bosilca
bosilca
Give me a week or so to add support for CUDA.
If you don't add all the OPAL objectification you only have to free a single object, the initial json object via `json_value_free`. Simpler and cleaner, no need for extra reference...
To state this differently instead of trusting to user to release a single object once (aka. the main json object she obtained from the load function), you trust them to...
I think we can solve our disagreement if instead of `opal_json_t**` the json API takes an `opal_json_t*`object. This way the user will always be in charge of the life expectancy...
This discussion is not grounded on current OMPI code. The `--disable-visibility` configure option, does little to the building environment except preventing our *_DECLSPEC from being useful (it is defined to...
So we are all looking into this. So much wasted time ! I recall we were running a checker (at the end of debug builds or something) to make sure...
The reason we forcefully make some symbols externally visible has nothing to do with debugging. Without these symbols being visible our MCA components would not be able to access them...
@jsquyres helped me find it. It's in `config/find_common_symbol` and it is called at the end of Makefile.am but it does not do everything I said. But you can use something...
If I'm not mistaken long ago it was the first step toward integrating what would become cheetah or hcoll into OMPI.
How this compares with https://github.com/open-mpi/ompi/pull/10470 ?