fixing errors with pddrive_ABglobal

xiaoyeli · Jan 9, 2025 · 46a2e83 · 46a2e83
1 parent 95c8828
commit 46a2e83
Show file tree

Hide file tree

Showing 18 changed files with 56 additions and 18 deletions.
diff --git a/SRC/complex16/pzdistribute.c b/SRC/complex16/pzdistribute.c
@@ -2325,7 +2325,7 @@ if ( !iam) printf(".. Construct Reduce tree for U: %.2f\t\n", t);
 		for(int ii=0; ii<nsupers; ii++)
 			supernodeMask[ii]=1;
 		ztrs_compute_communication_structure(options, n, LUstruct,
-					ScalePermstruct, supernodeMask, grid);
+					supernodeMask, grid);
 		SUPERLU_FREE(supernodeMask);
 	}
 

diff --git a/SRC/complex16/pzdistribute3d.c b/SRC/complex16/pzdistribute3d.c
@@ -1126,13 +1126,13 @@ pzdistribute3d(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 		// /* Compute communication structure for trisolve. */ 
 		if (get_new3dsolve()){
 			ztrs_compute_communication_structure(options, n, LUstruct,
-						ScalePermstruct, trf3Dpart->supernodeMask, grid);
+						trf3Dpart->supernodeMask, grid);
 		}else{
 			int* supernodeMask = int32Malloc_dist(nsupers);
 			for(int ii=0; ii<nsupers; ii++)
 				supernodeMask[ii]=1;
 			ztrs_compute_communication_structure(options, n, LUstruct,
-						ScalePermstruct, supernodeMask, grid);
+						supernodeMask, grid);
 			SUPERLU_FREE(supernodeMask);
 		}
 	}

diff --git a/SRC/complex16/pzgstrs3d.c b/SRC/complex16/pzgstrs3d.c
@@ -150,7 +150,6 @@ doublecomplex *zready_x, *zready_lsum;
 // #endif
 
 int ztrs_compute_communication_structure(superlu_dist_options_t *options, int_t n, zLUstruct_t * LUstruct,
-                           zScalePermstruct_t * ScalePermstruct,
                            int* supernodeMask, gridinfo_t *grid)
 {
     Glu_persist_t *Glu_persist = LUstruct->Glu_persist;

diff --git a/SRC/complex16/pzsymbfact_distdata.c b/SRC/complex16/pzsymbfact_distdata.c
@@ -3319,7 +3319,7 @@ doublecomplex *dense, *dense_col; /* SPA */
     for(int ii=0; ii<nsupers; ii++)
       supernodeMask[ii]=1;
     ztrs_compute_communication_structure(options, n, LUstruct,
-          ScalePermstruct, supernodeMask, grid);
+          supernodeMask, grid);
     SUPERLU_FREE(supernodeMask);
   }
 

diff --git a/SRC/complex16/zdistribute.c b/SRC/complex16/zdistribute.c
@@ -1978,6 +1978,20 @@ if ( !iam) printf(".. Construct Reduce tree for U: %.2f\t\n", t);
 
     } /* else fact != SamePattern_SameRowPerm */
 
+	if ( options->Fact != SamePattern_SameRowPerm ) {
+		// /* Flatten L metadata into one buffer. */
+		pzflatten_LDATA(options, n, LUstruct, grid);
+
+		// /* Compute communication structure for trisolve. */ 
+		int* supernodeMask = int32Malloc_dist(nsupers);
+		for(int ii=0; ii<nsupers; ii++)
+			supernodeMask[ii]=1;
+		ztrs_compute_communication_structure(options, n, LUstruct,
+					supernodeMask, grid);
+		SUPERLU_FREE(supernodeMask);
+	}
+
+
 #if ( DEBUGlevel>=1 )
     /* Memory allocated but not freed:
        ilsum, fmod, fsendx_plist, bmod, bsendx_plist  */

diff --git a/SRC/double/ddistribute.c b/SRC/double/ddistribute.c
@@ -1979,6 +1979,21 @@ if ( !iam) printf(".. Construct Reduce tree for U: %.2f\t\n", t);
 
