+/* Copyright (c) 2013-2023. The SimGrid Team. All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
/* extracted from mpig_myreduce.c with
:3,$s/MPL/MPI/g and :%s/\\\\$/ \\/ */
# define SCR_LNG_OPTIM(bytelng) 128 + ((bytelng+127)/256) * 256;
/* = 16 + multiple of 32 doubles*/
-# define REDUCE_LIMITS /* values are lower limits for count arg. */ \
- /* routine = reduce allreduce */ \
- /* size = 2, 3,2**n,other 2, 3,2**n,other */ \
- static int Lsh[2][4]={{ 896,1728, 576, 736},{ 448,1280, 512, 512}}; \
- static int Lin[2][4]={{ 896,1728, 576, 736},{ 448,1280, 512, 512}}; \
- static int Llg[2][4]={{ 896,1728, 576, 736},{ 448,1280, 512, 512}}; \
- static int Lfp[2][4]={{ 896,1728, 576, 736},{ 448,1280, 512, 512}}; \
- static int Ldb[2][4]={{ 896,1728, 576, 736},{ 448,1280, 512, 512}}; \
- static int Lby[2][4]={{ 896,1728, 576, 736},{ 448,1280, 512, 512}};
+#define REDUCE_LIMITS /* values are lower limits for count arg. */ \
+ /* routine = reduce allreduce */ \
+ /* size = 2, 3,2**n,other 2, 3,2**n,other */ \
+ static int Lsh[2][4] = {{896, 1728, 576, 736}, {448, 1280, 512, 512}}; \
+ static int Lin[2][4] = {{896, 1728, 576, 736}, {448, 1280, 512, 512}}; \
+ static int Llg[2][4] = {{896, 1728, 576, 736}, {448, 1280, 512, 512}}; \
+ static int Lfp[2][4] = {{896, 1728, 576, 736}, {448, 1280, 512, 512}}; \
+ static int Ldb[2][4] = {{896, 1728, 576, 736}, {448, 1280, 512, 512}}; \
+ static int Lby[2][4] = {{896, 1728, 576, 736}, {448, 1280, 512, 512}};
#endif
#ifdef REDUCE_NEW_ALWAYS
# undef REDUCE_LIMITS
-# define REDUCE_LIMITS /* values are lower limits for count arg. */ \
- /* routine = reduce allreduce */ \
- /* size = 2, 3,2**n,other 2, 3,2**n,other */ \
- static int Lsh[2][4]={{ 1, 1, 1, 1},{ 1, 1, 1, 1}}; \
- static int Lin[2][4]={{ 1, 1, 1, 1},{ 1, 1, 1, 1}}; \
- static int Llg[2][4]={{ 1, 1, 1, 1},{ 1, 1, 1, 1}}; \
- static int Lfp[2][4]={{ 1, 1, 1, 1},{ 1, 1, 1, 1}}; \
- static int Ldb[2][4]={{ 1, 1, 1, 1},{ 1, 1, 1, 1}}; \
- static int Lby[2][4]={{ 1, 1, 1, 1},{ 1, 1, 1, 1}};
+#define REDUCE_LIMITS /* values are lower limits for count arg. */ \
+ /* routine = reduce allreduce */ \
+ /* size = 2, 3,2**n,other 2, 3,2**n,other */ \
+ static int Lsh[2][4] = {{1, 1, 1, 1}, {1, 1, 1, 1}}; \
+ static int Lin[2][4] = {{1, 1, 1, 1}, {1, 1, 1, 1}}; \
+ static int Llg[2][4] = {{1, 1, 1, 1}, {1, 1, 1, 1}}; \
+ static int Lfp[2][4] = {{1, 1, 1, 1}, {1, 1, 1, 1}}; \
+ static int Ldb[2][4] = {{1, 1, 1, 1}, {1, 1, 1, 1}}; \
+ static int Lby[2][4] = {{1, 1, 1, 1}, {1, 1, 1, 1}};
#endif
/* Fast reduce and allreduce algorithm for longer buffers and predefined
operations.
- This algorithm is explaned with the example of 13 nodes.
+ This algorithm is explained with the example of 13 nodes.
