,15 @@ fi
 done


-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for __gmpf_init in -lgmp" >&5
-$as_echo_n "checking for __gmpf_init in -lgmp... " >&6; }
-if ${ac_cv_lib_gmp___gmpf_init+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
-  ac_check_lib_save_LIBS=$LIBS
-LIBS="-lgmp  $LIBS"
-cat confdefs.h - <<_ACEOF >conftest.$ac_ext
-/* end confdefs.h.  */
-
-/* Override any GCC internal prototype to avoid an error.
-   Use char because int might match the return type of a GCC
-   builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
-char __gmpf_init ();
-int
-main ()
-{
-return __gmpf_init ();
-  ;
-  return 0;
-}
-_ACEOF
-if ac_fn_cxx_try_link "$LINENO"; then :
-  ac_cv_lib_gmp___gmpf_init=yes
-else
-  ac_cv_lib_gmp___gmpf_init=no
-fi
-rm -f core conftest.err conftest.$ac_objext \
-    conftest$ac_exeext conftest.$ac_ext
-LIBS=$ac_check_lib_save_LIBS
-fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_gmp___gmpf_init" >&5
-$as_echo "$ac_cv_lib_gmp___gmpf_init" >&6; }
-if test "x$ac_cv_lib_gmp___gmpf_init" = xyes; then :
-  cat >>confdefs.h <<_ACEOF
-#define HAVE_LIBGMP 1
-_ACEOF
-
-  LIBS="-lgmp $LIBS"
-
-else
-  as_fn_error $? "GNU Multiple Precision GMP library was not found in your system.
-Please install or provide the correct path to your installation
-Info at: http://www.gmplib.org" "$LINENO" 5
-fi
-
-
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for mpfr_init in -lmpfr" >&5
-$as_echo_n "checking for mpfr_init in -lmpfr... " >&6; }
-if ${ac_cv_lib_mpfr_mpfr_init+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
-  ac_check_lib_save_LIBS=$LIBS
-LIBS="-lmpfr  $LIBS"
-cat confdefs.h - <<_ACEOF >conftest.$ac_ext
-/* end confdefs.h.  */
-
-/* Override any GCC internal prototype to avoid an error.
-   Use char because int might match the return type of a GCC
-   builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
-char mpfr_init ();
-int
-main ()
-{
-return mpfr_init ();
-  ;
-  return 0;
-}
-_ACEOF
-if ac_fn_cxx_try_link "$LINENO"; then :
-  ac_cv_lib_mpfr_mpfr_init=yes
-else
-  ac_cv_lib_mpfr_mpfr_init=no
-fi
-rm -f core conftest.err conftest.$ac_objext \
-    conftest$ac_exeext conftest.$ac_ext
-LIBS=$ac_check_lib_save_LIBS
-fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_mpfr_mpfr_init" >&5
-$as_echo "$ac_cv_lib_mpfr_mpfr_init" >&6; }
-if test "x$ac_cv_lib_mpfr_mpfr_init" = xyes; then :
-  cat >>confdefs.h <<_ACEOF
-#define HAVE_LIBMPFR 1
-_ACEOF
-
-  LIBS="-lmpfr $LIBS"
-
-else
-  as_fn_error $? "GNU Multiple Precision MPFR library was not found in your system.
-Please install or provide the correct path to your installation
-Info at: http://www.mpfr.org/" "$LINENO" 5
-fi
+#AC_CHECK_LIB([gmp],[__gmpf_init],,
+#        [AC_MSG_ERROR(GNU Multiple Precision GMP library was not found in your system.
+#Please install or provide the correct path to your installation
+#Info at: http://www.gmplib.org)])

+#AC_CHECK_LIB([mpfr],[mpfr_init],,
+#        [AC_MSG_ERROR(GNU Multiple Precision MPFR library was not found in your system.
+#Please install or provide the correct path to your installation
+#Info at: http://www.mpfr.org/)])

