Reference-ScaLAPACK · davidedelvento · Oct 20, 2025 · Oct 20, 2025 · Nov 3, 2025 · Nov 3, 2025
diff --git a/BLACS/SRC/BI_Arecv.c b/BLACS/SRC/BI_Arecv.c
@@ -3,9 +3,9 @@
 void BI_Arecv(BLACSCONTEXT *ctxt, Int src, Int msgid, BLACBUFF *bp)
 {
    Int i, info;
-   MpiInt errclass;
+   int errclass;
 
-   info=MPI_Irecv(bp->Buff, bp->N, bp->dtype, src, msgid, ctxt->scp->comm,
+   info=_MPI_Irecv(bp->Buff, bp->N, bp->dtype, src, msgid, ctxt->scp->comm,
                 &bp->Aops[bp->nAops]);
    while(info != MPI_SUCCESS)
    {
@@ -22,7 +22,7 @@ void BI_Arecv(BLACSCONTEXT *ctxt, Int src, Int msgid, BLACBUFF *bp)
 "MPI error %d assumed to mean out of non-blocking resources on call to MPI_Irecv",
                      info);
 #endif
-      info=MPI_Irecv(bp->Buff, bp->N, bp->dtype, src, msgid, ctxt->scp->comm,
+      info=_MPI_Irecv(bp->Buff, bp->N, bp->dtype, src, msgid, ctxt->scp->comm,
                    &bp->Aops[bp->nAops]);
    }
    bp->nAops++;

diff --git a/BLACS/SRC/BI_Asend.c b/BLACS/SRC/BI_Asend.c
@@ -3,9 +3,9 @@
 void BI_Asend(BLACSCONTEXT *ctxt, Int dest, Int msgid, BLACBUFF *bp)
 {
    Int i, info;
-   MpiInt errclass;
+   int errclass;
 
-   info=MPI_Isend(bp->Buff, bp->N, bp->dtype, dest, msgid, ctxt->scp->comm,
+   info=_MPI_Isend(bp->Buff, bp->N, bp->dtype, dest, msgid, ctxt->scp->comm,
                 &bp->Aops[bp->nAops]);
    while(info != MPI_SUCCESS)
    {
@@ -22,7 +22,7 @@ void BI_Asend(BLACSCONTEXT *ctxt, Int dest, Int msgid, BLACBUFF *bp)
 "MPI error %d assumed to mean out of non-blocking resources on call to MPI_Isend",
                         info);
 #endif
-      info=MPI_Isend(bp->Buff, bp->N, bp->dtype, dest, msgid, ctxt->scp->comm,
+      info=_MPI_Isend(bp->Buff, bp->N, bp->dtype, dest, msgid, ctxt->scp->comm,
                    &bp->Aops[bp->nAops]);
    }
    bp->nAops++;

diff --git a/BLACS/SRC/BI_BuffIsFree.c b/BLACS/SRC/BI_BuffIsFree.c
@@ -6,7 +6,7 @@ Int BI_BuffIsFree(BLACBUFF *bp, Int Wait)
  *  wait for all async. operations to complete.
  */
 {
-   MpiInt i, info;
+   int i, info;
    extern MPI_Status *BI_Stats;
    extern Int BI_Np;
 

diff --git a/BLACS/SRC/BI_GetMpiGeType.c b/BLACS/SRC/BI_GetMpiGeType.c
@@ -1,6 +1,6 @@
 #include "Bdef.h"
 MPI_Datatype BI_GetMpiGeType(BLACSCONTEXT *ctxt, Int m, Int n, Int lda,
-                                MPI_Datatype Dtype, Int *N)
+                                MPI_Datatype Dtype, MpiInt *N)
 {
    Int info;
    MPI_Datatype GeType;

diff --git a/BLACS/SRC/BI_GetMpiTrType.c b/BLACS/SRC/BI_GetMpiTrType.c
@@ -1,8 +1,8 @@
 #include "Bdef.h"
 
 MPI_Datatype BI_GetMpiTrType(BLACSCONTEXT *ctxt, char uplo, char diag,
-                                Int m, Int n, Int lda, MPI_Datatype Dtype,
-                                Int *N)
+                                MpiInt m, MpiInt n, Int lda, MPI_Datatype Dtype,
+                                MpiInt *N)
 {
    BLACBUFF *BI_GetBuff(Int);
    MPI_Datatype TrType;
@@ -116,7 +116,7 @@ MPI_Datatype BI_GetMpiTrType(BLACSCONTEXT *ctxt, char uplo, char diag,
    }
 #endif
 