The nodes are numbered 0, 1, 2, ... 12.
The sendbuf content is a, b, c, ... m.
The buffer array is notated with ABCDEFGH, this means that
Exa.: size=13 ==> n=3, r=5 (i.e. size == 13 == 2**n+r == 2**3 + 5)
- The algoritm needs for the execution of one Colls::reduce
+ The algorithm needs for the execution of one colls::reduce
- for r==0
exec_time = n*(L1+L2) + buf_lng * (1-1/2**n) * (T1 + T2 + O/d)
7: { [(a+b)+(c+d)] + [(e+f)+(g+h)] } + { [(i+j)+k] + [l+m] } for H
-For Colls::allreduce:
+For colls::allreduce:
------------------
Step 6.1)
on all nodes 0..12
-For Colls::reduce:
+For colls::reduce:
---------------
Step 6.0)
2) This line shows the limit for the count argument.
If count < limit then the vendor protocol is used,
otherwise the new protocol is used (see variable Ldb).
- 3) These lines show the bandwidth (=bufer length / execution time)
+ 3) These lines show the bandwidth (= buffer length / execution time)
for both protocols.
- 4) This line shows that the limit is choosen well if the ratio is
- between 0.95 (loosing 5% for buffer length near and >=limit)
+ 4) This line shows that the limit is chosen well if the ratio is
+ between 0.95 (losing 5% for buffer length near and >=limit)
and 1.10 (not gaining 10% for buffer length near and <limit).
5) This line shows that the new protocol is 2..7 times faster
for long counts.
#ifdef REDUCE_LIMITS
#ifdef USE_Irecv
-#define MPI_I_Sendrecv(sb,sc,sd,dest,st,rb,rc,rd,source,rt,comm,stat) \
- { MPI_Request req; \
- req=Request::irecv(rb,rc,rd,source,rt,comm); \
- Request::send(sb,sc,sd,dest,st,comm); \
- Request::wait(&req,stat); \
- }
+#define MPI_I_Sendrecv(sb, sc, sd, dest, st, rb, rc, rd, source, rt, comm, stat) \
+ { \
+ MPI_Request req; \
+ req = Request::irecv(rb, rc, rd, source, rt, comm); \
+ Request::send(sb, sc, sd, dest, st, comm); \
+ Request::wait(&req, stat); \
+ }
#else
#ifdef USE_Isend
-#define MPI_I_Sendrecv(sb,sc,sd,dest,st,rb,rc,rd,source,rt,comm,stat) \
- { MPI_Request req; \
- req=mpi_mpi_isend(sb,sc,sd,dest,st,comm); \
- Request::recv(rb,rc,rd,source,rt,comm,stat); \
- Request::wait(&req,stat); \
- }
+#define MPI_I_Sendrecv(sb, sc, sd, dest, st, rb, rc, rd, source, rt, comm, stat) \
+ { \
+ MPI_Request req; \
+ req = mpi_mpi_isend(sb, sc, sd, dest, st, comm); \
+ Request::recv(rb, rc, rd, source, rt, comm, stat); \
+ Request::wait(&req, stat); \
+ }
#else
-#define MPI_I_Sendrecv(sb,sc,sd,dest,st,rb,rc,rd,source,rt,comm,stat) \
- Request::sendrecv(sb,sc,sd,dest,st,rb,rc,rd,source,rt,comm,stat)
+#define MPI_I_Sendrecv(sb, sc, sd, dest, st, rb, rc, rd, source, rt, comm, stat) \
+ Request::sendrecv(sb, sc, sd, dest, st, rb, rc, rd, source, rt, comm, stat)
#endif
#endif
MPIM_LXOR,
MPIM_BXOR
};
-#define MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_TYPE,TYPE) \
-static void MPI_I_do_op_TYPE(TYPE* b1,TYPE* b2,TYPE* rslt, int cnt,MPIM_Op op)\
-{ int i; \
- switch (op) { \
- case MPIM_MAX : \
- for(i=0;i<cnt;i++) rslt[i] = (b1[i]>b2[i]?b1[i]:b2[i]); break; \
- case MPIM_MIN : \
- for(i=0;i<cnt;i++) rslt[i] = (b1[i]<b2[i]?b1[i]:b2[i]); break; \
- case MPIM_SUM :for(i=0;i<cnt;i++) rslt[i] = b1[i] + b2[i]; break; \
- case MPIM_PROD:for(i=0;i<cnt;i++) rslt[i] = b1[i] * b2[i]; break; \
