[willend]

Fixes #392. This ports mccode-dev/McCode#2808 (fix for mccode-dev/McCode#2807) to mccode-antlr's code generator, so the generated C matches McCode's classic cogen.c.in.

Warning

Results change for every KISS-RNG run and for every instrument with SPLIT. They will not be bit-identical to earlier mccode-antlr output. The new output matches McCode's classic generator with #2808.

Changes

• Per-ray seeding (c_raytrace.py, in both raytrace_all and the funnel loop): srandom(_hash((pidx+1)*(seed+1))) becomes srandom(_hash(_hash(seed) + cloop*gpu_innerloop + pidx)), and the per-batch seed = seed + gpu_innerloop is removed. The old key was a product, so different seeds and different MPI ranks shared rays, and the key wrapped at 2^32 on MSVC.
• _uid is now cloop*gpu_innerloop + pidx, plus mpi_node_rank * loops*gpu_innerloop under MPI. This makes it unique across GPU batches. On CPU the values are unchanged.
• randstate[RANDSTATE_LEN] replaces the hard-coded randstate[7] in the particle struct (c_header.py). Unlike what Port per-ray RNG seeding and _uid fix from McCode cogen (mccode-dev/McCode#2807) #392 assumed, mccode-antlr's RNG_ALG block did not define RANDSTATE_LEN; the runtime header that defines it comes after the struct. So this PR adds #define RANDSTATE_LEN 1 / 7 to that block, as cogen.c.in does. Older runtimes (v3.8.7) define the same value, which is a legal identical redefinition.
• SPLIT bug: each SPLIT child used to restore only randstate[0], so KISS children restarted words 1-6 from the same state and got correlated streams. All RANDSTATE_LEN words are now backed up and restored, with the same loop and variable names as cogen.c.in.

sort_absorb_last() (FUNNEL + SPLIT) lives in the shared runtime mccode-r.c, so mccode-antlr picks it up with the McCode release.

Tested

• pytest with MCCODEANTLR_MCCODE_POOCH__TAG=v3.8.7 and with latest (v3.9.2), C compile tests included: 1267 passed, 10 skipped (MPI, OpenACC, mcpl, build123d, matplotlib). No test needed changing.
• Test_RNG_rand01: generated with this branch, compiled with the McCode #2808 runtime (786934b; its mccode-r.c / mccode-r.h.in changes spliced into the embedded runtime), and run with -s 1000 Ncount=1000. The md5 of the non-# lines of rngout.dat is 6c203d69205b0656c1391555e6068c93, the same as McCode's classic generator.
• Unittest_SPLIT with SPLITS=10, -n 1e5: compiles and runs, PSD_I=1, PSD_N=1e6, as the instrument's example expects.

Not tested

• MPI and OpenACC/GPU builds (no toolchain here), the funnel loop at runtime, McXtrace, and MSVC.
• No statistical test of the SPLIT children's streams in mccode-antlr. McCode validated the seeding against a Mersenne Twister referee and with BigCrush.

🤖 Generated with Claude Code

[willend]

@g5t parallel to McCode PR mccode-dev/McCode#2808

[g5t]

A spot check looks like this is identical to your changes in McStas 🤞

[willend]

@g5t cool, thanks. I will now proceed to merge on the McStas side also.

I have adapted my CI to take mccode-antlr from main, should give us a clearer picture of the state of everything going forward.