1 /* smpi_datatype.cpp -- MPI primitives to handle datatypes */
2 /* Copyright (c) 2009-2020. The SimGrid Team. All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 #include "simgrid/modelchecker.h"
9 #include "smpi_datatype_derived.hpp"
10 #include "smpi_op.hpp"
11 #include "src/instr/instr_private.hpp"
12 #include "src/smpi/include/smpi_actor.hpp"
16 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_datatype, smpi, "Logging specific to SMPI (datatype)");
18 static std::unordered_map<std::string, simgrid::smpi::Datatype*> id2type_lookup;
20 #define CREATE_MPI_DATATYPE(name, id, type) \
21 static simgrid::smpi::Datatype _XBT_CONCAT(mpi_, name)((char*)_XBT_STRINGIFY(name), (id), sizeof(type), /* size */ \
23 sizeof(type), /* ub = lb + size */ \
24 DT_FLAG_BASIC /* flags */ \
26 const MPI_Datatype name = &_XBT_CONCAT(mpi_, name);
28 #define CREATE_MPI_DATATYPE_NULL(name, id) \
29 static simgrid::smpi::Datatype _XBT_CONCAT(mpi_, name)((char*)_XBT_STRINGIFY(name), (id), 0, /* size */ \
31 0, /* ub = lb + size */ \
32 DT_FLAG_BASIC /* flags */ \
34 const MPI_Datatype name = &_XBT_CONCAT(mpi_, name);
36 // Predefined data types
37 CREATE_MPI_DATATYPE(MPI_CHAR, 2, char)
38 CREATE_MPI_DATATYPE(MPI_SHORT, 3, short)
39 CREATE_MPI_DATATYPE(MPI_INT, 1, int)
40 CREATE_MPI_DATATYPE(MPI_LONG, 4, long)
41 CREATE_MPI_DATATYPE(MPI_LONG_LONG, 7, long long)
42 CREATE_MPI_DATATYPE(MPI_SIGNED_CHAR, 8, signed char)
43 CREATE_MPI_DATATYPE(MPI_UNSIGNED_CHAR, 9, unsigned char)
44 CREATE_MPI_DATATYPE(MPI_UNSIGNED_SHORT, 10, unsigned short)
45 CREATE_MPI_DATATYPE(MPI_UNSIGNED, 11, unsigned int)
46 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG, 12, unsigned long)
47 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG_LONG, 13, unsigned long long)
48 CREATE_MPI_DATATYPE(MPI_FLOAT, 5, float)
49 CREATE_MPI_DATATYPE(MPI_DOUBLE, 0, double)
50 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, 14, long double)
51 CREATE_MPI_DATATYPE(MPI_WCHAR, 15, wchar_t)
52 CREATE_MPI_DATATYPE(MPI_C_BOOL, 16, bool)
53 CREATE_MPI_DATATYPE(MPI_BYTE, 6, int8_t)
54 CREATE_MPI_DATATYPE(MPI_INT8_T, 17, int8_t)
55 CREATE_MPI_DATATYPE(MPI_INT16_T, 18, int16_t)
56 CREATE_MPI_DATATYPE(MPI_INT32_T, 19, int32_t)
57 CREATE_MPI_DATATYPE(MPI_INT64_T, 20, int64_t)
58 CREATE_MPI_DATATYPE(MPI_UINT8_T, 21, uint8_t)
59 CREATE_MPI_DATATYPE(MPI_UINT16_T, 22, uint16_t)
60 CREATE_MPI_DATATYPE(MPI_UINT32_T, 23, uint32_t)
61 CREATE_MPI_DATATYPE(MPI_UINT64_T, 24, uint64_t)
62 CREATE_MPI_DATATYPE(MPI_C_FLOAT_COMPLEX, 25, float _Complex)
63 CREATE_MPI_DATATYPE(MPI_C_DOUBLE_COMPLEX, 26, double _Complex)
64 CREATE_MPI_DATATYPE(MPI_C_LONG_DOUBLE_COMPLEX, 27, long double _Complex)
65 CREATE_MPI_DATATYPE(MPI_AINT, 28, MPI_Aint)
66 CREATE_MPI_DATATYPE(MPI_OFFSET, 29, MPI_Offset)
68 CREATE_MPI_DATATYPE(MPI_FLOAT_INT, 30, float_int)
69 CREATE_MPI_DATATYPE(MPI_LONG_INT, 31, long_int)
70 CREATE_MPI_DATATYPE(MPI_DOUBLE_INT, 32, double_int)
71 CREATE_MPI_DATATYPE(MPI_SHORT_INT, 33, short_int)
72 CREATE_MPI_DATATYPE(MPI_2INT, 34, int_int)
73 CREATE_MPI_DATATYPE(MPI_2FLOAT, 35, float_float)
74 CREATE_MPI_DATATYPE(MPI_2DOUBLE, 36, double_double)
75 CREATE_MPI_DATATYPE(MPI_2LONG, 37, long_long)
77 CREATE_MPI_DATATYPE(MPI_REAL, 38, float)
78 CREATE_MPI_DATATYPE(MPI_REAL4, 39, float)
79 CREATE_MPI_DATATYPE(MPI_REAL8, 40, double)
80 CREATE_MPI_DATATYPE(MPI_REAL16, 41, long double)