- case MPIM_LAND:for(i=0;i<cnt;i++) rslt[i] = b1[i] && b2[i]; break; \
- case MPIM_LOR :for(i=0;i<cnt;i++) rslt[i] = b1[i] || b2[i]; break; \
- case MPIM_LXOR:for(i=0;i<cnt;i++) rslt[i] = b1[i] != b2[i]; break; \
- case MPIM_BAND:for(i=0;i<cnt;i++) rslt[i] = b1[i] & b2[i]; break; \
- case MPIM_BOR :for(i=0;i<cnt;i++) rslt[i] = b1[i] | b2[i]; break; \
- case MPIM_BXOR:for(i=0;i<cnt;i++) rslt[i] = b1[i] ^ b2[i]; break; \
- default: break; \
- } \
-}
+#define MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_TYPE, TYPE) \
+ static void MPI_I_do_op_TYPE(TYPE* b1, TYPE* b2, TYPE* rslt, int cnt, MPIM_Op op) \
+ { \
+ int i; \
+ switch (op) { \
+ case MPIM_MAX: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = (b1[i] > b2[i] ? b1[i] : b2[i]); \
+ break; \
+ case MPIM_MIN: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = (b1[i] < b2[i] ? b1[i] : b2[i]); \
+ break; \
+ case MPIM_SUM: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] + b2[i]; \
+ break; \
+ case MPIM_PROD: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] * b2[i]; \
+ break; \
+ case MPIM_LAND: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] && b2[i]; \
+ break; \
+ case MPIM_LOR: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] || b2[i]; \
+ break; \
+ case MPIM_LXOR: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] != b2[i]; \
+ break; \
+ case MPIM_BAND: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] & b2[i]; \
+ break; \
+ case MPIM_BOR: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] | b2[i]; \
+ break; \
+ case MPIM_BXOR: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] ^ b2[i]; \
+ break; \
+ default: \
+ break; \
+ } \
+ }
-#define MPI_I_DO_OP_FP(MPI_I_do_op_TYPE,TYPE) \
-static void MPI_I_do_op_TYPE(TYPE* b1,TYPE* b2,TYPE* rslt, int cnt,MPIM_Op op) \
-{ int i; \
- switch (op) { \
- case MPIM_MAX : \
- for(i=0;i<cnt;i++) rslt[i] = (b1[i]>b2[i]?b1[i]:b2[i]); break; \
- case MPIM_MIN : \
- for(i=0;i<cnt;i++) rslt[i] = (b1[i]<b2[i]?b1[i]:b2[i]); break; \
- case MPIM_SUM :for(i=0;i<cnt;i++) rslt[i] = b1[i] + b2[i]; break; \
- case MPIM_PROD:for(i=0;i<cnt;i++) rslt[i] = b1[i] * b2[i]; break; \
- default: break; \
- } \
-}
+#define MPI_I_DO_OP_FP(MPI_I_do_op_TYPE, TYPE) \
+ static void MPI_I_do_op_TYPE(TYPE* b1, TYPE* b2, TYPE* rslt, int cnt, MPIM_Op op) \
+ { \
+ int i; \
+ switch (op) { \
+ case MPIM_MAX: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = (b1[i] > b2[i] ? b1[i] : b2[i]); \
+ break; \
+ case MPIM_MIN: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = (b1[i] < b2[i] ? b1[i] : b2[i]); \
+ break; \
+ case MPIM_SUM: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] + b2[i]; \
+ break; \
+ case MPIM_PROD: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] * b2[i]; \
+ break; \
+ default: \
+ break; \
+ } \
+ }
-#define MPI_I_DO_OP_BYTE(MPI_I_do_op_TYPE,TYPE) \
-static void MPI_I_do_op_TYPE(TYPE* b1,TYPE* b2,TYPE* rslt, int cnt,MPIM_Op op)\
-{ int i; \
- switch (op) { \
- case MPIM_BAND:for(i=0;i<cnt;i++) rslt[i] = b1[i] & b2[i]; break; \
- case MPIM_BOR :for(i=0;i<cnt;i++) rslt[i] = b1[i] | b2[i]; break; \
- case MPIM_BXOR:for(i=0;i<cnt;i++) rslt[i] = b1[i] ^ b2[i]; break; \
- default: break; \
