xref: /petsc/src/vec/is/sf/impls/basic/sfpack.c (revision 2c71b3e237ead271e4f3aa1505f92bf476e3413d)
120c24465SJunchao Zhang #include "petsc/private/sfimpl.h"
240e23c03SJunchao Zhang #include <../src/vec/is/sf/impls/basic/sfpack.h>
340e23c03SJunchao Zhang #include <../src/vec/is/sf/impls/basic/sfbasic.h>
440e23c03SJunchao Zhang 
57fd2d3dbSJunchao Zhang /* This is a C file that contains packing facilities, with dispatches to device if enabled. */
67fd2d3dbSJunchao Zhang 
740e23c03SJunchao Zhang /*
840e23c03SJunchao Zhang  * MPI_Reduce_local is not really useful because it can't handle sparse data and it vectorizes "in the wrong direction",
940e23c03SJunchao Zhang  * therefore we pack data types manually. This file defines packing routines for the standard data types.
1040e23c03SJunchao Zhang  */
1140e23c03SJunchao Zhang 
12cd620004SJunchao Zhang #define CPPJoin4(a,b,c,d)  a##_##b##_##c##_##d
1340e23c03SJunchao Zhang 
14cd620004SJunchao Zhang /* Operations working like s += t */
15cd620004SJunchao Zhang #define OP_BINARY(op,s,t)   do {(s) = (s) op (t);  } while (0)      /* binary ops in the middle such as +, *, && etc. */
16cd620004SJunchao Zhang #define OP_FUNCTION(op,s,t) do {(s) = op((s),(t)); } while (0)      /* ops like a function, such as PetscMax, PetscMin */
17cd620004SJunchao Zhang #define OP_LXOR(op,s,t)     do {(s) = (!(s)) != (!(t));} while (0)  /* logical exclusive OR */
18cd620004SJunchao Zhang #define OP_ASSIGN(op,s,t)   do {(s) = (t);} while (0)
19cd620004SJunchao Zhang /* Ref MPI MAXLOC */
20cd620004SJunchao Zhang #define OP_XLOC(op,s,t) \
21cd620004SJunchao Zhang   do {                                       \
22cd620004SJunchao Zhang     if ((s).u == (t).u) (s).i = PetscMin((s).i,(t).i); \
23cd620004SJunchao Zhang     else if (!((s).u op (t).u)) s = t;           \
24cd620004SJunchao Zhang   } while (0)
2540e23c03SJunchao Zhang 
2640e23c03SJunchao Zhang /* DEF_PackFunc - macro defining a Pack routine
2740e23c03SJunchao Zhang 
2840e23c03SJunchao Zhang    Arguments of the macro:
29b23bfdefSJunchao Zhang    +Type      Type of the basic data in an entry, i.e., int, PetscInt, PetscReal etc. It is not the type of an entry.
30fcc7397dSJunchao Zhang    .BS        Block size for vectorization. It is a factor of bsz.
31b23bfdefSJunchao Zhang    -EQ        (bs == BS) ? 1 : 0. EQ is a compile-time const to help compiler optimizations. See below.
3240e23c03SJunchao Zhang 
3340e23c03SJunchao Zhang    Arguments of the Pack routine:
34cd620004SJunchao Zhang    +count     Number of indices in idx[].
35fcc7397dSJunchao Zhang    .start     When opt and idx are NULL, it means indices are contiguous & start is the first index; otherwise, not used.
36fcc7397dSJunchao Zhang    .opt       Per-pack optimization plan. NULL means no such plan.
37fcc7397dSJunchao Zhang    .idx       Indices of entries to packed.
38eb02082bSJunchao Zhang    .link      Provide a context for the current call, such as link->bs, number of basic types in an entry. Ex. if unit is MPI_2INT, then bs=2 and the basic type is int.
39cd620004SJunchao Zhang    .unpacked  Address of the unpacked data. The entries will be packed are unpacked[idx[i]],for i in [0,count).
40cd620004SJunchao Zhang    -packed    Address of the packed data.
4140e23c03SJunchao Zhang  */
42b23bfdefSJunchao Zhang #define DEF_PackFunc(Type,BS,EQ) \
43fcc7397dSJunchao Zhang   static PetscErrorCode CPPJoin4(Pack,Type,BS,EQ)(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,const void *unpacked,void *packed) \
44b23bfdefSJunchao Zhang   {                                                                                                          \
4540e23c03SJunchao Zhang     PetscErrorCode ierr;                                                                                     \
46b23bfdefSJunchao Zhang     const Type     *u = (const Type*)unpacked,*u2;                                                           \
47b23bfdefSJunchao Zhang     Type           *p = (Type*)packed,*p2;                                                                   \
48fcc7397dSJunchao Zhang     PetscInt       i,j,k,X,Y,r,bs=link->bs;                                                                  \
49fcc7397dSJunchao Zhang     const PetscInt M = (EQ) ? 1 : bs/BS; /* If EQ, then M=1 enables compiler's const-propagation */          \
50b23bfdefSJunchao Zhang     const PetscInt MBS = M*BS; /* MBS=bs. We turn MBS into a compile time const when EQ=1. */                \
5140e23c03SJunchao Zhang     PetscFunctionBegin;                                                                                      \
52fcc7397dSJunchao Zhang     if (!idx) {ierr = PetscArraycpy(p,u+start*MBS,MBS*count);CHKERRQ(ierr);}/* idx[] are contiguous */       \
53fcc7397dSJunchao Zhang     else if (opt) { /* has optimizations available */                                                        \
54fcc7397dSJunchao Zhang       p2 = p;                                                                                                \
55fcc7397dSJunchao Zhang       for (r=0; r<opt->n; r++) {                                                                             \
56fcc7397dSJunchao Zhang         u2 = u + opt->start[r]*MBS;                                                                          \
57fcc7397dSJunchao Zhang         X  = opt->X[r];                                                                                      \
58fcc7397dSJunchao Zhang         Y  = opt->Y[r];                                                                                      \
59fcc7397dSJunchao Zhang         for (k=0; k<opt->dz[r]; k++)                                                                         \
60fcc7397dSJunchao Zhang           for (j=0; j<opt->dy[r]; j++) {                                                                     \
61fcc7397dSJunchao Zhang             ierr = PetscArraycpy(p2,u2+(X*Y*k+X*j)*MBS,opt->dx[r]*MBS);CHKERRQ(ierr);                        \
62fcc7397dSJunchao Zhang             p2  += opt->dx[r]*MBS;                                                                           \
63fcc7397dSJunchao Zhang           }                                                                                                  \
64fcc7397dSJunchao Zhang       }                                                                                                      \
65fcc7397dSJunchao Zhang     } else {                                                                                                 \
66b23bfdefSJunchao Zhang       for (i=0; i<count; i++)                                                                                \
67eb02082bSJunchao Zhang         for (j=0; j<M; j++)     /* Decent compilers should eliminate this loop when M = const 1 */           \
68eb02082bSJunchao Zhang           for (k=0; k<BS; k++)  /* Compiler either unrolls (BS=1) or vectorizes (BS=2,4,8,etc) this loop */  \
69b23bfdefSJunchao Zhang             p[i*MBS+j*BS+k] = u[idx[i]*MBS+j*BS+k];                                                          \
7040e23c03SJunchao Zhang     }                                                                                                        \
7140e23c03SJunchao Zhang     PetscFunctionReturn(0);                                                                                  \
7240e23c03SJunchao Zhang   }
7340e23c03SJunchao Zhang 
74cd620004SJunchao Zhang /* DEF_Action - macro defining a UnpackAndInsert routine that unpacks data from a contiguous buffer
75cd620004SJunchao Zhang                 and inserts into a sparse array.
7640e23c03SJunchao Zhang 
7740e23c03SJunchao Zhang    Arguments:
78b23bfdefSJunchao Zhang   .Type       Type of the data
7940e23c03SJunchao Zhang   .BS         Block size for vectorization
80b23bfdefSJunchao Zhang   .EQ        (bs == BS) ? 1 : 0. EQ is a compile-time const.
8140e23c03SJunchao Zhang 
8240e23c03SJunchao Zhang   Notes:
8340e23c03SJunchao Zhang    This macro is not combined with DEF_ActionAndOp because we want to use memcpy in this macro.
8440e23c03SJunchao Zhang  */
85cd620004SJunchao Zhang #define DEF_UnpackFunc(Type,BS,EQ)               \
86fcc7397dSJunchao Zhang   static PetscErrorCode CPPJoin4(UnpackAndInsert,Type,BS,EQ)(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,void *unpacked,const void *packed) \
87b23bfdefSJunchao Zhang   {                                                                                                          \
8840e23c03SJunchao Zhang     PetscErrorCode ierr;                                                                                     \
89b23bfdefSJunchao Zhang     Type           *u = (Type*)unpacked,*u2;                                                                 \
90fcc7397dSJunchao Zhang     const Type     *p = (const Type*)packed;                                                                 \
91fcc7397dSJunchao Zhang     PetscInt       i,j,k,X,Y,r,bs=link->bs;                                                                  \
92fcc7397dSJunchao Zhang     const PetscInt M = (EQ) ? 1 : bs/BS; /* If EQ, then M=1 enables compiler's const-propagation */          \
93b23bfdefSJunchao Zhang     const PetscInt MBS = M*BS; /* MBS=bs. We turn MBS into a compile time const when EQ=1. */                \
9440e23c03SJunchao Zhang     PetscFunctionBegin;                                                                                      \
95b23bfdefSJunchao Zhang     if (!idx) {                                                                                              \
96fcc7397dSJunchao Zhang       u += start*MBS;                                                                                        \
97fcc7397dSJunchao Zhang       if (u != p) {ierr = PetscArraycpy(u,p,count*MBS);CHKERRQ(ierr);}                                       \
98fcc7397dSJunchao Zhang     } else if (opt) { /* has optimizations available */                                                      \
99fcc7397dSJunchao Zhang       for (r=0; r<opt->n; r++) {                                                                             \
100fcc7397dSJunchao Zhang         u2 = u + opt->start[r]*MBS;                                                                          \
101fcc7397dSJunchao Zhang         X  = opt->X[r];                                                                                      \
102fcc7397dSJunchao Zhang         Y  = opt->Y[r];                                                                                      \
103fcc7397dSJunchao Zhang         for (k=0; k<opt->dz[r]; k++)                                                                         \
104fcc7397dSJunchao Zhang           for (j=0; j<opt->dy[r]; j++) {                                                                     \
105fcc7397dSJunchao Zhang             ierr = PetscArraycpy(u2+(X*Y*k+X*j)*MBS,p,opt->dx[r]*MBS);CHKERRQ(ierr);                         \
106fcc7397dSJunchao Zhang             p   += opt->dx[r]*MBS;                                                                           \
107fcc7397dSJunchao Zhang           }                                                                                                  \
108fcc7397dSJunchao Zhang       }                                                                                                      \
109fcc7397dSJunchao Zhang     } else {                                                                                                 \
110b23bfdefSJunchao Zhang       for (i=0; i<count; i++)                                                                                \
111b23bfdefSJunchao Zhang         for (j=0; j<M; j++)                                                                                  \
112cd620004SJunchao Zhang           for (k=0; k<BS; k++) u[idx[i]*MBS+j*BS+k] = p[i*MBS+j*BS+k];                                       \
11340e23c03SJunchao Zhang     }                                                                                                        \
11440e23c03SJunchao Zhang     PetscFunctionReturn(0);                                                                                  \
11540e23c03SJunchao Zhang   }
11640e23c03SJunchao Zhang 
117cd620004SJunchao Zhang /* DEF_UnpackAndOp - macro defining a UnpackAndOp routine where Op should not be Insert
11840e23c03SJunchao Zhang 
11940e23c03SJunchao Zhang    Arguments:
120cd620004SJunchao Zhang   +Opname     Name of the Op, such as Add, Mult, LAND, etc.
121b23bfdefSJunchao Zhang   .Type       Type of the data
12240e23c03SJunchao Zhang   .BS         Block size for vectorization
123b23bfdefSJunchao Zhang   .EQ         (bs == BS) ? 1 : 0. EQ is a compile-time const.
124cd620004SJunchao Zhang   .Op         Operator for the op, such as +, *, &&, ||, PetscMax, PetscMin, etc.
125cd620004SJunchao Zhang   .OpApply    Macro defining application of the op. Could be OP_BINARY, OP_FUNCTION, OP_LXOR
12640e23c03SJunchao Zhang  */
127cd620004SJunchao Zhang #define DEF_UnpackAndOp(Type,BS,EQ,Opname,Op,OpApply) \
128fcc7397dSJunchao Zhang   static PetscErrorCode CPPJoin4(UnpackAnd##Opname,Type,BS,EQ)(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,void *unpacked,const void *packed) \
129b23bfdefSJunchao Zhang   {                                                                                                          \
130cd620004SJunchao Zhang     Type           *u = (Type*)unpacked,*u2;                                                                 \
131fcc7397dSJunchao Zhang     const Type     *p = (const Type*)packed;                                                                 \
132fcc7397dSJunchao Zhang     PetscInt       i,j,k,X,Y,r,bs=link->bs;                                                                  \
133fcc7397dSJunchao Zhang     const PetscInt M = (EQ) ? 1 : bs/BS; /* If EQ, then M=1 enables compiler's const-propagation */          \
134b23bfdefSJunchao Zhang     const PetscInt MBS = M*BS; /* MBS=bs. We turn MBS into a compile time const when EQ=1. */                \
13540e23c03SJunchao Zhang     PetscFunctionBegin;                                                                                      \
136b23bfdefSJunchao Zhang     if (!idx) {                                                                                              \
137fcc7397dSJunchao Zhang       u += start*MBS;                                                                                        \
138cd620004SJunchao Zhang       for (i=0; i<count; i++)                                                                                \
139cd620004SJunchao Zhang         for (j=0; j<M; j++)                                                                                  \
140cd620004SJunchao Zhang           for (k=0; k<BS; k++)                                                                               \
141cd620004SJunchao Zhang             OpApply(Op,u[i*MBS+j*BS+k],p[i*MBS+j*BS+k]);                                                     \
142fcc7397dSJunchao Zhang     } else if (opt) { /* idx[] has patterns */                                                               \
143fcc7397dSJunchao Zhang       for (r=0; r<opt->n; r++) {                                                                             \
144fcc7397dSJunchao Zhang         u2 = u + opt->start[r]*MBS;                                                                          \
145fcc7397dSJunchao Zhang         X  = opt->X[r];                                                                                      \
146fcc7397dSJunchao Zhang         Y  = opt->Y[r];                                                                                      \
147fcc7397dSJunchao Zhang         for (k=0; k<opt->dz[r]; k++)                                                                         \
148fcc7397dSJunchao Zhang           for (j=0; j<opt->dy[r]; j++) {                                                                     \
149fcc7397dSJunchao Zhang             for (i=0; i<opt->dx[r]*MBS; i++) OpApply(Op,u2[(X*Y*k+X*j)*MBS+i],p[i]);                         \
150fcc7397dSJunchao Zhang             p += opt->dx[r]*MBS;                                                                             \
151fcc7397dSJunchao Zhang           }                                                                                                  \
152fcc7397dSJunchao Zhang       }                                                                                                      \
153fcc7397dSJunchao Zhang     } else {                                                                                                 \
154cd620004SJunchao Zhang       for (i=0; i<count; i++)                                                                                \
155cd620004SJunchao Zhang         for (j=0; j<M; j++)                                                                                  \
156cd620004SJunchao Zhang           for (k=0; k<BS; k++)                                                                               \
157cd620004SJunchao Zhang             OpApply(Op,u[idx[i]*MBS+j*BS+k],p[i*MBS+j*BS+k]);                                                \
158cd620004SJunchao Zhang     }                                                                                                        \
159cd620004SJunchao Zhang     PetscFunctionReturn(0);                                                                                  \
160cd620004SJunchao Zhang   }
161cd620004SJunchao Zhang 
162cd620004SJunchao Zhang #define DEF_FetchAndOp(Type,BS,EQ,Opname,Op,OpApply) \
163fcc7397dSJunchao Zhang   static PetscErrorCode CPPJoin4(FetchAnd##Opname,Type,BS,EQ)(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,void *unpacked,void *packed) \
164cd620004SJunchao Zhang   {                                                                                                          \
165fcc7397dSJunchao Zhang     Type           *u = (Type*)unpacked,*p = (Type*)packed,tmp;                                              \
166fcc7397dSJunchao Zhang     PetscInt       i,j,k,r,l,bs=link->bs;                                                                    \
167fcc7397dSJunchao Zhang     const PetscInt M = (EQ) ? 1 : bs/BS;                                                                     \
168fcc7397dSJunchao Zhang     const PetscInt MBS = M*BS;                                                                               \
169cd620004SJunchao Zhang     PetscFunctionBegin;                                                                                      \
170fcc7397dSJunchao Zhang     for (i=0; i<count; i++) {                                                                                \
171fcc7397dSJunchao Zhang       r = (!idx ? start+i : idx[i])*MBS;                                                                     \
172fcc7397dSJunchao Zhang       l = i*MBS;                                                                                             \
173b23bfdefSJunchao Zhang       for (j=0; j<M; j++)                                                                                    \
174b23bfdefSJunchao Zhang         for (k=0; k<BS; k++) {                                                                               \