 # Check whether --enable-simd was given.
 if test "${enable_simd+set}" = set; then :
--- a/configure.ac
+++ b/configure.ac
@ -55,15 +55,15 @@ echo :::::::::::::::::::::::::::::::::::::::::::

 AC_CHECK_FUNCS([gettimeofday])

-AC_CHECK_LIB([gmp],[__gmpf_init],,
-        [AC_MSG_ERROR(GNU Multiple Precision GMP library was not found in your system.
-Please install or provide the correct path to your installation
-Info at: http://www.gmplib.org)])
+#AC_CHECK_LIB([gmp],[__gmpf_init],,
+#        [AC_MSG_ERROR(GNU Multiple Precision GMP library was not found in your system.
+#Please install or provide the correct path to your installation
+#Info at: http://www.gmplib.org)])

-AC_CHECK_LIB([mpfr],[mpfr_init],,
-        [AC_MSG_ERROR(GNU Multiple Precision MPFR library was not found in your system.
-Please install or provide the correct path to your installation
-Info at: http://www.mpfr.org/)])
+#AC_CHECK_LIB([mpfr],[mpfr_init],,
+#        [AC_MSG_ERROR(GNU Multiple Precision MPFR library was not found in your system.
+#Please install or provide the correct path to your installation
+#Info at: http://www.mpfr.org/)])

 AC_ARG_ENABLE([simd],[AC_HELP_STRING([--enable-simd=SSE4|AVX|AVX2|AVX512|MIC],\
 	[Select instructions to be SSE4.0, AVX 1.0, AVX 2.0+FMA, AVX 512, MIC])],\
--- a/lib/Config.h.in
+++ b/lib/Config.h.in
@ -56,12 +56,6 @@
 /* Define to 1 if you have the <inttypes.h> header file. */
 #undef HAVE_INTTYPES_H

-/* Define to 1 if you have the `gmp' library (-lgmp). */
-#undef HAVE_LIBGMP
-
-/* Define to 1 if you have the `mpfr' library (-lmpfr). */
-#undef HAVE_LIBMPFR
-
 /* Define to 1 if you have the <malloc.h> header file. */
 #undef HAVE_MALLOC_H

--- a/lib/algorithms/approx/Remez.h
+++ b/lib/algorithms/approx/Remez.h
@ -17,7 +17,8 @@

 #include <stddef.h>

-#include <algorithms/approx/bigfloat.h>
+//#include <algorithms/approx/bigfloat.h>
+#include <algorithms/approx/bigfloat_double.h>

 #define JMAX 10000 //Maximum number of iterations of Newton's approximation
 #define SUM_MAX 10 // Maximum number of terms in exponential
--- a/lib/qcd/action/Actions.h
+++ b/lib/qcd/action/Actions.h
@ -54,10 +54,10 @@ typedef WilsonGaugeAction<LatticeGaugeFieldD>    WilsonGaugeActionD;
 ////////////////////////////////////////////////////////////////////////////////////////////////////

 #define FermOpTemplateInstantiate(A) \
-  template class A<GparityWilsonImplF>;		\
-  template class A<GparityWilsonImplD>;		\
  template class A<WilsonImplF>;		\
-  template class A<WilsonImplD>;
+  template class A<WilsonImplD>; 
+  //  template class A<GparityWilsonImplF>;	\
+  //  template class A<GparityWilsonImplD>;		

 ////////////////////////////////////////////
 // Fermion operators / actions
@ -79,8 +79,8 @@ typedef WilsonGaugeAction<LatticeGaugeFieldD>    WilsonGaugeActionD;
 #include <qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h>

 #include <qcd/action/fermion/ContinuedFractionFermion5D.h>               // Continued fraction
-#include <qcd/action/fermion/OverlapWilsonContFracTanhFermion.h>
-#include <qcd/action/fermion/OverlapWilsonContFracZolotarevFermion.h>
+#include <qcd/action/fermion/OverlapWilsonContfracTanhFermion.h>
+#include <qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h>

