qphix_interface work:

1) pass parameters to qphix from input file, new "BeginExternalInverter QPHIX"
   structure implemented for this purpose (warning: no correctness checks here)

2) move _initQphix call into invert_eo_qphix

3) add sample file to call qphix Wilson inverter

4) use angled brackets to include qphix headers

caveats: single precision results in a segfault, only the Wilson operator is currently supported

Author: kostrzewa

invert_eo.c

@@ -102,7 +102,8 @@ int invert_eo(spinor * const Even_new, spinor * const Odd_new,
     return invert_eo_qphix(Even_new, Odd_new, Even, Odd,
                            precision, max_iter,
                            solver_flag, rel_prec,
-                           solver_params, compression);
+                           solver_params, 
+                           compression);
   }
 #endif
 

qphix_interface.cpp

@@ -70,12 +70,12 @@ extern "C" {
 #include <cfloat>
 #include <cstdlib>
 #include <cstring>
-#include "qphix/blas_new_c.h"
-#include "qphix/invcg.h"
-#include "qphix/invbicgstab.h"
-#include "qphix/print_utils.h"
-#include "qphix/qphix_config.h"
-#include "qphix/wilson.h"
+#include <qphix/blas_new_c.h>
+#include <qphix/invcg.h>
+#include <qphix/invbicgstab.h>
+#include <qphix/print_utils.h>
+#include <qphix/qphix_config.h>
+#include <qphix/wilson.h>
 using namespace std;
 using namespace QPhiX;
 
@@ -118,6 +118,15 @@ using namespace QPhiX;
 #include <qmp.h>
 #endif
 
+int qphix_input_By;
+int qphix_input_Bz;
+int qphix_input_NCores;
+int qphix_input_Sy;
+int qphix_input_Sz;
+int qphix_input_PadXY;
+int qphix_input_PadXYZ;
+int qphix_input_MinCt;
+
 int By;
 int Bz;
 int NCores;
@@ -128,7 +137,7 @@ int PadXYZ;
 int MinCt;
 int N_simt;
 bool compress12;
-QphixPrec precision;
+QphixPrec qphix_precision;
 
 int subLattSize[4];
 int lattSize[4];
@@ -174,9 +183,10 @@ void _initQphix(int argc, char **argv, int By_, int Bz_, int NCores_, int Sy_, i
   MinCt = MinCt_;
   N_simt = Sy_ * Sz_;
   compress12 = c12;
-  precision = precision_;
+  qphix_precision = precision_;
 
-  omp_set_num_threads(NCores * Sy * Sz);
+  // this is called within tmLQCD already, it should not be called here
+  //omp_set_num_threads(NCores * Sy * Sz);
 
 #ifdef QPHIX_QMP_COMMS
   // Declare the logical topology
@@ -564,11 +574,7 @@ int invert_eo_qphix_helper(spinor * const tmlqcd_even_out,
    *                      *
   ************************/
 
-  // FIXME: This should be called properly somewhere else
-  _initQphix(0, nullptr, /* By = */ 2, /* Bz = */ 2, /* Ncores = */ 1,
-             /* Sy = */ 1, /* Sz = */ 1,
-             /* PadXY = */ 1, /* PadXYZ = */ 1, /* MinCt = */ 1,
-             /* compress12 = */ 1, QPHIX_DOUBLE_PREC);
+  // _initQphix should have been called at this point
 
   masterPrintf("# VECLEN = %d, SOALEN = %d\n", V, S);
   masterPrintf("# Declared QMP Topology: %d %d %d %d\n",
@@ -860,7 +866,7 @@ int invert_eo_qphix_helper(spinor * const tmlqcd_even_out,
 
