MPI version #47
MPI version #47
Conversation
Added policy's to GooPdf. Added python script to find optimal group/grain size for a given problem.
…transform_reduce'.
…or OMP to not use goofit_policy
CMakeLists.txt
Outdated
| @@ -8,6 +8,9 @@ if(EXISTS ${LOC_PATH}) | |||
| message(FATAL_ERROR "You cannot build in the source directory (or any directory with a CMakeLists.txt file). Please make a build subdirectory.") | |||
| endif() | |||
|
|
|||
| set(OR_GROUPSIZE 128) | |||
| set(OR_GRAINSIZE 7) | |||
There was a problem hiding this comment.
This is not enough to get values into C++. You also need to set up a C++ file. See Version.h.
There was a problem hiding this comment.
Ahh, I see. Do you recommend that I make a 'ThrustOverride.h.in' to set these values?
Maybe it would be better to modify these two lines as follows?
add_definitions ("-DOR_GROUPSIZE 128")
add_definitions ("-DOR_GRAINSIZE 128")Either way, the code will need to be recompiled fully.
There was a problem hiding this comment.
I would probably use the CMake.in, but either would work. I'd make them CMake options, too:
set(GOOFIT_MPI_OR_GROUPSIZE "128" CACHE STRING "The group size for goofit MPI")
...There was a problem hiding this comment.
Don't worry about full recompiles, they will happen...
| #include <thrust/reduce.h> | ||
| #include <thrust/detail/seq.h> | ||
| #include <thrust/detail/temporary_array.h> | ||
| #include <thrust/system/cuda/detail/bulk.h> |
There was a problem hiding this comment.
These break the OMP build (for example, on Travis), since detail/cuda expects cuda to be available.
|
Deleting Makefile.common was unintentional, I can add it back in until CMake is hard-required. |
|
If you add me to your repository, I can help, or you can make a branch in GooFit official (master is protected, but making a new branch should be available to you). |
|
Good to know about making branches in GooFit official. I have not pushed my committed changes yet, I am still testing them out right now. I think you can checkout the branch I'm working on by creating a new branch, then calling 'git pull https://github.com/galapaegos/GooFit.git dev' into that branch for my current changes, at least that is how I've been getting the changes for rebasing. I am currently testing that setting OMP / MPI is working together, and that the override variables are passed properly. |
…I_CXX_COMPILE_FLAGS
Added CUDA override for group/grain size Fixed code to use THRUST_DEVICE methods for checking CUDA vs OMP versions
|
Are you ready to merge? You can still work on it in the |
|
Yes, ready to go! |
The MPI version divides the data based on the number of MPI processes used, so we need to use 'm_iEventsPerTask' when making thrust calls on the whole data.