     } /* else fact != SamePattern_SameRowPerm */
 
+
+	if ( options->Fact != SamePattern_SameRowPerm ) {
+		// /* Flatten L metadata into one buffer. */
+		pdflatten_LDATA(options, n, LUstruct, grid);
+
+		// /* Compute communication structure for trisolve. */ 
+		int* supernodeMask = int32Malloc_dist(nsupers);
+		for(int ii=0; ii<nsupers; ii++)
+			supernodeMask[ii]=1;
+		dtrs_compute_communication_structure(options, n, LUstruct,
+					supernodeMask, grid);
+		SUPERLU_FREE(supernodeMask);
+	}
+
+
 #if ( DEBUGlevel>=1 )
     /* Memory allocated but not freed:
        ilsum, fmod, fsendx_plist, bmod, bsendx_plist  */

diff --git a/SRC/double/pddistribute.c b/SRC/double/pddistribute.c
@@ -2327,7 +2327,7 @@ if ( !iam) printf(".. Construct Reduce tree for U: %.2f\t\n", t);
 		for(int ii=0; ii<nsupers; ii++)
 			supernodeMask[ii]=1;
 		dtrs_compute_communication_structure(options, n, LUstruct,
-					ScalePermstruct, supernodeMask, grid);
+					supernodeMask, grid);
 		SUPERLU_FREE(supernodeMask);
 	}
 

diff --git a/SRC/double/pddistribute3d.c b/SRC/double/pddistribute3d.c
@@ -2459,13 +2459,13 @@ pddistribute3d(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
         // /* Compute communication structure for trisolve. */ 
         if (get_new3dsolve()){
             dtrs_compute_communication_structure(options, n, LUstruct,
-                        ScalePermstruct, trf3Dpart->supernodeMask, grid);
+                        trf3Dpart->supernodeMask, grid);
         }else{
             int* supernodeMask = int32Malloc_dist(nsupers);
             for(int ii=0; ii<nsupers; ii++)
                 supernodeMask[ii]=1;
             dtrs_compute_communication_structure(options, n, LUstruct,
-                        ScalePermstruct, supernodeMask, grid);
+                        supernodeMask, grid);
             SUPERLU_FREE(supernodeMask);
         }
     }

diff --git a/SRC/double/pdgstrs3d.c b/SRC/double/pdgstrs3d.c
@@ -151,7 +151,6 @@ double *dready_x, *dready_lsum;
 // #endif
 
 int dtrs_compute_communication_structure(superlu_dist_options_t *options, int_t n, dLUstruct_t * LUstruct,
-                           dScalePermstruct_t * ScalePermstruct,
                            int* supernodeMask, gridinfo_t *grid)
 {
     Glu_persist_t *Glu_persist = LUstruct->Glu_persist;

diff --git a/SRC/double/pdsymbfact_distdata.c b/SRC/double/pdsymbfact_distdata.c
@@ -3317,7 +3317,7 @@ double *dense, *dense_col; /* SPA */
     for(int ii=0; ii<nsupers; ii++)
       supernodeMask[ii]=1;
     dtrs_compute_communication_structure(options, n, LUstruct,
-          ScalePermstruct, supernodeMask, grid);
+          supernodeMask, grid);
     SUPERLU_FREE(supernodeMask);
   }
 

diff --git a/SRC/include/superlu_ddefs.h b/SRC/include/superlu_ddefs.h
@@ -773,7 +773,6 @@ extern int_t dleafForestForwardSolve3d(superlu_dist_options_t *options, int_t tr
 
 
 extern int dtrs_compute_communication_structure(superlu_dist_options_t *options, int_t n, dLUstruct_t * LUstruct,
-                           dScalePermstruct_t * ScalePermstruct,
                            int* supernodeMask, gridinfo_t *grid);
 extern int_t dreduceSolvedX_newsolve(int_t treeId, int_t sender, int_t receiver, double* x, int nrhs,
                       dtrf3Dpartition_t*  trf3Dpartition, dLUstruct_t* LUstruct, gridinfo3d_t* grid3d, double* recvbuf, xtrsTimer_t *xtrsTimer);

diff --git a/SRC/include/superlu_sdefs.h b/SRC/include/superlu_sdefs.h
@@ -773,7 +773,6 @@ extern int_t sleafForestForwardSolve3d(superlu_dist_options_t *options, int_t tr
 
 
 extern int strs_compute_communication_structure(superlu_dist_options_t *options, int_t n, sLUstruct_t * LUstruct,
-                           sScalePermstruct_t * ScalePermstruct,
                            int* supernodeMask, gridinfo_t *grid);
 extern int_t sreduceSolvedX_newsolve(int_t treeId, int_t sender, int_t receiver, float* x, int nrhs,
                       strf3Dpartition_t*  trf3Dpartition, sLUstruct_t* LUstruct, gridinfo3d_t* grid3d, float* recvbuf, xtrsTimer_t *xtrsTimer);

