Merge branch 'master', remote-tracking branch 'origin' into omf4

Makefile.in

@@ -38,8 +38,8 @@ COMPILE = ${CC} ${DEFS} ${INCLUDES} -o $@ ${CFLAGS}
 
 SMODULES = 
 
-MODULES = read_input gamma hybrid_update measure_gauge_action start \
-  measure_oriented_plaquettes \
+MODULES = read_input gamma measure_gauge_action start \
+	measure_oriented_plaquettes \
 	expo get_staples update_backward_gauge \
 	measure_rectangles get_rectangle_staples  \
 	test/check_geometry test/check_xchange \

Ptilde_nd.c

@@ -295,16 +295,12 @@ void degree_of_Ptilde(int * _degree, double ** coefs,
       printf("# NDPOLY Acceptance Polynomial: coef[degree] = %e\n", (*coefs)[degree]);
     }
     if(sum < acc) { 
-/*     if(fabs(*coefs[degree]) < acc) { */
-      if((g_proc_id == g_stdio_proc) && (g_debug_level > 1)) {
-        printf(" sum %e, coef %e\n", sum, (*coefs)[degree]);
-      }
       break;
     }
     degree= (int)(degree*1.2);
   }
 
-  if(g_debug_level > 0) {
+  if(g_debug_level > 2) {
     /* Ptilde P S P  Ptilde X - X */
     /* for random spinor X        */
     random_spinor_field_eo(ss, repro, RN_GAUSS);

chebyshev_polynomial_nd.c

@@ -139,35 +139,35 @@ void degree_of_polynomial_nd(int * _degree_of_p, double ** coefs,
     fflush(stdout);
   }
 
-  /* Here we check the accuracy */
-  Ptilde_ndpsi(&auxs[0], &auxc[0], *coefs, degree_of_p, &ss[0], &sc[0], Qsq);
-  Qsq(&aux2s[0], &aux2c[0], &auxs[0], &auxc[0]);
-  Ptilde_ndpsi(&auxs[0], &auxc[0], *coefs, degree_of_p, &aux2s[0], &aux2c[0], Qsq);
-
-  diff(&aux2s[0],&auxs[0],&ss[0],VOLUME/2);
-  temp=square_norm(&aux2s[0],VOLUME/2, 1)/square_norm(&ss[0],VOLUME/2, 1)/4.0;
-
-  diff(&aux2c[0],&auxc[0],&sc[0],VOLUME/2);
-  temp2 = square_norm(&aux2c[0],VOLUME/2, 1)/square_norm(&sc[0],VOLUME/2, 1)/4.0;
-
-  if(g_epsbar == 0.){ 
-    temp2 = 0.0;
-  }
-
-  if(g_proc_id == g_stdio_proc && g_debug_level > 0){
-    /* this is || (P S P - 1)X ||^2 /|| 2X ||^2 */
-    /* where X is a random spinor field         */
-    printf("# NDPOLY MD Polynomial: relative squared accuracy in components:\n# UP=%e  DN=%e \n", temp, temp2);
-    fflush(stdout);
-  }
-
   if(g_debug_level > 1) {
+    /* Here we check the accuracy */
+    Ptilde_ndpsi(&auxs[0], &auxc[0], *coefs, degree_of_p, &ss[0], &sc[0], Qsq);
+    Qsq(&aux2s[0], &aux2c[0], &auxs[0], &auxc[0]);
+    Ptilde_ndpsi(&auxs[0], &auxc[0], *coefs, degree_of_p, &aux2s[0], &aux2c[0], Qsq);
+
+    diff(&aux2s[0],&auxs[0],&ss[0],VOLUME/2);
+    temp=square_norm(&aux2s[0],VOLUME/2, 1)/square_norm(&ss[0],VOLUME/2, 1)/4.0;
+
+    diff(&aux2c[0],&auxc[0],&sc[0],VOLUME/2);
+    temp2 = square_norm(&aux2c[0],VOLUME/2, 1)/square_norm(&sc[0],VOLUME/2, 1)/4.0;
+
+    if(g_epsbar == 0.){ 
+      temp2 = 0.0;
+    }
+
+    if(g_proc_id == g_stdio_proc){
+      /* this is || (P S P - 1)X ||^2 /|| 2X ||^2 */
+      /* where X is a random spinor field         */
+      printf("# NDPOLY MD Polynomial: relative squared accuracy in components:\n# UP=%e  DN=%e \n", temp, temp2);
+      fflush(stdout);
+    }
+
     temp = cheb_eval(degree_of_p, *coefs, EVMin);
     temp *= EVMin;
     temp *= cheb_eval(degree_of_p, *coefs, EVMin);
     temp = 0.5*fabs(temp - 1);
     if(g_proc_id == g_stdio_proc) {
-      printf("# PHMC: Delta_IR at s=%f:    | P s_low P - 1 |/2 = %e \n", EVMin, temp);
+      printf("# NDPOLY MD Polynomial: Delta_IR at s=%f:    | P s_low P - 1 |/2 = %e \n", EVMin, temp);
     }
   }
   /* RECALL THAT WE NEED AN EVEN DEGREE !!!! */