- } \
-}
+#define MPI_I_DO_OP_BYTE(MPI_I_do_op_TYPE, TYPE) \
+ static void MPI_I_do_op_TYPE(TYPE* b1, TYPE* b2, TYPE* rslt, int cnt, MPIM_Op op) \
+ { \
+ int i; \
+ switch (op) { \
+ case MPIM_BAND: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] & b2[i]; \
+ break; \
+ case MPIM_BOR: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] | b2[i]; \
+ break; \
+ case MPIM_BXOR: \
+ for (i = 0; i < cnt; i++) \
+ rslt[i] = b1[i] ^ b2[i]; \
+ break; \
+ default: \
+ break; \
+ } \
+ }
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_short, short)
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_int, int)
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_long, long)
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_ushort, unsigned short)
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_uint, unsigned int)
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_ulong, unsigned long)
-MPI_I_DO_OP_C_INTEGER( MPI_I_do_op_ulonglong, unsigned long long)
-MPI_I_DO_OP_FP( MPI_I_do_op_float, float)
-MPI_I_DO_OP_FP( MPI_I_do_op_double, double)
-MPI_I_DO_OP_BYTE( MPI_I_do_op_byte, char)
-
-#define MPI_I_DO_OP_CALL(MPI_I_do_op_TYPE,TYPE) \
- MPI_I_do_op_TYPE ((TYPE*)b1, (TYPE*)b2, (TYPE*)rslt, cnt, op); break;
-
-static void MPI_I_do_op(void* b1, void* b2, void* rslt, int cnt,
- MPIM_Datatype datatype, MPIM_Op op)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_short, short)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_int, int)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_long, long)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_ushort, unsigned short)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_uint, unsigned int)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_ulong, unsigned long)
+MPI_I_DO_OP_C_INTEGER(MPI_I_do_op_ulonglong, unsigned long long)
+MPI_I_DO_OP_FP(MPI_I_do_op_float, float)
+MPI_I_DO_OP_FP(MPI_I_do_op_double, double)
+MPI_I_DO_OP_BYTE(MPI_I_do_op_byte, char)
+
+#define MPI_I_DO_OP_CALL(MPI_I_do_op_TYPE, TYPE) \
+ MPI_I_do_op_TYPE((TYPE*)b1, (TYPE*)b2, (TYPE*)rslt, cnt, op); \
+ break;
+
+static void MPI_I_do_op(void* b1, void* b2, void* rslt, int cnt, MPIM_Datatype datatype, MPIM_Op op)
{
- switch (datatype) {
- case MPIM_SHORT : MPI_I_DO_OP_CALL(MPI_I_do_op_short, short)
- case MPIM_INT : MPI_I_DO_OP_CALL(MPI_I_do_op_int, int)
- case MPIM_LONG : MPI_I_DO_OP_CALL(MPI_I_do_op_long, long)
- case MPIM_UNSIGNED_SHORT:
- MPI_I_DO_OP_CALL(MPI_I_do_op_ushort, unsigned short)
- case MPIM_UNSIGNED:
- MPI_I_DO_OP_CALL(MPI_I_do_op_uint, unsigned int)
- case MPIM_UNSIGNED_LONG:
- MPI_I_DO_OP_CALL(MPI_I_do_op_ulong, unsigned long)
- case MPIM_UNSIGNED_LONG_LONG:
- MPI_I_DO_OP_CALL(MPI_I_do_op_ulonglong, unsigned long long)
- case MPIM_FLOAT : MPI_I_DO_OP_CALL(MPI_I_do_op_float, float)
- case MPIM_DOUBLE: MPI_I_DO_OP_CALL(MPI_I_do_op_double, double)
- case MPIM_BYTE : MPI_I_DO_OP_CALL(MPI_I_do_op_byte, char)
- }
+ switch (datatype) {
+ case MPIM_SHORT:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_short, short)
+ case MPIM_INT:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_int, int)
+ case MPIM_LONG:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_long, long)