175fcc7397dSJunchao Zhang           tmp = u[r+j*BS+k];                                                                                 \
176fcc7397dSJunchao Zhang           OpApply(Op,u[r+j*BS+k],p[l+j*BS+k]);                                                               \
177fcc7397dSJunchao Zhang           p[l+j*BS+k] = tmp;                                                                                 \
178cd620004SJunchao Zhang         }                                                                                                    \
179cd620004SJunchao Zhang     }                                                                                                        \
180cd620004SJunchao Zhang     PetscFunctionReturn(0);                                                                                  \
181cd620004SJunchao Zhang   }
182cd620004SJunchao Zhang 
183cd620004SJunchao Zhang #define DEF_ScatterAndOp(Type,BS,EQ,Opname,Op,OpApply) \
184fcc7397dSJunchao Zhang   static PetscErrorCode CPPJoin4(ScatterAnd##Opname,Type,BS,EQ)(PetscSFLink link,PetscInt count,PetscInt srcStart,PetscSFPackOpt srcOpt,const PetscInt *srcIdx,const void *src,PetscInt dstStart,PetscSFPackOpt dstOpt,const PetscInt *dstIdx,void *dst) \
185cd620004SJunchao Zhang   {                                                                                                          \
186fcc7397dSJunchao Zhang     PetscErrorCode ierr;                                                                                     \
187fcc7397dSJunchao Zhang     const Type     *u = (const Type*)src;                                                                    \
188fcc7397dSJunchao Zhang     Type           *v = (Type*)dst;                                                                          \
189fcc7397dSJunchao Zhang     PetscInt       i,j,k,s,t,X,Y,bs = link->bs;                                                              \
190cd620004SJunchao Zhang     const PetscInt M = (EQ) ? 1 : bs/BS;                                                                     \
191cd620004SJunchao Zhang     const PetscInt MBS = M*BS;                                                                               \
192cd620004SJunchao Zhang     PetscFunctionBegin;                                                                                      \
193fcc7397dSJunchao Zhang     if (!srcIdx) { /* src is contiguous */                                                                   \
194fcc7397dSJunchao Zhang       u += srcStart*MBS;                                                                                     \
195fcc7397dSJunchao Zhang       ierr = CPPJoin4(UnpackAnd##Opname,Type,BS,EQ)(link,count,dstStart,dstOpt,dstIdx,dst,u);CHKERRQ(ierr);  \
196fcc7397dSJunchao Zhang     } else if (srcOpt && !dstIdx) { /* src is 3D, dst is contiguous */                                       \
197fcc7397dSJunchao Zhang       u += srcOpt->start[0]*MBS;                                                                             \
198fcc7397dSJunchao Zhang       v += dstStart*MBS;                                                                                     \
199fcc7397dSJunchao Zhang       X  = srcOpt->X[0]; Y = srcOpt->Y[0];                                                                   \
200fcc7397dSJunchao Zhang       for (k=0; k<srcOpt->dz[0]; k++)                                                                        \
201fcc7397dSJunchao Zhang         for (j=0; j<srcOpt->dy[0]; j++) {                                                                    \
202fcc7397dSJunchao Zhang           for (i=0; i<srcOpt->dx[0]*MBS; i++) OpApply(Op,v[i],u[(X*Y*k+X*j)*MBS+i]);                         \
203fcc7397dSJunchao Zhang           v += srcOpt->dx[0]*MBS;                                                                            \
204fcc7397dSJunchao Zhang         }                                                                                                    \
205fcc7397dSJunchao Zhang     } else { /* all other cases */                                                                           \
206fcc7397dSJunchao Zhang       for (i=0; i<count; i++) {                                                                              \
207fcc7397dSJunchao Zhang         s = (!srcIdx ? srcStart+i : srcIdx[i])*MBS;                                                          \
208fcc7397dSJunchao Zhang         t = (!dstIdx ? dstStart+i : dstIdx[i])*MBS;                                                          \
209cd620004SJunchao Zhang         for (j=0; j<M; j++)                                                                                  \
210fcc7397dSJunchao Zhang           for (k=0; k<BS; k++) OpApply(Op,v[t+j*BS+k],u[s+j*BS+k]);                                          \
211fcc7397dSJunchao Zhang       }                                                                                                      \
212cd620004SJunchao Zhang     }                                                                                                        \
213cd620004SJunchao Zhang     PetscFunctionReturn(0);                                                                                  \
214cd620004SJunchao Zhang   }
215cd620004SJunchao Zhang 
216cd620004SJunchao Zhang #define DEF_FetchAndOpLocal(Type,BS,EQ,Opname,Op,OpApply) \
217fcc7397dSJunchao Zhang   static PetscErrorCode CPPJoin4(FetchAnd##Opname##Local,Type,BS,EQ)(PetscSFLink link,PetscInt count,PetscInt rootstart,PetscSFPackOpt rootopt,const PetscInt *rootidx,void *rootdata,PetscInt leafstart,PetscSFPackOpt leafopt,const PetscInt *leafidx,const void *leafdata,void *leafupdate) \
218cd620004SJunchao Zhang   {                                                                                                          \
219fcc7397dSJunchao Zhang     Type           *rdata = (Type*)rootdata,*lupdate = (Type*)leafupdate;                                    \
220fcc7397dSJunchao Zhang     const Type     *ldata = (const Type*)leafdata;                                                           \
221fcc7397dSJunchao Zhang     PetscInt       i,j,k,r,l,bs = link->bs;                                                                  \
222cd620004SJunchao Zhang     const PetscInt M = (EQ) ? 1 : bs/BS;                                                                     \
223cd620004SJunchao Zhang     const PetscInt MBS = M*BS;                                                                               \
224cd620004SJunchao Zhang     PetscFunctionBegin;                                                                                      \
225fcc7397dSJunchao Zhang     for (i=0; i<count; i++) {                                                                                \
226fcc7397dSJunchao Zhang       r = (rootidx ? rootidx[i] : rootstart+i)*MBS;                                                          \
227fcc7397dSJunchao Zhang       l = (leafidx ? leafidx[i] : leafstart+i)*MBS;                                                          \
228cd620004SJunchao Zhang       for (j=0; j<M; j++)                                                                                    \
229cd620004SJunchao Zhang         for (k=0; k<BS; k++) {                                                                               \
230fcc7397dSJunchao Zhang           lupdate[l+j*BS+k] = rdata[r+j*BS+k];                                                               \
231fcc7397dSJunchao Zhang           OpApply(Op,rdata[r+j*BS+k],ldata[l+j*BS+k]);                                                       \
23240e23c03SJunchao Zhang         }                                                                                                    \
23340e23c03SJunchao Zhang     }                                                                                                        \
23440e23c03SJunchao Zhang     PetscFunctionReturn(0);                                                                                  \
23540e23c03SJunchao Zhang   }
23640e23c03SJunchao Zhang 
237b23bfdefSJunchao Zhang /* Pack, Unpack/Fetch ops */
238b23bfdefSJunchao Zhang #define DEF_Pack(Type,BS,EQ)                                                      \
239b23bfdefSJunchao Zhang   DEF_PackFunc(Type,BS,EQ)                                                        \
240cd620004SJunchao Zhang   DEF_UnpackFunc(Type,BS,EQ)                                                      \
241cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,Insert,=,OP_ASSIGN)                                 \
242cd620004SJunchao Zhang   static void CPPJoin4(PackInit_Pack,Type,BS,EQ)(PetscSFLink link) {              \
243eb02082bSJunchao Zhang     link->h_Pack            = CPPJoin4(Pack,           Type,BS,EQ);               \
244eb02082bSJunchao Zhang     link->h_UnpackAndInsert = CPPJoin4(UnpackAndInsert,Type,BS,EQ);               \
245cd620004SJunchao Zhang     link->h_ScatterAndInsert= CPPJoin4(ScatterAndInsert,Type,BS,EQ);              \
24640e23c03SJunchao Zhang   }
24740e23c03SJunchao Zhang 
248b23bfdefSJunchao Zhang /* Add, Mult ops */
249b23bfdefSJunchao Zhang #define DEF_Add(Type,BS,EQ)                                                       \
250cd620004SJunchao Zhang   DEF_UnpackAndOp    (Type,BS,EQ,Add, +,OP_BINARY)                                \
251cd620004SJunchao Zhang   DEF_UnpackAndOp    (Type,BS,EQ,Mult,*,OP_BINARY)                                \
252cd620004SJunchao Zhang   DEF_FetchAndOp     (Type,BS,EQ,Add, +,OP_BINARY)                                \
253cd620004SJunchao Zhang   DEF_ScatterAndOp   (Type,BS,EQ,Add, +,OP_BINARY)                                \
254cd620004SJunchao Zhang   DEF_ScatterAndOp   (Type,BS,EQ,Mult,*,OP_BINARY)                                \
255cd620004SJunchao Zhang   DEF_FetchAndOpLocal(Type,BS,EQ,Add, +,OP_BINARY)                                \
256cd620004SJunchao Zhang   static void CPPJoin4(PackInit_Add,Type,BS,EQ)(PetscSFLink link) {               \
257eb02082bSJunchao Zhang     link->h_UnpackAndAdd     = CPPJoin4(UnpackAndAdd,    Type,BS,EQ);             \
258eb02082bSJunchao Zhang     link->h_UnpackAndMult    = CPPJoin4(UnpackAndMult,   Type,BS,EQ);             \
259eb02082bSJunchao Zhang     link->h_FetchAndAdd      = CPPJoin4(FetchAndAdd,     Type,BS,EQ);             \
260cd620004SJunchao Zhang     link->h_ScatterAndAdd    = CPPJoin4(ScatterAndAdd,   Type,BS,EQ);             \
261cd620004SJunchao Zhang     link->h_ScatterAndMult   = CPPJoin4(ScatterAndMult,  Type,BS,EQ);             \
262cd620004SJunchao Zhang     link->h_FetchAndAddLocal = CPPJoin4(FetchAndAddLocal,Type,BS,EQ);             \
26340e23c03SJunchao Zhang   }
26440e23c03SJunchao Zhang 
265b23bfdefSJunchao Zhang /* Max, Min ops */
266b23bfdefSJunchao Zhang #define DEF_Cmp(Type,BS,EQ)                                                       \
267cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,Max,PetscMax,OP_FUNCTION)                           \
268cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,Min,PetscMin,OP_FUNCTION)                           \
269cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,Max,PetscMax,OP_FUNCTION)                           \
270cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,Min,PetscMin,OP_FUNCTION)                           \
271cd620004SJunchao Zhang   static void CPPJoin4(PackInit_Compare,Type,BS,EQ)(PetscSFLink link) {           \
272eb02082bSJunchao Zhang     link->h_UnpackAndMax    = CPPJoin4(UnpackAndMax,   Type,BS,EQ);               \
273eb02082bSJunchao Zhang     link->h_UnpackAndMin    = CPPJoin4(UnpackAndMin,   Type,BS,EQ);               \
274cd620004SJunchao Zhang     link->h_ScatterAndMax   = CPPJoin4(ScatterAndMax,  Type,BS,EQ);               \
275cd620004SJunchao Zhang     link->h_ScatterAndMin   = CPPJoin4(ScatterAndMin,  Type,BS,EQ);               \
276b23bfdefSJunchao Zhang   }
277b23bfdefSJunchao Zhang 
278b23bfdefSJunchao Zhang /* Logical ops.
279cd620004SJunchao Zhang   The operator in OP_LXOR should be empty but is ||. It is not used. Put here to avoid
28040e23c03SJunchao Zhang   the compilation warning "empty macro arguments are undefined in ISO C90"
28140e23c03SJunchao Zhang  */
282b23bfdefSJunchao Zhang #define DEF_Log(Type,BS,EQ)                                                       \
283cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,LAND,&&,OP_BINARY)                                  \
284cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,LOR, ||,OP_BINARY)                                  \
285cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,LXOR,||, OP_LXOR)                                   \
286cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,LAND,&&,OP_BINARY)                                  \
287cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,LOR, ||,OP_BINARY)                                  \
288cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,LXOR,||, OP_LXOR)                                   \
289cd620004SJunchao Zhang   static void CPPJoin4(PackInit_Logical,Type,BS,EQ)(PetscSFLink link) {           \
290eb02082bSJunchao Zhang     link->h_UnpackAndLAND   = CPPJoin4(UnpackAndLAND, Type,BS,EQ);                \
291eb02082bSJunchao Zhang     link->h_UnpackAndLOR    = CPPJoin4(UnpackAndLOR,  Type,BS,EQ);                \
292eb02082bSJunchao Zhang     link->h_UnpackAndLXOR   = CPPJoin4(UnpackAndLXOR, Type,BS,EQ);                \
293cd620004SJunchao Zhang     link->h_ScatterAndLAND  = CPPJoin4(ScatterAndLAND,Type,BS,EQ);                \
294cd620004SJunchao Zhang     link->h_ScatterAndLOR   = CPPJoin4(ScatterAndLOR, Type,BS,EQ);                \
295cd620004SJunchao Zhang     link->h_ScatterAndLXOR  = CPPJoin4(ScatterAndLXOR,Type,BS,EQ);                \
29640e23c03SJunchao Zhang   }
29740e23c03SJunchao Zhang 
298b23bfdefSJunchao Zhang /* Bitwise ops */
299b23bfdefSJunchao Zhang #define DEF_Bit(Type,BS,EQ)                                                       \
300cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,BAND,&,OP_BINARY)                                   \
301cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,BOR, |,OP_BINARY)                                   \
302cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,BXOR,^,OP_BINARY)                                   \
303cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,BAND,&,OP_BINARY)                                   \
304cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,BOR, |,OP_BINARY)                                   \
305cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,BXOR,^,OP_BINARY)                                   \
306cd620004SJunchao Zhang   static void CPPJoin4(PackInit_Bitwise,Type,BS,EQ)(PetscSFLink link) {           \
307eb02082bSJunchao Zhang     link->h_UnpackAndBAND   = CPPJoin4(UnpackAndBAND, Type,BS,EQ);                \
308eb02082bSJunchao Zhang     link->h_UnpackAndBOR    = CPPJoin4(UnpackAndBOR,  Type,BS,EQ);                \
309eb02082bSJunchao Zhang     link->h_UnpackAndBXOR   = CPPJoin4(UnpackAndBXOR, Type,BS,EQ);                \
310cd620004SJunchao Zhang     link->h_ScatterAndBAND  = CPPJoin4(ScatterAndBAND,Type,BS,EQ);                \
311cd620004SJunchao Zhang     link->h_ScatterAndBOR   = CPPJoin4(ScatterAndBOR, Type,BS,EQ);                \
312cd620004SJunchao Zhang     link->h_ScatterAndBXOR  = CPPJoin4(ScatterAndBXOR,Type,BS,EQ);                \
31340e23c03SJunchao Zhang   }
31440e23c03SJunchao Zhang 
315cd620004SJunchao Zhang /* Maxloc, Minloc ops */
316cd620004SJunchao Zhang #define DEF_Xloc(Type,BS,EQ)                                                      \
317cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,Max,>,OP_XLOC)                                      \
318cd620004SJunchao Zhang   DEF_UnpackAndOp (Type,BS,EQ,Min,<,OP_XLOC)                                      \
319cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,Max,>,OP_XLOC)                                      \
320cd620004SJunchao Zhang   DEF_ScatterAndOp(Type,BS,EQ,Min,<,OP_XLOC)                                      \
321cd620004SJunchao Zhang   static void CPPJoin4(PackInit_Xloc,Type,BS,EQ)(PetscSFLink link) {              \
322cd620004SJunchao Zhang     link->h_UnpackAndMaxloc  = CPPJoin4(UnpackAndMax, Type,BS,EQ);                \
323cd620004SJunchao Zhang     link->h_UnpackAndMinloc  = CPPJoin4(UnpackAndMin, Type,BS,EQ);                \
324cd620004SJunchao Zhang     link->h_ScatterAndMaxloc = CPPJoin4(ScatterAndMax,Type,BS,EQ);                \
325cd620004SJunchao Zhang     link->h_ScatterAndMinloc = CPPJoin4(ScatterAndMin,Type,BS,EQ);                \
32640e23c03SJunchao Zhang   }
32740e23c03SJunchao Zhang 
328b23bfdefSJunchao Zhang #define DEF_IntegerType(Type,BS,EQ)                                               \
329b23bfdefSJunchao Zhang   DEF_Pack(Type,BS,EQ)                                                            \
330b23bfdefSJunchao Zhang   DEF_Add(Type,BS,EQ)                                                             \
331b23bfdefSJunchao Zhang   DEF_Cmp(Type,BS,EQ)                                                             \
332b23bfdefSJunchao Zhang   DEF_Log(Type,BS,EQ)                                                             \
333b23bfdefSJunchao Zhang   DEF_Bit(Type,BS,EQ)                                                             \
334cd620004SJunchao Zhang   static void CPPJoin4(PackInit_IntegerType,Type,BS,EQ)(PetscSFLink link) {       \
335b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Pack,Type,BS,EQ)(link);                                     \
336b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Add,Type,BS,EQ)(link);                                      \
337b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Compare,Type,BS,EQ)(link);                                  \
338b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Logical,Type,BS,EQ)(link);                                  \
339b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Bitwise,Type,BS,EQ)(link);                                  \
34040e23c03SJunchao Zhang   }
34140e23c03SJunchao Zhang 
342b23bfdefSJunchao Zhang #define DEF_RealType(Type,BS,EQ)                                                  \
343b23bfdefSJunchao Zhang   DEF_Pack(Type,BS,EQ)                                                            \
344b23bfdefSJunchao Zhang   DEF_Add(Type,BS,EQ)                                                             \
345b23bfdefSJunchao Zhang   DEF_Cmp(Type,BS,EQ)                                                             \
346cd620004SJunchao Zhang   static void CPPJoin4(PackInit_RealType,Type,BS,EQ)(PetscSFLink link) {          \
347b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Pack,Type,BS,EQ)(link);                                     \
348b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Add,Type,BS,EQ)(link);                                      \
349b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Compare,Type,BS,EQ)(link);                                  \
350b23bfdefSJunchao Zhang   }
35140e23c03SJunchao Zhang 
35240e23c03SJunchao Zhang #if defined(PETSC_HAVE_COMPLEX)
353b23bfdefSJunchao Zhang #define DEF_ComplexType(Type,BS,EQ)                                               \
354b23bfdefSJunchao Zhang   DEF_Pack(Type,BS,EQ)                                                            \