configure.in

@@ -117,7 +117,7 @@ if test $enable_qpx = yes; then
   AC_MSG_RESULT(yes)
   AC_DEFINE(BGQ,1,Compile with QPX intrinsics)
   AC_MSG_NOTICE([Compiling with QPX intrinsics on BGQ, enabling compiler optimizations for XLC.])
-  OPTARGS="-O3 -qstrict=all -qtune=qp -qarch=qp -qmaxmem=-1"
+  OPTARGS="-O2 -qstrict=all -qtune=qp -qarch=qp -qmaxmem=-1"
   SOPTARGS="$OPTARGS" 
 else
   AC_MSG_RESULT(no)
@@ -302,15 +302,17 @@ dnl for the case of other configure scripts
 dnl AC_CONFIG_SUBDIRS( rng )
 
 dnl check for clock_gettime and set correct library flag if one is required
-dnl (as done by AC_CHECK_LIB)
+dnl (this is done by AC_CHECK_LIB)
 AC_CHECK_FUNCS(clock_gettime, [], [AC_CHECK_LIB(rt, clock_gettime)])
 
-
-dnl with glibc and c99 mode the timespec required for clock_gettime is only
-dnl available in POSIX 199309L compatibility mode when clock_gettime is in librt
-if test "$ac_cv_lib_rt_clock_gettime" = "yes"; then
+dnl in principle clock_gettime and CLOCK_MONOTONIC/CLOCK_REALTIME should be available
+dnl only when using POSIX 199309, we set this explicitly here
+dnl this should not cause problems on any relatively modern (post y2k) machine!
+if ( test "$ac_cv_lib_rt_clock_gettime" = "yes" || test "$ac_cv_func_clock_gettime" = "yes" ); then
   AC_DEFINE(HAVE_CLOCK_GETTIME,1)
-  CFLAGS="$CFLAGS -D_POSIX_C_SOURCE=199309L"
+dnl  we set this in gettime.c explicitly for the time being 
+dnl  due to endian problem on BG/Q
+dnl  CFLAGS="$CFLAGS -D_POSIX_C_SOURCE=199309L"
   AC_MSG_NOTICE([Instructing the compiler to use POSIX 199309L])
 fi
 
@@ -623,7 +625,7 @@ elif test "$host_cpu" = "powerpc64" && test "$host_vendor" = "unknown" && test "
     DEBUGFLAG="$DEBUGFLAG -qfullpath"
   fi
 
