1 /* Copyright (c) 2009-2017. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "src/smpi/private.h"
7 #include "src/smpi/private.hpp"
8 #include "src/smpi/smpi_datatype.hpp"
9 #include "src/smpi/smpi_op.hpp"
10 #include "src/smpi/smpi_process.hpp"
12 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_op, smpi, "Logging specific to SMPI (op)");
14 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
15 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
16 #define SUM_OP(a, b) (b) += (a)
17 #define PROD_OP(a, b) (b) *= (a)
18 #define LAND_OP(a, b) (b) = (a) && (b)
19 #define LOR_OP(a, b) (b) = (a) || (b)
20 #define LXOR_OP(a, b) (b) = (not(a) && (b)) || ((a) && not(b))
21 #define BAND_OP(a, b) (b) &= (a)
22 #define BOR_OP(a, b) (b) |= (a)
23 #define BXOR_OP(a, b) (b) ^= (a)
24 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : ((a.value) == (b.value) ? ((a.index) < (b.index) ? (a) : (b)) : (a))
25 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : ((a.value) == (b.value) ? ((a.index) < (b.index) ? (a) : (b)) : (b))
27 #define APPLY_FUNC(a, b, length, type, func) \
30 type* x = (type*)(a); \
31 type* y = (type*)(b); \
32 for(i = 0; i < *(length); i++) { \
37 #define APPLY_OP_LOOP(dtype, type, op) \
38 if (*datatype == dtype) {\
39 APPLY_FUNC(a, b, length, type, op)\
43 #define APPLY_BASIC_OP_LOOP(op)\
44 APPLY_OP_LOOP(MPI_CHAR, char,op)\
45 APPLY_OP_LOOP(MPI_SHORT, short,op)\
46 APPLY_OP_LOOP(MPI_INT, int,op)\
47 APPLY_OP_LOOP(MPI_LONG, long,op)\
48 APPLY_OP_LOOP(MPI_LONG_LONG, long long,op)\
49 APPLY_OP_LOOP(MPI_SIGNED_CHAR, signed char,op)\
50 APPLY_OP_LOOP(MPI_UNSIGNED_CHAR, unsigned char,op)\
51 APPLY_OP_LOOP(MPI_UNSIGNED_SHORT, unsigned short,op)\
52 APPLY_OP_LOOP(MPI_UNSIGNED, unsigned int,op)\
53 APPLY_OP_LOOP(MPI_UNSIGNED_LONG, unsigned long,op)\
54 APPLY_OP_LOOP(MPI_UNSIGNED_LONG_LONG, unsigned long long,op)\
55 APPLY_OP_LOOP(MPI_WCHAR, wchar_t,op)\
56 APPLY_OP_LOOP(MPI_BYTE, int8_t,op)\
57 APPLY_OP_LOOP(MPI_INT8_T, int8_t,op)\
58 APPLY_OP_LOOP(MPI_INT16_T, int16_t,op)\
59 APPLY_OP_LOOP(MPI_INT32_T, int32_t,op)\
60 APPLY_OP_LOOP(MPI_INT64_T, int64_t,op)\
61 APPLY_OP_LOOP(MPI_UINT8_T, uint8_t,op)\
62 APPLY_OP_LOOP(MPI_UINT16_T, uint16_t,op)\
63 APPLY_OP_LOOP(MPI_UINT32_T, uint32_t,op)\
64 APPLY_OP_LOOP(MPI_UINT64_T, uint64_t,op)\
65 APPLY_OP_LOOP(MPI_AINT, MPI_Aint,op)\
66 APPLY_OP_LOOP(MPI_OFFSET, MPI_Offset,op)\
67 APPLY_OP_LOOP(MPI_INTEGER1, int,op)\
68 APPLY_OP_LOOP(MPI_INTEGER2, int16_t,op)\
69 APPLY_OP_LOOP(MPI_INTEGER4, int32_t,op)\
70 APPLY_OP_LOOP(MPI_INTEGER8, int64_t,op)
72 #define APPLY_BOOL_OP_LOOP(op)\
73 APPLY_OP_LOOP(MPI_C_BOOL, bool,op)
75 #define APPLY_FLOAT_OP_LOOP(op)\
76 APPLY_OP_LOOP(MPI_FLOAT, float,op)\
77 APPLY_OP_LOOP(MPI_DOUBLE, double,op)\
78 APPLY_OP_LOOP(MPI_LONG_DOUBLE, long double,op)\
79 APPLY_OP_LOOP(MPI_REAL, float,op)\
80 APPLY_OP_LOOP(MPI_REAL4, float,op)\
81 APPLY_OP_LOOP(MPI_REAL8, float,op)\
82 APPLY_OP_LOOP(MPI_REAL16, double,op)
84 #define APPLY_COMPLEX_OP_LOOP(op)\
85 APPLY_OP_LOOP(MPI_C_FLOAT_COMPLEX, float _Complex,op)\
86 APPLY_OP_LOOP(MPI_C_DOUBLE_COMPLEX, double _Complex,op)\
87 APPLY_OP_LOOP(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex,op)
89 #define APPLY_PAIR_OP_LOOP(op)\
90 APPLY_OP_LOOP(MPI_FLOAT_INT, float_int,op)\
91 APPLY_OP_LOOP(MPI_LONG_INT, long_int,op)\
92 APPLY_OP_LOOP(MPI_DOUBLE_INT, double_int,op)\
93 APPLY_OP_LOOP(MPI_SHORT_INT, short_int,op)\
94 APPLY_OP_LOOP(MPI_2INT, int_int,op)\
95 APPLY_OP_LOOP(MPI_2FLOAT, float_float,op)\
96 APPLY_OP_LOOP(MPI_2DOUBLE, double_double,op)\
97 APPLY_OP_LOOP(MPI_LONG_DOUBLE_INT, long_double_int,op)\
98 APPLY_OP_LOOP(MPI_2LONG, long_long,op)
100 #define APPLY_END_OP_LOOP(op)\
102 xbt_die("Failed to apply " #op " to type %s", (*datatype)->name());\
105 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
107 APPLY_BASIC_OP_LOOP(MAX_OP)
108 APPLY_FLOAT_OP_LOOP(MAX_OP)
109 APPLY_END_OP_LOOP(MAX_OP)
