1 /* Copyright (c) 2009-2021. 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. */
8 #include "smpi_datatype.hpp"
9 #include "src/smpi/include/smpi_actor.hpp"
11 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_op, smpi, "Logging specific to SMPI (op)");
13 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
14 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
15 #define SUM_OP(a, b) (b) += (a)
16 #define SUM_OP_COMPLEX(a, b) \
18 ((b).value) += ((a).value); \
19 ((b).index) += ((a).index); \
21 #define PROD_OP(a, b) (b) *= (a)
22 #define PROD_OP_COMPLEX(a, b) \
24 ((b).value) *= ((a).value); \
25 ((b).index) *= ((a).index); \
27 #define LAND_OP(a, b) (b) = (a) && (b)
28 #define LOR_OP(a, b) (b) = (a) || (b)
29 #define LXOR_OP(a, b) (b) = bool(a) != bool(b)
30 #define BAND_OP(a, b) (b) &= (a)
31 #define BOR_OP(a, b) (b) |= (a)
32 #define BXOR_OP(a, b) (b) ^= (a)
33 #define MAXLOC_OP(a, b) \
34 (b) = ((a).value) < ((b).value) ? (b) : (((a).value) == ((b).value) ? (((a).index) < ((b).index) ? (a) : (b)) : (a))
35 #define MINLOC_OP(a, b) \
36 (b) = ((a).value) < ((b).value) ? (a) : (((a).value) == ((b).value) ? (((a).index) < ((b).index) ? (a) : (b)) : (b))
38 #define APPLY_FUNC(a, b, length, type, func) \
41 type* x = (type*)(a); \
42 type* y = (type*)(b); \
43 for(i = 0; i < *(length); i++) { \
48 #define APPLY_OP_LOOP(dtype, type, op) \
49 if (*datatype == (dtype)) { \
50 APPLY_FUNC(a, b, length, type, op) \
53 #define APPLY_BASIC_OP_LOOP(op)\
54 APPLY_OP_LOOP(MPI_CHAR, char,op)\
55 APPLY_OP_LOOP(MPI_SHORT, short,op)\
56 APPLY_OP_LOOP(MPI_INT, int,op)\
57 APPLY_OP_LOOP(MPI_LONG, long,op)\
58 APPLY_OP_LOOP(MPI_LONG_LONG, long long,op)\
59 APPLY_OP_LOOP(MPI_SIGNED_CHAR, signed char,op)\
60 APPLY_OP_LOOP(MPI_UNSIGNED_CHAR, unsigned char,op)\
61 APPLY_OP_LOOP(MPI_UNSIGNED_SHORT, unsigned short,op)\
62 APPLY_OP_LOOP(MPI_UNSIGNED, unsigned int,op)\
63 APPLY_OP_LOOP(MPI_UNSIGNED_LONG, unsigned long,op)\
64 APPLY_OP_LOOP(MPI_UNSIGNED_LONG_LONG, unsigned long long,op)\
65 APPLY_OP_LOOP(MPI_WCHAR, wchar_t,op)\
66 APPLY_OP_LOOP(MPI_INT8_T, int8_t,op)\
67 APPLY_OP_LOOP(MPI_INT16_T, int16_t,op)\
68 APPLY_OP_LOOP(MPI_INT32_T, int32_t,op)\
69 APPLY_OP_LOOP(MPI_INT64_T, int64_t,op)\
70 APPLY_OP_LOOP(MPI_UINT8_T, uint8_t,op)\
71 APPLY_OP_LOOP(MPI_UINT16_T, uint16_t,op)\
72 APPLY_OP_LOOP(MPI_UINT32_T, uint32_t,op)\
73 APPLY_OP_LOOP(MPI_UINT64_T, uint64_t,op)\
74 APPLY_OP_LOOP(MPI_AINT, MPI_Aint,op)\
75 APPLY_OP_LOOP(MPI_OFFSET, MPI_Offset,op)\
76 APPLY_OP_LOOP(MPI_INTEGER1, int,op)\
77 APPLY_OP_LOOP(MPI_INTEGER2, int16_t,op)\
78 APPLY_OP_LOOP(MPI_INTEGER4, int32_t,op)\
79 APPLY_OP_LOOP(MPI_INTEGER8, int64_t,op)\
80 APPLY_OP_LOOP(MPI_COUNT, long long,op)
83 #define APPLY_BOOL_OP_LOOP(op)\
84 APPLY_OP_LOOP(MPI_C_BOOL, bool,op)
86 #define APPLY_BYTE_OP_LOOP(op)\
87 APPLY_OP_LOOP(MPI_BYTE, int8_t,op)
89 #define APPLY_FLOAT_OP_LOOP(op)\