-  OPTARGS="$OPTARGS -O3"
+  OPTARGS="$OPTARGS"
   SOPTARGS="$OPTARGS"
   if test "$CCDEP" = "gcc"; then
     DEPFLAGS="-MM"
@@ -710,8 +712,8 @@ elif test "$host_cpu" = "powerpc" && test "$host_vendor" = "ibm" && test "$host_
 elif test "$host_cpu" = "powerpc" && test "$host_vendor" = "ibm"; then
   CFLAGS="$CFLAGS -q64 -qsrcmsg"
   LDFLAGS="$LDFLAGS -q64"
-  OPTARGS="-O3 -qhot"
-  SOPTARGS="-O3 -qhot"
+  OPTARGS="-O2"
+  SOPTARGS="-O2"
   DEBUG_FLAG="-qfullpath -g"
   if test "$CCDEP" = "gcc"; then
     DEPFLAGS="-MM"

gettime.c

@@ -22,6 +22,13 @@
 #ifdef HAVE_CONFIG_H
 # include<config.h>
 #endif
+#ifdef HAVE_CLOCK_GETTIME
+#  ifndef _POSIX_C_SOURCE
+#    define _POSIX_C_SOURCE 199309L
+#  endif
+#  include <sys/time.h>
+#  include <bits/time.h>
+#endif
 #include <time.h>
 #if (defined BGL && !defined BGP)
 #  include <rts.h>

monomial/Makefile.in

@@ -32,7 +32,7 @@ COMPILE = ${CC} $(DEFS) ${INCLUDES} ${CFLAGS}
 LIBRARIES = libmonomial
 libmonomial_TARGETS =  nddetratio_monomial monomial det_monomial detratio_monomial \
 	gauge_monomial ndpoly_monomial clover_trlog_monomial cloverdet_monomial cloverdetratio_monomial \
-	clovernd_trlog_monomial poly_monomial cloverndpoly_monomial
+	clovernd_trlog_monomial poly_monomial cloverndpoly_monomial moment_energy
 
 
 libmonomial_STARGETS = 

monomial/clover_trlog_monomial.c

@@ -33,6 +33,7 @@
 #include "operator/clover_leaf.h"
 #include "monomial/monomial.h"
 #include "operator/Hopping_Matrix.h"
+#include "gettime.h"
 #include "clover_trlog_monomial.h"
 
 void clover_trlog_derivative(const int id, hamiltonian_field_t * const hf) {
@@ -48,27 +49,43 @@ void clover_trlog_derivative(const int id, hamiltonian_field_t * const hf) {
 
 void clover_trlog_heatbath(const int id, hamiltonian_field_t * const hf) {
   monomial * mnl = &monomial_list[id];
+  double atime, etime;
+  atime = gettime();
   mnl->energy0 = 0.;
 
   init_sw_fields();
   sw_term( (const su3**) hf->gaugefield, mnl->kappa, mnl->c_sw); 
   /*compute the contribution from the clover trlog term */
   mnl->energy0 = -sw_trace(EO, mnl->mu);
-  if(g_proc_id == 0 && g_debug_level > 3) {
-    printf("called clover_trlog_heatbath for id %d E = %e\n", id, mnl->energy0);
+  etime = gettime();
+  if(g_proc_id == 0) {
+    if(g_debug_level > 1) {
+      printf("# Time for %s monomial heatbath: %e s\n", mnl->name, etime-atime);
+    }
+    if(g_debug_level > 3) {
+      printf("called clover_trlog_heatbath for id %d E = %e\n", id, mnl->energy0);
+    }
   }
   return;
 }
 
 double clover_trlog_acc(const int id, hamiltonian_field_t * const hf) {
   monomial * mnl = &monomial_list[id];
+  double atime, etime;
+  atime = gettime();
   mnl->energy1 = 0.;
   sw_term( (const su3**) hf->gaugefield, mnl->kappa, mnl->c_sw); 
   /*compute the contribution from the clover trlog term */
   mnl->energy1 = -sw_trace(EO, mnl->mu);
+  etime = gettime();
   if(g_proc_id == 0 && g_debug_level > 3) {
-    printf("called clover_trlog_acc for id %d dH = %1.10e\n", 
-	   id, mnl->energy1 - mnl->energy0);
+    if(g_debug_level > 1) {
+      printf("# Time for %s monomial acc step: %e s\n", mnl->name, etime-atime);
+    }
+    if(g_debug_level > 3) {
+      printf("called clover_trlog_acc for id %d dH = %1.10e\n", 
+	     id, mnl->energy1 - mnl->energy0);
+    }
   }
   return(mnl->energy1 - mnl->energy0);
 }