355b23bfdefSJunchao Zhang   DEF_Add(Type,BS,EQ)                                                             \
356cd620004SJunchao Zhang   static void CPPJoin4(PackInit_ComplexType,Type,BS,EQ)(PetscSFLink link) {       \
357b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Pack,Type,BS,EQ)(link);                                     \
358b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Add,Type,BS,EQ)(link);                                      \
359b23bfdefSJunchao Zhang   }
36040e23c03SJunchao Zhang #endif
361b23bfdefSJunchao Zhang 
362b23bfdefSJunchao Zhang #define DEF_DumbType(Type,BS,EQ)                                                  \
363b23bfdefSJunchao Zhang   DEF_Pack(Type,BS,EQ)                                                            \
364cd620004SJunchao Zhang   static void CPPJoin4(PackInit_DumbType,Type,BS,EQ)(PetscSFLink link) {          \
365b23bfdefSJunchao Zhang     CPPJoin4(PackInit_Pack,Type,BS,EQ)(link);                                     \
366b23bfdefSJunchao Zhang   }
367b23bfdefSJunchao Zhang 
368b23bfdefSJunchao Zhang /* Maxloc, Minloc */
369cd620004SJunchao Zhang #define DEF_PairType(Type,BS,EQ)                                                  \
370cd620004SJunchao Zhang   DEF_Pack(Type,BS,EQ)                                                            \
371cd620004SJunchao Zhang   DEF_Xloc(Type,BS,EQ)                                                            \
372cd620004SJunchao Zhang   static void CPPJoin4(PackInit_PairType,Type,BS,EQ)(PetscSFLink link) {          \
373cd620004SJunchao Zhang     CPPJoin4(PackInit_Pack,Type,BS,EQ)(link);                                     \
374cd620004SJunchao Zhang     CPPJoin4(PackInit_Xloc,Type,BS,EQ)(link);                                     \
375b23bfdefSJunchao Zhang   }
376b23bfdefSJunchao Zhang 
377b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,1,1) /* unit = 1 MPIU_INT  */
378b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,2,1) /* unit = 2 MPIU_INTs */
379b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,4,1) /* unit = 4 MPIU_INTs */
380b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,8,1) /* unit = 8 MPIU_INTs */
381b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,1,0) /* unit = 1*n MPIU_INTs, n>1 */
382b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,2,0) /* unit = 2*n MPIU_INTs, n>1 */
383b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,4,0) /* unit = 4*n MPIU_INTs, n>1 */
384b23bfdefSJunchao Zhang DEF_IntegerType(PetscInt,8,0) /* unit = 8*n MPIU_INTs, n>1. Routines with bigger BS are tried first. */
385b23bfdefSJunchao Zhang 
386b23bfdefSJunchao Zhang #if defined(PETSC_USE_64BIT_INDICES) /* Do not need (though it is OK) to generate redundant functions if PetscInt is int */
387b23bfdefSJunchao Zhang DEF_IntegerType(int,1,1)
388b23bfdefSJunchao Zhang DEF_IntegerType(int,2,1)
389b23bfdefSJunchao Zhang DEF_IntegerType(int,4,1)
390b23bfdefSJunchao Zhang DEF_IntegerType(int,8,1)
391b23bfdefSJunchao Zhang DEF_IntegerType(int,1,0)
392b23bfdefSJunchao Zhang DEF_IntegerType(int,2,0)
393b23bfdefSJunchao Zhang DEF_IntegerType(int,4,0)
394b23bfdefSJunchao Zhang DEF_IntegerType(int,8,0)
395b23bfdefSJunchao Zhang #endif
396b23bfdefSJunchao Zhang 
397b23bfdefSJunchao Zhang /* The typedefs are used to get a typename without space that CPPJoin can handle */
398b23bfdefSJunchao Zhang typedef signed char SignedChar;
399b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,1,1)
400b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,2,1)
401b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,4,1)
402b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,8,1)
403b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,1,0)
404b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,2,0)
405b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,4,0)
406b23bfdefSJunchao Zhang DEF_IntegerType(SignedChar,8,0)
407b23bfdefSJunchao Zhang 
408b23bfdefSJunchao Zhang typedef unsigned char UnsignedChar;
409b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,1,1)
410b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,2,1)
411b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,4,1)
412b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,8,1)
413b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,1,0)
414b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,2,0)
415b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,4,0)
416b23bfdefSJunchao Zhang DEF_IntegerType(UnsignedChar,8,0)
417b23bfdefSJunchao Zhang 
418b23bfdefSJunchao Zhang DEF_RealType(PetscReal,1,1)
419b23bfdefSJunchao Zhang DEF_RealType(PetscReal,2,1)
420b23bfdefSJunchao Zhang DEF_RealType(PetscReal,4,1)
421b23bfdefSJunchao Zhang DEF_RealType(PetscReal,8,1)
422b23bfdefSJunchao Zhang DEF_RealType(PetscReal,1,0)
423b23bfdefSJunchao Zhang DEF_RealType(PetscReal,2,0)
424b23bfdefSJunchao Zhang DEF_RealType(PetscReal,4,0)
425b23bfdefSJunchao Zhang DEF_RealType(PetscReal,8,0)
426b23bfdefSJunchao Zhang 
427b23bfdefSJunchao Zhang #if defined(PETSC_HAVE_COMPLEX)
428b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,1,1)
429b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,2,1)
430b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,4,1)
431b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,8,1)
432b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,1,0)
433b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,2,0)
434b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,4,0)
435b23bfdefSJunchao Zhang DEF_ComplexType(PetscComplex,8,0)
436b23bfdefSJunchao Zhang #endif
437b23bfdefSJunchao Zhang 
438cd620004SJunchao Zhang #define PairType(Type1,Type2) Type1##_##Type2
439cd620004SJunchao Zhang typedef struct {int u; int i;}           PairType(int,int);
440cd620004SJunchao Zhang typedef struct {PetscInt u; PetscInt i;} PairType(PetscInt,PetscInt);
441cd620004SJunchao Zhang DEF_PairType(PairType(int,int),1,1)
442cd620004SJunchao Zhang DEF_PairType(PairType(PetscInt,PetscInt),1,1)
443b23bfdefSJunchao Zhang 
444b23bfdefSJunchao Zhang /* If we don't know the basic type, we treat it as a stream of chars or ints */
445b23bfdefSJunchao Zhang DEF_DumbType(char,1,1)
446b23bfdefSJunchao Zhang DEF_DumbType(char,2,1)
447b23bfdefSJunchao Zhang DEF_DumbType(char,4,1)
448b23bfdefSJunchao Zhang DEF_DumbType(char,1,0)
449b23bfdefSJunchao Zhang DEF_DumbType(char,2,0)
450b23bfdefSJunchao Zhang DEF_DumbType(char,4,0)
451b23bfdefSJunchao Zhang 
452eb02082bSJunchao Zhang typedef int DumbInt; /* To have a different name than 'int' used above. The name is used to make routine names. */
453b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,1,1)
454b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,2,1)
455b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,4,1)
456b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,8,1)
457b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,1,0)
458b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,2,0)
459b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,4,0)
460b23bfdefSJunchao Zhang DEF_DumbType(DumbInt,8,0)
46140e23c03SJunchao Zhang 
46240e23c03SJunchao Zhang #if !defined(PETSC_HAVE_MPI_TYPE_DUP)
4639fbee547SJacob Faibussowitsch static inline int MPI_Type_dup(MPI_Datatype datatype,MPI_Datatype *newtype)
46440e23c03SJunchao Zhang {
46540e23c03SJunchao Zhang   int ierr;
46640e23c03SJunchao Zhang   ierr = MPI_Type_contiguous(1,datatype,newtype); if (ierr) return ierr;
46740e23c03SJunchao Zhang   ierr = MPI_Type_commit(newtype); if (ierr) return ierr;
46840e23c03SJunchao Zhang   return MPI_SUCCESS;
46940e23c03SJunchao Zhang }
47040e23c03SJunchao Zhang #endif
47140e23c03SJunchao Zhang 
47271438e86SJunchao Zhang PetscErrorCode PetscSFLinkDestroy(PetscSF sf,PetscSFLink link)
47340e23c03SJunchao Zhang {
47440e23c03SJunchao Zhang   PetscErrorCode    ierr;
475cd620004SJunchao Zhang   PetscSF_Basic     *bas = (PetscSF_Basic*)sf->data;
47671438e86SJunchao Zhang   PetscInt          i,nreqs = (bas->nrootreqs+sf->nleafreqs)*8;
47771438e86SJunchao Zhang 
47871438e86SJunchao Zhang   PetscFunctionBegin;
47971438e86SJunchao Zhang   /* Destroy device-specific fields */
48071438e86SJunchao Zhang   if (link->deviceinited) {ierr = (*link->Destroy)(sf,link);CHKERRQ(ierr);}
48171438e86SJunchao Zhang 
48271438e86SJunchao Zhang   /* Destroy host related fields */
48371438e86SJunchao Zhang   if (!link->isbuiltin) {ierr = MPI_Type_free(&link->unit);CHKERRMPI(ierr);}
48471438e86SJunchao Zhang   if (!link->use_nvshmem) {
48571438e86SJunchao Zhang     for (i=0; i<nreqs; i++) { /* Persistent reqs must be freed. */
48671438e86SJunchao Zhang       if (link->reqs[i] != MPI_REQUEST_NULL) {ierr = MPI_Request_free(&link->reqs[i]);CHKERRMPI(ierr);}
48771438e86SJunchao Zhang     }
48871438e86SJunchao Zhang     ierr = PetscFree(link->reqs);CHKERRQ(ierr);
48971438e86SJunchao Zhang     for (i=PETSCSF_LOCAL; i<=PETSCSF_REMOTE; i++) {
49071438e86SJunchao Zhang       ierr = PetscFree(link->rootbuf_alloc[i][PETSC_MEMTYPE_HOST]);CHKERRQ(ierr);
49171438e86SJunchao Zhang       ierr = PetscFree(link->leafbuf_alloc[i][PETSC_MEMTYPE_HOST]);CHKERRQ(ierr);
49271438e86SJunchao Zhang     }
49371438e86SJunchao Zhang   }
49471438e86SJunchao Zhang   ierr = PetscFree(link);CHKERRQ(ierr);
49571438e86SJunchao Zhang   PetscFunctionReturn(0);
49671438e86SJunchao Zhang }
49771438e86SJunchao Zhang 
49871438e86SJunchao Zhang PetscErrorCode PetscSFLinkCreate(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,const void *leafdata,MPI_Op op,PetscSFOperation sfop,PetscSFLink *mylink)
49971438e86SJunchao Zhang {
50071438e86SJunchao Zhang   PetscErrorCode    ierr;
501cd620004SJunchao Zhang 
502cd620004SJunchao Zhang   PetscFunctionBegin;
503cd620004SJunchao Zhang   ierr = PetscSFSetErrorOnUnsupportedOverlap(sf,unit,rootdata,leafdata);CHKERRQ(ierr);
50471438e86SJunchao Zhang  #if defined(PETSC_HAVE_NVSHMEM)
50571438e86SJunchao Zhang   {
50671438e86SJunchao Zhang     PetscBool use_nvshmem;
50771438e86SJunchao Zhang     ierr = PetscSFLinkNvshmemCheck(sf,rootmtype,rootdata,leafmtype,leafdata,&use_nvshmem);CHKERRQ(ierr);
50871438e86SJunchao Zhang     if (use_nvshmem) {
50971438e86SJunchao Zhang       ierr = PetscSFLinkCreate_NVSHMEM(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,sfop,mylink);CHKERRQ(ierr);
51071438e86SJunchao Zhang       PetscFunctionReturn(0);
511cd620004SJunchao Zhang     }
512cd620004SJunchao Zhang   }
5137fd2d3dbSJunchao Zhang  #endif
51471438e86SJunchao Zhang   ierr = PetscSFLinkCreate_MPI(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,sfop,mylink);CHKERRQ(ierr);
515cd620004SJunchao Zhang   PetscFunctionReturn(0);
516cd620004SJunchao Zhang }
517cd620004SJunchao Zhang 
518cd620004SJunchao Zhang /* Return root/leaf buffers and MPI requests attached to the link for MPI communication in the given direction.
519cd620004SJunchao Zhang    If the sf uses persistent requests and the requests have not been initialized, then initialize them.
520cd620004SJunchao Zhang */
521cd620004SJunchao Zhang PetscErrorCode PetscSFLinkGetMPIBuffersAndRequests(PetscSF sf,PetscSFLink link,PetscSFDirection direction,void **rootbuf, void **leafbuf,MPI_Request **rootreqs,MPI_Request **leafreqs)
522cd620004SJunchao Zhang {
523cd620004SJunchao Zhang   PetscErrorCode       ierr;
524cd620004SJunchao Zhang   PetscSF_Basic        *bas = (PetscSF_Basic*)sf->data;
525c87b50c4SJunchao Zhang   PetscInt             i,j,cnt,nrootranks,ndrootranks,nleafranks,ndleafranks;
526cd620004SJunchao Zhang   const PetscInt       *rootoffset,*leafoffset;
527cd620004SJunchao Zhang   MPI_Aint             disp;
528cd620004SJunchao Zhang   MPI_Comm             comm = PetscObjectComm((PetscObject)sf);
529cd620004SJunchao Zhang   MPI_Datatype         unit = link->unit;
530cd620004SJunchao Zhang   const PetscMemType   rootmtype_mpi = link->rootmtype_mpi,leafmtype_mpi = link->leafmtype_mpi; /* Used to select buffers passed to MPI */
531cd620004SJunchao Zhang   const PetscInt       rootdirect_mpi = link->rootdirect_mpi,leafdirect_mpi = link->leafdirect_mpi;
532cd620004SJunchao Zhang 
533cd620004SJunchao Zhang   PetscFunctionBegin;
534cd620004SJunchao Zhang   /* Init persistent MPI requests if not yet. Currently only SFBasic uses persistent MPI */
535cd620004SJunchao Zhang   if (sf->persistent) {
536cd620004SJunchao Zhang     if (rootreqs && bas->rootbuflen[PETSCSF_REMOTE] && !link->rootreqsinited[direction][rootmtype_mpi][rootdirect_mpi]) {
537cd620004SJunchao Zhang       ierr = PetscSFGetRootInfo_Basic(sf,&nrootranks,&ndrootranks,NULL,&rootoffset,NULL);CHKERRQ(ierr);
538cd620004SJunchao Zhang       if (direction == PETSCSF_LEAF2ROOT) {
539cd620004SJunchao Zhang         for (i=ndrootranks,j=0; i<nrootranks; i++,j++) {
540cd620004SJunchao Zhang           disp = (rootoffset[i] - rootoffset[ndrootranks])*link->unitbytes;
541c87b50c4SJunchao Zhang           cnt  = rootoffset[i+1]-rootoffset[i];
542c87b50c4SJunchao Zhang           ierr = MPIU_Recv_init(link->rootbuf[PETSCSF_REMOTE][rootmtype_mpi]+disp,cnt,unit,bas->iranks[i],link->tag,comm,link->rootreqs[direction][rootmtype_mpi][rootdirect_mpi]+j);CHKERRMPI(ierr);
543cd620004SJunchao Zhang         }
544cd620004SJunchao Zhang       } else { /* PETSCSF_ROOT2LEAF */
545cd620004SJunchao Zhang         for (i=ndrootranks,j=0; i<nrootranks; i++,j++) {
546cd620004SJunchao Zhang           disp = (rootoffset[i] - rootoffset[ndrootranks])*link->unitbytes;
547c87b50c4SJunchao Zhang           cnt  = rootoffset[i+1]-rootoffset[i];
548c87b50c4SJunchao Zhang           ierr = MPIU_Send_init(link->rootbuf[PETSCSF_REMOTE][rootmtype_mpi]+disp,cnt,unit,bas->iranks[i],link->tag,comm,link->rootreqs[direction][rootmtype_mpi][rootdirect_mpi]+j);CHKERRMPI(ierr);
549cd620004SJunchao Zhang         }
550cd620004SJunchao Zhang       }
551cd620004SJunchao Zhang       link->rootreqsinited[direction][rootmtype_mpi][rootdirect_mpi] = PETSC_TRUE;
552cd620004SJunchao Zhang     }
553cd620004SJunchao Zhang 
554cd620004SJunchao Zhang     if (leafreqs && sf->leafbuflen[PETSCSF_REMOTE] && !link->leafreqsinited[direction][leafmtype_mpi][leafdirect_mpi]) {
555cd620004SJunchao Zhang       ierr = PetscSFGetLeafInfo_Basic(sf,&nleafranks,&ndleafranks,NULL,&leafoffset,NULL,NULL);CHKERRQ(ierr);
556cd620004SJunchao Zhang       if (direction == PETSCSF_LEAF2ROOT) {
557cd620004SJunchao Zhang         for (i=ndleafranks,j=0; i<nleafranks; i++,j++) {
558cd620004SJunchao Zhang           disp = (leafoffset[i] - leafoffset[ndleafranks])*link->unitbytes;
559c87b50c4SJunchao Zhang           cnt  = leafoffset[i+1]-leafoffset[i];
560c87b50c4SJunchao Zhang           ierr = MPIU_Send_init(link->leafbuf[PETSCSF_REMOTE][leafmtype_mpi]+disp,cnt,unit,sf->ranks[i],link->tag,comm,link->leafreqs[direction][leafmtype_mpi][leafdirect_mpi]+j);CHKERRMPI(ierr);
561cd620004SJunchao Zhang         }
562cd620004SJunchao Zhang       } else { /* PETSCSF_ROOT2LEAF */
563cd620004SJunchao Zhang         for (i=ndleafranks,j=0; i<nleafranks; i++,j++) {
564cd620004SJunchao Zhang           disp = (leafoffset[i] - leafoffset[ndleafranks])*link->unitbytes;
565c87b50c4SJunchao Zhang           cnt  = leafoffset[i+1]-leafoffset[i];
566c87b50c4SJunchao Zhang           ierr = MPIU_Recv_init(link->leafbuf[PETSCSF_REMOTE][leafmtype_mpi]+disp,cnt,unit,sf->ranks[i],link->tag,comm,link->leafreqs[direction][leafmtype_mpi][leafdirect_mpi]+j);CHKERRMPI(ierr);
567cd620004SJunchao Zhang         }
568cd620004SJunchao Zhang       }
569cd620004SJunchao Zhang       link->leafreqsinited[direction][leafmtype_mpi][leafdirect_mpi] = PETSC_TRUE;
570cd620004SJunchao Zhang     }
571cd620004SJunchao Zhang   }
572cd620004SJunchao Zhang   if (rootbuf)  *rootbuf  = link->rootbuf[PETSCSF_REMOTE][rootmtype_mpi];
573cd620004SJunchao Zhang   if (leafbuf)  *leafbuf  = link->leafbuf[PETSCSF_REMOTE][leafmtype_mpi];
574cd620004SJunchao Zhang   if (rootreqs) *rootreqs = link->rootreqs[direction][rootmtype_mpi][rootdirect_mpi];
575cd620004SJunchao Zhang   if (leafreqs) *leafreqs = link->leafreqs[direction][leafmtype_mpi][leafdirect_mpi];
576cd620004SJunchao Zhang   PetscFunctionReturn(0);
577cd620004SJunchao Zhang }
578cd620004SJunchao Zhang 
579cd620004SJunchao Zhang PetscErrorCode PetscSFLinkGetInUse(PetscSF sf,MPI_Datatype unit,const void *rootdata,const void *leafdata,PetscCopyMode cmode,PetscSFLink *mylink)
580cd620004SJunchao Zhang {
581cd620004SJunchao Zhang   PetscErrorCode    ierr;
582cd620004SJunchao Zhang   PetscSFLink       link,*p;
58340e23c03SJunchao Zhang   PetscSF_Basic     *bas=(PetscSF_Basic*)sf->data;
58440e23c03SJunchao Zhang 
58540e23c03SJunchao Zhang   PetscFunctionBegin;
58640e23c03SJunchao Zhang   /* Look for types in cache */
58740e23c03SJunchao Zhang   for (p=&bas->inuse; (link=*p); p=&link->next) {
58840e23c03SJunchao Zhang     PetscBool match;
58940e23c03SJunchao Zhang     ierr = MPIPetsc_Type_compare(unit,link->unit,&match);CHKERRQ(ierr);
590637e6665SJunchao Zhang     if (match && (rootdata == link->rootdata) && (leafdata == link->leafdata)) {
59140e23c03SJunchao Zhang       switch (cmode) {
59240e23c03SJunchao Zhang       case PETSC_OWN_POINTER: *p = link->next; break; /* Remove from inuse list */
59340e23c03SJunchao Zhang       case PETSC_USE_POINTER: break;
59440e23c03SJunchao Zhang       default: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"invalid cmode");
59540e23c03SJunchao Zhang       }
59640e23c03SJunchao Zhang       *mylink = link;
59740e23c03SJunchao Zhang       PetscFunctionReturn(0);
59840e23c03SJunchao Zhang     }
59940e23c03SJunchao Zhang   }
60040e23c03SJunchao Zhang   SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Could not find pack");
60140e23c03SJunchao Zhang }
60240e23c03SJunchao Zhang 
60371438e86SJunchao Zhang PetscErrorCode PetscSFLinkReclaim(PetscSF sf,PetscSFLink *mylink)
60440e23c03SJunchao Zhang {
60540e23c03SJunchao Zhang   PetscSF_Basic     *bas = (PetscSF_Basic*)sf->data;
60671438e86SJunchao Zhang   PetscSFLink       link = *mylink;
60740e23c03SJunchao Zhang 
60840e23c03SJunchao Zhang   PetscFunctionBegin;
60971438e86SJunchao Zhang   link->rootdata = NULL;
61071438e86SJunchao Zhang   link->leafdata = NULL;
61171438e86SJunchao Zhang   link->next     = bas->avail;
61271438e86SJunchao Zhang   bas->avail     = link;
61371438e86SJunchao Zhang   *mylink        = NULL;
614eb02082bSJunchao Zhang   PetscFunctionReturn(0);
615eb02082bSJunchao Zhang }
616eb02082bSJunchao Zhang 
6179d1c8addSJunchao Zhang /* Error out on unsupported overlapped communications */
618cd620004SJunchao Zhang PetscErrorCode PetscSFSetErrorOnUnsupportedOverlap(PetscSF sf,MPI_Datatype unit,const void *rootdata,const void *leafdata)
6199d1c8addSJunchao Zhang {
6209d1c8addSJunchao Zhang   PetscErrorCode    ierr;
621cd620004SJunchao Zhang   PetscSFLink       link,*p;
6229d1c8addSJunchao Zhang   PetscSF_Basic     *bas = (PetscSF_Basic*)sf->data;
6239d1c8addSJunchao Zhang   PetscBool         match;
6249d1c8addSJunchao Zhang 
6259d1c8addSJunchao Zhang   PetscFunctionBegin;
626b458e8f1SJose E. Roman   if (PetscDefined(USE_DEBUG)) {
62718fb5014SJunchao Zhang     /* Look up links in use and error out if there is a match. When both rootdata and leafdata are NULL, ignore
62818fb5014SJunchao Zhang        the potential overlapping since this process does not participate in communication. Overlapping is harmless.
