h4 fatbin build, 2026-10-16¶
The build log of mmaemu/backends/cuda/bin/h4.fatbin, the fast CUDA kernel of the hopper preset (kernel v4's
arithmetic with 32-term blocks; mmaemu/backends/cuda/csrc/h4.cu). It was written by
python tools/build_fatbin.py --release --modules h4 --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; the entries of the five frozen release modules
(v1, v4, v9, tc3, probes; their build) and of h1
(its build) are unchanged: tools/build_fatbin.py --check passes for all
seven. A second build gave a byte-identical fatbin (sha256 7e1f96c3…) and the same build report.
- build_report.json: the nvcc command (SASS sm_80 to sm_120, PTX compute_80; like v4, h4 needs
the bf16x2 FMA of sm_80) and the ptxas register, stack and spill counts of every architecture: 248 to 255
registers (
__launch_bounds__(256, 1)) and 0 spill bytes everywhere. There are no compiler warnings and no forbidden PTX. - cuobjdump.txt:
cuobjdump --list-elf --list-ptxof the fatbin. The architectures equal the requested ones.