81 CREATE_MPI_DATATYPE_NULL(MPI_DATATYPE_NULL, -1)
82 CREATE_MPI_DATATYPE(MPI_COMPLEX8, 42, float_float)
83 CREATE_MPI_DATATYPE(MPI_COMPLEX16, 43, double_double)
84 CREATE_MPI_DATATYPE(MPI_COMPLEX32, 44, double_double)
85 CREATE_MPI_DATATYPE(MPI_INTEGER1, 45, int)
86 CREATE_MPI_DATATYPE(MPI_INTEGER2, 46, int16_t)
87 CREATE_MPI_DATATYPE(MPI_INTEGER4, 47, int32_t)
88 CREATE_MPI_DATATYPE(MPI_INTEGER8, 48, int64_t)
89 CREATE_MPI_DATATYPE(MPI_INTEGER16, 49, integer128_t)
91 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, 50, long_double_int)
93 CREATE_MPI_DATATYPE_NULL(MPI_UB, 51)
94 CREATE_MPI_DATATYPE_NULL(MPI_LB, 52)
95 CREATE_MPI_DATATYPE(MPI_PACKED, 53, char)
97 CREATE_MPI_DATATYPE(MPI_PTR, 54, void*)
98 CREATE_MPI_DATATYPE(MPI_COUNT, 55, long long)
104 std::unordered_map<int, smpi_key_elem> Datatype::keyvals_; // required by the Keyval class implementation
105 int Datatype::keyval_id_=0; // required by the Keyval class implementation
106 Datatype::Datatype(int ident, int size, MPI_Aint lb, MPI_Aint ub, int flags) : Datatype(size, lb, ub, flags)
108 id = std::to_string(ident);
111 Datatype::Datatype(int size, MPI_Aint lb, MPI_Aint ub, int flags) : size_(size), lb_(lb), ub_(ub), flags_(flags)
115 MC_ignore(&(refcount_), sizeof(refcount_));
119 // for predefined types, so refcount_ = 0.
120 Datatype::Datatype(char* name, int ident, int size, MPI_Aint lb, MPI_Aint ub, int flags)
121 : name_(name), id(std::to_string(ident)), size_(size), lb_(lb), ub_(ub), flags_(flags), refcount_(0)
123 id2type_lookup.insert({id, this});
126 MC_ignore(&(refcount_), sizeof(refcount_));
130 Datatype::Datatype(Datatype* datatype, int* ret)
131 : size_(datatype->size_), lb_(datatype->lb_), ub_(datatype->ub_), flags_(datatype->flags_)
133 flags_ &= ~DT_FLAG_PREDEFINED;
136 if (not datatype->attributes()->empty()) {
139 for (auto const& it : *(datatype->attributes())) {
140 smpi_key_elem elem = keyvals_.at(it.first);
141 if (elem != nullptr){
142 if( elem->copy_fn.type_copy_fn != MPI_NULL_COPY_FN &&
143 elem->copy_fn.type_copy_fn != MPI_TYPE_DUP_FN)
144 *ret = elem->copy_fn.type_copy_fn(datatype, it.first, elem->extra_state, it.second, &value_out, &flag);
145 else if ( elem->copy_fn.type_copy_fn_fort != MPI_NULL_COPY_FN &&
146 (*(int*)*elem->copy_fn.type_copy_fn_fort) != 1){
147 value_out=(int*)xbt_malloc(sizeof(int));
148 elem->copy_fn.type_copy_fn_fort(datatype, it.first, elem->extra_state, it.second, value_out, &flag,ret);
150 if (*ret != MPI_SUCCESS) {
153 if(elem->copy_fn.type_copy_fn == MPI_TYPE_DUP_FN ||
154 ((elem->copy_fn.type_copy_fn_fort != MPI_NULL_COPY_FN) && (*(int*)*elem->copy_fn.type_copy_fn_fort == 1))){
156 attributes()->insert({it.first, it.second});
159 attributes()->insert({it.first, value_out});
166 Datatype::~Datatype()
168 xbt_assert(refcount_ >= 0);
170 if(flags_ & DT_FLAG_PREDEFINED)
173 //if still used, mark for deletion
175 flags_ |=DT_FLAG_DESTROYED;
179 cleanup_attr<Datatype>();
190 MC_ignore(&(refcount_), sizeof(refcount_));
194 void Datatype::unref(MPI_Datatype datatype)
196 if (datatype->refcount_ > 0)
197 datatype->refcount_--;
199 if (datatype->refcount_ == 0 && not(datatype->flags_ & DT_FLAG_PREDEFINED))
204 MC_ignore(&(datatype->refcount_), sizeof(datatype->refcount_));
208 void Datatype::commit()
210 flags_ |= DT_FLAG_COMMITED;
213 bool Datatype::is_valid(){
214 return (flags_ & DT_FLAG_COMMITED);