-   i=MPI_Type_indexed(n, len, disp, Dtype, &TrType);
+   i=_MPI_Type_indexed(n, len, disp, Dtype, &TrType);
    i=MPI_Type_commit(&TrType);
    return(TrType);
 }
diff --git a/BLACS/SRC/BI_Pack.c b/BLACS/SRC/BI_Pack.c
@@ -2,7 +2,7 @@
 BLACBUFF *BI_Pack(BLACSCONTEXT *ctxt,BVOID *A,BLACBUFF *bp,MPI_Datatype Dtype)
 {
    BLACBUFF *BI_GetBuff(Int);
-   MpiInt i, info, one=1;
+   int i, info, one=1;
    MPI_Aint eltsiz;
 #ifdef ZeroByteTypeBug
    char *cptr;
@@ -38,6 +38,7 @@ BLACBUFF *BI_Pack(BLACSCONTEXT *ctxt,BVOID *A,BLACBUFF *bp,MPI_Datatype Dtype)
 #endif
    if (bp == NULL)
    {
+      // MPI_Pack_size_c exists but no need to invoke it for a count of one==1
       info=MPI_Pack_size(one, Dtype, ctxt->scp->comm, &i);
       bp = BI_GetBuff(i);
    }

diff --git a/BLACS/SRC/BI_TransUserComm.c b/BLACS/SRC/BI_TransUserComm.c
@@ -1,12 +1,12 @@
 #include "Bdef.h"
 
-MPI_Comm BI_TransUserComm(Int Ucomm, Int Np, Int *pmap)
+MPI_Comm BI_TransUserComm(Int Ucomm, int Np, Int *pmap)
 {
    MPI_Comm bcomm, ucomm;
    MPI_Group bgrp, ugrp;
    Int i;
 
-   MpiInt *mpmap = (MpiInt *)malloc(Np * sizeof(MpiInt));
+   int *mpmap = (int *)malloc(Np * sizeof(int));
    for (i=0; i<Np; i++) mpmap[i] = pmap[i];
 
    ucomm = MPI_Comm_f2c(Ucomm);

diff --git a/BLACS/SRC/BI_Unpack.c b/BLACS/SRC/BI_Unpack.c
@@ -2,7 +2,7 @@
 
 void BI_Unpack(BLACSCONTEXT *ctxt, BVOID *A, BLACBUFF *bp, MPI_Datatype Dtype)
 {
-   MpiInt i=0, info, one=1;
+   int i=0, info, one=1;
 
 /*
  * Some versions of mpich and its derivitives cannot handle 0 byte typedefs,
@@ -11,6 +11,9 @@ void BI_Unpack(BLACSCONTEXT *ctxt, BVOID *A, BLACBUFF *bp, MPI_Datatype Dtype)
 #ifdef ZeroByteTypeBug
    if (Dtype == MPI_BYTE) return;
 #endif
+   // TODO
+   // MPI_Unpack_c exists but no need to invoke it for a count of one==1
+   // however not clear how to handle bp->Len which may be 64-bit, see BLACS/SRC/Bdef.h:54
    info=MPI_Unpack(bp->Buff, bp->Len, &i, A, one, Dtype, ctxt->scp->comm);
    info=MPI_Type_free(&Dtype);
 }
diff --git a/BLACS/SRC/BI_cMPI_sum.c b/BLACS/SRC/BI_cMPI_sum.c
@@ -1,6 +1,6 @@
 #include "Bdef.h"
 void BI_cMPI_sum(void *in, void *inout, MpiInt *N, MPI_Datatype *dtype)
 {
-   void BI_cvvsum(Int, char *, char *);
+   void BI_cvvsum(MpiInt, char *, char *);
    BI_cvvsum(*N, inout, in);
 }
diff --git a/BLACS/SRC/BI_cvvamx.c b/BLACS/SRC/BI_cvvamx.c
@@ -1,10 +1,10 @@
 #include "Bdef.h"
-void BI_cvvamx(Int N, char *vec1, char *vec2)
+void BI_cvvamx(MpiInt N, char *vec1, char *vec2)
 {
    SCOMPLEX *v1=(SCOMPLEX*)vec1, *v2=(SCOMPLEX*)vec2;
    float diff;
    BI_DistType *dist1, *dist2;
-   Int i, k;
+   MpiInt i, k;
 