62918fb5014SJunchao Zhang     */
630637e6665SJunchao Zhang     if (rootdata || leafdata) {
6319d1c8addSJunchao Zhang       for (p=&bas->inuse; (link=*p); p=&link->next) {
6329d1c8addSJunchao Zhang         ierr = MPIPetsc_Type_compare(unit,link->unit,&match);CHKERRQ(ierr);
633*2c71b3e2SJacob Faibussowitsch         PetscCheckFalse(match && (rootdata == link->rootdata) && (leafdata == link->leafdata),PETSC_COMM_SELF,PETSC_ERR_SUP,"Overlapped PetscSF with the same rootdata(%p), leafdata(%p) and data type. Undo the overlapping to avoid the error.",rootdata,leafdata);
6349d1c8addSJunchao Zhang       }
63518fb5014SJunchao Zhang     }
636b458e8f1SJose E. Roman   }
6379d1c8addSJunchao Zhang   PetscFunctionReturn(0);
6389d1c8addSJunchao Zhang }
6399d1c8addSJunchao Zhang 
64020c24465SJunchao Zhang static PetscErrorCode PetscSFLinkMemcpy_Host(PetscSFLink link,PetscMemType dstmtype,void* dst,PetscMemType srcmtype,const void*src,size_t n)
64120c24465SJunchao Zhang {
64220c24465SJunchao Zhang   PetscFunctionBegin;
64320c24465SJunchao Zhang   if (n) {PetscErrorCode ierr = PetscMemcpy(dst,src,n);CHKERRQ(ierr);}
64420c24465SJunchao Zhang   PetscFunctionReturn(0);
64520c24465SJunchao Zhang }
64620c24465SJunchao Zhang 
647cd620004SJunchao Zhang PetscErrorCode PetscSFLinkSetUp_Host(PetscSF sf,PetscSFLink link,MPI_Datatype unit)
64840e23c03SJunchao Zhang {
64940e23c03SJunchao Zhang   PetscErrorCode ierr;
650b23bfdefSJunchao Zhang   PetscInt       nSignedChar=0,nUnsignedChar=0,nInt=0,nPetscInt=0,nPetscReal=0;
651b23bfdefSJunchao Zhang   PetscBool      is2Int,is2PetscInt;
65240e23c03SJunchao Zhang   PetscMPIInt    ni,na,nd,combiner;
65340e23c03SJunchao Zhang #if defined(PETSC_HAVE_COMPLEX)
654b23bfdefSJunchao Zhang   PetscInt       nPetscComplex=0;
65540e23c03SJunchao Zhang #endif
65640e23c03SJunchao Zhang 
65740e23c03SJunchao Zhang   PetscFunctionBegin;
658b23bfdefSJunchao Zhang   ierr = MPIPetsc_Type_compare_contig(unit,MPI_SIGNED_CHAR,  &nSignedChar);CHKERRQ(ierr);
659b23bfdefSJunchao Zhang   ierr = MPIPetsc_Type_compare_contig(unit,MPI_UNSIGNED_CHAR,&nUnsignedChar);CHKERRQ(ierr);
660b23bfdefSJunchao Zhang   /* MPI_CHAR is treated below as a dumb type that does not support reduction according to MPI standard */
661b23bfdefSJunchao Zhang   ierr = MPIPetsc_Type_compare_contig(unit,MPI_INT,  &nInt);CHKERRQ(ierr);
662b23bfdefSJunchao Zhang   ierr = MPIPetsc_Type_compare_contig(unit,MPIU_INT, &nPetscInt);CHKERRQ(ierr);
663b23bfdefSJunchao Zhang   ierr = MPIPetsc_Type_compare_contig(unit,MPIU_REAL,&nPetscReal);CHKERRQ(ierr);
66440e23c03SJunchao Zhang #if defined(PETSC_HAVE_COMPLEX)
665b23bfdefSJunchao Zhang   ierr = MPIPetsc_Type_compare_contig(unit,MPIU_COMPLEX,&nPetscComplex);CHKERRQ(ierr);
66640e23c03SJunchao Zhang #endif
66740e23c03SJunchao Zhang   ierr = MPIPetsc_Type_compare(unit,MPI_2INT,&is2Int);CHKERRQ(ierr);
66840e23c03SJunchao Zhang   ierr = MPIPetsc_Type_compare(unit,MPIU_2INT,&is2PetscInt);CHKERRQ(ierr);
669b23bfdefSJunchao Zhang   /* TODO: shaell we also handle Fortran MPI_2REAL? */
670ffc4695bSBarry Smith   ierr = MPI_Type_get_envelope(unit,&ni,&na,&nd,&combiner);CHKERRMPI(ierr);
6715ad15460SJunchao Zhang   link->isbuiltin = (combiner == MPI_COMBINER_NAMED) ? PETSC_TRUE : PETSC_FALSE; /* unit is MPI builtin */
672b23bfdefSJunchao Zhang   link->bs = 1; /* default */
67340e23c03SJunchao Zhang 
674eb02082bSJunchao Zhang   if (is2Int) {
675cd620004SJunchao Zhang     PackInit_PairType_int_int_1_1(link);
676eb02082bSJunchao Zhang     link->bs        = 1;
677eb02082bSJunchao Zhang     link->unitbytes = 2*sizeof(int);
6785ad15460SJunchao Zhang     link->isbuiltin = PETSC_TRUE; /* unit is PETSc builtin */
679eb02082bSJunchao Zhang     link->basicunit = MPI_2INT;
6805ad15460SJunchao Zhang     link->unit      = MPI_2INT;
681eb02082bSJunchao Zhang   } else if (is2PetscInt) { /* TODO: when is2PetscInt and nPetscInt=2, we don't know which path to take. The two paths support different ops. */
682cd620004SJunchao Zhang     PackInit_PairType_PetscInt_PetscInt_1_1(link);
683eb02082bSJunchao Zhang     link->bs        = 1;
684eb02082bSJunchao Zhang     link->unitbytes = 2*sizeof(PetscInt);
685eb02082bSJunchao Zhang     link->basicunit = MPIU_2INT;
6865ad15460SJunchao Zhang     link->isbuiltin = PETSC_TRUE; /* unit is PETSc builtin */
6875ad15460SJunchao Zhang     link->unit      = MPIU_2INT;
688eb02082bSJunchao Zhang   } else if (nPetscReal) {
689b23bfdefSJunchao Zhang     if      (nPetscReal == 8) PackInit_RealType_PetscReal_8_1(link); else if (nPetscReal%8 == 0) PackInit_RealType_PetscReal_8_0(link);
690b23bfdefSJunchao Zhang     else if (nPetscReal == 4) PackInit_RealType_PetscReal_4_1(link); else if (nPetscReal%4 == 0) PackInit_RealType_PetscReal_4_0(link);
691b23bfdefSJunchao Zhang     else if (nPetscReal == 2) PackInit_RealType_PetscReal_2_1(link); else if (nPetscReal%2 == 0) PackInit_RealType_PetscReal_2_0(link);
692b23bfdefSJunchao Zhang     else if (nPetscReal == 1) PackInit_RealType_PetscReal_1_1(link); else if (nPetscReal%1 == 0) PackInit_RealType_PetscReal_1_0(link);
693b23bfdefSJunchao Zhang     link->bs        = nPetscReal;
694eb02082bSJunchao Zhang     link->unitbytes = nPetscReal*sizeof(PetscReal);
69540e23c03SJunchao Zhang     link->basicunit = MPIU_REAL;
6965ad15460SJunchao Zhang     if (link->bs == 1) {link->isbuiltin = PETSC_TRUE; link->unit = MPIU_REAL;}
697b23bfdefSJunchao Zhang   } else if (nPetscInt) {
698b23bfdefSJunchao Zhang     if      (nPetscInt == 8) PackInit_IntegerType_PetscInt_8_1(link); else if (nPetscInt%8 == 0) PackInit_IntegerType_PetscInt_8_0(link);
699b23bfdefSJunchao Zhang     else if (nPetscInt == 4) PackInit_IntegerType_PetscInt_4_1(link); else if (nPetscInt%4 == 0) PackInit_IntegerType_PetscInt_4_0(link);
700b23bfdefSJunchao Zhang     else if (nPetscInt == 2) PackInit_IntegerType_PetscInt_2_1(link); else if (nPetscInt%2 == 0) PackInit_IntegerType_PetscInt_2_0(link);
701b23bfdefSJunchao Zhang     else if (nPetscInt == 1) PackInit_IntegerType_PetscInt_1_1(link); else if (nPetscInt%1 == 0) PackInit_IntegerType_PetscInt_1_0(link);
702b23bfdefSJunchao Zhang     link->bs        = nPetscInt;
703eb02082bSJunchao Zhang     link->unitbytes = nPetscInt*sizeof(PetscInt);
704b23bfdefSJunchao Zhang     link->basicunit = MPIU_INT;
7055ad15460SJunchao Zhang     if (link->bs == 1) {link->isbuiltin = PETSC_TRUE; link->unit = MPIU_INT;}
706b23bfdefSJunchao Zhang #if defined(PETSC_USE_64BIT_INDICES)
707b23bfdefSJunchao Zhang   } else if (nInt) {
708b23bfdefSJunchao Zhang     if      (nInt == 8) PackInit_IntegerType_int_8_1(link); else if (nInt%8 == 0) PackInit_IntegerType_int_8_0(link);
709b23bfdefSJunchao Zhang     else if (nInt == 4) PackInit_IntegerType_int_4_1(link); else if (nInt%4 == 0) PackInit_IntegerType_int_4_0(link);
710b23bfdefSJunchao Zhang     else if (nInt == 2) PackInit_IntegerType_int_2_1(link); else if (nInt%2 == 0) PackInit_IntegerType_int_2_0(link);
711b23bfdefSJunchao Zhang     else if (nInt == 1) PackInit_IntegerType_int_1_1(link); else if (nInt%1 == 0) PackInit_IntegerType_int_1_0(link);
712b23bfdefSJunchao Zhang     link->bs        = nInt;
713eb02082bSJunchao Zhang     link->unitbytes = nInt*sizeof(int);
714b23bfdefSJunchao Zhang     link->basicunit = MPI_INT;
7155ad15460SJunchao Zhang     if (link->bs == 1) {link->isbuiltin = PETSC_TRUE; link->unit = MPI_INT;}
716b23bfdefSJunchao Zhang #endif
717b23bfdefSJunchao Zhang   } else if (nSignedChar) {
718b23bfdefSJunchao Zhang     if      (nSignedChar == 8) PackInit_IntegerType_SignedChar_8_1(link); else if (nSignedChar%8 == 0) PackInit_IntegerType_SignedChar_8_0(link);
719b23bfdefSJunchao Zhang     else if (nSignedChar == 4) PackInit_IntegerType_SignedChar_4_1(link); else if (nSignedChar%4 == 0) PackInit_IntegerType_SignedChar_4_0(link);
720b23bfdefSJunchao Zhang     else if (nSignedChar == 2) PackInit_IntegerType_SignedChar_2_1(link); else if (nSignedChar%2 == 0) PackInit_IntegerType_SignedChar_2_0(link);
721b23bfdefSJunchao Zhang     else if (nSignedChar == 1) PackInit_IntegerType_SignedChar_1_1(link); else if (nSignedChar%1 == 0) PackInit_IntegerType_SignedChar_1_0(link);
722b23bfdefSJunchao Zhang     link->bs        = nSignedChar;
723eb02082bSJunchao Zhang     link->unitbytes = nSignedChar*sizeof(SignedChar);
724b23bfdefSJunchao Zhang     link->basicunit = MPI_SIGNED_CHAR;
7255ad15460SJunchao Zhang     if (link->bs == 1) {link->isbuiltin = PETSC_TRUE; link->unit = MPI_SIGNED_CHAR;}
726b23bfdefSJunchao Zhang   }  else if (nUnsignedChar) {
727b23bfdefSJunchao Zhang     if      (nUnsignedChar == 8) PackInit_IntegerType_UnsignedChar_8_1(link); else if (nUnsignedChar%8 == 0) PackInit_IntegerType_UnsignedChar_8_0(link);
728b23bfdefSJunchao Zhang     else if (nUnsignedChar == 4) PackInit_IntegerType_UnsignedChar_4_1(link); else if (nUnsignedChar%4 == 0) PackInit_IntegerType_UnsignedChar_4_0(link);
729b23bfdefSJunchao Zhang     else if (nUnsignedChar == 2) PackInit_IntegerType_UnsignedChar_2_1(link); else if (nUnsignedChar%2 == 0) PackInit_IntegerType_UnsignedChar_2_0(link);
730b23bfdefSJunchao Zhang     else if (nUnsignedChar == 1) PackInit_IntegerType_UnsignedChar_1_1(link); else if (nUnsignedChar%1 == 0) PackInit_IntegerType_UnsignedChar_1_0(link);
731b23bfdefSJunchao Zhang     link->bs        = nUnsignedChar;
732eb02082bSJunchao Zhang     link->unitbytes = nUnsignedChar*sizeof(UnsignedChar);
733b23bfdefSJunchao Zhang     link->basicunit = MPI_UNSIGNED_CHAR;
7345ad15460SJunchao Zhang     if (link->bs == 1) {link->isbuiltin = PETSC_TRUE; link->unit = MPI_UNSIGNED_CHAR;}
73540e23c03SJunchao Zhang #if defined(PETSC_HAVE_COMPLEX)
736b23bfdefSJunchao Zhang   } else if (nPetscComplex) {
737b23bfdefSJunchao Zhang     if      (nPetscComplex == 8) PackInit_ComplexType_PetscComplex_8_1(link); else if (nPetscComplex%8 == 0) PackInit_ComplexType_PetscComplex_8_0(link);
738b23bfdefSJunchao Zhang     else if (nPetscComplex == 4) PackInit_ComplexType_PetscComplex_4_1(link); else if (nPetscComplex%4 == 0) PackInit_ComplexType_PetscComplex_4_0(link);
739b23bfdefSJunchao Zhang     else if (nPetscComplex == 2) PackInit_ComplexType_PetscComplex_2_1(link); else if (nPetscComplex%2 == 0) PackInit_ComplexType_PetscComplex_2_0(link);
740b23bfdefSJunchao Zhang     else if (nPetscComplex == 1) PackInit_ComplexType_PetscComplex_1_1(link); else if (nPetscComplex%1 == 0) PackInit_ComplexType_PetscComplex_1_0(link);
741b23bfdefSJunchao Zhang     link->bs        = nPetscComplex;
742eb02082bSJunchao Zhang     link->unitbytes = nPetscComplex*sizeof(PetscComplex);
74340e23c03SJunchao Zhang     link->basicunit = MPIU_COMPLEX;
7445ad15460SJunchao Zhang     if (link->bs == 1) {link->isbuiltin = PETSC_TRUE; link->unit = MPIU_COMPLEX;}
74540e23c03SJunchao Zhang #endif
74640e23c03SJunchao Zhang   } else {
747b23bfdefSJunchao Zhang     MPI_Aint lb,nbyte;
748ffc4695bSBarry Smith     ierr = MPI_Type_get_extent(unit,&lb,&nbyte);CHKERRMPI(ierr);
749*2c71b3e2SJacob Faibussowitsch     PetscCheckFalse(lb != 0,PETSC_COMM_SELF,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld",(long)lb);
750eb02082bSJunchao Zhang     if (nbyte % sizeof(int)) { /* If the type size is not multiple of int */
751eb02082bSJunchao Zhang       if      (nbyte == 4) PackInit_DumbType_char_4_1(link); else if (nbyte%4 == 0) PackInit_DumbType_char_4_0(link);
752eb02082bSJunchao Zhang       else if (nbyte == 2) PackInit_DumbType_char_2_1(link); else if (nbyte%2 == 0) PackInit_DumbType_char_2_0(link);
753eb02082bSJunchao Zhang       else if (nbyte == 1) PackInit_DumbType_char_1_1(link); else if (nbyte%1 == 0) PackInit_DumbType_char_1_0(link);
754eb02082bSJunchao Zhang       link->bs        = nbyte;
755b23bfdefSJunchao Zhang       link->unitbytes = nbyte;
756b23bfdefSJunchao Zhang       link->basicunit = MPI_BYTE;
75740e23c03SJunchao Zhang     } else {
758eb02082bSJunchao Zhang       nInt = nbyte / sizeof(int);
759eb02082bSJunchao Zhang       if      (nInt == 8) PackInit_DumbType_DumbInt_8_1(link); else if (nInt%8 == 0) PackInit_DumbType_DumbInt_8_0(link);
760eb02082bSJunchao Zhang       else if (nInt == 4) PackInit_DumbType_DumbInt_4_1(link); else if (nInt%4 == 0) PackInit_DumbType_DumbInt_4_0(link);
761eb02082bSJunchao Zhang       else if (nInt == 2) PackInit_DumbType_DumbInt_2_1(link); else if (nInt%2 == 0) PackInit_DumbType_DumbInt_2_0(link);
762eb02082bSJunchao Zhang       else if (nInt == 1) PackInit_DumbType_DumbInt_1_1(link); else if (nInt%1 == 0) PackInit_DumbType_DumbInt_1_0(link);
763eb02082bSJunchao Zhang       link->bs        = nInt;
764b23bfdefSJunchao Zhang       link->unitbytes = nbyte;
76540e23c03SJunchao Zhang       link->basicunit = MPI_INT;
76640e23c03SJunchao Zhang     }
7675ad15460SJunchao Zhang     if (link->isbuiltin) link->unit = unit;
76840e23c03SJunchao Zhang   }
769b23bfdefSJunchao Zhang 
770ffc4695bSBarry Smith   if (!link->isbuiltin) {ierr = MPI_Type_dup(unit,&link->unit);CHKERRMPI(ierr);}
77120c24465SJunchao Zhang 
77220c24465SJunchao Zhang   link->Memcpy = PetscSFLinkMemcpy_Host;
77340e23c03SJunchao Zhang   PetscFunctionReturn(0);
77440e23c03SJunchao Zhang }
77540e23c03SJunchao Zhang 
776fcc7397dSJunchao Zhang PetscErrorCode PetscSFLinkGetUnpackAndOp(PetscSFLink link,PetscMemType mtype,MPI_Op op,PetscBool atomic,PetscErrorCode (**UnpackAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*))
77740e23c03SJunchao Zhang {
77840e23c03SJunchao Zhang   PetscFunctionBegin;
77940e23c03SJunchao Zhang   *UnpackAndOp = NULL;
78071438e86SJunchao Zhang   if (PetscMemTypeHost(mtype)) {
78183df288dSJunchao Zhang     if      (op == MPI_REPLACE)               *UnpackAndOp = link->h_UnpackAndInsert;
782eb02082bSJunchao Zhang     else if (op == MPI_SUM || op == MPIU_SUM) *UnpackAndOp = link->h_UnpackAndAdd;
783eb02082bSJunchao Zhang     else if (op == MPI_PROD)                  *UnpackAndOp = link->h_UnpackAndMult;
784eb02082bSJunchao Zhang     else if (op == MPI_MAX || op == MPIU_MAX) *UnpackAndOp = link->h_UnpackAndMax;