monomial/cloverdet_monomial.c

@@ -31,6 +31,7 @@
 #include "ranlxd.h"
 #include "sse.h"
 #include "start.h"
+#include "gettime.h"
 #include "linalg_eo.h"
 #include "deriv_Sb.h"
 #include "gamma.h"
@@ -50,7 +51,8 @@
 
 void cloverdet_derivative(const int id, hamiltonian_field_t * const hf) {
   monomial * mnl = &monomial_list[id];
-
+  double atime, etime;
+  atime = gettime();
   for(int i = 0; i < VOLUME; i++) { 
     for(int mu = 0; mu < 4; mu++) { 
       _su3_zero(swm[i][mu]);
@@ -128,14 +130,20 @@ void cloverdet_derivative(const int id, hamiltonian_field_t * const hf) {
   g_mu = g_mu1;
   g_mu3 = 0.;
   boundary(g_kappa);
-
+  etime = gettime();
+  if(g_debug_level > 1 && g_proc_id == 0) {
+    printf("# Time for %s monomial derivative: %e s\n", mnl->name, etime-atime);
+  }
   return;
 }
 
 
 void cloverdet_heatbath(const int id, hamiltonian_field_t * const hf) {
 
   monomial * mnl = &monomial_list[id];
+  double atime, etime;
+  atime = gettime();
+
   g_mu = mnl->mu;
   g_mu3 = mnl->rho;
   g_c_sw = mnl->c_sw;
@@ -159,8 +167,14 @@ void cloverdet_heatbath(const int id, hamiltonian_field_t * const hf) {
   g_mu = g_mu1;
   g_mu3 = 0.;
   boundary(g_kappa);
-  if(g_proc_id == 0 && g_debug_level > 3) {
-    printf("called cloverdet_heatbath for id %d %d\n", id, mnl->even_odd_flag);
+  etime = gettime();
+  if(g_proc_id == 0) {
+    if(g_debug_level > 1) {
+      printf("# Time for %s monomial heatbath: %e s\n", mnl->name, etime-atime);
+    }
+    if(g_debug_level > 3) {
+      printf("called cloverdet_heatbath for id %d energy %f\n", id, mnl->energy0);
+    }
   }
   return;
 }
@@ -169,6 +183,8 @@ void cloverdet_heatbath(const int id, hamiltonian_field_t * const hf) {
 double cloverdet_acc(const int id, hamiltonian_field_t * const hf) {
   monomial * mnl = &monomial_list[id];
   int save_sloppy = g_sloppy_precision_flag;
+  double atime, etime;
+  atime = gettime();
 
   g_mu = mnl->mu;
   g_mu3 = mnl->rho;
@@ -192,9 +208,15 @@ double cloverdet_acc(const int id, hamiltonian_field_t * const hf) {
   g_mu = g_mu1;
   g_mu3 = 0.;
   boundary(g_kappa);
-  if(g_proc_id == 0 && g_debug_level > 3) {
-    printf("called cloverdet_acc for id %d %d dH = %1.10e\n", 
-	   id, mnl->even_odd_flag, mnl->energy1 - mnl->energy0);
+  etime = gettime();
+  if(g_proc_id == 0) {
+    if(g_debug_level > 1) {
+      printf("# Time for %s monomial acc step: %e s\n", mnl->name, etime-atime);
+    }
+    if(g_debug_level > 3) {
+      printf("called cloverdet_acc for id %d dH = %1.10e\n", 
+	     id, mnl->energy1 - mnl->energy0);
+    }
   }
   return(mnl->energy1 - mnl->energy0);
 }