Remove MPI from serial ADIOS interface #258
Conversation
| @@ -1121,7 +1121,7 @@ TEST_CASE( "no_serial_hdf5", "[serial][hdf5]" ) | |||
| REQUIRE(true); | |||
| } | |||
| #endif | |||
| #if openPMD_HAVE_ADIOS1 | |||
| #if openPMD_HAVE_ADIOS1 && !openPMD_HAVE_MPI | |||
There was a problem hiding this comment.
You can revert this change when #254 is ready.
There was a problem hiding this comment.
uff, I rebased on your branch and applied the mentioned changes.
I won't get a running solution with our new approach unless I do explicit symbol hiding/exposing in the two transalation units for adios, since otherwise symbols like adios_open and common_adios_open will be mismatched at link and runtime.
Can do, but this will exclude pgic++ at least since not all compilers know symbol hiding. harsh world...
Or so some symbol renaming with --prefix-symbols via objcpy... gosh!
There was a problem hiding this comment.
We could maybe go for dlopen() for one ADIOS part xD
| { | ||
| char const* env = std::getenv(key.c_str()); | ||
| if( env != nullptr ) | ||
| /* create all files where ADIOS file creation has been deferred, but this has never been triggered |
There was a problem hiding this comment.
the comment is not very clear: what is "this" what has never been triggered and what is "this" that happens in the line below?
There was a problem hiding this comment.
True, the sentence really is a bit vague.
First "this" means actual file creation. The operation (file creation) has been deferred, but because no IO op that triggers execution of the deferred operation (file creation) has been scheduled across the whole Series, we have to do it manually.
Seconds "this" stands for the situation described above (i.e. the scenario in which no appropiate IO op has been scheduled).
|
really nice, thank you! |
| @@ -25,7 +25,6 @@ | |||
| #if openPMD_HAVE_ADIOS1 | |||
| # include "openPMD/IO/AbstractIOHandlerImpl.hpp" | |||
| # include <adios.h> | |||
There was a problem hiding this comment.
hm, adios.h also includes adios_mpi.h... argh.. but this calls the dummy... of have to check my setters more.
There was a problem hiding this comment.
I think you need adios.h and mpi.h in ParallelADIOS1...hpp only.
In the serial header, use adios_read.h and in the serial impl. mpidummy.h as you do.
To be able to compile both serial and parallel ADIOS1 backends, remove all traces of MPI from the serial interface. MPI_COMM_NULL is used in the implementation, but not exposed.
The adios.h header is not required for the interfaces. This might make separate compilation and multiple linking easier.
C0nsultant
force-pushed
the
topic/no_mpi_in_adios_interface
branch
from
d2cda3a to
439226b
Compare
To be able to compile both serial and parallel ADIOS1 backends, remove
all traces of MPI from the serial interface. MPI_COMM_NULL is used in
the implementation, but not exposed.
Suggested in #252 (comment), blocker for #254.