1 /* smpi_datatype.cpp -- MPI primitives to handle datatypes */
2 /* Copyright (c) 2009-2019. 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);
110 Datatype::Datatype(int size,MPI_Aint lb, MPI_Aint ub, int flags) : name_(nullptr), size_(size), lb_(lb), ub_(ub), flags_(flags), refcount_(1){
113 MC_ignore(&(refcount_), sizeof(refcount_));
117 //for predefined types, so in_use = 0.
118 Datatype::Datatype(char* name, int ident, int size, MPI_Aint lb, MPI_Aint ub, int flags)
119 : name_(name), id(std::to_string(ident)), size_(size), lb_(lb), ub_(ub), flags_(flags), refcount_(0)
121 id2type_lookup.insert({id, this});
124 MC_ignore(&(refcount_), sizeof(refcount_));
128 Datatype::Datatype(Datatype *datatype, int* ret) : name_(nullptr), size_(datatype->size_), lb_(datatype->lb_), ub_(datatype->ub_), flags_(datatype->flags_), refcount_(1)
130 flags_ &= ~DT_FLAG_PREDEFINED;
133 if (not datatype->attributes()->empty()) {
136 for (auto const& it : *(datatype->attributes())) {
137 smpi_key_elem elem = keyvals_.at(it.first);
138 if (elem != nullptr){
139 if( elem->copy_fn.type_copy_fn != MPI_NULL_COPY_FN &&
140 elem->copy_fn.type_copy_fn != MPI_TYPE_DUP_FN)
141 *ret = elem->copy_fn.type_copy_fn(datatype, it.first, elem->extra_state, it.second, &value_out, &flag);
142 else if ( elem->copy_fn.type_copy_fn_fort != MPI_NULL_COPY_FN &&
143 (*(int*)*elem->copy_fn.type_copy_fn_fort) != 1){
144 value_out=(int*)xbt_malloc(sizeof(int));
145 elem->copy_fn.type_copy_fn_fort(datatype, it.first, elem->extra_state, it.second, value_out, &flag,ret);
147 if (*ret != MPI_SUCCESS) {
150 if(elem->copy_fn.type_copy_fn == MPI_TYPE_DUP_FN ||
151 ((elem->copy_fn.type_copy_fn_fort != MPI_NULL_COPY_FN) && (*(int*)*elem->copy_fn.type_copy_fn_fort == 1))){
153 attributes()->insert({it.first, it.second});
156 attributes()->insert({it.first, value_out});
163 Datatype::~Datatype(){
164 xbt_assert(refcount_ >= 0);
166 if(flags_ & DT_FLAG_PREDEFINED)
169 //if still used, mark for deletion
171 flags_ |=DT_FLAG_DESTROYED;
175 cleanup_attr<Datatype>();
180 void Datatype::ref(){
186 MC_ignore(&(refcount_), sizeof(refcount_));
190 void Datatype::unref(MPI_Datatype datatype)
192 if (datatype->refcount_ > 0)
193 datatype->refcount_--;
195 if (datatype->refcount_ == 0 && not(datatype->flags_ & DT_FLAG_PREDEFINED))
200 MC_ignore(&(datatype->refcount_), sizeof(datatype->refcount_));
204 void Datatype::commit()
206 flags_ |= DT_FLAG_COMMITED;
209 bool Datatype::is_valid(){
210 return (flags_ & DT_FLAG_COMMITED);
213 bool Datatype::is_basic()
215 return (flags_ & DT_FLAG_BASIC);
218 bool Datatype::is_replayable()
220 return (simgrid::instr::trace_format == simgrid::instr::TraceFormat::Ti) &&
221 ((this == MPI_BYTE) || (this == MPI_DOUBLE) || (this == MPI_INT) || (this == MPI_CHAR) ||
222 (this == MPI_SHORT) || (this == MPI_LONG) || (this == MPI_FLOAT));
225 MPI_Datatype Datatype::decode(const std::string& datatype_id)
227 return id2type_lookup.find(datatype_id)->second;