    k = N * sizeof(SCOMPLEX);
    i = k % sizeof(BI_DistType);

diff --git a/BLACS/SRC/BI_cvvamx2.c b/BLACS/SRC/BI_cvvamx2.c
@@ -1,7 +1,7 @@
 #include "Bdef.h"
-void BI_cvvamx2(Int N, char *vec1, char *vec2)
+void BI_cvvamx2(MpiInt N, char *vec1, char *vec2)
 {
-   Int r, i;
+   MpiInt r, i;
    float *v1=(float*)vec1, *v2=(float*)vec2;
    float diff;
 

diff --git a/BLACS/SRC/BI_cvvsum.c b/BLACS/SRC/BI_cvvsum.c
@@ -1,8 +1,8 @@
 #include "Bdef.h"
-void BI_cvvsum(Int N, char *vec1, char *vec2)
+void BI_cvvsum(MpiInt N, char *vec1, char *vec2)
 {
    float *v1=(float*)vec1, *v2=(float*)vec2;
-   Int k;
+   MpiInt k; // size_t might be more appropriate, as in BI_cvvsum.c BI_dvvsum.c BI_ivvsum.c BI_svvsum.c BI_zvvsum.c
    N *=2;
    for (k=0; k < N; k++) v1[k] += v2[k];
 }
diff --git a/BLACS/SRC/BI_dvvsum.c b/BLACS/SRC/BI_dvvsum.c
@@ -1,7 +1,7 @@
 #include "Bdef.h"
-void BI_dvvsum(Int N, char *vec1, char *vec2)
+void BI_dvvsum(MpiInt N, char *vec1, char *vec2)
 {
    double *v1=(double*)vec1, *v2=(double*)vec2;
-   Int k;
+   MpiInt k; // size_t might be more appropriate, as in BI_cvvsum.c BI_dvvsum.c BI_ivvsum.c BI_svvsum.c BI_zvvsum.c
    for (k=0; k < N; k++) v1[k] += v2[k];
 }
diff --git a/BLACS/SRC/BI_ivvamn.c b/BLACS/SRC/BI_ivvamn.c
@@ -1,10 +1,10 @@
 #include "Bdef.h"
-void BI_ivvamn(Int N, char *vec1, char *vec2)
+void BI_ivvamn(MpiInt N, char *vec1, char *vec2)
 {
    Int *v1=(Int*)vec1, *v2=(Int*)vec2;
    Int diff;
    BI_DistType *dist1, *dist2;
-   Int i, k;
+   MpiInt i, k;
 
    k = N * sizeof(Int);
    i = k % sizeof(BI_DistType);

diff --git a/BLACS/SRC/BI_ivvamn2.c b/BLACS/SRC/BI_ivvamn2.c
@@ -1,8 +1,8 @@
 #include "Bdef.h"
 
-void BI_ivvamn2(Int N, char *vec1, char *vec2)
+void BI_ivvamn2(MpiInt N, char *vec1, char *vec2)
 {
-   Int k;
+   MpiInt k;
    Int *v1=(Int*)vec1, *v2=(Int*)vec2;
    Int diff;
 

