h1 fatbin build, 2026-10-15

The build log of mmaemu/backends/cuda/bin/h1.fatbin, the CUDA kernel of the hopper preset (32-term blocks; mmaemu/backends/cuda/csrc/h1.cu). It was written by python tools/build_fatbin.py --release --modules h1 --out <dir> with nvcc 12.8.61 and the release flags (-O3 --fmad=false), as in docs/release.md, section 2. Only this module was built. Its manifest entry was added to mmaemu/backends/cuda/bin/manifest.json, and the entries of the five frozen release modules (v1, v4, v9, tc3, probes; their build) are unchanged: tools/build_fatbin.py --check passes for all six. A second build gave a byte-identical fatbin (sha256 261a7605…).

Commit 3d7dc66 changed only a comment in h1.cu (the unroll note now states the register counts below). The rebuild was byte-identical (sha256 261a7605…, same build report and cuobjdump listing), so only the manifest's source_sha256 for h1 changed (1b5e42b9…).

• build_report.json: the nvcc command (SASS sm_75 to sm_120, PTX compute_75) and the ptxas register, stack and spill counts of every architecture: 63 to 96 registers and 0 spill bytes everywhere. The product loop is unrolled 16-way, because a full 32-way unroll spilled 16 bytes on sm_80. There are no compiler warnings and no forbidden PTX.
• cuobjdump.txt: cuobjdump --list-elf --list-ptx of the fatbin. The architectures equal the requested ones.