230 void Datatype::addflag(int flag){
234 int Datatype::extent(MPI_Aint * lb, MPI_Aint * extent){
240 void Datatype::get_name(char* name, int* length){
242 *length = strlen(name_);
243 strncpy(name, name_, *length+1);
249 void Datatype::set_name(const char* name){
250 if(name_!=nullptr && (flags_ & DT_FLAG_PREDEFINED) == 0)
252 name_ = xbt_strdup(name);
255 int Datatype::pack(const void* inbuf, int incount, void* outbuf, int outcount, int* position, MPI_Comm)
257 if (outcount - *position < incount*static_cast<int>(size_))
258 return MPI_ERR_OTHER;
259 Datatype::copy(inbuf, incount, this, static_cast<char*>(outbuf) + *position, outcount, MPI_CHAR);
260 *position += incount * size_;
264 int Datatype::unpack(const void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Comm)
266 if (outcount*static_cast<int>(size_)> insize)
267 return MPI_ERR_OTHER;
268 Datatype::copy(static_cast<const char*>(inbuf) + *position, insize, MPI_CHAR, outbuf, outcount, this);
269 *position += outcount * size_;
273 int Datatype::copy(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
274 void *recvbuf, int recvcount, MPI_Datatype recvtype){
276 // FIXME Handle the case of a partial shared malloc.
278 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
279 smpi_switch_data_segment(simgrid::s4u::Actor::self());
281 /* First check if we really have something to do */
283 std::vector<std::pair<size_t, size_t>> private_blocks;
284 if(smpi_is_shared(sendbuf,private_blocks,&offset)
285 && (private_blocks.size()==1
286 && (private_blocks[0].second - private_blocks[0].first)==(unsigned long)(sendcount * sendtype->get_extent()))){
287 XBT_VERB("sendbuf is shared. Ignoring copies");
290 if(smpi_is_shared(recvbuf,private_blocks,&offset)
291 && (private_blocks.size()==1
292 && (private_blocks[0].second - private_blocks[0].first)==(unsigned long)(recvcount * recvtype->get_extent()))){
293 XBT_VERB("recvbuf is shared. Ignoring copies");
297 if (recvcount > 0 && recvbuf != sendbuf) {
298 sendcount *= sendtype->size();
299 recvcount *= recvtype->size();
300 int count = sendcount < recvcount ? sendcount : recvcount;
301 XBT_DEBUG("Copying %d bytes from %p to %p", count, sendbuf, recvbuf);
302 if (not(sendtype->flags() & DT_FLAG_DERIVED) && not(recvtype->flags() & DT_FLAG_DERIVED)) {
303 if (not smpi_process()->replaying())
304 memcpy(recvbuf, sendbuf, count);
305 } else if (not(sendtype->flags() & DT_FLAG_DERIVED)) {
306 recvtype->unserialize(sendbuf, recvbuf, count / recvtype->size(), MPI_REPLACE);
307 } else if (not(recvtype->flags() & DT_FLAG_DERIVED)) {
308 sendtype->serialize(sendbuf, recvbuf, count / sendtype->size());
311 void * buf_tmp = xbt_malloc(count);
313 sendtype->serialize( sendbuf, buf_tmp,count/sendtype->size());
314 recvtype->unserialize( buf_tmp, recvbuf,count/recvtype->size(), MPI_REPLACE);
320 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
323 //Default serialization method : memcpy.