diff --git a/BLACS/SRC/BI_ivvsum.c b/BLACS/SRC/BI_ivvsum.c
@@ -1,7 +1,7 @@
 #include "Bdef.h"
-void BI_ivvsum(Int N, char *vec1, char *vec2)
+void BI_ivvsum(MpiInt N, char *vec1, char *vec2)
 {
    Int *v1=(Int*)vec1, *v2=(Int*)vec2;
-   Int k;
+   MpiInt k; // size_t might be more appropriate, as in BI_cvvsum.c BI_dvvsum.c BI_ivvsum.c BI_svvsum.c BI_zvvsum.c
    for (k=0; k < N; k++) v1[k] += v2[k];
 }
diff --git a/BLACS/SRC/BI_svvsum.c b/BLACS/SRC/BI_svvsum.c
@@ -1,7 +1,7 @@
 #include "Bdef.h"
-void BI_svvsum(Int N, char *vec1, char *vec2)
+void BI_svvsum(MpiInt N, char *vec1, char *vec2)
 {
    float *v1=(float*)vec1, *v2=(float*)vec2;
-   Int k;
+   MpiInt k; // size_t might be more appropriate, as in BI_cvvsum.c BI_dvvsum.c BI_ivvsum.c BI_svvsum.c BI_zvvsum.c
    for (k=0; k < N; k++) v1[k] += v2[k];
 }
diff --git a/BLACS/SRC/BI_zvvsum.c b/BLACS/SRC/BI_zvvsum.c
@@ -1,8 +1,8 @@
 #include "Bdef.h"
-void BI_zvvsum(Int N, char *vec1, char *vec2)
+void BI_zvvsum(MpiInt N, char *vec1, char *vec2)
 {
    double *v1=(double*)vec1, *v2=(double*)vec2;
-   Int k;
+   MpiInt k; // size_t might be more appropriate, as in BI_cvvsum.c BI_dvvsum.c BI_ivvsum.c BI_svvsum.c BI_zvvsum.c
    N *=2;
    for (k=0; k < N; k++) v1[k] += v2[k];
 }
diff --git a/BLACS/SRC/Bconfig.h b/BLACS/SRC/Bconfig.h
@@ -21,9 +21,107 @@
 #ifndef Int
 #define Int int
 #endif
-#ifndef MpiInt
+
+/* If somebody is building with BigMPI
+ * give it priority compared to other
+ * solutions since otherwise they would
+ * not have bothered to configure BLACS
+ * with BigMPI
+ */
+#ifdef BIGMPI
+#define MpiInt MPI_Count
+inline int _MPI_Isend(const void *buf, MPI_Count count, MPI_Datatype datatype, int dest,
+              int tag, MPI_Comm comm, MPI_Request *request) {
+	return MPIX_Isend_x(buf, count, datatype, dest, tag, comm, request);
+}
+inline int _MPI_Irecv(void *buf, MPI_Count count, MPI_Datatype datatype,
+		int source, int tag, MPI_Comm comm, MPI_Request *request) {
+	return MPIX_Irecv_x (buf, count, datatype, source, tag, comm, request);
+}
+inline int _MPI_Op_create(MPI_User_function_c *user_fn, int commute, MPI_Op *op) {
+	return BigMPI_Op_create (user_fn, commute, op);
+}
+inline int _MPI_Type_create_struct(MPI_Count count,
+                             const MPI_Count array_of_blocklengths[],
+                             const MPI_Count array_of_displacements[],
+                             const MPI_Datatype array_of_types[],
+                             MPI_Datatype *newtype) {
+	// TODO not sure what to do here
+}
+inline int _MPI_Type_indexed(MPI_Count count, const MPI_Count array_of_blocklengths[],
+                       const MPI_Count array_of_displacements[],
+                       MPI_Datatype oldtype, MPI_Datatype *newtype) {
+	// TODO not sure what to do here
+}
+#else // no BIGMPI
+/* If there is no BigMPI, but there is
+ * an MPI implementation v4.x use
+ * that for large tranfers. Caveat:
+ * there are some incomplete v4.x
+ * implementations around which will fail.
+ * Hopefully they will be fixed
+ * before this goes to production
+ */
+#if MPI_VERSION==4
+#define MpiInt MPI_Count
+inline int _MPI_Isend(const void *buf, MPI_Count count, MPI_Datatype datatype, int dest,
+              int tag, MPI_Comm comm, MPI_Request *request) {
+	return MPI_Isend_c(buf, count, datatype, dest, tag, comm, request);
+}
+inline int _MPI_Irecv(void *buf, MPI_Count count, MPI_Datatype datatype,
+		int source, int tag, MPI_Comm comm, MPI_Request *request) {
+	return MPI_Irecv_c (buf, count, datatype, source, tag, comm, request);
+}
+inline int _MPI_Op_create(MPI_User_function_c *user_fn, int commute, MPI_Op *op) {
+	return MPI_Op_create_c (user_fn, commute, op);
+}
+inline int _MPI_Type_create_struct(MPI_Count count,
+                             const MPI_Count array_of_blocklengths[],
+                             const MPI_Count array_of_displacements[],
+                             const MPI_Datatype array_of_types[],
+                             MPI_Datatype *newtype) {
+	return MPI_Type_create_struct_c(count, array_of_blocklengths,
+                             array_of_displacements, array_of_types,
+                             newtype);
+}
+inline int _MPI_Type_indexed(MPI_Count count, const MPI_Count array_of_blocklengths[],
+                       const MPI_Count array_of_displacements[],
+                       MPI_Datatype oldtype, MPI_Datatype *newtype) {
+	return MPI_Type_indexed_c(count, array_of_blocklengths, array_of_displacements,
+                       oldtype, newtype);
+}
+/* Otherwise default to non-big
+ * transfers
+ */
+#else // no MPI v4
 #define MpiInt int
-#endif
+inline int _MPI_Isend(const void *buf, int count, MPI_Datatype datatype, int dest,
+              int tag, MPI_Comm comm, MPI_Request *request) {
+	return MPI_Isend(buf, count, datatype, dest, tag, comm, request);
+}
+inline int _MPI_Irecv(void *buf, int count, MPI_Datatype datatype,
+		int source, int tag, MPI_Comm comm, MPI_Request *request) {
+	return MPI_Irecv (buf, count, datatype, source, tag, comm, request);
+}
+inline int _MPI_Op_create(MPI_User_function *user_fn, int commute, MPI_Op *op) {
+	return MPI_Op_create (user_fn, commute, op);
+}
+inline int _MPI_Type_create_struct(int count, const int array_of_blocklengths[],
+                           const MPI_Aint array_of_displacements[],
+                           const MPI_Datatype array_of_types[],
+                           MPI_Datatype *newtype)
+	return MPI_Type_create_struct(count, array_of_blocklengths,
+                             array_of_displacements, array_of_types,
+                             newtype);
+}
+int MPI_Type_indexed(int count, const int array_of_blocklengths[],
+                     const int array_of_displacements[], MPI_Datatype oldtype,
+                     MPI_Datatype *newtype) {
+	return MPI_Type_indexed(count, array_of_blocklengths, array_of_displacements,
+                       oldtype, newtype);
+}
+#endif // MPI version
+#endif // BIGMPI
 
