Conversation
valassi
left a comment
There was a problem hiding this comment.
Hi @oliviermattelaer this is essentially good for me, thanks.
One point: in PR #798, the avx_level and cudacpp_backend are merged: the AVX is included in the backend (because cpp by itself does not exist, it is always either cppnone, cppsse4, etc: and cpp defaults to cppauto i.e. let MG choose).
Another point: I would suggest to add aslo helinl and hrdcod here, see #700. And when we do that, I would be much more consistent in the naming, and name all these cudacppp options with a prefix: so in teh runcard we would have cudacpp_backend, cudacpp_fptype, cudacpp_helinl, cudacpp_hrdcod. These four variables only make sense for the cudacpp plugin.
Anyway, feel free to merge and we discuss these points another time. Thanks!
epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/launch_plugin.py
Outdated
Show resolved
Hide resolved
avoid to put back avx_level
…graph5#851 patches - only codelogs change anyway
…_type' to 'cudacpp_fptype' for runcard options (madgraph5#700, madgraph5#788, madgraph5#835) (Older commit on gg_tt.mad, later cherry-picked fixing AVX conflicts, and later backported to CODEGEN too)
change default fptype to "m" (second try)
Note that the value from the run_card is written inside the make_opts file for the moment.
I can write it in another file if needed by the refactorization of the matrix-element but for the current master branch this is at least not needed (I can do it in the associated branch if needed)