785eb02082bSJunchao Zhang     else if (op == MPI_MIN || op == MPIU_MIN) *UnpackAndOp = link->h_UnpackAndMin;
786eb02082bSJunchao Zhang     else if (op == MPI_LAND)                  *UnpackAndOp = link->h_UnpackAndLAND;
787eb02082bSJunchao Zhang     else if (op == MPI_BAND)                  *UnpackAndOp = link->h_UnpackAndBAND;
788eb02082bSJunchao Zhang     else if (op == MPI_LOR)                   *UnpackAndOp = link->h_UnpackAndLOR;
789eb02082bSJunchao Zhang     else if (op == MPI_BOR)                   *UnpackAndOp = link->h_UnpackAndBOR;
790eb02082bSJunchao Zhang     else if (op == MPI_LXOR)                  *UnpackAndOp = link->h_UnpackAndLXOR;
791eb02082bSJunchao Zhang     else if (op == MPI_BXOR)                  *UnpackAndOp = link->h_UnpackAndBXOR;
792eb02082bSJunchao Zhang     else if (op == MPI_MAXLOC)                *UnpackAndOp = link->h_UnpackAndMaxloc;
793eb02082bSJunchao Zhang     else if (op == MPI_MINLOC)                *UnpackAndOp = link->h_UnpackAndMinloc;
794eb02082bSJunchao Zhang   }
7957fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
79671438e86SJunchao Zhang   else if (PetscMemTypeDevice(mtype) && !atomic) {
79783df288dSJunchao Zhang     if      (op == MPI_REPLACE)               *UnpackAndOp = link->d_UnpackAndInsert;
798eb02082bSJunchao Zhang     else if (op == MPI_SUM || op == MPIU_SUM) *UnpackAndOp = link->d_UnpackAndAdd;
799eb02082bSJunchao Zhang     else if (op == MPI_PROD)                  *UnpackAndOp = link->d_UnpackAndMult;
800eb02082bSJunchao Zhang     else if (op == MPI_MAX || op == MPIU_MAX) *UnpackAndOp = link->d_UnpackAndMax;
801eb02082bSJunchao Zhang     else if (op == MPI_MIN || op == MPIU_MIN) *UnpackAndOp = link->d_UnpackAndMin;
802eb02082bSJunchao Zhang     else if (op == MPI_LAND)                  *UnpackAndOp = link->d_UnpackAndLAND;
803eb02082bSJunchao Zhang     else if (op == MPI_BAND)                  *UnpackAndOp = link->d_UnpackAndBAND;
804eb02082bSJunchao Zhang     else if (op == MPI_LOR)                   *UnpackAndOp = link->d_UnpackAndLOR;
805eb02082bSJunchao Zhang     else if (op == MPI_BOR)                   *UnpackAndOp = link->d_UnpackAndBOR;
806eb02082bSJunchao Zhang     else if (op == MPI_LXOR)                  *UnpackAndOp = link->d_UnpackAndLXOR;
807eb02082bSJunchao Zhang     else if (op == MPI_BXOR)                  *UnpackAndOp = link->d_UnpackAndBXOR;
808eb02082bSJunchao Zhang     else if (op == MPI_MAXLOC)                *UnpackAndOp = link->d_UnpackAndMaxloc;
809eb02082bSJunchao Zhang     else if (op == MPI_MINLOC)                *UnpackAndOp = link->d_UnpackAndMinloc;
81071438e86SJunchao Zhang   } else if (PetscMemTypeDevice(mtype) && atomic) {
81183df288dSJunchao Zhang     if      (op == MPI_REPLACE)               *UnpackAndOp = link->da_UnpackAndInsert;
812eb02082bSJunchao Zhang     else if (op == MPI_SUM || op == MPIU_SUM) *UnpackAndOp = link->da_UnpackAndAdd;
813eb02082bSJunchao Zhang     else if (op == MPI_PROD)                  *UnpackAndOp = link->da_UnpackAndMult;
814eb02082bSJunchao Zhang     else if (op == MPI_MAX || op == MPIU_MAX) *UnpackAndOp = link->da_UnpackAndMax;
815eb02082bSJunchao Zhang     else if (op == MPI_MIN || op == MPIU_MIN) *UnpackAndOp = link->da_UnpackAndMin;
816eb02082bSJunchao Zhang     else if (op == MPI_LAND)                  *UnpackAndOp = link->da_UnpackAndLAND;
817eb02082bSJunchao Zhang     else if (op == MPI_BAND)                  *UnpackAndOp = link->da_UnpackAndBAND;
818eb02082bSJunchao Zhang     else if (op == MPI_LOR)                   *UnpackAndOp = link->da_UnpackAndLOR;
819eb02082bSJunchao Zhang     else if (op == MPI_BOR)                   *UnpackAndOp = link->da_UnpackAndBOR;
820eb02082bSJunchao Zhang     else if (op == MPI_LXOR)                  *UnpackAndOp = link->da_UnpackAndLXOR;
821eb02082bSJunchao Zhang     else if (op == MPI_BXOR)                  *UnpackAndOp = link->da_UnpackAndBXOR;
822eb02082bSJunchao Zhang     else if (op == MPI_MAXLOC)                *UnpackAndOp = link->da_UnpackAndMaxloc;
823eb02082bSJunchao Zhang     else if (op == MPI_MINLOC)                *UnpackAndOp = link->da_UnpackAndMinloc;
824eb02082bSJunchao Zhang   }
825eb02082bSJunchao Zhang #endif
82640e23c03SJunchao Zhang   PetscFunctionReturn(0);
82740e23c03SJunchao Zhang }
82840e23c03SJunchao Zhang 
829fcc7397dSJunchao Zhang PetscErrorCode PetscSFLinkGetScatterAndOp(PetscSFLink link,PetscMemType mtype,MPI_Op op,PetscBool atomic,PetscErrorCode (**ScatterAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*))
83040e23c03SJunchao Zhang {
83140e23c03SJunchao Zhang   PetscFunctionBegin;
832cd620004SJunchao Zhang   *ScatterAndOp = NULL;
83371438e86SJunchao Zhang   if (PetscMemTypeHost(mtype)) {
83483df288dSJunchao Zhang     if      (op == MPI_REPLACE)               *ScatterAndOp = link->h_ScatterAndInsert;
835cd620004SJunchao Zhang     else if (op == MPI_SUM || op == MPIU_SUM) *ScatterAndOp = link->h_ScatterAndAdd;
836cd620004SJunchao Zhang     else if (op == MPI_PROD)                  *ScatterAndOp = link->h_ScatterAndMult;
837cd620004SJunchao Zhang     else if (op == MPI_MAX || op == MPIU_MAX) *ScatterAndOp = link->h_ScatterAndMax;
838cd620004SJunchao Zhang     else if (op == MPI_MIN || op == MPIU_MIN) *ScatterAndOp = link->h_ScatterAndMin;
839cd620004SJunchao Zhang     else if (op == MPI_LAND)                  *ScatterAndOp = link->h_ScatterAndLAND;
840cd620004SJunchao Zhang     else if (op == MPI_BAND)                  *ScatterAndOp = link->h_ScatterAndBAND;
841cd620004SJunchao Zhang     else if (op == MPI_LOR)                   *ScatterAndOp = link->h_ScatterAndLOR;
842cd620004SJunchao Zhang     else if (op == MPI_BOR)                   *ScatterAndOp = link->h_ScatterAndBOR;
843cd620004SJunchao Zhang     else if (op == MPI_LXOR)                  *ScatterAndOp = link->h_ScatterAndLXOR;
844cd620004SJunchao Zhang     else if (op == MPI_BXOR)                  *ScatterAndOp = link->h_ScatterAndBXOR;
845cd620004SJunchao Zhang     else if (op == MPI_MAXLOC)                *ScatterAndOp = link->h_ScatterAndMaxloc;
846cd620004SJunchao Zhang     else if (op == MPI_MINLOC)                *ScatterAndOp = link->h_ScatterAndMinloc;
847eb02082bSJunchao Zhang   }
8487fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
84971438e86SJunchao Zhang   else if (PetscMemTypeDevice(mtype) && !atomic) {
85083df288dSJunchao Zhang     if      (op == MPI_REPLACE)               *ScatterAndOp = link->d_ScatterAndInsert;
851cd620004SJunchao Zhang     else if (op == MPI_SUM || op == MPIU_SUM) *ScatterAndOp = link->d_ScatterAndAdd;
852cd620004SJunchao Zhang     else if (op == MPI_PROD)                  *ScatterAndOp = link->d_ScatterAndMult;
853cd620004SJunchao Zhang     else if (op == MPI_MAX || op == MPIU_MAX) *ScatterAndOp = link->d_ScatterAndMax;
854cd620004SJunchao Zhang     else if (op == MPI_MIN || op == MPIU_MIN) *ScatterAndOp = link->d_ScatterAndMin;
855cd620004SJunchao Zhang     else if (op == MPI_LAND)                  *ScatterAndOp = link->d_ScatterAndLAND;
856cd620004SJunchao Zhang     else if (op == MPI_BAND)                  *ScatterAndOp = link->d_ScatterAndBAND;
857cd620004SJunchao Zhang     else if (op == MPI_LOR)                   *ScatterAndOp = link->d_ScatterAndLOR;
858cd620004SJunchao Zhang     else if (op == MPI_BOR)                   *ScatterAndOp = link->d_ScatterAndBOR;
859cd620004SJunchao Zhang     else if (op == MPI_LXOR)                  *ScatterAndOp = link->d_ScatterAndLXOR;
860cd620004SJunchao Zhang     else if (op == MPI_BXOR)                  *ScatterAndOp = link->d_ScatterAndBXOR;
861cd620004SJunchao Zhang     else if (op == MPI_MAXLOC)                *ScatterAndOp = link->d_ScatterAndMaxloc;
862cd620004SJunchao Zhang     else if (op == MPI_MINLOC)                *ScatterAndOp = link->d_ScatterAndMinloc;
86371438e86SJunchao Zhang   } else if (PetscMemTypeDevice(mtype) && atomic) {
86483df288dSJunchao Zhang     if      (op == MPI_REPLACE)               *ScatterAndOp = link->da_ScatterAndInsert;
865cd620004SJunchao Zhang     else if (op == MPI_SUM || op == MPIU_SUM) *ScatterAndOp = link->da_ScatterAndAdd;
866cd620004SJunchao Zhang     else if (op == MPI_PROD)                  *ScatterAndOp = link->da_ScatterAndMult;
867cd620004SJunchao Zhang     else if (op == MPI_MAX || op == MPIU_MAX) *ScatterAndOp = link->da_ScatterAndMax;
868cd620004SJunchao Zhang     else if (op == MPI_MIN || op == MPIU_MIN) *ScatterAndOp = link->da_ScatterAndMin;
869cd620004SJunchao Zhang     else if (op == MPI_LAND)                  *ScatterAndOp = link->da_ScatterAndLAND;
870cd620004SJunchao Zhang     else if (op == MPI_BAND)                  *ScatterAndOp = link->da_ScatterAndBAND;
871cd620004SJunchao Zhang     else if (op == MPI_LOR)                   *ScatterAndOp = link->da_ScatterAndLOR;
872cd620004SJunchao Zhang     else if (op == MPI_BOR)                   *ScatterAndOp = link->da_ScatterAndBOR;
873cd620004SJunchao Zhang     else if (op == MPI_LXOR)                  *ScatterAndOp = link->da_ScatterAndLXOR;
874cd620004SJunchao Zhang     else if (op == MPI_BXOR)                  *ScatterAndOp = link->da_ScatterAndBXOR;
875cd620004SJunchao Zhang     else if (op == MPI_MAXLOC)                *ScatterAndOp = link->da_ScatterAndMaxloc;
876cd620004SJunchao Zhang     else if (op == MPI_MINLOC)                *ScatterAndOp = link->da_ScatterAndMinloc;
877eb02082bSJunchao Zhang   }
878eb02082bSJunchao Zhang #endif
879cd620004SJunchao Zhang   PetscFunctionReturn(0);
880cd620004SJunchao Zhang }
881cd620004SJunchao Zhang 
882fcc7397dSJunchao Zhang PetscErrorCode PetscSFLinkGetFetchAndOp(PetscSFLink link,PetscMemType mtype,MPI_Op op,PetscBool atomic,PetscErrorCode (**FetchAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,void*))
883cd620004SJunchao Zhang {
884cd620004SJunchao Zhang   PetscFunctionBegin;
885cd620004SJunchao Zhang   *FetchAndOp = NULL;
886*2c71b3e2SJacob Faibussowitsch   PetscCheckFalse(op != MPI_SUM && op != MPIU_SUM,PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for MPI_Op in FetchAndOp");
88771438e86SJunchao Zhang   if (PetscMemTypeHost(mtype)) *FetchAndOp = link->h_FetchAndAdd;
8887fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
88971438e86SJunchao Zhang   else if (PetscMemTypeDevice(mtype) && !atomic) *FetchAndOp = link->d_FetchAndAdd;
89071438e86SJunchao Zhang   else if (PetscMemTypeDevice(mtype) && atomic)  *FetchAndOp = link->da_FetchAndAdd;
891cd620004SJunchao Zhang #endif
892cd620004SJunchao Zhang   PetscFunctionReturn(0);
893cd620004SJunchao Zhang }
894cd620004SJunchao Zhang 
895fcc7397dSJunchao Zhang PetscErrorCode PetscSFLinkGetFetchAndOpLocal(PetscSFLink link,PetscMemType mtype,MPI_Op op,PetscBool atomic,PetscErrorCode (**FetchAndOpLocal)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*))
896cd620004SJunchao Zhang {
897cd620004SJunchao Zhang   PetscFunctionBegin;
898cd620004SJunchao Zhang   *FetchAndOpLocal = NULL;
899*2c71b3e2SJacob Faibussowitsch   PetscCheckFalse(op != MPI_SUM && op != MPIU_SUM,PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for MPI_Op in FetchAndOp");
90071438e86SJunchao Zhang   if (PetscMemTypeHost(mtype)) *FetchAndOpLocal = link->h_FetchAndAddLocal;
9017fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
90271438e86SJunchao Zhang   else if (PetscMemTypeDevice(mtype) && !atomic) *FetchAndOpLocal = link->d_FetchAndAddLocal;
90371438e86SJunchao Zhang   else if (PetscMemTypeDevice(mtype) && atomic)  *FetchAndOpLocal = link->da_FetchAndAddLocal;
904cd620004SJunchao Zhang #endif
905cd620004SJunchao Zhang   PetscFunctionReturn(0);
906cd620004SJunchao Zhang }
907cd620004SJunchao Zhang 
9089fbee547SJacob Faibussowitsch static inline PetscErrorCode PetscSFLinkLogFlopsAfterUnpackRootData(PetscSF sf,PetscSFLink link,PetscSFScope scope,MPI_Op op)
909cd620004SJunchao Zhang {
910cd620004SJunchao Zhang   PetscErrorCode ierr;
911cd620004SJunchao Zhang   PetscLogDouble flops;
912cd620004SJunchao Zhang   PetscSF_Basic  *bas = (PetscSF_Basic*)sf->data;
913cd620004SJunchao Zhang 
914cd620004SJunchao Zhang   PetscFunctionBegin;
91583df288dSJunchao Zhang   if (op != MPI_REPLACE && link->basicunit == MPIU_SCALAR) { /* op is a reduction on PetscScalars */
916cd620004SJunchao Zhang     flops = bas->rootbuflen[scope]*link->bs; /* # of roots in buffer x # of scalars in unit */
9177fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
91871438e86SJunchao Zhang     if (PetscMemTypeDevice(link->rootmtype)) {ierr = PetscLogGpuFlops(flops);CHKERRQ(ierr);} else
919cd620004SJunchao Zhang #endif
920cd620004SJunchao Zhang     {ierr = PetscLogFlops(flops);CHKERRQ(ierr);}
921cd620004SJunchao Zhang   }
922cd620004SJunchao Zhang   PetscFunctionReturn(0);
923cd620004SJunchao Zhang }
924cd620004SJunchao Zhang 
9259fbee547SJacob Faibussowitsch static inline PetscErrorCode PetscSFLinkLogFlopsAfterUnpackLeafData(PetscSF sf,PetscSFLink link,PetscSFScope scope,MPI_Op op)
926cd620004SJunchao Zhang {
927cd620004SJunchao Zhang   PetscLogDouble flops;
928cd620004SJunchao Zhang   PetscErrorCode ierr;
929cd620004SJunchao Zhang 
930cd620004SJunchao Zhang   PetscFunctionBegin;
93183df288dSJunchao Zhang   if (op != MPI_REPLACE && link->basicunit == MPIU_SCALAR) { /* op is a reduction on PetscScalars */
932cd620004SJunchao Zhang     flops = sf->leafbuflen[scope]*link->bs; /* # of roots in buffer x # of scalars in unit */
9337fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
93471438e86SJunchao Zhang     if (PetscMemTypeDevice(link->leafmtype)) {ierr = PetscLogGpuFlops(flops);CHKERRQ(ierr);} else
935cd620004SJunchao Zhang #endif
936cd620004SJunchao Zhang     {ierr = PetscLogFlops(flops);CHKERRQ(ierr);}
937cd620004SJunchao Zhang   }
938cd620004SJunchao Zhang   PetscFunctionReturn(0);
939cd620004SJunchao Zhang }
940cd620004SJunchao Zhang 
941cd620004SJunchao Zhang /* When SF could not find a proper UnpackAndOp() from link, it falls back to MPI_Reduce_local.