 // Template wrapper for the Dslash operator call-able from C code
 void D_psi_qphix(spinor* tmlqcd_out, const spinor* tmlqcd_in) {
-  if (precision == QPHIX_DOUBLE_PREC) {
+  if (qphix_precision == QPHIX_DOUBLE_PREC) {
     if (QPHIX_SOALEN > VECLEN_DP) {
       masterPrintf("SOALEN=%d is greater than the double prec VECLEN=%d\n", QPHIX_SOALEN,
                    VECLEN_DP);
@@ -872,7 +878,7 @@ void D_psi_qphix(spinor* tmlqcd_out, const spinor* tmlqcd_in) {
     } else {
       D_psi<double, VECLEN_DP, QPHIX_SOALEN, false>(tmlqcd_out, tmlqcd_in);
     }
-  } else if (precision == QPHIX_FLOAT_PREC) {
+  } else if (qphix_precision == QPHIX_FLOAT_PREC) {
     if (QPHIX_SOALEN > VECLEN_SP) {
       masterPrintf("SOALEN=%d is greater than the single prec VECLEN=%d\n", QPHIX_SOALEN,
                    VECLEN_SP);
@@ -886,7 +892,7 @@ void D_psi_qphix(spinor* tmlqcd_out, const spinor* tmlqcd_in) {
     }
   }
 #if defined(QPHIX_MIC_SOURCE)
-  else if (precision == QPHIX_HALF_PREC) {
+  else if (qphix_precision == QPHIX_HALF_PREC) {
     if (QPHIX_SOALEN > VECLEN_HP) {
       masterPrintf("SOALEN=%d is greater than the single prec VECLEN=%d\n", QPHIX_SOALEN,
                    VECLEN_SP);
@@ -924,6 +930,11 @@ int invert_eo_qphix(spinor * const Even_new,
     }
     masterPrintf("# INITIALIZING QPHIX SOLVER\n");
     masterPrintf("# USING DOUBLE PRECISION\n");
+    _initQphix(0, nullptr, qphix_input_By, qphix_input_Bz, qphix_input_NCores,
+           qphix_input_Sy, qphix_input_Sz,
+           qphix_input_PadXY, qphix_input_PadXYZ, qphix_input_MinCt,
+           compression, QPHIX_DOUBLE_PREC);
+    
     if (compress12) {
       return invert_eo_qphix_helper<double, VECLEN_DP, QPHIX_SOALEN, true>
         (Even_new,
@@ -949,16 +960,25 @@ int invert_eo_qphix(spinor * const Even_new,
          solver_params,
          compression);
     }
+#ifdef QPHIX_MIC_SOURCE
   } else if (precision < rsdTarget<float>::value) {
+#else
+  } else {
+#endif
     if (QPHIX_SOALEN > VECLEN_SP) {
       masterPrintf("SOALEN=%d is greater than the single prec VECLEN=%d\n", QPHIX_SOALEN,
                    VECLEN_SP);
       abort();
     }
     masterPrintf("# INITIALIZING QPHIX SOLVER\n");
     masterPrintf("# USING SINGLE PRECISION\n");
+    _initQphix(0, nullptr, qphix_input_By, qphix_input_Bz, qphix_input_NCores,
+           qphix_input_Sy, qphix_input_Sz,
+           qphix_input_PadXY, qphix_input_PadXYZ, qphix_input_MinCt,
+           compression, QPHIX_FLOAT_PREC);
+
     if (compress12) {
-      return invert_eo_qphix_helper<float, VECLEN_DP, QPHIX_SOALEN, true>
+      return invert_eo_qphix_helper<float, VECLEN_SP, QPHIX_SOALEN, true>
         (Even_new,
          Odd_new,
          Even,
@@ -970,7 +990,7 @@ int invert_eo_qphix(spinor * const Even_new,
          solver_params,
          compression);
     } else {
-      return invert_eo_qphix_helper<float, VECLEN_DP, QPHIX_SOALEN, false>
+      return invert_eo_qphix_helper<float, VECLEN_SP, QPHIX_SOALEN, false>
         (Even_new,
          Odd_new,
          Even,
@@ -992,8 +1012,13 @@ int invert_eo_qphix(spinor * const Even_new,
     }
     masterPrintf("# INITIALIZING QPHIX SOLVER\n");
     masterPrintf("# USING HALF PRECISION\n");
+    _initQphix(0, nullptr, qphix_input_By, qphix_input_Bz, qphix_input_NCores,
+           qphix_input_Sy, qphix_input_Sz,
+           qphix_input_PadXY, qphix_input_PadXYZ, qphix_input_MinCt,
+           compression, QPHIX_HALF_PREC);
+    
     if (compress12) {
-      return invert_eo_qphix_helper<half, VECLEN_DP, QPHIX_SOALEN, true>
+      return invert_eo_qphix_helper<half, VECLEN_SP, QPHIX_SOALEN, true>
         (Even_new,
          Odd_new,
          Even,
@@ -1005,7 +1030,7 @@ int invert_eo_qphix(spinor * const Even_new,
          solver_params,
          compression);
     } else {
-      return invert_eo_qphix_helper<half, VECLEN_DP, QPHIX_SOALEN, false>
+      return invert_eo_qphix_helper<half, VECLEN_SP, QPHIX_SOALEN, false>
         (Even_new,
          Odd_new,
          Even,