324 void Datatype::serialize(const void* noncontiguous_buf, void* contiguous_buf, int count)
326 char* contiguous_buf_char = static_cast<char*>(contiguous_buf);
327 const char* noncontiguous_buf_char = static_cast<const char*>(noncontiguous_buf)+lb_;
328 memcpy(contiguous_buf_char, noncontiguous_buf_char, count*size_);
331 void Datatype::unserialize(const void* contiguous_buf, void *noncontiguous_buf, int count, MPI_Op op){
332 const char* contiguous_buf_char = static_cast<const char*>(contiguous_buf);
333 char* noncontiguous_buf_char = static_cast<char*>(noncontiguous_buf)+lb_;
336 op->apply( contiguous_buf_char, noncontiguous_buf_char, &n, this);
339 int Datatype::create_contiguous(int count, MPI_Datatype old_type, MPI_Aint lb, MPI_Datatype* new_type){
340 if(old_type->flags_ & DT_FLAG_DERIVED){
341 //handle this case as a hvector with stride equals to the extent of the datatype
342 return create_hvector(count, 1, old_type->get_extent(), old_type, new_type);
345 *new_type = new Type_Contiguous(count * old_type->size(), lb, lb + count * old_type->size(),
346 DT_FLAG_DERIVED, count, old_type);
348 *new_type = new Datatype(count * old_type->size(), lb, lb + count * old_type->size(),0);
352 int Datatype::create_vector(int count, int block_length, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
361 ub=((count-1)*stride+block_length-1)*old_type->get_extent()+old_type->ub();
363 if(old_type->flags() & DT_FLAG_DERIVED || stride != block_length){
364 *new_type = new Type_Vector(count * (block_length) * old_type->size(), lb, ub,
365 DT_FLAG_DERIVED, count, block_length, stride, old_type);
368 /* in this situation the data are contiguous thus it's not required to serialize and unserialize it*/
369 *new_type = new Datatype(count * block_length * old_type->size(), 0, ((count -1) * stride + block_length)*
370 old_type->size(), DT_FLAG_CONTIGUOUS);
377 int Datatype::create_hvector(int count, int block_length, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
386 ub=((count-1)*stride)+(block_length-1)*old_type->get_extent()+old_type->ub();
388 if(old_type->flags() & DT_FLAG_DERIVED || stride != block_length*old_type->get_extent()){
389 *new_type = new Type_Hvector(count * (block_length) * old_type->size(), lb, ub,
390 DT_FLAG_DERIVED, count, block_length, stride, old_type);
393 /* in this situation the data are contiguous thus it's not required to serialize and unserialize it*/
394 *new_type = new Datatype(count * block_length * old_type->size(), 0, count * block_length * old_type->size(), DT_FLAG_CONTIGUOUS);
400 int Datatype::create_indexed(int count, const int* block_lengths, const int* indices, MPI_Datatype old_type, MPI_Datatype* new_type){
402 bool contiguous=true;
406 lb=indices[0]*old_type->get_extent();
407 ub=indices[0]*old_type->get_extent() + block_lengths[0]*old_type->ub();
410 for (int i = 0; i < count; i++) {
411 if (block_lengths[i] < 0)
413 size += block_lengths[i];
415 if(indices[i]*old_type->get_extent()+old_type->lb()<lb)
416 lb = indices[i]*old_type->get_extent()+old_type->lb();
417 if(indices[i]*old_type->get_extent()+block_lengths[i]*old_type->ub()>ub)
418 ub = indices[i]*old_type->get_extent()+block_lengths[i]*old_type->ub();
420 if ( (i< count -1) && (indices[i]+block_lengths[i] != indices[i+1]) )
423 if(old_type->flags_ & DT_FLAG_DERIVED)
426 if (not contiguous) {
427 *new_type = new Type_Indexed(size * old_type->size(),lb,ub,
428 DT_FLAG_DERIVED|DT_FLAG_DATA, count, block_lengths, indices, old_type);
430 Datatype::create_contiguous(size, old_type, lb, new_type);
435 int Datatype::create_hindexed(int count, const int* block_lengths, const MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type){