diff --git a/SRC/include/superlu_zdefs.h b/SRC/include/superlu_zdefs.h
@@ -773,7 +773,6 @@ extern int_t zleafForestForwardSolve3d(superlu_dist_options_t *options, int_t tr
 
 
 extern int ztrs_compute_communication_structure(superlu_dist_options_t *options, int_t n, zLUstruct_t * LUstruct,
-                           zScalePermstruct_t * ScalePermstruct,
                            int* supernodeMask, gridinfo_t *grid);
 extern int_t zreduceSolvedX_newsolve(int_t treeId, int_t sender, int_t receiver, doublecomplex* x, int nrhs,
                       ztrf3Dpartition_t*  trf3Dpartition, zLUstruct_t* LUstruct, gridinfo3d_t* grid3d, doublecomplex* recvbuf, xtrsTimer_t *xtrsTimer);

diff --git a/SRC/single/psdistribute.c b/SRC/single/psdistribute.c
@@ -2327,7 +2327,7 @@ if ( !iam) printf(".. Construct Reduce tree for U: %.2f\t\n", t);
 		for(int ii=0; ii<nsupers; ii++)
 			supernodeMask[ii]=1;
 		strs_compute_communication_structure(options, n, LUstruct,
-					ScalePermstruct, supernodeMask, grid);
+					supernodeMask, grid);
 		SUPERLU_FREE(supernodeMask);
 	}
 

diff --git a/SRC/single/psdistribute3d.c b/SRC/single/psdistribute3d.c
@@ -1128,13 +1128,13 @@ psdistribute3d(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 		// /* Compute communication structure for trisolve. */ 
 		if (get_new3dsolve()){
 			strs_compute_communication_structure(options, n, LUstruct,
-						ScalePermstruct, trf3Dpart->supernodeMask, grid);
+						 trf3Dpart->supernodeMask, grid);
 		}else{
 			int* supernodeMask = int32Malloc_dist(nsupers);
 			for(int ii=0; ii<nsupers; ii++)
 				supernodeMask[ii]=1;
 			strs_compute_communication_structure(options, n, LUstruct,
-						ScalePermstruct, supernodeMask, grid);
+						supernodeMask, grid);
 			SUPERLU_FREE(supernodeMask);
 		}
 	}

diff --git a/SRC/single/psgstrs3d.c b/SRC/single/psgstrs3d.c
@@ -151,7 +151,6 @@ float *sready_x, *sready_lsum;
 // #endif
 
 int strs_compute_communication_structure(superlu_dist_options_t *options, int_t n, sLUstruct_t * LUstruct,
-                           sScalePermstruct_t * ScalePermstruct,
                            int* supernodeMask, gridinfo_t *grid)
 {
     Glu_persist_t *Glu_persist = LUstruct->Glu_persist;

diff --git a/SRC/single/pssymbfact_distdata.c b/SRC/single/pssymbfact_distdata.c
@@ -3321,7 +3321,7 @@ float *dense, *dense_col; /* SPA */
     for(int ii=0; ii<nsupers; ii++)
       supernodeMask[ii]=1;
     strs_compute_communication_structure(options, n, LUstruct,
-          ScalePermstruct, supernodeMask, grid);
+          supernodeMask, grid);
     SUPERLU_FREE(supernodeMask);
   }
 

diff --git a/SRC/single/sdistribute.c b/SRC/single/sdistribute.c
@@ -1979,6 +1979,21 @@ if ( !iam) printf(".. Construct Reduce tree for U: %.2f\t\n", t);
 
     } /* else fact != SamePattern_SameRowPerm */
 
+
+	if ( options->Fact != SamePattern_SameRowPerm ) {
+		// /* Flatten L metadata into one buffer. */
+		psflatten_LDATA(options, n, LUstruct, grid);
+
+		// /* Compute communication structure for trisolve. */ 
+		int* supernodeMask = int32Malloc_dist(nsupers);
+		for(int ii=0; ii<nsupers; ii++)
+			supernodeMask[ii]=1;
+		strs_compute_communication_structure(options, n, LUstruct,
+					supernodeMask, grid);
+		SUPERLU_FREE(supernodeMask);
+	}
+
+
 #if ( DEBUGlevel>=1 )
     /* Memory allocated but not freed:
        ilsum, fmod, fsendx_plist, bmod, bsendx_plist  */