 #include <qcd/action/fermion/PartialFractionFermion5D.h>                 // Partial fraction
 #include <qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h>
--- a/lib/qcd/action/fermion/WilsonKernels.h
+++ b/lib/qcd/action/fermion/WilsonKernels.h
@ -28,18 +28,31 @@ namespace Grid {
     void DiracOptDhopDir(CartesianStencil &st,DoubledGaugeField &U,
 			  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
 			  int sF,int sU,const FermionField &in, FermionField &out,int dirdisp,int gamma);
+#define HANDOPT
+#ifdef HANDOPT
+     void DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+			       std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
+			       int sF,int sU,const FermionField &in, FermionField &out);
+
+     void DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+				  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
+				  int sF,int sU,const FermionField &in, FermionField &out);
+#else

     void DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
 			       std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
-			       int sF,int sU,const FermionField &in, FermionField &out){
+			       int sF,int sU,const FermionField &in, FermionField &out)
+     {
       DiracOptDhopSite(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
     }

     void DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
 				  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
-				  int sF,int sU,const FermionField &in, FermionField &out){
+				  int sF,int sU,const FermionField &in, FermionField &out)
+     {
       DiracOptDhopSiteDag(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
     }
+#endif

     WilsonKernels(const ImplParams &p= ImplParams()) : Base(p) {};

--- a/lib/qcd/action/fermion/WilsonKernelsHand.cc
+++ b/lib/qcd/action/fermion/WilsonKernelsHand.cc
@ -3,7 +3,7 @@
 #define REGISTER

 #define LOAD_CHIMU \
-  const vSpinColourVector & ref (in._odata[offset]);	\
+  const SiteSpinor & ref (in._odata[offset]);	\
    Chimu_00=ref()(0)(0);\
    Chimu_01=ref()(0)(1);\
    Chimu_02=ref()(0)(2);\
@ -18,7 +18,7 @@
    Chimu_32=ref()(3)(2);

 #define LOAD_CHI\
-  const vHalfSpinColourVector &ref(buf[offset]);	\
+  const SiteHalfSpinor &ref(buf[offset]);	\
    Chi_00 = ref()(0)(0);\
    Chi_01 = ref()(0)(1);\
    Chi_02 = ref()(0)(2);\
@ -280,11 +280,11 @@
 namespace Grid {
 namespace QCD {

-#if 0
-template<class Simd>
-void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+#ifdef HANDOPT
+template<class Impl>
+void WilsonKernels<Impl >::DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
 					       std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
-					       int sF,int sU,const FermionField &in, FermionField &out)
+					       int ss,int sU,const FermionField &in, FermionField &out)
 {
  REGISTER Simd result_00; // 12 regs on knc
  REGISTER Simd result_01;
@ -339,14 +339,14 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
 #define Chimu_32 UChi_12


+  StencilEntry *SE;
  int offset,local,perm, ptype;
-  int ss=sF;
  
  // Xp
-  offset = st._offsets [Xp][ss];
-  local  = st._is_local[Xp][ss];
-  perm   = st._permute[Xp][ss];
-  ptype  = st._permute_type[Xp];
+  SE=st.GetEntry(ptype,Xp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -364,10 +364,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  XP_RECON;

  // Yp
-  offset = st._offsets [Yp][ss];
-  local  = st._is_local[Yp][ss];
-  perm   = st._permute[Yp][ss];
-  ptype  = st._permute_type[Yp];
+  SE=st.GetEntry(ptype,Yp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -385,10 +385,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &


  // Zp
-  offset = st._offsets [Zp][ss];
-  local  = st._is_local[Zp][ss];
-  perm   = st._permute[Zp][ss];
-  ptype  = st._permute_type[Zp];
+  SE=st.GetEntry(ptype,Zp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -405,10 +405,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  ZP_RECON_ACCUM;