112 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
114 APPLY_BASIC_OP_LOOP(MIN_OP)
115 APPLY_FLOAT_OP_LOOP(MIN_OP)
116 APPLY_END_OP_LOOP(MIN_OP)
119 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
121 APPLY_BASIC_OP_LOOP(SUM_OP)
122 APPLY_FLOAT_OP_LOOP(SUM_OP)
123 APPLY_COMPLEX_OP_LOOP(SUM_OP)
124 APPLY_END_OP_LOOP(SUM_OP)
127 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
129 APPLY_BASIC_OP_LOOP(PROD_OP)
130 APPLY_FLOAT_OP_LOOP(PROD_OP)
131 APPLY_COMPLEX_OP_LOOP(PROD_OP)
132 APPLY_END_OP_LOOP(PROD_OP)
135 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
137 APPLY_BASIC_OP_LOOP(LAND_OP)
138 APPLY_BOOL_OP_LOOP(LAND_OP)
139 APPLY_END_OP_LOOP(LAND_OP)
142 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
144 APPLY_BASIC_OP_LOOP(LOR_OP)
145 APPLY_BOOL_OP_LOOP(LOR_OP)
146 APPLY_END_OP_LOOP(LOR_OP)
149 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
151 APPLY_BASIC_OP_LOOP(LXOR_OP)
152 APPLY_BOOL_OP_LOOP(LXOR_OP)
153 APPLY_END_OP_LOOP(LXOR_OP)
156 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
158 APPLY_BASIC_OP_LOOP(BAND_OP)
159 APPLY_BOOL_OP_LOOP(BAND_OP)
160 APPLY_END_OP_LOOP(BAND_OP)
163 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
165 APPLY_BASIC_OP_LOOP(BOR_OP)
166 APPLY_BOOL_OP_LOOP(BOR_OP)
167 APPLY_END_OP_LOOP(BOR_OP)
170 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
172 APPLY_BASIC_OP_LOOP(BXOR_OP)
173 APPLY_BOOL_OP_LOOP(BXOR_OP)
174 APPLY_END_OP_LOOP(BXOR_OP)
177 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
179 APPLY_PAIR_OP_LOOP(MINLOC_OP)
180 APPLY_END_OP_LOOP(MINLOC_OP)
183 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
185 APPLY_PAIR_OP_LOOP(MAXLOC_OP)
186 APPLY_END_OP_LOOP(MAXLOC_OP)
189 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
191 memcpy(b, a, *length * (*datatype)->size());
194 static void no_func(void *a, void *b, int *length, MPI_Datatype * datatype)
196 /* obviously a no-op */
199 #define CREATE_MPI_OP(name, func) \
200 static SMPI_Op mpi_##name (&(func) /* func */, true ); \
201 MPI_Op name = &mpi_##name;
203 CREATE_MPI_OP(MPI_MAX, max_func);
204 CREATE_MPI_OP(MPI_MIN, min_func);
205 CREATE_MPI_OP(MPI_SUM, sum_func);
206 CREATE_MPI_OP(MPI_PROD, prod_func);
207 CREATE_MPI_OP(MPI_LAND, land_func);
208 CREATE_MPI_OP(MPI_LOR, lor_func);
209 CREATE_MPI_OP(MPI_LXOR, lxor_func);
210 CREATE_MPI_OP(MPI_BAND, band_func);
211 CREATE_MPI_OP(MPI_BOR, bor_func);
212 CREATE_MPI_OP(MPI_BXOR, bxor_func);
213 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
214 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
215 CREATE_MPI_OP(MPI_REPLACE, replace_func);
216 CREATE_MPI_OP(MPI_NO_OP, no_func);
221 Op::Op(MPI_User_function * function, bool commutative) : func_(function), is_commutative_(commutative)
223 is_fortran_op_ = false;
226 bool Op::is_commutative()
228 return is_commutative_;
231 bool Op::is_fortran_op()
233 return is_fortran_op_;
236 void Op::set_fortran_op()
238 //tell that we were created from fortran, so we need to translate the type to fortran when called
239 is_fortran_op_ = true;
242 void Op::apply(void *invec, void *inoutvec, int *len, MPI_Datatype datatype)
244 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){//we need to switch as the called function may silently touch global variables
245 XBT_DEBUG("Applying operation, switch to the right data frame ");
246 smpi_switch_data_segment(smpi_process()->index());
249 if (not smpi_process()->replaying() && *len > 0) {
250 if (not is_fortran_op_)
251 this->func_(invec, inoutvec, len, &datatype);
253 XBT_DEBUG("Applying operation of length %d from %p and from/to %p", *len, invec, inoutvec);
254 int tmp = datatype->c2f();
255 /* Unfortunately, the C and Fortran version of the MPI standard do not agree on the type here,
256 thus the reinterpret_cast. */
257 this->func_(invec, inoutvec, len, reinterpret_cast<MPI_Datatype*>(&tmp) );
263 return static_cast<Op*>(F2C::f2c(id));