90 APPLY_OP_LOOP(MPI_FLOAT, float,op)\
91 APPLY_OP_LOOP(MPI_DOUBLE, double,op)\
92 APPLY_OP_LOOP(MPI_LONG_DOUBLE, long double,op)\
93 APPLY_OP_LOOP(MPI_REAL, float,op)\
94 APPLY_OP_LOOP(MPI_REAL4, float,op)\
95 APPLY_OP_LOOP(MPI_REAL8, double,op)\
96 APPLY_OP_LOOP(MPI_REAL16, long double,op)
98 #define APPLY_COMPLEX_OP_LOOP(op)\
99 APPLY_OP_LOOP(MPI_C_FLOAT_COMPLEX, float _Complex,op)\
100 APPLY_OP_LOOP(MPI_C_DOUBLE_COMPLEX, double _Complex,op)\
101 APPLY_OP_LOOP(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex,op)
103 #define APPLY_PAIR_OP_LOOP(op)\
104 APPLY_OP_LOOP(MPI_FLOAT_INT, float_int,op)\
105 APPLY_OP_LOOP(MPI_LONG_INT, long_int,op)\
106 APPLY_OP_LOOP(MPI_DOUBLE_INT, double_int,op)\
107 APPLY_OP_LOOP(MPI_SHORT_INT, short_int,op)\
108 APPLY_OP_LOOP(MPI_2INT, int_int,op)\
109 APPLY_OP_LOOP(MPI_2FLOAT, float_float,op)\
110 APPLY_OP_LOOP(MPI_2DOUBLE, double_double,op)\
111 APPLY_OP_LOOP(MPI_LONG_DOUBLE_INT, long_double_int,op)\
112 APPLY_OP_LOOP(MPI_2LONG, long_long,op)\
113 APPLY_OP_LOOP(MPI_COMPLEX8, float_float,op)\
114 APPLY_OP_LOOP(MPI_COMPLEX16, double_double,op)\
115 APPLY_OP_LOOP(MPI_COMPLEX32, double_double,op)
117 #define APPLY_END_OP_LOOP(op) \
119 xbt_die("Failed to apply " _XBT_STRINGIFY(op) " to type %s", (*datatype)->name()); \
122 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
124 APPLY_BASIC_OP_LOOP(MAX_OP)
125 APPLY_FLOAT_OP_LOOP(MAX_OP)
126 APPLY_END_OP_LOOP(MAX_OP)
129 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
131 APPLY_BASIC_OP_LOOP(MIN_OP)
132 APPLY_FLOAT_OP_LOOP(MIN_OP)
133 APPLY_END_OP_LOOP(MIN_OP)
136 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
138 APPLY_BASIC_OP_LOOP(SUM_OP)
139 APPLY_FLOAT_OP_LOOP(SUM_OP)
140 APPLY_COMPLEX_OP_LOOP(SUM_OP)
141 APPLY_PAIR_OP_LOOP(SUM_OP_COMPLEX)
142 APPLY_END_OP_LOOP(SUM_OP)
145 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
147 APPLY_BASIC_OP_LOOP(PROD_OP)
148 APPLY_FLOAT_OP_LOOP(PROD_OP)
149 APPLY_COMPLEX_OP_LOOP(PROD_OP)
150 APPLY_PAIR_OP_LOOP(PROD_OP_COMPLEX)
151 APPLY_END_OP_LOOP(PROD_OP)
154 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
156 APPLY_BASIC_OP_LOOP(LAND_OP)
157 APPLY_FLOAT_OP_LOOP(LAND_OP)
158 APPLY_BOOL_OP_LOOP(LAND_OP)
159 APPLY_END_OP_LOOP(LAND_OP)
162 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
164 APPLY_BASIC_OP_LOOP(LOR_OP)
165 APPLY_FLOAT_OP_LOOP(LOR_OP)
166 APPLY_BOOL_OP_LOOP(LOR_OP)
167 APPLY_END_OP_LOOP(LOR_OP)
170 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
172 APPLY_BASIC_OP_LOOP(LXOR_OP)
173 APPLY_FLOAT_OP_LOOP(LXOR_OP)
174 APPLY_BOOL_OP_LOOP(LXOR_OP)
175 APPLY_END_OP_LOOP(LXOR_OP)
178 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
180 APPLY_BASIC_OP_LOOP(BAND_OP)
181 APPLY_BOOL_OP_LOOP(BAND_OP)
182 APPLY_BYTE_OP_LOOP(BAND_OP)
183 APPLY_END_OP_LOOP(BAND_OP)
186 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
188 APPLY_BASIC_OP_LOOP(BOR_OP)
189 APPLY_BOOL_OP_LOOP(BOR_OP)