9424165533cSJose E. Roman   Input Parameters:
943cd620004SJunchao Zhang   +sf      - The StarForest
944cd620004SJunchao Zhang   .link    - The link
945cd620004SJunchao Zhang   .count   - Number of entries to unpack
946cd620004SJunchao Zhang   .start   - The first index, significent when indices=NULL
947cd620004SJunchao Zhang   .indices - Indices of entries in <data>. If NULL, it means indices are contiguous and the first is given in <start>
948cd620004SJunchao Zhang   .buf     - A contiguous buffer to unpack from
949cd620004SJunchao Zhang   -op      - Operation after unpack
950cd620004SJunchao Zhang 
9514165533cSJose E. Roman   Output Parameters:
952cd620004SJunchao Zhang   .data    - The data to unpack to
953cd620004SJunchao Zhang */
9549fbee547SJacob Faibussowitsch static inline PetscErrorCode PetscSFLinkUnpackDataWithMPIReduceLocal(PetscSF sf,PetscSFLink link,PetscInt count,PetscInt start,const PetscInt *indices,void *data,const void *buf,MPI_Op op)
955cd620004SJunchao Zhang {
956cd620004SJunchao Zhang   PetscFunctionBegin;
957cd620004SJunchao Zhang #if defined(PETSC_HAVE_MPI_REDUCE_LOCAL)
958cd620004SJunchao Zhang   {
959cd620004SJunchao Zhang     PetscErrorCode ierr;
960cd620004SJunchao Zhang     PetscInt       i;
961cd620004SJunchao Zhang     if (indices) {
962cd620004SJunchao Zhang       /* Note we use link->unit instead of link->basicunit. When op can be mapped to MPI_SUM etc, it operates on
963cd620004SJunchao Zhang          basic units of a root/leaf element-wisely. Otherwise, it is meant to operate on a whole root/leaf.
964cd620004SJunchao Zhang       */
965ffc4695bSBarry Smith       for (i=0; i<count; i++) {ierr = MPI_Reduce_local((const char*)buf+i*link->unitbytes,(char*)data+indices[i]*link->unitbytes,1,link->unit,op);CHKERRMPI(ierr);}
966cd620004SJunchao Zhang     } else {
967c87b50c4SJunchao Zhang       ierr = MPIU_Reduce_local(buf,(char*)data+start*link->unitbytes,count,link->unit,op);CHKERRMPI(ierr);
968cd620004SJunchao Zhang     }
969cd620004SJunchao Zhang   }
970b458e8f1SJose E. Roman   PetscFunctionReturn(0);
971cd620004SJunchao Zhang #else
972cd620004SJunchao Zhang   SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No unpacking reduction operation for this MPI_Op");
973cd620004SJunchao Zhang #endif
974cd620004SJunchao Zhang }
975cd620004SJunchao Zhang 
9769fbee547SJacob Faibussowitsch static inline PetscErrorCode PetscSFLinkScatterDataWithMPIReduceLocal(PetscSF sf,PetscSFLink link,PetscInt count,PetscInt srcStart,const PetscInt *srcIdx,const void *src,PetscInt dstStart,const PetscInt *dstIdx,void *dst,MPI_Op op)
977cd620004SJunchao Zhang {
978cd620004SJunchao Zhang   PetscFunctionBegin;
979cd620004SJunchao Zhang #if defined(PETSC_HAVE_MPI_REDUCE_LOCAL)
980cd620004SJunchao Zhang   {
981cd620004SJunchao Zhang     PetscErrorCode ierr;
982cd620004SJunchao Zhang     PetscInt       i,disp;
983fcc7397dSJunchao Zhang     if (!srcIdx) {
984fcc7397dSJunchao Zhang       ierr = PetscSFLinkUnpackDataWithMPIReduceLocal(sf,link,count,dstStart,dstIdx,dst,(const char*)src+srcStart*link->unitbytes,op);CHKERRQ(ierr);
985fcc7397dSJunchao Zhang     } else {
986cd620004SJunchao Zhang       for (i=0; i<count; i++) {
987fcc7397dSJunchao Zhang         disp = dstIdx? dstIdx[i] : dstStart + i;
988c87b50c4SJunchao Zhang         ierr = MPIU_Reduce_local((const char*)src+srcIdx[i]*link->unitbytes,(char*)dst+disp*link->unitbytes,1,link->unit,op);CHKERRMPI(ierr);
989fcc7397dSJunchao Zhang       }
990cd620004SJunchao Zhang     }
991cd620004SJunchao Zhang   }
992b458e8f1SJose E. Roman   PetscFunctionReturn(0);
993cd620004SJunchao Zhang #else
994cd620004SJunchao Zhang   SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No unpacking reduction operation for this MPI_Op");
995cd620004SJunchao Zhang #endif
996cd620004SJunchao Zhang }
997cd620004SJunchao Zhang 
998cd620004SJunchao Zhang /*=============================================================================
999cd620004SJunchao Zhang               Pack/Unpack/Fetch/Scatter routines
1000cd620004SJunchao Zhang  ============================================================================*/
1001cd620004SJunchao Zhang 
1002cd620004SJunchao Zhang /* Pack rootdata to rootbuf
10034165533cSJose E. Roman   Input Parameters:
1004cd620004SJunchao Zhang   + sf       - The SF this packing works on.
1005cd620004SJunchao Zhang   . link     - It gives the memtype of the roots and also provides root buffer.
1006cd620004SJunchao Zhang   . scope    - PETSCSF_LOCAL or PETSCSF_REMOTE. Note SF has the ability to do local and remote communications separately.
1007cd620004SJunchao Zhang   - rootdata - Where to read the roots.
1008cd620004SJunchao Zhang 
1009cd620004SJunchao Zhang   Notes:
1010cd620004SJunchao Zhang   When rootdata can be directly used as root buffer, the routine is almost a no-op. After the call, root data is
101171438e86SJunchao Zhang   in a place where the underlying MPI is ready to access (use_gpu_aware_mpi or not)
1012cd620004SJunchao Zhang  */
101371438e86SJunchao Zhang PetscErrorCode PetscSFLinkPackRootData_Private(PetscSF sf,PetscSFLink link,PetscSFScope scope,const void *rootdata)
1014cd620004SJunchao Zhang {
1015cd620004SJunchao Zhang   PetscErrorCode   ierr;
1016cd620004SJunchao Zhang   const PetscInt   *rootindices = NULL;
1017cd620004SJunchao Zhang   PetscInt         count,start;
1018fcc7397dSJunchao Zhang   PetscErrorCode   (*Pack)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*) = NULL;
1019cd620004SJunchao Zhang   PetscMemType     rootmtype = link->rootmtype;
1020fcc7397dSJunchao Zhang   PetscSFPackOpt   opt = NULL;
1021fcc7397dSJunchao Zhang 
1022cd620004SJunchao Zhang   PetscFunctionBegin;
1023cd620004SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
102471438e86SJunchao Zhang   if (!link->rootdirect[scope]) { /* If rootdata works directly as rootbuf, skip packing */
1025fcc7397dSJunchao Zhang     ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,rootmtype,scope,&count,&start,&opt,&rootindices);CHKERRQ(ierr);
1026cd620004SJunchao Zhang     ierr = PetscSFLinkGetPack(link,rootmtype,&Pack);CHKERRQ(ierr);
1027fcc7397dSJunchao Zhang     ierr = (*Pack)(link,count,start,opt,rootindices,rootdata,link->rootbuf[scope][rootmtype]);CHKERRQ(ierr);
1028cd620004SJunchao Zhang   }
1029cd620004SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
1030cd620004SJunchao Zhang   PetscFunctionReturn(0);
1031cd620004SJunchao Zhang }
1032cd620004SJunchao Zhang 
1033cd620004SJunchao Zhang /* Pack leafdata to leafbuf */
103471438e86SJunchao Zhang PetscErrorCode PetscSFLinkPackLeafData_Private(PetscSF sf,PetscSFLink link,PetscSFScope scope,const void *leafdata)
1035cd620004SJunchao Zhang {
1036cd620004SJunchao Zhang   PetscErrorCode   ierr;
1037cd620004SJunchao Zhang   const PetscInt   *leafindices = NULL;
1038cd620004SJunchao Zhang   PetscInt         count,start;
1039fcc7397dSJunchao Zhang   PetscErrorCode   (*Pack)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*) = NULL;
1040cd620004SJunchao Zhang   PetscMemType     leafmtype = link->leafmtype;
1041fcc7397dSJunchao Zhang   PetscSFPackOpt   opt = NULL;
1042cd620004SJunchao Zhang 
1043cd620004SJunchao Zhang   PetscFunctionBegin;
1044cd620004SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
104571438e86SJunchao Zhang   if (!link->leafdirect[scope]) { /* If leafdata works directly as rootbuf, skip packing */
1046fcc7397dSJunchao Zhang     ierr = PetscSFLinkGetLeafPackOptAndIndices(sf,link,leafmtype,scope,&count,&start,&opt,&leafindices);CHKERRQ(ierr);
1047cd620004SJunchao Zhang     ierr = PetscSFLinkGetPack(link,leafmtype,&Pack);CHKERRQ(ierr);
1048fcc7397dSJunchao Zhang     ierr = (*Pack)(link,count,start,opt,leafindices,leafdata,link->leafbuf[scope][leafmtype]);CHKERRQ(ierr);
1049cd620004SJunchao Zhang   }
1050cd620004SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
1051cd620004SJunchao Zhang   PetscFunctionReturn(0);
1052cd620004SJunchao Zhang }
1053cd620004SJunchao Zhang 
105471438e86SJunchao Zhang /* Pack rootdata to rootbuf, which are in the same memory space */
105571438e86SJunchao Zhang PetscErrorCode PetscSFLinkPackRootData(PetscSF sf,PetscSFLink link,PetscSFScope scope,const void *rootdata)
105671438e86SJunchao Zhang {
105771438e86SJunchao Zhang   PetscErrorCode   ierr;
105871438e86SJunchao Zhang   PetscSF_Basic    *bas = (PetscSF_Basic*)sf->data;
105971438e86SJunchao Zhang 
106071438e86SJunchao Zhang   PetscFunctionBegin;
106171438e86SJunchao Zhang   if (scope == PETSCSF_REMOTE) { /* Sync the device if rootdata is not on petsc default stream */
106271438e86SJunchao Zhang     if (PetscMemTypeDevice(link->rootmtype) && link->SyncDevice && sf->unknown_input_stream) {ierr = (*link->SyncDevice)(link);CHKERRQ(ierr);}
106371438e86SJunchao Zhang     if (link->PrePack) {ierr = (*link->PrePack)(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr);} /* Used by SF nvshmem */
106471438e86SJunchao Zhang   }
106571438e86SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
106671438e86SJunchao Zhang   if (bas->rootbuflen[scope]) {ierr = PetscSFLinkPackRootData_Private(sf,link,scope,rootdata);CHKERRQ(ierr);}
106771438e86SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
106871438e86SJunchao Zhang   PetscFunctionReturn(0);
106971438e86SJunchao Zhang }
107071438e86SJunchao Zhang /* Pack leafdata to leafbuf, which are in the same memory space */
107171438e86SJunchao Zhang PetscErrorCode PetscSFLinkPackLeafData(PetscSF sf,PetscSFLink link,PetscSFScope scope,const void *leafdata)
107271438e86SJunchao Zhang {
107371438e86SJunchao Zhang   PetscErrorCode   ierr;
107471438e86SJunchao Zhang 
107571438e86SJunchao Zhang   PetscFunctionBegin;
107671438e86SJunchao Zhang   if (scope == PETSCSF_REMOTE) {
107771438e86SJunchao Zhang     if (PetscMemTypeDevice(link->leafmtype) && link->SyncDevice && sf->unknown_input_stream) {ierr = (*link->SyncDevice)(link);CHKERRQ(ierr);}
107871438e86SJunchao Zhang     if (link->PrePack) {ierr = (*link->PrePack)(sf,link,PETSCSF_LEAF2ROOT);CHKERRQ(ierr);}  /* Used by SF nvshmem */
107971438e86SJunchao Zhang   }
108071438e86SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
108171438e86SJunchao Zhang   if (sf->leafbuflen[scope]) {ierr = PetscSFLinkPackLeafData_Private(sf,link,scope,leafdata);CHKERRQ(ierr);}
108271438e86SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Pack,sf,0,0,0);CHKERRQ(ierr);
108371438e86SJunchao Zhang   PetscFunctionReturn(0);
108471438e86SJunchao Zhang }
108571438e86SJunchao Zhang 
108671438e86SJunchao Zhang PetscErrorCode PetscSFLinkUnpackRootData_Private(PetscSF sf,PetscSFLink link,PetscSFScope scope,void *rootdata,MPI_Op op)
1087cd620004SJunchao Zhang {
1088cd620004SJunchao Zhang   PetscErrorCode   ierr;
1089cd620004SJunchao Zhang   const PetscInt   *rootindices = NULL;
1090cd620004SJunchao Zhang   PetscInt         count,start;
1091cd620004SJunchao Zhang   PetscSF_Basic    *bas = (PetscSF_Basic*)sf->data;
1092fcc7397dSJunchao Zhang   PetscErrorCode   (*UnpackAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*) = NULL;
1093cd620004SJunchao Zhang   PetscMemType     rootmtype = link->rootmtype;
1094fcc7397dSJunchao Zhang   PetscSFPackOpt   opt = NULL;
1095cd620004SJunchao Zhang 
1096cd620004SJunchao Zhang   PetscFunctionBegin;
109771438e86SJunchao Zhang   if (!link->rootdirect[scope]) { /* If rootdata works directly as rootbuf, skip unpacking */
1098cd620004SJunchao Zhang     ierr = PetscSFLinkGetUnpackAndOp(link,rootmtype,op,bas->rootdups[scope],&UnpackAndOp);CHKERRQ(ierr);
1099cd620004SJunchao Zhang     if (UnpackAndOp) {
1100fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,rootmtype,scope,&count,&start,&opt,&rootindices);CHKERRQ(ierr);
1101fcc7397dSJunchao Zhang       ierr = (*UnpackAndOp)(link,count,start,opt,rootindices,rootdata,link->rootbuf[scope][rootmtype]);CHKERRQ(ierr);
1102cd620004SJunchao Zhang     } else {
1103fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,PETSC_MEMTYPE_HOST,scope,&count,&start,&opt,&rootindices);CHKERRQ(ierr);
1104cd620004SJunchao Zhang       ierr = PetscSFLinkUnpackDataWithMPIReduceLocal(sf,link,count,start,rootindices,rootdata,link->rootbuf[scope][rootmtype],op);CHKERRQ(ierr);
1105cd620004SJunchao Zhang     }
1106cd620004SJunchao Zhang   }
1107cd620004SJunchao Zhang   ierr = PetscSFLinkLogFlopsAfterUnpackRootData(sf,link,scope,op);CHKERRQ(ierr);
1108cd620004SJunchao Zhang   PetscFunctionReturn(0);
1109cd620004SJunchao Zhang }
1110cd620004SJunchao Zhang 
111171438e86SJunchao Zhang PetscErrorCode PetscSFLinkUnpackLeafData_Private(PetscSF sf,PetscSFLink link,PetscSFScope scope,void *leafdata,MPI_Op op)
1112cd620004SJunchao Zhang {
1113cd620004SJunchao Zhang   PetscErrorCode   ierr;
1114cd620004SJunchao Zhang   const PetscInt   *leafindices = NULL;
1115cd620004SJunchao Zhang   PetscInt         count,start;
1116fcc7397dSJunchao Zhang   PetscErrorCode   (*UnpackAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*) = NULL;
1117cd620004SJunchao Zhang   PetscMemType     leafmtype = link->leafmtype;
1118fcc7397dSJunchao Zhang   PetscSFPackOpt   opt = NULL;
1119cd620004SJunchao Zhang 
1120cd620004SJunchao Zhang   PetscFunctionBegin;
112171438e86SJunchao Zhang   if (!link->leafdirect[scope]) { /* If leafdata works directly as rootbuf, skip unpacking */
1122cd620004SJunchao Zhang     ierr = PetscSFLinkGetUnpackAndOp(link,leafmtype,op,sf->leafdups[scope],&UnpackAndOp);CHKERRQ(ierr);
1123cd620004SJunchao Zhang     if (UnpackAndOp) {