217 bool Datatype::is_basic()
219 return (flags_ & DT_FLAG_BASIC);
222 bool Datatype::is_replayable()
224 return (simgrid::instr::trace_format == simgrid::instr::TraceFormat::Ti) &&
225 ((this == MPI_BYTE) || (this == MPI_DOUBLE) || (this == MPI_INT) || (this == MPI_CHAR) ||
226 (this == MPI_SHORT) || (this == MPI_LONG) || (this == MPI_FLOAT));
229 MPI_Datatype Datatype::decode(const std::string& datatype_id)
231 return id2type_lookup.find(datatype_id)->second;
234 void Datatype::addflag(int flag){
238 int Datatype::extent(MPI_Aint * lb, MPI_Aint * extent){
244 void Datatype::get_name(char* name, int* length){
246 *length = strlen(name_);
247 strncpy(name, name_, *length+1);
253 void Datatype::set_name(const char* name){
254 if(name_!=nullptr && (flags_ & DT_FLAG_PREDEFINED) == 0)
256 name_ = xbt_strdup(name);
259 int Datatype::pack(const void* inbuf, int incount, void* outbuf, int outcount, int* position, const Comm*)
261 if (outcount - *position < incount*static_cast<int>(size_))
262 return MPI_ERR_OTHER;
263 Datatype::copy(inbuf, incount, this, static_cast<char*>(outbuf) + *position, outcount, MPI_CHAR);
264 *position += incount * size_;
268 int Datatype::unpack(const void* inbuf, int insize, int* position, void* outbuf, int outcount, const Comm*)
270 if (outcount*static_cast<int>(size_)> insize)
271 return MPI_ERR_OTHER;
272 Datatype::copy(static_cast<const char*>(inbuf) + *position, insize, MPI_CHAR, outbuf, outcount, this);
273 *position += outcount * size_;
277 int Datatype::copy(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
278 MPI_Datatype recvtype)
280 // FIXME Handle the case of a partial shared malloc.
282 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
283 smpi_switch_data_segment(simgrid::s4u::Actor::self());
285 /* First check if we really have something to do */
287 std::vector<std::pair<size_t, size_t>> private_blocks;
288 if(smpi_is_shared(sendbuf,private_blocks,&offset)
289 && (private_blocks.size()==1
290 && (private_blocks[0].second - private_blocks[0].first)==(unsigned long)(sendcount * sendtype->get_extent()))){
291 XBT_VERB("sendbuf is shared. Ignoring copies");
294 if(smpi_is_shared(recvbuf,private_blocks,&offset)
295 && (private_blocks.size()==1
296 && (private_blocks[0].second - private_blocks[0].first)==(unsigned long)(recvcount * recvtype->get_extent()))){
297 XBT_VERB("recvbuf is shared. Ignoring copies");
301 if (recvcount > 0 && recvbuf != sendbuf) {
302 sendcount *= sendtype->size();
303 recvcount *= recvtype->size();
304 int count = sendcount < recvcount ? sendcount : recvcount;
305 XBT_DEBUG("Copying %d bytes from %p to %p", count, sendbuf, recvbuf);
306 if (not(sendtype->flags() & DT_FLAG_DERIVED) && not(recvtype->flags() & DT_FLAG_DERIVED)) {
307 if (not smpi_process()->replaying())
308 memcpy(recvbuf, sendbuf, count);
309 } else if (not(sendtype->flags() & DT_FLAG_DERIVED)) {
310 recvtype->unserialize(sendbuf, recvbuf, count / recvtype->size(), MPI_REPLACE);
311 } else if (not(recvtype->flags() & DT_FLAG_DERIVED)) {
312 sendtype->serialize(sendbuf, recvbuf, count / sendtype->size());
314 void * buf_tmp = xbt_malloc(count);
316 sendtype->serialize( sendbuf, buf_tmp,count/sendtype->size());
317 recvtype->unserialize( buf_tmp, recvbuf,count/recvtype->size(), MPI_REPLACE);
323 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
326 //Default serialization method : memcpy.