190 APPLY_BYTE_OP_LOOP(BOR_OP)
191 APPLY_END_OP_LOOP(BOR_OP)
194 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
196 APPLY_BASIC_OP_LOOP(BXOR_OP)
197 APPLY_BOOL_OP_LOOP(BXOR_OP)
198 APPLY_BYTE_OP_LOOP(BXOR_OP)
199 APPLY_END_OP_LOOP(BXOR_OP)
202 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
204 APPLY_PAIR_OP_LOOP(MINLOC_OP)
205 APPLY_END_OP_LOOP(MINLOC_OP)
208 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
210 APPLY_PAIR_OP_LOOP(MAXLOC_OP)
211 APPLY_END_OP_LOOP(MAXLOC_OP)
214 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
216 memcpy(b, a, *length * (*datatype)->size());
219 static void no_func(void*, void*, int*, MPI_Datatype*)
221 /* obviously a no-op */
225 #define CREATE_MPI_OP(name, func, types) \
226 SMPI_Op _XBT_CONCAT(smpi_MPI_, name)(&(func) /* func */, true, true, types);
228 #define MAX_TYPES DT_FLAG_C_INTEGER|DT_FLAG_F_INTEGER|DT_FLAG_FP|DT_FLAG_MULTILANG
229 #define LAND_TYPES DT_FLAG_C_INTEGER|DT_FLAG_FP|DT_FLAG_LOGICAL|DT_FLAG_MULTILANG
230 #define BAND_TYPES DT_FLAG_C_INTEGER|DT_FLAG_F_INTEGER|DT_FLAG_BYTE|DT_FLAG_MULTILANG
232 CREATE_MPI_OP(MAX, max_func, MAX_TYPES)
233 CREATE_MPI_OP(MIN, min_func, MAX_TYPES)
234 CREATE_MPI_OP(SUM, sum_func, MAX_TYPES|DT_FLAG_COMPLEX)
235 CREATE_MPI_OP(PROD, prod_func, MAX_TYPES|DT_FLAG_COMPLEX)
236 CREATE_MPI_OP(LAND, land_func, LAND_TYPES)
237 CREATE_MPI_OP(LOR, lor_func, LAND_TYPES)
238 CREATE_MPI_OP(LXOR, lxor_func, LAND_TYPES)
239 CREATE_MPI_OP(BAND, band_func, BAND_TYPES)
240 CREATE_MPI_OP(BOR, bor_func, BAND_TYPES)
241 CREATE_MPI_OP(BXOR, bxor_func, BAND_TYPES)
242 CREATE_MPI_OP(MAXLOC, maxloc_func, DT_FLAG_REDUCTION)
243 CREATE_MPI_OP(MINLOC, minloc_func, DT_FLAG_REDUCTION)
244 CREATE_MPI_OP(REPLACE, replace_func, 0)
245 CREATE_MPI_OP(NO_OP, no_func, 0)
250 void Op::apply(const void* invec, void* inoutvec, const int* len, MPI_Datatype datatype) const
252 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
253 // we need to switch as the called function may silently touch global variables
254 XBT_DEBUG("Applying operation, switch to the right data frame ");
255 smpi_switch_data_segment(simgrid::s4u::Actor::self());
258 if (not smpi_process()->replaying() && *len > 0) {
259 if (not is_fortran_op_)
260 this->func_(const_cast<void*>(invec), inoutvec, const_cast<int*>(len), &datatype);
262 XBT_DEBUG("Applying operation of length %d from %p and from/to %p", *len, invec, inoutvec);
263 int tmp = datatype->c2f();
264 /* Unfortunately, the C and Fortran version of the MPI standard do not agree on the type here,
265 thus the reinterpret_cast. */
266 this->func_(const_cast<void*>(invec), inoutvec, const_cast<int*>(len), reinterpret_cast<MPI_Datatype*>(&tmp));
272 return static_cast<Op*>(F2C::f2c(id));
279 void Op::unref(MPI_Op* op){
280 if((*op)!=MPI_OP_NULL){
282 if ((*op)->refcount_ == 0 && not (*op)->is_predefined_){
283 F2C::free_f((*op)->c2f());