+/* Copyright (c) 2010-2021. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#ifndef SMPI_FILE_HPP_INCLUDED
+#define SMPI_FILE_HPP_INCLUDED
+#include "simgrid/plugins/file_system.h"
+#include "smpi_comm.hpp"
+#include "smpi_coll.hpp"
+#include "smpi_datatype.hpp"
+#include "smpi_errhandler.hpp"
+#include "smpi_info.hpp"
+#include <algorithm>
+
+XBT_LOG_EXTERNAL_CATEGORY(smpi_pmpi);
+
+namespace simgrid{
+namespace smpi{
+class File : public F2C{
+ MPI_Comm comm_;
+ int flags_;
+ simgrid::s4u::File* file_;
+ MPI_Info info_;
+ MPI_Offset* shared_file_pointer_;
+ s4u::MutexPtr shared_mutex_;
+ MPI_Win win_;
+ char* list_;
+ MPI_Errhandler errhandler_;
+ MPI_Datatype etype_;
+ MPI_Datatype filetype_;
+ std::string datarep_;
+
+ public:
+ File(MPI_Comm comm, const char *filename, int amode, MPI_Info info);
+ File(const File&) = delete;
+ File& operator=(const File&) = delete;
+ ~File() override;
+ int size() const;
+ int get_position(MPI_Offset* offset) const;
+ int get_position_shared(MPI_Offset* offset) const;
+ int flags() const;
+ MPI_Comm comm() const;
+ int sync();
+ int seek(MPI_Offset offset, int whence);
+ int seek_shared(MPI_Offset offset, int whence);
+ int set_view(MPI_Offset disp, MPI_Datatype etype, MPI_Datatype filetype, const char* datarep, const Info* info);
+ int get_view(MPI_Offset* disp, MPI_Datatype* etype, MPI_Datatype* filetype, char* datarep) const;
+ MPI_Info info();
+ void set_info( MPI_Info info);
+ static int read(MPI_File fh, void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ static int read_shared(MPI_File fh, void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ static int read_ordered(MPI_File fh, void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ static int write(MPI_File fh, void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ static int write_shared(MPI_File fh, const void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ static int write_ordered(MPI_File fh, const void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ template <int (*T)(MPI_File, void*, int, const Datatype*, MPI_Status*)>
+ int op_all(void* buf, int count, const Datatype* datatype, MPI_Status* status);
+ static int close(MPI_File *fh);
+ static int del(const char* filename, const Info* info);
+ MPI_Errhandler errhandler();
+ void set_errhandler( MPI_Errhandler errhandler);
+ static File* f2c(int id);
+};
+
+/* Read_all, Write_all : loosely based on */
+/* @article{Thakur:1996:ETM:245875.245879,*/
+/* author = {Thakur, Rajeev and Choudhary, Alok},*/
+/* title = {An Extended Two-phase Method for Accessing Sections of Out-of-core Arrays},*/
+/* journal = {Sci. Program.},*/
+/* issue_date = {Winter 1996},*/
+/* pages = {301--317},*/
+/* }*/
+template <int (*T)(MPI_File, void*, int, const Datatype*, MPI_Status*)>
+int File::op_all(void* buf, int count, const Datatype* datatype, MPI_Status* status)
+{
+ // get min and max offsets from everyone.
+ int size = comm_->size();
+ int rank = comm_->rank();
+ MPI_Offset min_offset = file_->tell();
+ MPI_Offset max_offset =
+ min_offset +
+ count * datatype->get_extent(); // cheating, as we don't care about exact data location, we can skip extent
+ std::vector<MPI_Offset> min_offsets(size);
+ std::vector<MPI_Offset> max_offsets(size);
+ simgrid::smpi::colls::allgather(&min_offset, 1, MPI_OFFSET, min_offsets.data(), 1, MPI_OFFSET, comm_);
+ simgrid::smpi::colls::allgather(&max_offset, 1, MPI_OFFSET, max_offsets.data(), 1, MPI_OFFSET, comm_);
+ MPI_Offset min = min_offset;
+ MPI_Offset max = max_offset;
+ MPI_Offset tot = 0;
+ int empty = 1;
+ for (int i = 0; i < size; i++) {
+ if (min_offsets[i] != max_offsets[i])
+ empty = 0;
+ tot += (max_offsets[i] - min_offsets[i]);
+ if (min_offsets[i] < min)
+ min = min_offsets[i];
+ if (max_offsets[i] > max)
+ max = max_offsets[i];
+ }
+
+ XBT_CDEBUG(smpi_pmpi, "my offsets to read : %lld:%lld, global min and max %lld:%lld", min_offset, max_offset, min,
+ max);
+ if (empty == 1) {
+ if (status != MPI_STATUS_IGNORE)
+ status->count = 0;
+ return MPI_SUCCESS;
+ }
+ MPI_Offset total = max - min;
+ if (total == tot && (datatype->flags() & DT_FLAG_CONTIGUOUS)) {
+ // contiguous. Just have each proc perform its read
+ if (status != MPI_STATUS_IGNORE)
+ status->count = count * datatype->size();
+ return T(this, buf, count, datatype, status);
+ }
+
+ // Interleaved case : How much do I need to read, and whom to send it ?
