(WIP) HELINL=L (L for linker) helas mode: pre-compile templates into separate .o object files (using RDC for CUDA; still missing HIP)#978

Draft

valassi wants to merge 56 commits intomadgraph5:masterfrom valassi:helas

+24,733-6,623

Commits on Aug 27, 2024

[helas] in gg_tt.mad, proof of concept for removing template/inline FFVs and for compiling them as separate object files (related to splitting kernels)
valassi
committed

Commits on Aug 28, 2024

[helas] in gg_tt.mad and CODEGEN, add comments in MemoryAccessGs.h and MemoryAccessMomenta.h
valassi
committed
[helas] in gg_tt.mad, compile HelAmps.o as a separate object file in the P subdirectory (depends on npar) - build succeeds for cpp, link fails for cuda
valassi
committed
[helas] in gg_tt.mad, add RDC to ensure that cuda builds succeed and cuda tests succeed
valassi
committed
[helas] in gg_tt.mad, avoid link warnings when using RDC
valassi
committed
[helas] in gg_tt.mad, clean up 'linked HelAmps' implementation: add option HELINL=L and '#ifdef MGONGPU_LINKER_HELAMPS'
valassi
committed
[helas] in tput/teeThroughputX.sh, print out the preliminary build time on each log
valassi
committed
[helas] in tput throughputX.sh and teeThroughputX.sh, add the -inlL and -inlLonly options
valassi
committed
[helas] in tput/allTees.sh, add 18 inlL tests
valassi
committed
[helas] in gg_tt.mad, fix clang formatting
valassi
committed
[helas] in gg_tt.mad, fix inlineHel=L printout in check_sa.cc
valassi
committed
[helas] in gg_tt.mad CPPProcess.cc and HelAmps_sm.h, move code around to ease code generation
valassi
committed
[helas] in gg_tt.mad cudacpp.mk, build HelAmps.o and use rdc=true only in the HELINL=L mode
valassi
committed
[helas] first quick tput test of ggtt including -inlL option: ok for c++, a factor 3 slower for cuda...
valassi
committed
[helas] in CODEGEN, complete the backport from gg_tt.mad of file templates in HELINL=L mode
valassi
committed
[helas] in CODEGEN model_handling.py, complete the backport from gg_tt.mad of HelAmps.h in HELINL=L mode
valassi
committed
[helas] in CODEGEN model_handling.py, complete the backport from gg_tt.mad of CPPProcess.cc in HELINL=L mode
valassi
committed
[helas] in gg_tt.mad, move HelAmps.cc to SubProcesses and link it in P* (the source is the same but it must be compiled in each P* separately)
valassi
committed
[helas] in CODEGEN and gg_tt.mad, fix HelAmps.cc in HELINL=L mode and complete its backport
valassi
committed
[helas] regenerate gg_tt.mad, check that all is ok (codegen for HELINL=L is complete)
valassi
committed
[helas] regenerate all processes with support for HELINL=L
valassi
committed
[helas] in tmad madX.sh and teeMadX.sh, add -inlonly and -inlLonly options
valassi
committed
[helas] add HelAmps.cc to all regenerated processes
valassi
committed
[helas] aborted tput test of ggttggg with all helinl values - inl1 build failed?
valassi
committed
[helas] rerun the ggttggg tput test only in inl0 mode - note that the build time is from cache
valassi
committed
[helas] manually fix the build time in the ggttggg tput test in inl0 mode (use that from the previous run, not from cache)
valassi
committed
[helas] first run of the ggttggg tput test in inlL mode - build is a factor x2 faster (c++? cuda?), runtime is 5-10% slower in C++, but 5-10% faster in cuda!?
valassi
committed

Commits on Aug 29, 2024

Commits on Sep 2, 2024

Commits on Sep 21, 2024

Commits on Sep 25, 2024

Commits on Oct 5, 2024