madmatrix GPU: running flavour couplings, no good helicity, stream order, unset selections

[oliviermattelaer]

Stacked on #244 (same GPU template files and GPU CI job list); retarget to main once #244 is merged.

Four bugs of the GPU backend of madmatrix (the CPU and SIMD backends were right), found by the review that followed the HIP cut crash (#254). Same class as that one: silent or lucky on one platform, a fault on another.

Fixes (backend/gpu, plus the generator)

1. Running flavour couplings (nDPF > 0) read and written out of bounds. The per-thread dpf_value (the alpha_s-dependent flavour couplings of this event, gathered per event) was accessed with CD_ACCESS, which on the device adds the offset of the event in the all-events coupling buffer. Every thread past the first event page went outside dpf_value, in the gather and in the vertex routines.
   • Affects MSSM SUSY-QCD (squark–quark–gluino GC_106/GC_110). Every subprocess of p p > go go has nDPF = 16, g g > go go included, because the count accumulates over subprocesses.
   • SM p p > t t~ has nDPF = 0, so the GPU CI never saw it.
   • The calls on flvCOUPs_dep are now instantiated with a new CDPF_ACCESS: CD_ACCESS on CPU/SIMD (unchanged), the trivial one-event access on the GPU.
   • The generator refuses a vertex that mixes such a coupling with another running one (one access type per routine). None does in MSSM p p > go go.
2. No good helicity aborted. For a subprocess whose |M|² vanishes identically (e.g. FCNC with zero Wilson coefficients), the helicity loop was launched with gridDim.y = 0, an invalid configuration, and checkGpu aborted. The loop is now skipped and |M|² = 0, as on the CPU.
3. Selected helicity and colour left unset for a vanishing or NaN |M|², so umami returned stale pool memory. They are now 0 (none, -1 out of umami) unless one is chosen, as the zero-initialised CPU buffers give.
4. Stream order. In async (umami) mode the memsets and the helicity, output and colour kernels went to the default stream, while everything else ran on the stream umami is given.
   • This was only correct because madspace never passes its stream (a madspace bug, fixed in a separate PR) and its streams are blocking.
   • With a non-blocking stream (a torch side stream) it races.
   • All of sigmaKin now runs on the given stream, and umami waits for its input copies before the helicity filtering of the first call (default stream).
   • Memset sizes are now computed in size_t.

GPU CI: new job madmatrix

madmatrix_gpu.sh / madmatrix_checks.py compare the GPU backend with the scalar CPU backend through umami. Each check runs in its own process, so an abort fails only that check:

• mssm_gpu_vs_cpu: p p > go go (both subprocesses), event by event, with a random alpha_s and flavour per event, plus an alpha_s² scaling check;
• zero_me: b b~ > ta+ ta- / z a with ymtau = 0 (no good helicity): |M|² = 0, helicity -1, colour none or the only flow;
• nonblocking_stream: g g > t t~ g on a non-blocking stream must match the default stream over 4 calls of 16384 events.

The umami ctypes code moves to umami_harness.py, shared with interference_checks.py. Its CPU buffers now keep the host arrays alive; temporaries were freed before the call, which this PR's checks would have hit.

Validation

• CPU: test_madmatrix_mssm_gogo (madmatrix vs Fortran standalone, MSSM p p > go go) passes. Unit tests .*madmatrix.*, .*cudacpp.*, .*flav.*: 59 OK.
• The check script runs locally on CPU backends. mssm_gpu_vs_cpu with scalar against scalar exercises the plumbing, the nDPF parsing and the alpha_s² scaling. zero_me on simd_128 passes: the CPU gives |M|² = 0 and helicity -1.
• GPU, all jobs green on both clusters (pp_ttx, pp_jj, interference, madmatrix):
  • manneback (CUDA): https://github.com/MadGraphTeam/MadGraph7/actions/runs/38046370291
  • lemaitre4 (HIP, MI300A): https://github.com/MadGraphTeam/MadGraph7/actions/runs/38046372350
  • mssm_gpu_vs_cpu: GPU = CPU event by event to 2.8e-15 (CUDA) and 1.5e-15 (HIP), nDPF = 16 in both subprocesses;
    zero_me: |M|^2 = 0, helicity -1, colour -1; nonblocking_stream: 4 calls of 16384 events identical to the default stream.