327 void Datatype::serialize(const void* noncontiguous_buf, void* contiguous_buf, int count)
329 char* contiguous_buf_char = static_cast<char*>(contiguous_buf);
330 const char* noncontiguous_buf_char = static_cast<const char*>(noncontiguous_buf)+lb_;
331 memcpy(contiguous_buf_char, noncontiguous_buf_char, count*size_);
334 void Datatype::unserialize(const void* contiguous_buf, void *noncontiguous_buf, int count, MPI_Op op){
335 const char* contiguous_buf_char = static_cast<const char*>(contiguous_buf);
336 char* noncontiguous_buf_char = static_cast<char*>(noncontiguous_buf)+lb_;
339 op->apply( contiguous_buf_char, noncontiguous_buf_char, &n, this);
342 int Datatype::create_contiguous(int count, MPI_Datatype old_type, MPI_Aint lb, MPI_Datatype* new_type){
343 if(old_type->flags_ & DT_FLAG_DERIVED){
344 //handle this case as a hvector with stride equals to the extent of the datatype
345 return create_hvector(count, 1, old_type->get_extent(), old_type, new_type);
348 *new_type = new Type_Contiguous(count * old_type->size(), lb, lb + count * old_type->size(),
349 DT_FLAG_DERIVED, count, old_type);
351 *new_type = new Datatype(count * old_type->size(), lb, lb + count * old_type->size(),0);
355 int Datatype::create_vector(int count, int block_length, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
364 ub=((count-1)*stride+block_length-1)*old_type->get_extent()+old_type->ub();
366 if(old_type->flags() & DT_FLAG_DERIVED || stride != block_length){
367 *new_type = new Type_Vector(count * (block_length) * old_type->size(), lb, ub,
368 DT_FLAG_DERIVED, count, block_length, stride, old_type);
371 /* in this situation the data are contiguous thus it's not required to serialize and unserialize it*/
372 *new_type = new Datatype(count * block_length * old_type->size(), 0, ((count -1) * stride + block_length)*
373 old_type->size(), DT_FLAG_CONTIGUOUS);
380 int Datatype::create_hvector(int count, int block_length, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
389 ub=((count-1)*stride)+(block_length-1)*old_type->get_extent()+old_type->ub();
391 if(old_type->flags() & DT_FLAG_DERIVED || stride != block_length*old_type->get_extent()){
392 *new_type = new Type_Hvector(count * (block_length) * old_type->size(), lb, ub,
393 DT_FLAG_DERIVED, count, block_length, stride, old_type);
396 /* in this situation the data are contiguous thus it's not required to serialize and unserialize it*/
397 *new_type = new Datatype(count * block_length * old_type->size(), 0, count * block_length * old_type->size(), DT_FLAG_CONTIGUOUS);
403 int Datatype::create_indexed(int count, const int* block_lengths, const int* indices, MPI_Datatype old_type, MPI_Datatype* new_type){
405 bool contiguous=true;
409 lb=indices[0]*old_type->get_extent();
410 ub=indices[0]*old_type->get_extent() + block_lengths[0]*old_type->ub();
413 for (int i = 0; i < count; i++) {
414 if (block_lengths[i] < 0)
416 size += block_lengths[i];
418 if(indices[i]*old_type->get_extent()+old_type->lb()<lb)
419 lb = indices[i]*old_type->get_extent()+old_type->lb();
420 if(indices[i]*old_type->get_extent()+block_lengths[i]*old_type->ub()>ub)
421 ub = indices[i]*old_type->get_extent()+block_lengths[i]*old_type->ub();
423 if ( (i< count -1) && (indices[i]+block_lengths[i] != indices[i+1]) )
426 if(old_type->flags_ & DT_FLAG_DERIVED)
429 if (not contiguous) {
430 *new_type = new Type_Indexed(size * old_type->size(),lb,ub,
431 DT_FLAG_DERIVED|DT_FLAG_DATA, count, block_lengths, indices, old_type);
433 Datatype::create_contiguous(size, old_type, lb, new_type);
438 int Datatype::create_hindexed(int count, const int* block_lengths, const MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type){