read_input.l

@@ -190,8 +190,19 @@ static inline void rmQuotes(char *str){
   double mu_Msap;
 
   int use_preconditioning;
+
+#ifdef TM_USE_QPHIX
+  extern int qphix_input_By;
+  extern int qphix_input_Bz;
+  extern int qphix_input_NCores;
+  extern int qphix_input_Sy;
+  extern int qphix_input_Sz;
+  extern int qphix_input_PadXY;
+  extern int qphix_input_PadXYZ;
+  extern int qphix_input_MinCt;
+#endif
+
   
-  int use_qudainverter;
 %}
 
 %option never-interactive
@@ -344,6 +355,8 @@ static inline void rmQuotes(char *str){
 %x LOWMEM
 %x SUBPROCESS 
 
+%x INITEXTERNALINVERTER
+
 %%
 ^SourceFilename{EQL}               BEGIN(SOURCEFILE);
 ^PropagatorFilename{EQL}           BEGIN(PROPFILE);
@@ -412,8 +425,6 @@ static inline void rmQuotes(char *str){
 ^NoSamples{EQL}                    BEGIN(NOSAMPLES);
 ^SplittedPropagator{EQL}           BEGIN(PROPSPLIT);
 ^UsePreconditioning{EQL}           BEGIN(PRECONDITIONING);
-^UseQudaInverter{EQL}              BEGIN(QUDAINVERTER);
-^UseQphixInverter{EQL}             BEGIN(QPHIXINVERTER);
 ^UseCompression{EQL}               BEGIN(COMPRESSION);
 
 ^BeginMeasurement{SPC}+            BEGIN(INITMEASUREMENT);
@@ -425,6 +436,8 @@ static inline void rmQuotes(char *str){
 ^BeginInt                          BEGIN(INITINTEGRATOR);
 ^BeginOperator{SPC}+               BEGIN(INITOPERATOR);
 
+^BeginExternalInverter{SPC}+       BEGIN(INITEXTERNALINVERTER);
+
 ^PropagatorComparison{EQL}         BEGIN(PCOMP);
 ^NbCoresPerNode{EQL}               BEGIN(NBCORES);
 
@@ -1081,6 +1094,61 @@ static inline void rmQuotes(char *str){
   }
 }
 
+<INITEXTERNALINVERTER>{TYPE} {
+  if(strcmp(yytext, "QPHIX")==0) {
+    if(myverbose) printf("Setting QPHIX external inverter parameters line %d\n", line_of_file);
+    BEGIN(QPHIXINVERTER);
+  } else {
+    printf("Error on line %d, inverter '%s' is unknown. Exiting!\n", line_of_file, yytext);
+    exit(1);
+  }
+}
+<QPHIXINVERTER>{
+  {SPC}*NCores{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_NCores = a;
+    if(myverbose) printf("  Ncores = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*By{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_By = a;
+    if(myverbose) printf("  By = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*Bz{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_Bz = a;
+    if(myverbose) printf("  Bz = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*Sy{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_Sy = a;
+    if(myverbose) printf("  Sy = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*Sz{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_Sz = a;
+    if(myverbose) printf("  Sz = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*PadXY{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_PadXY = a;
+    if(myverbose) printf("  PadXY = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*PadXYZ{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_PadXYZ = a;
+    if(myverbose) printf("  PadXYZ = %d, line %d\n", a, line_of_file );
+  }
+  {SPC}*MinCt{EQL}{DIGIT}+ {
+    sscanf(yytext, " %[a-zA-Z] = %d", name, &a);
+    qphix_input_MinCt = a;
+    if(myverbose) printf("  MinCt = %d, line %d\n", a, line_of_file );
+  }
+  ^EndExternalInverter{SPC}*  {
+    if(myverbose) printf("QPHIX external inverter parameters parsed on line %d\n\n", line_of_file);
+    BEGIN(0);
+  }
+}
 
 <CSWSOLVER>{
   increigcg {
@@ -2360,7 +2428,7 @@ static inline void rmQuotes(char *str){
   BEGIN(comment_caller);
 }
 
