int size = comm_->size();\r
int rank = comm_-> rank();\r
MPI_Offset min_offset = file_->tell();\r
- MPI_Offset max_offset = min_offset + count * datatype->size();//cheating, as we don't care about exact data location, we can skip extent\r
+ MPI_Offset max_offset = (min_offset + count * datatype->size());//cheating, as we don't care about exact data location, we can skip extent\r
MPI_Offset* min_offsets = xbt_new(MPI_Offset, size);\r
MPI_Offset* max_offsets = xbt_new(MPI_Offset, size);\r
simgrid::smpi::Colls::allgather(&min_offset, 1, MPI_OFFSET, min_offsets, 1, MPI_OFFSET, comm_);\r
MPI_Offset min=min_offset;\r
MPI_Offset max=max_offset;\r
MPI_Offset tot= 0;\r
+ int empty=1;\r
for(int i=0;i<size;i++){\r
+ if(min_offsets[i]!=max_offsets[i])\r
+ empty=0;\r
tot+=(max_offsets[i]-min_offsets[i]);\r
if(min_offsets[i]<min)\r
min=min_offsets[i];\r
if(max_offsets[i]>max)\r
max=max_offsets[i];\r
}\r
+ \r
+ XBT_DEBUG("my offsets to read : %lld:%lld, global min and max %lld:%lld", min_offset, max_offset, min, max);\r
+ if(empty==1){\r
+ status->count=0;\r
+ return MPI_SUCCESS;\r
+ }\r
MPI_Offset total = max-min;\r
if(total==tot && (datatype->flags() & DT_FLAG_CONTIGUOUS)){\r
//contiguous. Just have each proc perform its read\r
+ status->count=count * datatype->size();\r
return T(this,buf,count,datatype, status);\r
}\r
\r
//Interleaved case : How much do I need to read, and whom to send it ?\r
- MPI_Offset my_chunk_start=(max-min)/size*rank;\r
- MPI_Offset my_chunk_end=((max-min)/size*(rank+1))-1;\r
+ MPI_Offset my_chunk_start=(max-min+1)/size*rank;\r
+ MPI_Offset my_chunk_end=((max-min+1)/size*(rank+1));\r
+ XBT_DEBUG("my chunks to read : %lld:%lld", my_chunk_start, my_chunk_end);\r
int* send_sizes = xbt_new0(int, size);\r
int* recv_sizes = xbt_new(int, size);\r
int* send_disps = xbt_new(int, size);\r
for(int i=0;i<size;i++){\r
if((my_chunk_start>=min_offsets[i] && my_chunk_start < max_offsets[i])||\r
((my_chunk_end<=max_offsets[i]) && my_chunk_end> min_offsets[i])){\r
- send_sizes[i]=(std::min(max_offsets[i], my_chunk_end+1)-std::max(min_offsets[i], my_chunk_start));\r
+ send_sizes[i]=(std::min(max_offsets[i]-1, my_chunk_end-1)-std::max(min_offsets[i], my_chunk_start));\r
//store min and max offest to actually read\r
- min_offset=std::max(min_offsets[i], my_chunk_start);\r
- max_offset=std::min(max_offsets[i], my_chunk_end+1);\r
+ min_offset=std::min(min_offset, min_offsets[i]);\r
send_disps[i]=0;//send_sizes[i]; cheat to avoid issues when send>recv as we use recv buffer\r
total_sent+=send_sizes[i];\r
+ XBT_DEBUG("will have to send %d bytes to %d", send_sizes[i], i);\r
}\r
}\r
+ min_offset=std::max(min_offset, my_chunk_start);\r
\r
//merge the ranges of every process\r
std::vector<std::pair<MPI_Offset, MPI_Offset>> ranges;\r
else if (chunks[i].first > my_chunk_end)\r
continue;\r
else\r
- totreads += (std::min(chunks[i].second, my_chunk_end+1)-std::max(chunks[i].first, my_chunk_start));\r
+ totreads += (std::min(chunks[i].second, my_chunk_end-1)-std::max(chunks[i].first, my_chunk_start));\r
}\r
+ XBT_DEBUG("will have to access %lld from my chunk", totreads);\r
+\r
char* sendbuf= static_cast<char *>(smpi_get_tmp_sendbuffer(totreads));\r
\r
if(totreads>0){\r
//Set buf value to avoid copying dumb data\r
simgrid::smpi::Colls::alltoallv(sendbuf, send_sizes, send_disps, MPI_BYTE,\r
buf, recv_sizes, recv_disps, MPI_BYTE, comm_);\r
+ status->count=count * datatype->size();\r
smpi_free_tmp_buffer(sendbuf);\r
xbt_free(send_sizes);\r
xbt_free(recv_sizes);\r