1124fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetLeafPackOptAndIndices(sf,link,leafmtype,scope,&count,&start,&opt,&leafindices);CHKERRQ(ierr);
1125fcc7397dSJunchao Zhang       ierr = (*UnpackAndOp)(link,count,start,opt,leafindices,leafdata,link->leafbuf[scope][leafmtype]);CHKERRQ(ierr);
1126cd620004SJunchao Zhang     } else {
1127fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetLeafPackOptAndIndices(sf,link,PETSC_MEMTYPE_HOST,scope,&count,&start,&opt,&leafindices);CHKERRQ(ierr);
1128cd620004SJunchao Zhang       ierr = PetscSFLinkUnpackDataWithMPIReduceLocal(sf,link,count,start,leafindices,leafdata,link->leafbuf[scope][leafmtype],op);CHKERRQ(ierr);
1129cd620004SJunchao Zhang     }
1130cd620004SJunchao Zhang   }
1131cd620004SJunchao Zhang   ierr = PetscSFLinkLogFlopsAfterUnpackLeafData(sf,link,scope,op);CHKERRQ(ierr);
113271438e86SJunchao Zhang   PetscFunctionReturn(0);
113371438e86SJunchao Zhang }
113471438e86SJunchao Zhang /* Unpack rootbuf to rootdata, which are in the same memory space */
113571438e86SJunchao Zhang PetscErrorCode PetscSFLinkUnpackRootData(PetscSF sf,PetscSFLink link,PetscSFScope scope,void *rootdata,MPI_Op op)
113671438e86SJunchao Zhang {
113771438e86SJunchao Zhang   PetscErrorCode   ierr;
113871438e86SJunchao Zhang   PetscSF_Basic    *bas = (PetscSF_Basic*)sf->data;
113971438e86SJunchao Zhang 
114071438e86SJunchao Zhang   PetscFunctionBegin;
114171438e86SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Unpack,sf,0,0,0);CHKERRQ(ierr);
114271438e86SJunchao Zhang   if (bas->rootbuflen[scope]) {ierr = PetscSFLinkUnpackRootData_Private(sf,link,scope,rootdata,op);CHKERRQ(ierr);}
1143cd620004SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Unpack,sf,0,0,0);CHKERRQ(ierr);
114471438e86SJunchao Zhang   if (scope == PETSCSF_REMOTE) {
114571438e86SJunchao Zhang     if (link->PostUnpack) {ierr = (*link->PostUnpack)(sf,link,PETSCSF_LEAF2ROOT);CHKERRQ(ierr);}  /* Used by SF nvshmem */
114671438e86SJunchao Zhang     if (PetscMemTypeDevice(link->rootmtype) && link->SyncDevice && sf->unknown_input_stream) {ierr = (*link->SyncDevice)(link);CHKERRQ(ierr);}
114771438e86SJunchao Zhang   }
1148cd620004SJunchao Zhang   PetscFunctionReturn(0);
1149cd620004SJunchao Zhang }
1150cd620004SJunchao Zhang 
115171438e86SJunchao Zhang /* Unpack leafbuf to leafdata for remote (common case) or local (rare case when rootmtype != leafmtype) */
115271438e86SJunchao Zhang PetscErrorCode PetscSFLinkUnpackLeafData(PetscSF sf,PetscSFLink link,PetscSFScope scope,void *leafdata,MPI_Op op)
115371438e86SJunchao Zhang {
115471438e86SJunchao Zhang   PetscErrorCode   ierr;
115571438e86SJunchao Zhang 
115671438e86SJunchao Zhang   PetscFunctionBegin;
115771438e86SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Unpack,sf,0,0,0);CHKERRQ(ierr);
11581e1ea65dSPierre Jolivet   if (sf->leafbuflen[scope]) {ierr = PetscSFLinkUnpackLeafData_Private(sf,link,scope,leafdata,op);CHKERRQ(ierr);}
115971438e86SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Unpack,sf,0,0,0);CHKERRQ(ierr);
116071438e86SJunchao Zhang   if (scope == PETSCSF_REMOTE) {
116171438e86SJunchao Zhang     if (link->PostUnpack) {ierr = (*link->PostUnpack)(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr);}  /* Used by SF nvshmem */
116271438e86SJunchao Zhang     if (PetscMemTypeDevice(link->leafmtype) && link->SyncDevice && sf->unknown_input_stream) {ierr = (*link->SyncDevice)(link);CHKERRQ(ierr);}
116371438e86SJunchao Zhang   }
116471438e86SJunchao Zhang   PetscFunctionReturn(0);
116571438e86SJunchao Zhang }
116671438e86SJunchao Zhang 
116771438e86SJunchao Zhang /* FetchAndOp rootdata with rootbuf, it is a kind of Unpack on rootdata, except it also updates rootbuf */
116871438e86SJunchao Zhang PetscErrorCode PetscSFLinkFetchAndOpRemote(PetscSF sf,PetscSFLink link,void *rootdata,MPI_Op op)
1169cd620004SJunchao Zhang {
1170cd620004SJunchao Zhang   PetscErrorCode     ierr;
1171cd620004SJunchao Zhang   const PetscInt     *rootindices = NULL;
1172cd620004SJunchao Zhang   PetscInt           count,start;
1173cd620004SJunchao Zhang   PetscSF_Basic      *bas = (PetscSF_Basic*)sf->data;
1174fcc7397dSJunchao Zhang   PetscErrorCode     (*FetchAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,void*) = NULL;
1175cd620004SJunchao Zhang   PetscMemType       rootmtype = link->rootmtype;
1176fcc7397dSJunchao Zhang   PetscSFPackOpt     opt = NULL;
1177cd620004SJunchao Zhang 
1178cd620004SJunchao Zhang   PetscFunctionBegin;
1179cd620004SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_Unpack,sf,0,0,0);CHKERRQ(ierr);
118071438e86SJunchao Zhang   if (bas->rootbuflen[PETSCSF_REMOTE]) {
1181cd620004SJunchao Zhang     /* Do FetchAndOp on rootdata with rootbuf */
118271438e86SJunchao Zhang     ierr = PetscSFLinkGetFetchAndOp(link,rootmtype,op,bas->rootdups[PETSCSF_REMOTE],&FetchAndOp);CHKERRQ(ierr);
118371438e86SJunchao Zhang     ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,rootmtype,PETSCSF_REMOTE,&count,&start,&opt,&rootindices);CHKERRQ(ierr);
118471438e86SJunchao Zhang     ierr = (*FetchAndOp)(link,count,start,opt,rootindices,rootdata,link->rootbuf[PETSCSF_REMOTE][rootmtype]);CHKERRQ(ierr);
1185cd620004SJunchao Zhang   }
118671438e86SJunchao Zhang   ierr = PetscSFLinkLogFlopsAfterUnpackRootData(sf,link,PETSCSF_REMOTE,op);CHKERRQ(ierr);
1187cd620004SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_Unpack,sf,0,0,0);CHKERRQ(ierr);
1188cd620004SJunchao Zhang   PetscFunctionReturn(0);
1189cd620004SJunchao Zhang }
1190cd620004SJunchao Zhang 
119171438e86SJunchao Zhang PetscErrorCode PetscSFLinkScatterLocal(PetscSF sf,PetscSFLink link,PetscSFDirection direction,void *rootdata,void *leafdata,MPI_Op op)
1192cd620004SJunchao Zhang {
1193cd620004SJunchao Zhang   PetscErrorCode       ierr;
1194cd620004SJunchao Zhang   const PetscInt       *rootindices = NULL,*leafindices = NULL;
1195cd620004SJunchao Zhang   PetscInt             count,rootstart,leafstart;
1196cd620004SJunchao Zhang   PetscSF_Basic        *bas = (PetscSF_Basic*)sf->data;
1197fcc7397dSJunchao Zhang   PetscErrorCode       (*ScatterAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*) = NULL;
119871438e86SJunchao Zhang   PetscMemType         rootmtype = link->rootmtype,leafmtype = link->leafmtype,srcmtype,dstmtype;
1199fcc7397dSJunchao Zhang   PetscSFPackOpt       leafopt = NULL,rootopt = NULL;
120071438e86SJunchao Zhang   PetscInt             buflen = sf->leafbuflen[PETSCSF_LOCAL];
120171438e86SJunchao Zhang   char                 *srcbuf = NULL,*dstbuf = NULL;
120271438e86SJunchao Zhang   PetscBool            dstdups;
1203cd620004SJunchao Zhang 
1204362febeeSStefano Zampini   PetscFunctionBegin;
120571438e86SJunchao Zhang   if (!buflen) PetscFunctionReturn(0);
120671438e86SJunchao Zhang   if (rootmtype != leafmtype) { /* The cross memory space local scatter is done by pack, copy and unpack */
120771438e86SJunchao Zhang     if (direction == PETSCSF_ROOT2LEAF) {
1208cd620004SJunchao Zhang       ierr     = PetscSFLinkPackRootData(sf,link,PETSCSF_LOCAL,rootdata);CHKERRQ(ierr);
120971438e86SJunchao Zhang       srcmtype = rootmtype;
121071438e86SJunchao Zhang       srcbuf   = link->rootbuf[PETSCSF_LOCAL][rootmtype];
121171438e86SJunchao Zhang       dstmtype = leafmtype;
121271438e86SJunchao Zhang       dstbuf   = link->leafbuf[PETSCSF_LOCAL][leafmtype];
121371438e86SJunchao Zhang     } else {
121471438e86SJunchao Zhang       ierr     = PetscSFLinkPackLeafData(sf,link,PETSCSF_LOCAL,leafdata);CHKERRQ(ierr);
121571438e86SJunchao Zhang       srcmtype = leafmtype;
121671438e86SJunchao Zhang       srcbuf   = link->leafbuf[PETSCSF_LOCAL][leafmtype];
121771438e86SJunchao Zhang       dstmtype = rootmtype;
121871438e86SJunchao Zhang       dstbuf   = link->rootbuf[PETSCSF_LOCAL][rootmtype];
121971438e86SJunchao Zhang     }
122071438e86SJunchao Zhang     ierr = (*link->Memcpy)(link,dstmtype,dstbuf,srcmtype,srcbuf,buflen*link->unitbytes);CHKERRQ(ierr);
122171438e86SJunchao Zhang     /* If above is a device to host copy, we have to sync the stream before accessing the buffer on host */
122271438e86SJunchao Zhang     if (PetscMemTypeHost(dstmtype)) {ierr = (*link->SyncStream)(link);CHKERRQ(ierr);}
122371438e86SJunchao Zhang     if (direction == PETSCSF_ROOT2LEAF) {
1224cd620004SJunchao Zhang       ierr = PetscSFLinkUnpackLeafData(sf,link,PETSCSF_LOCAL,leafdata,op);CHKERRQ(ierr);
1225cd620004SJunchao Zhang     } else {
122671438e86SJunchao Zhang       ierr = PetscSFLinkUnpackRootData(sf,link,PETSCSF_LOCAL,rootdata,op);CHKERRQ(ierr);
122771438e86SJunchao Zhang     }
122871438e86SJunchao Zhang   } else {
122971438e86SJunchao Zhang     dstdups  = (direction == PETSCSF_ROOT2LEAF) ? sf->leafdups[PETSCSF_LOCAL] : bas->rootdups[PETSCSF_LOCAL];
123071438e86SJunchao Zhang     dstmtype = (direction == PETSCSF_ROOT2LEAF) ? link->leafmtype : link->rootmtype;
123171438e86SJunchao Zhang     ierr = PetscSFLinkGetScatterAndOp(link,dstmtype,op,dstdups,&ScatterAndOp);CHKERRQ(ierr);
1232cd620004SJunchao Zhang     if (ScatterAndOp) {
1233fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,rootmtype,PETSCSF_LOCAL,&count,&rootstart,&rootopt,&rootindices);CHKERRQ(ierr);
1234fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetLeafPackOptAndIndices(sf,link,leafmtype,PETSCSF_LOCAL,&count,&leafstart,&leafopt,&leafindices);CHKERRQ(ierr);
123571438e86SJunchao Zhang       if (direction == PETSCSF_ROOT2LEAF) {
1236fcc7397dSJunchao Zhang         ierr = (*ScatterAndOp)(link,count,rootstart,rootopt,rootindices,rootdata,leafstart,leafopt,leafindices,leafdata);CHKERRQ(ierr);
1237cd620004SJunchao Zhang       } else {
1238fcc7397dSJunchao Zhang         ierr = (*ScatterAndOp)(link,count,leafstart,leafopt,leafindices,leafdata,rootstart,rootopt,rootindices,rootdata);CHKERRQ(ierr);
123971438e86SJunchao Zhang       }
1240cd620004SJunchao Zhang     } else {
1241fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,PETSC_MEMTYPE_HOST,PETSCSF_LOCAL,&count,&rootstart,&rootopt,&rootindices);CHKERRQ(ierr);
1242fcc7397dSJunchao Zhang       ierr = PetscSFLinkGetLeafPackOptAndIndices(sf,link,PETSC_MEMTYPE_HOST,PETSCSF_LOCAL,&count,&leafstart,&leafopt,&leafindices);CHKERRQ(ierr);
124371438e86SJunchao Zhang       if (direction == PETSCSF_ROOT2LEAF) {
124471438e86SJunchao Zhang         ierr = PetscSFLinkScatterDataWithMPIReduceLocal(sf,link,count,rootstart,rootindices,rootdata,leafstart,leafindices,leafdata,op);CHKERRQ(ierr);
124571438e86SJunchao Zhang       } else {
1246fcc7397dSJunchao Zhang         ierr = PetscSFLinkScatterDataWithMPIReduceLocal(sf,link,count,leafstart,leafindices,leafdata,rootstart,rootindices,rootdata,op);CHKERRQ(ierr);
1247cd620004SJunchao Zhang       }
1248cd620004SJunchao Zhang     }
124971438e86SJunchao Zhang   }
1250cd620004SJunchao Zhang   PetscFunctionReturn(0);
1251cd620004SJunchao Zhang }
1252cd620004SJunchao Zhang 
1253cd620004SJunchao Zhang /* Fetch rootdata to leafdata and leafupdate locally */
1254cd620004SJunchao Zhang PetscErrorCode PetscSFLinkFetchAndOpLocal(PetscSF sf,PetscSFLink link,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op)
1255cd620004SJunchao Zhang {
1256cd620004SJunchao Zhang   PetscErrorCode       ierr;
1257cd620004SJunchao Zhang   const PetscInt       *rootindices = NULL,*leafindices = NULL;
1258cd620004SJunchao Zhang   PetscInt             count,rootstart,leafstart;
1259cd620004SJunchao Zhang   PetscSF_Basic        *bas = (PetscSF_Basic*)sf->data;
1260fcc7397dSJunchao Zhang   PetscErrorCode       (*FetchAndOpLocal)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*) = NULL;
1261cd620004SJunchao Zhang   const PetscMemType   rootmtype = link->rootmtype,leafmtype = link->leafmtype;
1262fcc7397dSJunchao Zhang   PetscSFPackOpt       leafopt = NULL,rootopt = NULL;
1263cd620004SJunchao Zhang 
1264cd620004SJunchao Zhang   PetscFunctionBegin;
1265cd620004SJunchao Zhang   if (!bas->rootbuflen[PETSCSF_LOCAL]) PetscFunctionReturn(0);
1266cd620004SJunchao Zhang   if (rootmtype != leafmtype) {
1267cd620004SJunchao Zhang     /* The local communication has to go through pack and unpack */
1268cd620004SJunchao Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Doing PetscSFFetchAndOp with rootdata and leafdata on opposite side of CPU and GPU");
1269cd620004SJunchao Zhang   } else {
1270fcc7397dSJunchao Zhang     ierr = PetscSFLinkGetRootPackOptAndIndices(sf,link,rootmtype,PETSCSF_LOCAL,&count,&rootstart,&rootopt,&rootindices);CHKERRQ(ierr);
1271fcc7397dSJunchao Zhang     ierr = PetscSFLinkGetLeafPackOptAndIndices(sf,link,leafmtype,PETSCSF_LOCAL,&count,&leafstart,&leafopt,&leafindices);CHKERRQ(ierr);
1272cd620004SJunchao Zhang     ierr = PetscSFLinkGetFetchAndOpLocal(link,rootmtype,op,bas->rootdups[PETSCSF_LOCAL],&FetchAndOpLocal);CHKERRQ(ierr);
1273fcc7397dSJunchao Zhang     ierr = (*FetchAndOpLocal)(link,count,rootstart,rootopt,rootindices,rootdata,leafstart,leafopt,leafindices,leafdata,leafupdate);CHKERRQ(ierr);
1274cd620004SJunchao Zhang   }
127540e23c03SJunchao Zhang   PetscFunctionReturn(0);
127640e23c03SJunchao Zhang }
127740e23c03SJunchao Zhang 
127840e23c03SJunchao Zhang /*
1279cd620004SJunchao Zhang   Create per-rank pack/unpack optimizations based on indice patterns
128040e23c03SJunchao Zhang 
128140e23c03SJunchao Zhang    Input Parameters:
1282fcc7397dSJunchao Zhang   +  n       - Number of destination ranks
1283eb02082bSJunchao Zhang   .  offset  - [n+1] For the i-th rank, its associated indices are idx[offset[i], offset[i+1]). offset[0] needs not to be 0.