+ case MPIM_UNSIGNED_SHORT:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_ushort, unsigned short)
+ case MPIM_UNSIGNED:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_uint, unsigned int)
+ case MPIM_UNSIGNED_LONG:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_ulong, unsigned long)
+ case MPIM_UNSIGNED_LONG_LONG:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_ulonglong, unsigned long long)
+ case MPIM_FLOAT:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_float, float)
+ case MPIM_DOUBLE:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_double, double)
+ case MPIM_BYTE:
+ MPI_I_DO_OP_CALL(MPI_I_do_op_byte, char)
+ }
}
REDUCE_LIMITS
-namespace simgrid{
-namespace smpi{
-static int MPI_I_anyReduce(void* Sendbuf, void* Recvbuf, int count, MPI_Datatype mpi_datatype, MPI_Op mpi_op, int root, MPI_Comm comm, int is_all)
+namespace simgrid::smpi {
+static int MPI_I_anyReduce(const void* Sendbuf, void* Recvbuf, int count, MPI_Datatype mpi_datatype, MPI_Op mpi_op,
+ int root, MPI_Comm comm, bool is_all)
{
char *scr1buf, *scr2buf, *scr3buf, *xxx, *sendbuf, *recvbuf;
int myrank, size, x_base, x_size, computed, idx;
else if(mpi_datatype==MPI_DOUBLE ) datatype=MPIM_DOUBLE;
else if(mpi_datatype==MPI_BYTE ) datatype=MPIM_BYTE;
else
- THROWF(arg_error,0, "reduce rab algorithm can't be used with this datatype ! ");
+ throw std::invalid_argument("reduce rab algorithm can't be used with this datatype!");
if (mpi_op==MPI_MAX ) op=MPIM_MAX;
else if(mpi_op==MPI_MIN ) op=MPIM_MIN;
# endif
n = 0; x_size = 1;
while (2*x_size <= size) { n++; x_size = x_size * 2; }
- /* x_sixe == 2**n */
+ /* x_size == 2**n */
r = size - x_size;
/*...step 2 */
if (mynewrank >= 0)
{ /* begin -- only for nodes with new rank */
-# define OLDRANK(new) ((new)==newroot ? root \
- : ((new)<r ? (new)*2 : (new)+r) )
+#define OLDRANK(new) ((new) == newroot ? root : ((new) < r ? (new) * 2 : (new) + r))
for(idx=n-1, x_base=x_size/2; idx>=0; idx--, x_base=x_base/2)
{
} /* new_prot */
/*otherwise:*/
if (is_all)
- return( Colls::allreduce(Sendbuf, Recvbuf, count, mpi_datatype, mpi_op, comm) );
+ return (colls::allreduce(Sendbuf, Recvbuf, count, mpi_datatype, mpi_op, comm));
else
- return( Colls::reduce(Sendbuf,Recvbuf, count,mpi_datatype,mpi_op, root, comm) );
+ return (colls::reduce(Sendbuf, Recvbuf, count, mpi_datatype, mpi_op, root, comm));
}
#endif /*REDUCE_LIMITS*/
-
-int Coll_reduce_rab::reduce(void* Sendbuf, void* Recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
+int reduce__rab(const void* Sendbuf, void* Recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
+ MPI_Comm comm)
{
- return( MPI_I_anyReduce(Sendbuf, Recvbuf, count, datatype, op, root, comm, 0) );
+ return MPI_I_anyReduce(Sendbuf, Recvbuf, count, datatype, op, root, comm, false);
}
-int Coll_allreduce_rab::allreduce(void* Sendbuf, void* Recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+int allreduce__rab(const void* Sendbuf, void* Recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
- return( MPI_I_anyReduce(Sendbuf, Recvbuf, count, datatype, op, -1, comm, 1) );
-}
-}
+ return MPI_I_anyReduce(Sendbuf, Recvbuf, count, datatype, op, -1, comm, true);
}
+} // namespace simgrid::smpi