-<INITMONOMIAL,DETMONOMIAL,CLDETMONOMIAL,CLDETRATMONOMIAL,CLDETRATRWMONOMIAL,NDPOLYMONOMIAL,NDRATMONOMIAL,NDRATCORMONOMIAL,NDCLRATMONOMIAL,NDCLRATCORMONOMIAL,CLPOLYMONOMIAL,GAUGEMONOMIAL,INTEGRATOR,INITINTEGRATOR,INITMEASUREMENT,PIONNORMMEAS,ONLINEMEAS,ORIENTEDPLAQUETTESMEAS,GRADIENTFLOWMEAS,INITOPERATOR,TMOP,DBTMOP,OVERLAPOP,WILSONOP,CLOVEROP,DBCLOVEROP,POLYMONOMIAL,PLOOP,INITGPU,GPU,RATMONOMIAL,RATCORMONOMIAL,CLRATMONOMIAL,CLRATCORMONOMIAL,INITDEFLATION,DEFLATION>{SPC}*\n   {
+<INITMONOMIAL,DETMONOMIAL,CLDETMONOMIAL,CLDETRATMONOMIAL,CLDETRATRWMONOMIAL,NDPOLYMONOMIAL,NDRATMONOMIAL,NDRATCORMONOMIAL,NDCLRATMONOMIAL,NDCLRATCORMONOMIAL,CLPOLYMONOMIAL,GAUGEMONOMIAL,INTEGRATOR,INITINTEGRATOR,INITMEASUREMENT,PIONNORMMEAS,ONLINEMEAS,ORIENTEDPLAQUETTESMEAS,GRADIENTFLOWMEAS,INITOPERATOR,TMOP,DBTMOP,OVERLAPOP,WILSONOP,CLOVEROP,DBCLOVEROP,POLYMONOMIAL,PLOOP,INITGPU,GPU,RATMONOMIAL,RATCORMONOMIAL,CLRATMONOMIAL,CLRATCORMONOMIAL,INITDEFLATION,DEFLATION,INITEXTERNALINVERTER,QPHIXINVERTER>{SPC}*\n   {
   line_of_file++;
 }
 <*>{SPC}*\n                       {

sample-input/sample.cg-wilson-qphix.input

@@ -0,0 +1,64 @@
+L=24
+T=48
+
+# number of MPI tasks in T direction is computed automatically
+ 
+# number of MPI tasks in Z direction
+nrzprocs=2
+
+# currently only two-dimensional parallelisation works properly
+# also, qphix is slower when parallelized in x or y (especially x is poblematic)
+nrxprocs=1
+nryprocs=1
+
+# OpenMP threads per MPI task
+# down below, Ncores * Sy * Sz = ompnumthreads
+ompnumthreads = 12
+
+DebugLevel = 3
+InitialStoreCounter = 1000
+Measurements = 1
+2kappamu = 0.0008238
+kappa = 0.1373
+csw = 1.57551
+BCAngleT = 0
+GaugeConfigInputFile = conf
+disableiochecks=yes
+UseEvenOdd = yes
+
+SourceType = point
+indices = 0
+ReadSource = no
+
+userelativeprecision = no
+
+BeginExternalInverter QPHIX
+  # physical cores per MPI task
+  NCores = 12
+  
+  # block sizes (see qphix papers for details)
+  By = 4
+  Bz = 4
+  MinCt = 2
+  
+  # thread geometry
+  # ompnumthreads = Ncores * Sy * Sz
+  # hyperthreads should be specified here
+  Sy = 1
+  Sz = 1
+  
+  # paddings in XY and XYZ blocks
+  PadXY = 0
+  PadXYZ = 0
+EndExternalInverter
+
+BeginOperator WILSON
+  kappa = 0.132
+  Solver = cg
+  # solverprecision must be <= 1e-13 such that the CG in double precision is launched
+  # the other floating point precisions (single, half) seem to segfault currently
+  SolverPrecision = 1e-17
+  MaxSolverIterations = 20000
+  useqphixinverter = yes
+EndOperator
+

test_Dslash.c

@@ -16,7 +16,7 @@
 #include <config.h>
 #endif
 #ifdef TM_USE_QPHIX
-#include "qphix/qphix_config.h"
+#include <qphix/qphix_config.h>
 #endif
 #include <float.h>
 #include <math.h>