Extends the references from Hopper-only to sm_89 / sm_90a / sm_100a / sm_120a, every row produced by compiling for that target with CUDA 12.8. Only sm_90 hardware was available, so 'yes' means it assembles -- a lower bound on availability, not a claim about speed.
wgmma is Hopper-only and does NOT assemble for Blackwell, which uses tcgen05 with an accumulator in tensor memory rather than registers. tcgen05 is datacenter-only (sm_100a yes, sm_120a no). Ada is an Ampere-class model with fp8 added: no TMA, clusters, stmatrix or setmaxnreg, and 99 KB opt-in shared memory against Hopper's 227 KB.
fp4/fp6/MX-scale need care: e2m1x2 packs into 8 bits and requires a .reg .b8 inside the asm block, and ue8m0x2 accepts .rz only. Written wrongly they report 'Arguments mismatch', which reads like the instruction is absent.
Adds the Ampere/Ada/Hopper/Blackwell tuning guides to the fetcher and a --matrix mode to check_toolchain.py.
agent/: greppable offline docs, and a chapter-level guide to NVIDIA's documentation
PDFs dropped. A 700-page reference you cannot grep is not useful offline, so the fetcher now emits greppable text -- 7 documents, ~11 MB instead of 36 MB, with code examples keeping their indentation.
ref/official-docs.md is now a working guide: what each official document contains chapter by chapter, with section numbers read out of the CUDA 12.8 documents themselves. Includes a routing table from question to chapter, and the two things not to look up on the web -- device properties (query them) and runtime API signatures (grep the installed headers).
search_docs.sh greps the fetched corpus and falls back to the CUDA headers. Verified with --network none: wgmma syntax, fence.proxy.async ordering, and the cuTensorMapEncodeTiled signature all resolve offline.
The documents themselves remain un-redistributable and are not included; each user fetches their own copy from NVIDIA under NVIDIA's terms.
agent/: cite and fetch NVIDIA's official docs (not redistributable)
Adds agent/ref/official-docs.md and agent/tools/fetch_nvidia_docs.sh.
NVIDIA's CUDA/PTX documentation is copyrighted and cannot be redistributed here: the CUDA EULA distributes only the runtime libraries enumerated in its Attachment A, which contains no documentation, and separately forbids making the SDK subject to a licence requiring it be redistributable at no charge. So this ships an index, a citation, and a fetcher -- each user downloads their own copy direct from NVIDIA under NVIDIA's terms.
The fetcher matches docs to the toolkit: docs.nvidia.com serves the newest CUDA, but CUDA 12.8 ships PTX ISA 8.7 rather than the 9.3 now published. Verified live: 8/8 documents fetch, 6 with PDFs.
agent/: verified PTX, CUDA C++, Triton and CuTe references
Adds agent/ref/ (ptx, cuda-cpp, triton, cute-cutlass) and agent/tools/check_toolchain.py.
Every instruction and API was compiled with the container's own nvcc on sm_90a / CUDA 12.8: 39/41 PTX instructions assemble (tcgen05 is Blackwell-only and recorded as such), 24/24 CUDA C++ constructs compile, and the Triton tables come from introspecting the installed 3.6.0.
Measured facts that are easy to get wrong: wgmma.m64nNk16.f32 needs exactly N/2 accumulator registers per thread, so an m64n256k16 tile spends 128 registers on the accumulator alone; and Triton 3.6 accepts cache_modifier='.cg' with eviction_policy='evict_first' but ptxas rejects the pair.
check_toolchain.py re-derives the tables on whatever GPU the agent actually has.
Add agent/: the kernel-optimization skill and four subagents
A Claude Code skill (7-step loop: harness -> roofline -> algorithm -> toolchain -> implement -> profile -> specialize -> iterate) plus kernel-algorithmist, kernel-implementer, kernel-profiler and kernel-verifier.
Tools are verified on device: roofline.py measures the machine instead of quoting a datasheet, profile.sh runs nsys-then-ncu and emits a diagnosis rather than a metric dump, dump_ir.sh extracts Triton/Inductor IR and SASS.
README.md has the install steps and the two things that bite: ncu needs SYS_ADMIN or it exits ERR_NVGPUCTRPERM, and counters are hints not verdicts -- cuBLAS's own GEMM profiles at 15% occupancy with 168 regs/thread while being essentially optimal.