1284b23bfdefSJunchao Zhang   -  idx     - [*]   Array storing indices
128540e23c03SJunchao Zhang 
128640e23c03SJunchao Zhang    Output Parameters:
1287cd620004SJunchao Zhang   +  opt     - Pack optimizations. NULL if no optimizations.
128840e23c03SJunchao Zhang */
1289cd620004SJunchao Zhang PetscErrorCode PetscSFCreatePackOpt(PetscInt n,const PetscInt *offset,const PetscInt *idx,PetscSFPackOpt *out)
129040e23c03SJunchao Zhang {
129140e23c03SJunchao Zhang   PetscErrorCode ierr;
1292fcc7397dSJunchao Zhang   PetscInt       r,p,start,i,j,k,dx,dy,dz,dydz,m,X,Y;
1293fcc7397dSJunchao Zhang   PetscBool      optimizable = PETSC_TRUE;
129440e23c03SJunchao Zhang   PetscSFPackOpt opt;
129540e23c03SJunchao Zhang 
129640e23c03SJunchao Zhang   PetscFunctionBegin;
1297fcc7397dSJunchao Zhang   ierr   = PetscMalloc1(1,&opt);CHKERRQ(ierr);
1298fcc7397dSJunchao Zhang   ierr   = PetscMalloc1(7*n+2,&opt->array);CHKERRQ(ierr);
1299fcc7397dSJunchao Zhang   opt->n      = opt->array[0] = n;
1300fcc7397dSJunchao Zhang   opt->offset = opt->array + 1;
1301fcc7397dSJunchao Zhang   opt->start  = opt->array + n   + 2;
1302fcc7397dSJunchao Zhang   opt->dx     = opt->array + 2*n + 2;
1303fcc7397dSJunchao Zhang   opt->dy     = opt->array + 3*n + 2;
1304fcc7397dSJunchao Zhang   opt->dz     = opt->array + 4*n + 2;
1305fcc7397dSJunchao Zhang   opt->X      = opt->array + 5*n + 2;
1306fcc7397dSJunchao Zhang   opt->Y      = opt->array + 6*n + 2;
1307fcc7397dSJunchao Zhang 
1308fcc7397dSJunchao Zhang   for (r=0; r<n; r++) { /* For each destination rank */
1309fcc7397dSJunchao Zhang     m     = offset[r+1] - offset[r]; /* Total number of indices for this rank. We want to see if m can be factored into dx*dy*dz */
1310fcc7397dSJunchao Zhang     p     = offset[r];
1311fcc7397dSJunchao Zhang     start = idx[p]; /* First index for this rank */
1312fcc7397dSJunchao Zhang     p++;
1313fcc7397dSJunchao Zhang 
1314fcc7397dSJunchao Zhang     /* Search in X dimension */
1315fcc7397dSJunchao Zhang     for (dx=1; dx<m; dx++,p++) {
1316fcc7397dSJunchao Zhang       if (start+dx != idx[p]) break;
1317b23bfdefSJunchao Zhang     }
1318b23bfdefSJunchao Zhang 
1319fcc7397dSJunchao Zhang     dydz = m/dx;
1320fcc7397dSJunchao Zhang     X    = dydz > 1 ? (idx[p]-start) : dx;
1321fcc7397dSJunchao Zhang     /* Not optimizable if m is not a multiple of dx, or some unrecognized pattern is found */
1322fcc7397dSJunchao Zhang     if (m%dx || X <= 0) {optimizable = PETSC_FALSE; goto finish;}
1323fcc7397dSJunchao Zhang     for (dy=1; dy<dydz; dy++) { /* Search in Y dimension */
1324fcc7397dSJunchao Zhang       for (i=0; i<dx; i++,p++) {
1325fcc7397dSJunchao Zhang         if (start+X*dy+i != idx[p]) {
1326fcc7397dSJunchao Zhang           if (i) {optimizable = PETSC_FALSE; goto finish;} /* The pattern is violated in the middle of an x-walk */
1327fcc7397dSJunchao Zhang           else goto Z_dimension;
132840e23c03SJunchao Zhang         }
132940e23c03SJunchao Zhang       }
133040e23c03SJunchao Zhang     }
133140e23c03SJunchao Zhang 
1332fcc7397dSJunchao Zhang Z_dimension:
1333fcc7397dSJunchao Zhang     dz = m/(dx*dy);
1334fcc7397dSJunchao Zhang     Y  = dz > 1 ? (idx[p]-start)/X : dy;
1335fcc7397dSJunchao Zhang     /* Not optimizable if m is not a multiple of dx*dy, or some unrecognized pattern is found */
1336fcc7397dSJunchao Zhang     if (m%(dx*dy) || Y <= 0) {optimizable = PETSC_FALSE; goto finish;}
1337fcc7397dSJunchao Zhang     for (k=1; k<dz; k++) { /* Go through Z dimension to see if remaining indices follow the pattern */
1338fcc7397dSJunchao Zhang       for (j=0; j<dy; j++) {
1339fcc7397dSJunchao Zhang         for (i=0; i<dx; i++,p++) {
1340fcc7397dSJunchao Zhang           if (start+X*Y*k+X*j+i != idx[p]) {optimizable = PETSC_FALSE; goto finish;}
134140e23c03SJunchao Zhang         }
134240e23c03SJunchao Zhang       }
134340e23c03SJunchao Zhang     }
1344fcc7397dSJunchao Zhang     opt->start[r] = start;
1345fcc7397dSJunchao Zhang     opt->dx[r]    = dx;
1346fcc7397dSJunchao Zhang     opt->dy[r]    = dy;
1347fcc7397dSJunchao Zhang     opt->dz[r]    = dz;
1348fcc7397dSJunchao Zhang     opt->X[r]     = X;
1349fcc7397dSJunchao Zhang     opt->Y[r]     = Y;
135040e23c03SJunchao Zhang   }
135140e23c03SJunchao Zhang 
1352fcc7397dSJunchao Zhang finish:
1353fcc7397dSJunchao Zhang   /* If not optimizable, free arrays to save memory */
1354fcc7397dSJunchao Zhang   if (!n || !optimizable) {
1355fcc7397dSJunchao Zhang     ierr = PetscFree(opt->array);CHKERRQ(ierr);
135640e23c03SJunchao Zhang     ierr = PetscFree(opt);CHKERRQ(ierr);
135740e23c03SJunchao Zhang     *out = NULL;
1358fcc7397dSJunchao Zhang   } else {
1359fcc7397dSJunchao Zhang     opt->offset[0] = 0;
1360fcc7397dSJunchao Zhang     for (r=0; r<n; r++) opt->offset[r+1] = opt->offset[r] + opt->dx[r]*opt->dy[r]*opt->dz[r];
1361fcc7397dSJunchao Zhang     *out = opt;
1362fcc7397dSJunchao Zhang   }
136340e23c03SJunchao Zhang   PetscFunctionReturn(0);
136440e23c03SJunchao Zhang }
136540e23c03SJunchao Zhang 
13669fbee547SJacob Faibussowitsch static inline PetscErrorCode PetscSFDestroyPackOpt(PetscSF sf,PetscMemType mtype,PetscSFPackOpt *out)
136740e23c03SJunchao Zhang {
136840e23c03SJunchao Zhang   PetscErrorCode ierr;
136940e23c03SJunchao Zhang   PetscSFPackOpt opt = *out;
137040e23c03SJunchao Zhang 
137140e23c03SJunchao Zhang   PetscFunctionBegin;
137240e23c03SJunchao Zhang   if (opt) {
137320c24465SJunchao Zhang     ierr = PetscSFFree(sf,mtype,opt->array);CHKERRQ(ierr);
137440e23c03SJunchao Zhang     ierr = PetscFree(opt);CHKERRQ(ierr);
137540e23c03SJunchao Zhang     *out = NULL;
137640e23c03SJunchao Zhang   }
137740e23c03SJunchao Zhang   PetscFunctionReturn(0);
137840e23c03SJunchao Zhang }
1379cd620004SJunchao Zhang 
1380cd620004SJunchao Zhang PetscErrorCode PetscSFSetUpPackFields(PetscSF sf)
1381cd620004SJunchao Zhang {
1382cd620004SJunchao Zhang   PetscErrorCode ierr;
1383cd620004SJunchao Zhang   PetscSF_Basic  *bas = (PetscSF_Basic*)sf->data;
1384cd620004SJunchao Zhang   PetscInt       i,j;
1385cd620004SJunchao Zhang 
1386cd620004SJunchao Zhang   PetscFunctionBegin;
1387cd620004SJunchao Zhang   /* [0] for PETSCSF_LOCAL and [1] for PETSCSF_REMOTE in the following */
1388cd620004SJunchao Zhang   for (i=0; i<2; i++) { /* Set defaults */
1389cd620004SJunchao Zhang     sf->leafstart[i]   = 0;
1390cd620004SJunchao Zhang     sf->leafcontig[i]  = PETSC_TRUE;
1391cd620004SJunchao Zhang     sf->leafdups[i]    = PETSC_FALSE;
1392cd620004SJunchao Zhang     bas->rootstart[i]  = 0;
1393cd620004SJunchao Zhang     bas->rootcontig[i] = PETSC_TRUE;
1394cd620004SJunchao Zhang     bas->rootdups[i]   = PETSC_FALSE;
1395cd620004SJunchao Zhang   }
1396cd620004SJunchao Zhang 
1397cd620004SJunchao Zhang   sf->leafbuflen[0] = sf->roffset[sf->ndranks];
1398cd620004SJunchao Zhang   sf->leafbuflen[1] = sf->roffset[sf->nranks] - sf->roffset[sf->ndranks];
1399cd620004SJunchao Zhang 
1400cd620004SJunchao Zhang   if (sf->leafbuflen[0]) sf->leafstart[0] = sf->rmine[0];
1401cd620004SJunchao Zhang   if (sf->leafbuflen[1]) sf->leafstart[1] = sf->rmine[sf->roffset[sf->ndranks]];
1402cd620004SJunchao Zhang 
1403cd620004SJunchao Zhang   /* Are leaf indices for self and remote contiguous? If yes, it is best for pack/unpack */
1404cd620004SJunchao Zhang   for (i=0; i<sf->roffset[sf->ndranks]; i++) { /* self */
1405cd620004SJunchao Zhang     if (sf->rmine[i] != sf->leafstart[0]+i) {sf->leafcontig[0] = PETSC_FALSE; break;}
1406cd620004SJunchao Zhang   }
1407cd620004SJunchao Zhang   for (i=sf->roffset[sf->ndranks],j=0; i<sf->roffset[sf->nranks]; i++,j++) { /* remote */
1408cd620004SJunchao Zhang     if (sf->rmine[i] != sf->leafstart[1]+j) {sf->leafcontig[1] = PETSC_FALSE; break;}
1409cd620004SJunchao Zhang   }
1410cd620004SJunchao Zhang 
1411cd620004SJunchao Zhang   /* If not, see if we can have per-rank optimizations by doing index analysis */
1412cd620004SJunchao Zhang   if (!sf->leafcontig[0]) {ierr = PetscSFCreatePackOpt(sf->ndranks,            sf->roffset,             sf->rmine, &sf->leafpackopt[0]);CHKERRQ(ierr);}
1413cd620004SJunchao Zhang   if (!sf->leafcontig[1]) {ierr = PetscSFCreatePackOpt(sf->nranks-sf->ndranks, sf->roffset+sf->ndranks, sf->rmine, &sf->leafpackopt[1]);CHKERRQ(ierr);}
1414cd620004SJunchao Zhang 
1415cd620004SJunchao Zhang   /* Are root indices for self and remote contiguous? */
1416cd620004SJunchao Zhang   bas->rootbuflen[0] = bas->ioffset[bas->ndiranks];
1417cd620004SJunchao Zhang   bas->rootbuflen[1] = bas->ioffset[bas->niranks] - bas->ioffset[bas->ndiranks];
1418cd620004SJunchao Zhang 
1419cd620004SJunchao Zhang   if (bas->rootbuflen[0]) bas->rootstart[0] = bas->irootloc[0];
1420cd620004SJunchao Zhang   if (bas->rootbuflen[1]) bas->rootstart[1] = bas->irootloc[bas->ioffset[bas->ndiranks]];
1421cd620004SJunchao Zhang 
1422cd620004SJunchao Zhang   for (i=0; i<bas->ioffset[bas->ndiranks]; i++) {
1423cd620004SJunchao Zhang     if (bas->irootloc[i] != bas->rootstart[0]+i) {bas->rootcontig[0] = PETSC_FALSE; break;}
1424cd620004SJunchao Zhang   }
1425cd620004SJunchao Zhang   for (i=bas->ioffset[bas->ndiranks],j=0; i<bas->ioffset[bas->niranks]; i++,j++) {
1426cd620004SJunchao Zhang     if (bas->irootloc[i] != bas->rootstart[1]+j) {bas->rootcontig[1] = PETSC_FALSE; break;}
1427cd620004SJunchao Zhang   }
1428cd620004SJunchao Zhang 
1429cd620004SJunchao Zhang   if (!bas->rootcontig[0]) {ierr = PetscSFCreatePackOpt(bas->ndiranks,              bas->ioffset,               bas->irootloc, &bas->rootpackopt[0]);CHKERRQ(ierr);}
1430cd620004SJunchao Zhang   if (!bas->rootcontig[1]) {ierr = PetscSFCreatePackOpt(bas->niranks-bas->ndiranks, bas->ioffset+bas->ndiranks, bas->irootloc, &bas->rootpackopt[1]);CHKERRQ(ierr);}
1431cd620004SJunchao Zhang 
14327fd2d3dbSJunchao Zhang  #if defined(PETSC_HAVE_DEVICE)
1433cd620004SJunchao Zhang     /* Check dups in indices so that CUDA unpacking kernels can use cheaper regular instructions instead of atomics when they know there are no data race chances */
1434013b3241SStefano Zampini   if (PetscDefined(HAVE_DEVICE)) {
1435013b3241SStefano Zampini     PetscBool ismulti = (sf->multi == sf) ? PETSC_TRUE : PETSC_FALSE;
1436013b3241SStefano Zampini     if (!sf->leafcontig[0]  && !ismulti) {ierr = PetscCheckDupsInt(sf->leafbuflen[0],  sf->rmine,                                 &sf->leafdups[0]);CHKERRQ(ierr);}
1437013b3241SStefano Zampini     if (!sf->leafcontig[1]  && !ismulti) {ierr = PetscCheckDupsInt(sf->leafbuflen[1],  sf->rmine+sf->roffset[sf->ndranks],        &sf->leafdups[1]);CHKERRQ(ierr);}
1438013b3241SStefano Zampini     if (!bas->rootcontig[0] && !ismulti) {ierr = PetscCheckDupsInt(bas->rootbuflen[0], bas->irootloc,                             &bas->rootdups[0]);CHKERRQ(ierr);}
1439013b3241SStefano Zampini     if (!bas->rootcontig[1] && !ismulti) {ierr = PetscCheckDupsInt(bas->rootbuflen[1], bas->irootloc+bas->ioffset[bas->ndiranks], &bas->rootdups[1]);CHKERRQ(ierr);}
1440013b3241SStefano Zampini   }
1441cd620004SJunchao Zhang #endif
1442cd620004SJunchao Zhang   PetscFunctionReturn(0);
1443cd620004SJunchao Zhang }
1444cd620004SJunchao Zhang 
1445cd620004SJunchao Zhang PetscErrorCode PetscSFResetPackFields(PetscSF sf)
1446cd620004SJunchao Zhang {
1447cd620004SJunchao Zhang   PetscErrorCode ierr;
1448cd620004SJunchao Zhang   PetscSF_Basic  *bas = (PetscSF_Basic*)sf->data;
1449cd620004SJunchao Zhang   PetscInt       i;
1450cd620004SJunchao Zhang 
1451cd620004SJunchao Zhang   PetscFunctionBegin;
1452cd620004SJunchao Zhang   for (i=PETSCSF_LOCAL; i<=PETSCSF_REMOTE; i++) {
145320c24465SJunchao Zhang     ierr = PetscSFDestroyPackOpt(sf,PETSC_MEMTYPE_HOST,&sf->leafpackopt[i]);CHKERRQ(ierr);
145420c24465SJunchao Zhang     ierr = PetscSFDestroyPackOpt(sf,PETSC_MEMTYPE_HOST,&bas->rootpackopt[i]);CHKERRQ(ierr);
14557fd2d3dbSJunchao Zhang    #if defined(PETSC_HAVE_DEVICE)
145620c24465SJunchao Zhang     ierr = PetscSFDestroyPackOpt(sf,PETSC_MEMTYPE_DEVICE,&sf->leafpackopt_d[i]);CHKERRQ(ierr);
145720c24465SJunchao Zhang     ierr = PetscSFDestroyPackOpt(sf,PETSC_MEMTYPE_DEVICE,&bas->rootpackopt_d[i]);CHKERRQ(ierr);
14587fd2d3dbSJunchao Zhang    #endif
1459cd620004SJunchao Zhang   }
1460cd620004SJunchao Zhang   PetscFunctionReturn(0);
1461cd620004SJunchao Zhang }
1462