437 bool contiguous=true;
441 lb=indices[0] + old_type->lb();
442 ub=indices[0] + block_lengths[0]*old_type->ub();
444 for (int i = 0; i < count; i++) {
445 if (block_lengths[i] < 0)
447 size += block_lengths[i];
449 if(indices[i]+old_type->lb()<lb)
450 lb = indices[i]+old_type->lb();
451 if(indices[i]+block_lengths[i]*old_type->ub()>ub)
452 ub = indices[i]+block_lengths[i]*old_type->ub();
454 if ( (i< count -1) && (indices[i]+block_lengths[i]*(static_cast<int>(old_type->size())) != indices[i+1]) )
457 if (old_type->flags_ & DT_FLAG_DERIVED || lb!=0)
460 if (not contiguous) {
461 *new_type = new Type_Hindexed(size * old_type->size(),lb,ub,
462 DT_FLAG_DERIVED|DT_FLAG_DATA, count, block_lengths, indices, old_type);
464 Datatype::create_contiguous(size, old_type, lb, new_type);
469 int Datatype::create_struct(int count, const int* block_lengths, const MPI_Aint* indices, const MPI_Datatype* old_types, MPI_Datatype* new_type){
471 bool contiguous=true;
476 lb=indices[0] + old_types[0]->lb();
477 ub=indices[0] + block_lengths[0]*old_types[0]->ub();
479 bool forced_lb=false;
480 bool forced_ub=false;
481 for (int i = 0; i < count; i++) {
482 if (block_lengths[i]<0)
484 if (old_types[i]->flags_ & DT_FLAG_DERIVED)
487 size += block_lengths[i]*old_types[i]->size();
488 if (old_types[i]==MPI_LB){
492 if (old_types[i]==MPI_UB){
497 if (not forced_lb && indices[i] + old_types[i]->lb() < lb)
499 if (not forced_ub && indices[i] + block_lengths[i] * old_types[i]->ub() > ub)
500 ub = indices[i]+block_lengths[i]*old_types[i]->ub();
502 if ( (i< count -1) && (indices[i]+block_lengths[i]*static_cast<int>(old_types[i]->size()) != indices[i+1]) )
505 if (not contiguous) {
506 *new_type = new Type_Struct(size, lb,ub, DT_FLAG_DERIVED|DT_FLAG_DATA,
507 count, block_lengths, indices, old_types);
509 Datatype::create_contiguous(size, MPI_CHAR, lb, new_type);
514 int Datatype::create_subarray(int ndims, const int* array_of_sizes,
515 const int* array_of_subsizes, const int* array_of_starts,
516 int order, MPI_Datatype oldtype, MPI_Datatype *newtype){
519 for (int i = 0; i < ndims; i++) {
520 if (array_of_subsizes[i] > array_of_sizes[i]){
521 XBT_WARN("subarray : array_of_subsizes > array_of_sizes for dim %d",i);
524 if (array_of_starts[i] + array_of_subsizes[i] > array_of_sizes[i]){
525 XBT_WARN("subarray : array_of_starts + array_of_subsizes > array_of_sizes for dim %d",i);
530 MPI_Aint extent = oldtype->get_extent();
535 if( order==MPI_ORDER_C ) {
545 MPI_Aint size = (MPI_Aint)array_of_sizes[i] * (MPI_Aint)array_of_sizes[i+step];
546 MPI_Aint lb = (MPI_Aint)array_of_starts[i] + (MPI_Aint)array_of_starts[i+step] *(MPI_Aint)array_of_sizes[i];
548 create_vector( array_of_subsizes[i+step], array_of_subsizes[i], array_of_sizes[i],
553 for( i += 2 * step; i != end; i += step ) {
554 create_hvector( array_of_subsizes[i], 1, size * extent,
557 lb += size * array_of_starts[i];
558 size *= array_of_sizes[i];
562 MPI_Aint lbs[1] = {lb * extent};
564 //handle LB and UB with a resized call
565 create_hindexed( 1, sizes, lbs, tmp, newtype);
569 create_resized(tmp, 0, extent, newtype);
575 int Datatype::create_resized(MPI_Datatype oldtype,MPI_Aint lb, MPI_Aint extent, MPI_Datatype *newtype){
576 int blocks[3] = {1, 1, 1};
577 MPI_Aint disps[3] = {lb, 0, lb + extent};
578 MPI_Datatype types[3] = {MPI_LB, oldtype, MPI_UB};
580 *newtype = new simgrid::smpi::Type_Struct(oldtype->size(), lb, lb + extent, DT_FLAG_DERIVED, 3, blocks, disps, types);
582 (*newtype)->addflag(~DT_FLAG_COMMITED);
586 Datatype* Datatype::f2c(int id){
587 return static_cast<Datatype*>(F2C::f2c(id));