  // Tp
-  offset = st._offsets [Tp][ss];
-  local  = st._is_local[Tp][ss];
-  perm   = st._permute[Tp][ss];
-  ptype  = st._permute_type[Tp];
+  SE=st.GetEntry(ptype,Tp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -425,10 +425,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  TP_RECON_ACCUM;
  
  // Xm
-  offset = st._offsets [Xm][ss];
-  local  = st._is_local[Xm][ss];
-  perm   = st._permute[Xm][ss];
-  ptype  = st._permute_type[Xm];
+  SE=st.GetEntry(ptype,Xm,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -445,10 +445,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  XM_RECON_ACCUM;
  
  // Ym
-  offset = st._offsets [Ym][ss];
-  local  = st._is_local[Ym][ss];
-  perm   = st._permute[Ym][ss];
-  ptype  = st._permute_type[Ym];
+  SE=st.GetEntry(ptype,Ym,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -465,10 +465,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  YM_RECON_ACCUM;

  // Zm
-  offset = st._offsets [Zm][ss];
-  local  = st._is_local[Zm][ss];
-  perm   = st._permute[Zm][ss];
-  ptype  = st._permute_type[Zm];
+  SE=st.GetEntry(ptype,Zm,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;

  if ( local ) {
    LOAD_CHIMU;
@ -485,10 +485,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  ZM_RECON_ACCUM;

  // Tm
-  offset = st._offsets [Tm][ss];
-  local  = st._is_local[Tm][ss];
-  perm   = st._permute[Tm][ss];
-  ptype  = st._permute_type[Tm];
+  SE=st.GetEntry(ptype,Tm,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;

  if ( local ) {
    LOAD_CHIMU;
@ -505,7 +505,7 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  TM_RECON_ACCUM;

  {
-    vSpinColourVector & ref (out._odata[ss]);
+    SiteSpinor & ref (out._odata[ss]);
    vstream(ref()(0)(0),result_00*(-0.5));
    vstream(ref()(0)(1),result_01*(-0.5));
    vstream(ref()(0)(2),result_02*(-0.5));
@ -521,10 +521,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSite(CartesianStencil &
  }
 }

-template<class Simd>
-void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
-							      std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
-							      int ss,int sU,const FermionField &in, FermionField &out)
+template<class Impl>
+void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+						   std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  &buf,
+						   int ss,int sU,const FermionField &in, FermionField &out)
 {
  REGISTER Simd result_00; // 12 regs on knc
  REGISTER Simd result_01;
@ -580,12 +580,13 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci


  int offset,local,perm, ptype;
+  StencilEntry *SE;

  // Xp
-  offset = st._offsets [Xp][ss];
-  local  = st._is_local[Xp][ss];
-  perm   = st._permute[Xp][ss];
-  ptype  = st._permute_type[Xp];
+  SE=st.GetEntry(ptype,Xp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -602,10 +603,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
  XM_RECON;
  
  // Yp
-  offset = st._offsets [Yp][ss];
-  local  = st._is_local[Yp][ss];
-  perm   = st._permute[Yp][ss];
-  ptype  = st._permute_type[Yp];
+  SE=st.GetEntry(ptype,Yp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -623,10 +624,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci


  // Zp
-  offset = st._offsets [Zp][ss];
-  local  = st._is_local[Zp][ss];
-  perm   = st._permute[Zp][ss];
-  ptype  = st._permute_type[Zp];
+  SE=st.GetEntry(ptype,Zp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -643,10 +644,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
  ZM_RECON_ACCUM;

  // Tp
-  offset = st._offsets [Tp][ss];
-  local  = st._is_local[Tp][ss];
-  perm   = st._permute[Tp][ss];
-  ptype  = st._permute_type[Tp];
+  SE=st.GetEntry(ptype,Tp,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -663,10 +664,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
  TM_RECON_ACCUM;
  
