Backend separation

[stloufra]

Summary

Physically separates backend-owned driver code (cpu / simd / gpu) out of the madmatrix generator's shared and per-process-generated files into backend/{cpu,simd,gpu}/, so a given backend's code lives in its own directory instead of being selected at compile time via #ifdef MGONGPUCPP_GPUIMPL / MGONGPU_CPPSIMD.

What changed

• New layout: SigmaKin, color_sum, CrossSectionKernels, MatrixElementKernels, EventStatistics, umami, the 14 MemoryAccess*.h files, and mgOnGpuConfig/Cxtypes/Fptypes/Vectors.h now live under backend/{cpu,simd,gpu}/, sourced by the Makefile via -I/vpath instead of being copied per process.
• New per-process files: ColorData.h (merges colour channel information + the color matrix), ProcessData.h (merges contexpr process data e.g. nwf, extended with the helicity/flavor tables that used to be inline in CPPProcess.cc), ProcessTables.h (flavor-coupling / broken-symmetry tables) — single source of truth read by both CPPProcess and backend-owned code.
• Shrunk CPPProcess.cc: all funcs related to SigmaKin computation moved to /backend/Sigmakin fpeEnable()/getCompiler() moved to backend files; per-parameter imperative literal-assignment code replaced by a generated pointer-to-member table plus one generic gatherFptype/gatherCxtype loop written once in Parameters.h.
• Deleted dead code: Bridge.h, BridgeKernels.*, fbridge.*, fsampler.cc (zero remaining references);
• **GpuAbstraction.h/GpuRuntime.h made gpu-only.
• Namespace unification: mg5amcCpu/mg5amcGpu → one madmatrix namespace (the split is now physical, not namespace-based)
• Backend call sites, no more #ifdef at the point of use:
  • createNormalizedColorMatrix(): no-op inline on cpu/simd, real gpuMemcpyToSymbol on gpu.
  • ColorData.h's device→host copy of channel2iconfig removed entirely (its only consumer, runTest.cc, isn't wired into any export path — kept a local #ifndef/#else there since it can't get per-process %-substituted literals the way ColorData.h can for now).
• Removed backend/{cpu,gpu}/mgOnGpuVectors.h entirely — neither backend has real C++ vector types; their scalar-only content (neppV, fpternary/cxternary/maskand, the _sv typedefs, cxabs2) merged into mgOnGpuCxtypes.h. The handful of shared, non-per-backend .inc files that need this now pick the right header via __has_include. Stripped the remaining always-true #ifdef MGONGPU_CPPSIMD guards (and the dead code blocks they wrapped) from backend/simd/mgOnGpuVectors.h.
• Standalone files are guarded for output standalone_mg7 only
• USEBUILDDIR defaults to 1 for cleaner builds
• rambo used in standalone mode for both massive and massless version as well as the random number kernels live in ./src/rambo and is sourced when P1_* is compiled

Net effect

• defined MGONGPUCPP_GPUIMPL now survives only for poping diagnostics while carrying hip/cuda diversification and in the standalone specific files (rambo/check_sa)
• defined MGONGPU_CPPSIMD now survives only in the /backend/simd as carry for the sims lane size and in the HelAmps_sm.h polarisation functions for masking paths of conditional code

Test

• Full 4-backend (cppnone/cppsse4/cppavx2/cuda) × 3-process (gg_ttxgg, uu_huugg,epem_mupmum) bitwise match on our a100 machine (GCC + GPU) and Apple M4 Pro (Clang) checking 64 MatrixElement and MeanMatrixElemValue against the pre-refactor baseline commit (1a4f00df67, the brach from commit on main) using check_sa -p -v 256 64 1

Refs

• Summary presentation
  backend_separation.pdf
• Script for performing previous check
  check_backend_regression.py