440 bool contiguous=true;
444 lb=indices[0] + old_type->lb();
445 ub=indices[0] + block_lengths[0]*old_type->ub();
447 for (int i = 0; i < count; i++) {
448 if (block_lengths[i] < 0)
450 size += block_lengths[i];
452 if(indices[i]+old_type->lb()<lb)
453 lb = indices[i]+old_type->lb();
454 if(indices[i]+block_lengths[i]*old_type->ub()>ub)
455 ub = indices[i]+block_lengths[i]*old_type->ub();
457 if ( (i< count -1) && (indices[i]+block_lengths[i]*(static_cast<int>(old_type->size())) != indices[i+1]) )
460 if (old_type->flags_ & DT_FLAG_DERIVED || lb!=0)
463 if (not contiguous) {
464 *new_type = new Type_Hindexed(size * old_type->size(),lb,ub,
465 DT_FLAG_DERIVED|DT_FLAG_DATA, count, block_lengths, indices, old_type);
467 Datatype::create_contiguous(size, old_type, lb, new_type);
472 int Datatype::create_struct(int count, const int* block_lengths, const MPI_Aint* indices, const MPI_Datatype* old_types, MPI_Datatype* new_type){
474 bool contiguous=true;
479 lb=indices[0] + old_types[0]->lb();
480 ub=indices[0] + block_lengths[0]*old_types[0]->ub();
482 bool forced_lb=false;
483 bool forced_ub=false;
484 for (int i = 0; i < count; i++) {
485 if (block_lengths[i]<0)
487 if (old_types[i]->flags_ & DT_FLAG_DERIVED)
490 size += block_lengths[i]*old_types[i]->size();
491 if (old_types[i]==MPI_LB){
495 if (old_types[i]==MPI_UB){
500 if (not forced_lb && indices[i] + old_types[i]->lb() < lb)
502 if (not forced_ub && indices[i] + block_lengths[i] * old_types[i]->ub() > ub)
503 ub = indices[i]+block_lengths[i]*old_types[i]->ub();
505 if ( (i< count -1) && (indices[i]+block_lengths[i]*static_cast<int>(old_types[i]->size()) != indices[i+1]) )
508 if (not contiguous) {
509 *new_type = new Type_Struct(size, lb,ub, DT_FLAG_DERIVED|DT_FLAG_DATA,
510 count, block_lengths, indices, old_types);
512 Datatype::create_contiguous(size, MPI_CHAR, lb, new_type);
517 int Datatype::create_subarray(int ndims, const int* array_of_sizes,
518 const int* array_of_subsizes, const int* array_of_starts,
519 int order, MPI_Datatype oldtype, MPI_Datatype *newtype){
522 for (int i = 0; i < ndims; i++) {
523 if (array_of_subsizes[i] > array_of_sizes[i]){
524 XBT_WARN("subarray : array_of_subsizes > array_of_sizes for dim %d",i);
527 if (array_of_starts[i] + array_of_subsizes[i] > array_of_sizes[i]){
528 XBT_WARN("subarray : array_of_starts + array_of_subsizes > array_of_sizes for dim %d",i);
533 MPI_Aint extent = oldtype->get_extent();
538 if( order==MPI_ORDER_C ) {
548 MPI_Aint size = (MPI_Aint)array_of_sizes[i] * (MPI_Aint)array_of_sizes[i+step];
549 MPI_Aint lb = (MPI_Aint)array_of_starts[i] + (MPI_Aint)array_of_starts[i+step] *(MPI_Aint)array_of_sizes[i];
551 create_vector( array_of_subsizes[i+step], array_of_subsizes[i], array_of_sizes[i],
556 for( i += 2 * step; i != end; i += step ) {
557 create_hvector( array_of_subsizes[i], 1, size * extent,
560 lb += size * array_of_starts[i];
561 size *= array_of_sizes[i];
565 MPI_Aint lbs[1] = {lb * extent};
567 //handle LB and UB with a resized call
568 create_hindexed( 1, sizes, lbs, tmp, newtype);
572 create_resized(tmp, 0, extent, newtype);
578 int Datatype::create_resized(MPI_Datatype oldtype,MPI_Aint lb, MPI_Aint extent, MPI_Datatype *newtype){
579 int blocks[3] = {1, 1, 1};
580 MPI_Aint disps[3] = {lb, 0, lb + extent};
581 MPI_Datatype types[3] = {MPI_LB, oldtype, MPI_UB};
583 *newtype = new simgrid::smpi::Type_Struct(oldtype->size(), lb, lb + extent, DT_FLAG_DERIVED, 3, blocks, disps, types);
585 (*newtype)->addflag(~DT_FLAG_COMMITED);
589 Datatype* Datatype::f2c(int id)
591 return static_cast<Datatype*>(F2C::f2c(id));
594 } // namespace simgrid