#include "smpi_coll.hpp"\r
#include "smpi_datatype.hpp"\r
#include "smpi_info.hpp"\r
+#include "smpi_win.hpp"\r
#include "smpi_file.hpp"\r
+#include "smpi_status.hpp"\r
#include "simgrid/plugins/file_system.h"\r
\r
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_io, smpi, "Logging specific to SMPI (RMA operations)");\r
+#define FP_SIZE sizeof(MPI_Offset)\r
\r
\r
namespace simgrid{\r
namespace smpi{\r
- File::File(MPI_Comm comm, char *filename, int amode, MPI_Info info): comm_(comm), flags_(amode), info_(info){\r
+\r
+ File::File(MPI_Comm comm, char *filename, int amode, MPI_Info info): comm_(comm), flags_(amode), info_(info), shared_file_pointer_(0) {\r
file_= new simgrid::s4u::File(filename, nullptr);\r
+ list_=nullptr;\r
+ if (comm_->rank() == 0) {\r
+ int size= comm_->size() + FP_SIZE;\r
+ list_ = new char[size];\r
+ memset(list_, 0, size);\r
+ win_=new Win(list_, size, 1, MPI_INFO_NULL, comm_);\r
+ }else{\r
+ win_=new Win(list_, 0, 1, MPI_INFO_NULL, comm_);\r
+ }\r
}\r
- \r
+\r
File::~File(){\r
delete file_;\r
}\r
return MPI_SUCCESS;\r
}\r
\r
+ int File::get_position_shared(MPI_Offset* offset){\r
+ lock();\r
+ *offset=shared_file_pointer_;\r
+ unlock();\r
+ return MPI_SUCCESS;\r
+ }\r
+\r
int File::seek(MPI_Offset offset, int whence){\r
switch(whence){\r
case(MPI_SEEK_SET):\r
}\r
return MPI_SUCCESS;\r
}\r
- \r
+\r
+ int File::seek_shared(MPI_Offset offset, int whence){\r
+ lock();\r
+ seek(offset,whence);\r
+ shared_file_pointer_=file_->tell();\r
+ unlock();\r
+ return MPI_SUCCESS;\r
+ }\r
+\r
int File::read(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
//get position first as we may be doing non contiguous reads and it will probably be updated badly\r
MPI_Offset position = fh->file_->tell();\r
fh->file_->seek(position+movesize, SEEK_SET);\r
}\r
XBT_DEBUG("Position after read in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());\r
+ status->count=count*datatype->size();\r
+ return MPI_SUCCESS;\r
+ }\r
+\r
+ int File::read_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+ fh->lock();\r
+ fh->seek(fh->shared_file_pointer_,MPI_SEEK_SET);\r
+ read(fh, buf, count, datatype, status);\r
+ fh->shared_file_pointer_=fh->file_->tell();\r
+ fh->unlock();\r
return MPI_SUCCESS;\r
}\r
\r
+ int File::read_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+ //0 needs to get the shared pointer value\r
+ if(fh->comm_->rank()==0){\r
+ fh->lock();\r
+ fh->unlock();\r
+ }else{\r
+ fh->shared_file_pointer_=count*datatype->size();\r
+ }\r
+ MPI_Offset result;\r
+ simgrid::smpi::Colls::scan(&(fh->shared_file_pointer_), &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);\r
+ fh->seek(result, MPI_SEEK_SET);\r
+ int ret = fh->op_all<simgrid::smpi::File::read>(buf, count, datatype, status);\r
+ if(fh->comm_->rank()==fh->comm_->size()-1){\r
+ fh->lock();\r
+ fh->unlock();\r
+ }\r
+ char c;\r
+ simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);\r
+ return ret;\r
+ }\r