+ MPI_Offset my_chunk_start = (max - min + 1) / size * rank;
+ MPI_Offset my_chunk_end = ((max - min + 1) / size * (rank + 1));
+ XBT_CDEBUG(smpi_pmpi, "my chunks to read : %lld:%lld", my_chunk_start, my_chunk_end);
+ std::vector<int> send_sizes(size);
+ std::vector<int> recv_sizes(size);
+ std::vector<int> send_disps(size);
+ std::vector<int> recv_disps(size);
+ int total_sent = 0;
+ for (int i = 0; i < size; i++) {
+ send_sizes[i] = 0;
+ send_disps[i] = 0; // cheat to avoid issues when send>recv as we use recv buffer
+ if ((my_chunk_start >= min_offsets[i] && my_chunk_start < max_offsets[i]) ||
+ ((my_chunk_end <= max_offsets[i]) && my_chunk_end > min_offsets[i])) {
+ send_sizes[i] = (std::min(max_offsets[i] - 1, my_chunk_end - 1) - std::max(min_offsets[i], my_chunk_start));
+ // store min and max offset to actually read
+ min_offset = std::min(min_offset, min_offsets[i]);
+ total_sent += send_sizes[i];
+ XBT_CDEBUG(smpi_pmpi, "will have to send %d bytes to %d", send_sizes[i], i);
+ }
+ }
+ min_offset = std::max(min_offset, my_chunk_start);
+
+ // merge the ranges of every process
+ std::vector<std::pair<MPI_Offset, MPI_Offset>> ranges;
+ for (int i = 0; i < size; ++i)
+ ranges.emplace_back(min_offsets[i], max_offsets[i]);
+ std::sort(ranges.begin(), ranges.end());
+ std::vector<std::pair<MPI_Offset, MPI_Offset>> chunks;
+ chunks.push_back(ranges[0]);
+
+ unsigned int nchunks = 0;
+ unsigned int i = 1;
+ while (i < ranges.size()) {
+ if (ranges[i].second > chunks[nchunks].second) {
+ // else range included - ignore
+ if (ranges[i].first > chunks[nchunks].second) {
+ // new disjoint range
+ chunks.push_back(ranges[i]);
+ nchunks++;
+ } else {
+ // merge ranges
+ chunks[nchunks].second = ranges[i].second;
+ }
+ }
+ i++;
+ }
+ // what do I need to read ?
+ MPI_Offset totreads = 0;
+ for (i = 0; i < chunks.size(); i++) {
+ if (chunks[i].second < my_chunk_start)
+ continue;
+ else if (chunks[i].first > my_chunk_end)
+ continue;
+ else
+ totreads += (std::min(chunks[i].second, my_chunk_end - 1) - std::max(chunks[i].first, my_chunk_start));
+ }
+ XBT_CDEBUG(smpi_pmpi, "will have to access %lld from my chunk", totreads);
+
+ unsigned char* sendbuf = smpi_get_tmp_sendbuffer(total_sent);
+
+ if (totreads > 0) {
+ seek(min_offset, MPI_SEEK_SET);
+ T(this, sendbuf, totreads / datatype->size(), datatype, status);
+ }
+ simgrid::smpi::colls::alltoall(send_sizes.data(), 1, MPI_INT, recv_sizes.data(), 1, MPI_INT, comm_);
+ int total_recv = 0;
+ for (int i = 0; i < size; i++) {
+ recv_disps[i] = total_recv;
+ total_recv += recv_sizes[i];
+ }
+ // Set buf value to avoid copying dumb data
+ simgrid::smpi::colls::alltoallv(sendbuf, send_sizes.data(), send_disps.data(), MPI_BYTE, buf, recv_sizes.data(),
+ recv_disps.data(), MPI_BYTE, comm_);
+ if (status != MPI_STATUS_IGNORE)
+ status->count = count * datatype->size();
+ smpi_free_tmp_buffer(sendbuf);
+ return MPI_SUCCESS;
+}
+}
+}
+
+#endif