  // Xm
-  offset = st._offsets [Xm][ss];
-  local  = st._is_local[Xm][ss];
-  perm   = st._permute[Xm][ss];
-  ptype  = st._permute_type[Xm];
+  SE=st.GetEntry(ptype,Xm,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -684,10 +685,10 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
  
  
  // Ym
-  offset = st._offsets [Ym][ss];
-  local  = st._is_local[Ym][ss];
-  perm   = st._permute[Ym][ss];
-  ptype  = st._permute_type[Ym];
+  SE=st.GetEntry(ptype,Ym,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
  
  if ( local ) {
    LOAD_CHIMU;
@ -704,11 +705,11 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
  YP_RECON_ACCUM;

  // Zm
-  offset = st._offsets [Zm][ss];
-  local  = st._is_local[Zm][ss];
-  perm   = st._permute[Zm][ss];
-  ptype  = st._permute_type[Zm];
-
+  SE=st.GetEntry(ptype,Zm,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
+  
  if ( local ) {
    LOAD_CHIMU;
    ZP_PROJ;
@ -724,11 +725,11 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
  ZP_RECON_ACCUM;

  // Tm
-  offset = st._offsets [Tm][ss];
-  local  = st._is_local[Tm][ss];
-  perm   = st._permute[Tm][ss];
-  ptype  = st._permute_type[Tm];
-
+  SE=st.GetEntry(ptype,Tm,ss);
+  offset = SE->_offset;
+  local  = SE->_is_local;
+  perm   = SE->_permute;
+  
  if ( local ) {
    LOAD_CHIMU;
    TP_PROJ;
@ -759,5 +760,6 @@ void WilsonKernels<WilsonImpl<Simd,3> >::DiracOptHandDhopSiteDag(CartesianStenci
    vstream(ref()(3)(2),result_32*(-0.5));
  }
 }
+  FermOpTemplateInstantiate(WilsonKernels);
 #endif
 }}
--- a/tests/Make.inc
+++ b/tests/Make.inc
@ -1,5 +1,5 @@

-bin_PROGRAMS = Test_GaugeAction Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_hdcr Test_gamma Test_gparity Test_gpwilson_even_odd Test_hmc_EODWFRatio Test_hmc_EOWilsonFermionGauge Test_hmc_EOWilsonRatio Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_hmc_WilsonRatio Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_remez Test_rhmc_EOWilson1p1 Test_rhmc_EOWilsonRatio Test_rhmc_Wilson1p1 Test_rhmc_WilsonRatio Test_rng Test_rng_fixed Test_serialisation Test_simd Test_stencil Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi
+bin_PROGRAMS = Test_GaugeAction Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_hdcr Test_gamma Test_hmc_EODWFRatio Test_hmc_EOWilsonFermionGauge Test_hmc_EOWilsonRatio Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_hmc_WilsonRatio Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_remez Test_rhmc_EOWilson1p1 Test_rhmc_EOWilsonRatio Test_rhmc_Wilson1p1 Test_rhmc_WilsonRatio Test_rng Test_rng_fixed Test_serialisation Test_simd Test_stencil Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi


 Test_GaugeAction_SOURCES=Test_GaugeAction.cc
@ -90,12 +90,12 @@ Test_gamma_SOURCES=Test_gamma.cc
 Test_gamma_LDADD=-lGrid


-Test_gparity_SOURCES=Test_gparity.cc
-Test_gparity_LDADD=-lGrid
+#Test_gparity_SOURCES=Test_gparity.cc
+#Test_gparity_LDADD=-lGrid


-Test_gpwilson_even_odd_SOURCES=Test_gpwilson_even_odd.cc
-Test_gpwilson_even_odd_LDADD=-lGrid
+#Test_gpwilson_even_odd_SOURCES=Test_gpwilson_even_odd.cc
+#Test_gpwilson_even_odd_LDADD=-lGrid


 Test_hmc_EODWFRatio_SOURCES=Test_hmc_EODWFRatio.cc