\r
int File::write(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
//get position first as we may be doing non contiguous reads and it will probably be updated badly\r
fh->file_->seek(position+movesize, SEEK_SET);\r
}\r
XBT_DEBUG("Position after write in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());\r
+ status->count=count*datatype->size();\r
return MPI_SUCCESS;\r
}\r
\r
+ int File::write_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+ fh->lock();\r
+ fh->seek(fh->shared_file_pointer_,MPI_SEEK_SET);\r
+ write(fh, buf, count, datatype, status);\r
+ fh->shared_file_pointer_=fh->file_->tell();\r
+ fh->unlock();\r
+ return MPI_SUCCESS;\r
+ }\r
+\r
+ int File::write_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+ //0 needs to get the shared pointer value\r
+ if(fh->comm_->rank()==0){\r
+ fh->lock();\r
+ fh->unlock();\r
+ }else{\r
+ fh->shared_file_pointer_=count*datatype->size();\r
+ }\r
+ MPI_Offset result;\r
+ simgrid::smpi::Colls::scan(&(fh->shared_file_pointer_), &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);\r
+ fh->seek(result, MPI_SEEK_SET);\r
+ int ret = fh->op_all<simgrid::smpi::File::write>(buf, count, datatype, status);\r
+ if(fh->comm_->rank()==fh->comm_->size()-1){\r
+ fh->lock();\r
+ fh->unlock();\r
+ }\r
+ char c;\r
+ simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);\r
+ return ret;\r
+ }\r
+\r
int File::size(){\r
return file_->size();\r
}\r
//no idea\r
return simgrid::smpi::Colls::barrier(comm_);\r
}\r
+\r
+ int File::lock()\r
+{\r
+ int rank = comm_->rank();\r
+ int size = comm_->size();\r
+ char waitlist[size];\r
+ char lock = 1;\r
+ int tag=444;\r
+ int i;\r
+ win_->lock(MPI_LOCK_EXCLUSIVE, 0, 0);\r
+ win_->put(&lock, 1, MPI_CHAR, 0, FP_SIZE+rank, 1, MPI_CHAR);\r
+ win_->get(waitlist, size, MPI_CHAR, 0, FP_SIZE, size, MPI_CHAR);\r
+ win_->get(&shared_file_pointer_ , 1 , MPI_OFFSET , 0 , 0, 1, MPI_OFFSET);\r
+ win_->unlock(0);\r
+ for (i = 0; i < size; i++) {\r
+ if (waitlist[i] == 1 && i != rank) {\r
+ // wait for the lock\r
+ MPI_Recv(&lock, 1, MPI_CHAR, MPI_ANY_SOURCE, tag, comm_, MPI_STATUS_IGNORE);\r
+ break;\r
+ }\r
+ }\r
+ return 0;\r
+}\r
+\r
+int File::unlock()\r
+{\r
+ int rank = comm_->rank();\r
+ int size = comm_->size();\r
+ char waitlist[size];\r
+ char lock = 0;\r
+ int tag=444;\r
+ int i, next;\r
+ win_->lock(MPI_LOCK_EXCLUSIVE, 0, 0);\r
+ win_->put(&lock, 1, MPI_CHAR, 0, FP_SIZE+rank, 1, MPI_CHAR);\r
+ win_->get(waitlist, size, MPI_CHAR, 0, FP_SIZE, size, MPI_CHAR);\r
+ shared_file_pointer_=file_->tell();\r
+ win_->put(&shared_file_pointer_, 1 , MPI_OFFSET , 0 , 0, 1, MPI_OFFSET);\r
+\r
+ win_->unlock(0);\r
+ next = (rank + 1 + size) % size;\r
+ for (i = 0; i < size; i++, next = (next + 1) % size) {\r
+ if (waitlist[next] == 1) {\r
+ MPI_Send(&lock, 1, MPI_CHAR, next, tag, comm_);\r
+ break;\r
+ }\r
+ }\r
+ return 0;\r
+}\r
+\r
MPI_Info File::info(){\r
if(info_== MPI_INFO_NULL)\r
info_ = new Info();\r
