Heterogeneous MadGraph: parallel CPU+GPU executions

[valassi]

See a more complete description in #85

I created a simple prototype. The point here was mainly a proof of concept, also trying to sort out the build (which may be useful for addressing #83).

The current prototype runs exactly the same number of events with exactly the same random numbers in parallel on CPU (with OMP threads) and on GPU. Both computations give the same sets of events, which are not yet combined. As the GPU is much faster, essentiually the net effect is a computation that lasts as long as the CPU version, but does double events (because the same events are also on the GPU), so the throghput doubles.

This clearly needs a lot more work (especially the optimization is tricky), but it's a useful prrof of concept.

PS I forgot to mention: this includes and supersedes #82. As in that one, the build of runTest (#83) is disabled here. I would suggest however to fix it after including these changes, which give a possible direction for how to combine cuda and c++ modules. One option for #83 is to keep a single runTest.exe as it is now, but make it much clearer that real modules are either c++ with gcc or cuda with nvcc, and then only thin layers integrating both are compiled with nvcc. Essentially, I only had to add -lgomp to build the combined module, probably the same is ok for runTest.

[valassi]

I have just pushed more changes.

• I added a quick hack to process 70 times more events on the GPU, and use differenr random seeds
• I do not compute combined "physics" (average ME)
• I fixed the calculation of the combined throughput, one must add thoughputs, not times and events (the wall time is the same)

More is needed eventually, but I think this can be considered for merging. I remove the draft status.

This is an example

****************************************************************************
----------------------------------------------------------------------------
(GPU) TotalEventsComputed        = 367001600
(GPU) TotalTime[Rnd+Rmb+ME] (123)= ( 5.835859e+00                 )  sec
(GPU) TotalTime[Rambo+ME]    (23)= ( 5.378664e+00                 )  sec
(GPU) TotalTime[RndNumGen]    (1)= ( 4.571957e-01                 )  sec
(GPU) TotalTime[Rambo]        (2)= ( 4.823339e+00                 )  sec
(GPU) TotalTime[MatrixElems]  (3)= ( 5.553249e-01                 )  sec
----------------------------------------------------------------------------
(GPU) TotalEventsComputed        = 367001600
(GPU) EvtsPerSec[Rnd+Rmb+ME](123)= ( 6.288733e+07                 )  sec^-1
(GPU) EvtsPerSec[Rmb+ME]     (23)= ( 6.823286e+07                 )  sec^-1
(GPU) EvtsPerSec[MatrixElems] (3)= ( 6.608772e+08                 )  sec^-1
****************************************************************************
----------------------------------------------------------------------------
(CPU) TotalEventsComputed        = 5242880
(CPU) TotalTime[Rnd+Rmb+ME] (123)= ( 5.245397e+00                 )  sec
(CPU) TotalTime[Rambo+ME]    (23)= ( 4.964864e+00                 )  sec
(CPU) TotalTime[RndNumGen]    (1)= ( 2.805332e-01                 )  sec
(CPU) TotalTime[Rambo]        (2)= ( 1.005407e+00                 )  sec
(CPU) TotalTime[MatrixElems]  (3)= ( 3.959456e+00                 )  sec
----------------------------------------------------------------------------
(CPU) TotalEventsComputed        = 5242880
(CPU) EvtsPerSec[Rnd+Rmb+ME](123)= ( 9.995201e+05                 )  sec^-1
(CPU) EvtsPerSec[Rmb+ME]     (23)= ( 1.055997e+06                 )  sec^-1
(CPU) EvtsPerSec[MatrixElems] (3)= ( 1.324141e+06                 )  sec^-1
****************************************************************************
(HET) TotalEventsComputed        = 372244480
(HET) EvtsPerSec[Rnd+Rmb+ME](123)= ( 6.388685e+07                 )  sec^-1
(HET) EvtsPerSec[Rmb+ME]     (23)= ( 6.928886e+07                 )  sec^-1
(HET) EvtsPerSec[MatrixElems] (3)= ( 6.622013e+08                 )  sec^-1
****************************************************************************

[valassi]

PS Consider approving/merging #84 first (OMP multi threading), as that is included here too.

[valassi]

Hm I deleted and recreated the branch, I thought this would be picked up. I am reopening this as

git push -f origin 9026441:het
[reopen PR]
git push -f origin 52edc92:het

[valassi]

This PR is now obsolete and I will close it.

It is replaced by PR #159.

I copied a few relevant comments to the general issue #85.