 /*
  * These macros define the naming strategy needed for a fortran

diff --git a/BLACS/SRC/Bdef.h b/BLACS/SRC/Bdef.h
@@ -58,14 +58,14 @@ struct bLaCbUfF
    Int nAops;              /* number of asynchronous operations out of buff */
    MPI_Request *Aops;   /* list of async. operations out of buff */
    MPI_Datatype dtype;  /* data type of buffer */
-   Int N;                  /* number of elements of data type in buff */
+   MpiInt N;               /* number of elements of data type in buff */
    BLACBUFF *prev, *next;  /* pointer to the other BLACBUFF in queue */
 };
 
 /*
  * Pointer to the combine's vector-vector functions
  */
-typedef void (*VVFUNPTR)(Int, char *, char *);
+typedef void (*VVFUNPTR)(MpiInt, char *, char *);
 typedef void (*SDRVPTR)(BLACSCONTEXT *, Int, Int, BLACBUFF *);
 
 

diff --git a/BLACS/SRC/blacs_get_.c b/BLACS/SRC/blacs_get_.c
@@ -7,7 +7,7 @@ F_VOID_FUNC blacs_get_(Int *ConTxt, Int *what, Int *val)
 {
    Int Csys2blacs_handle(MPI_Comm);
    Int ierr, *iptr;
-   MpiInt flag;
+   int flag;
    Int comm;
    BLACSCONTEXT *ctxt;
 

diff --git a/BLACS/SRC/blacs_map_.c b/BLACS/SRC/blacs_map_.c
@@ -13,7 +13,7 @@ F_VOID_FUNC blacs_gridmap_(Int *ConTxt, Int *usermap, Int *ldup, Int *nprow0,
    MPI_Comm Cblacs2sys_handle(Int BlacsCtxt);
    MPI_Comm BI_TransUserComm(Int, Int, Int *);
 
-   MpiInt Iam;
+   int Iam;
    Int info, i, j, *iptr;
    Int myrow, mycol, nprow, npcol, Ng;
    BLACSCONTEXT *ctxt, **tCTxts;
@@ -53,7 +53,7 @@ F_VOID_FUNC blacs_gridmap_(Int *ConTxt, Int *usermap, Int *ldup, Int *nprow0,
       for (i=0; i < nprow; i++) iptr[i*npcol+j] = usermap[j*Mpval(ldup)+i];
    }
 #if (INTFACE == C_CALL)
-   MpiInt *miptr = (MpiInt *) malloc(Ng*sizeof(MpiInt));
+   int *miptr = (int *) malloc(Ng*sizeof(int));
    for (j=0; j < Ng; j++) miptr[j] = iptr[j];
    tcomm = Cblacs2sys_handle(*ConTxt);
    MPI_Comm_group(tcomm, &grp);           /* find input comm's group */