[DUMP] sm_86: kernel 01 (RTX 3090, Ampere)

[swesty]

Target SM version

sm_86 (RTX 3090)

GPU model

NVIDIA GeForce RTX 3090

CUDA toolkit version

12.8 (nvcc V12.8.93)

NVIDIA driver version

580.159.03

Which kernel chapter

corpus/basics/01_vector_add

Additional context

Hi — I have an RTX 3090 (Ampere, sm_86) available and noticed sm_86 isn't yet covered anywhere in the corpus (only sm_89 and sm_120 are present). I'd like to contribute sm_86 dumps, starting with kernel 01.

Environment is ready and the pipeline is verified — I already produced a clean sm_86 dump of kernel 01 locally to confirm it builds:

• Pop!_OS, driver 580.159.03
• CUDA toolkit 12.8 (nvcc V12.8.93), gcc 11.4
• cuobjdump + Nsight Compute (ncu 2025.1.1) available
• source commit: 0790169

A few things I'd like you to confirm before I do the work, to keep dumps comparable:

1. Compile flags. I used nvcc -arch=sm_86 <kernel>.cu (plus -ptx / -cubin). Do you want any specific flags (e.g. -O, -lineinfo, --generate-line-info, a fixed -arch/-code pairing) to match how the sm_120/sm_89 dumps were made?
2. Toolkit version. Mine is 12.8; your reference env shows 13.2. sm_86 is fully supported on 12.8 — is that acceptable, or do you want a specific toolkit?
3. Layout. CONTRIBUTING.md documents dumps/sm_XX.sass + sm_XX.meta.md, but the repo currently uses an sm_89/ subfolder with NN.ptx, NN.sass, env.txt. Which layout should I follow for sm_86?
4. NCU. Reproducibility requirements list an NCU report, but the existing sm_89 dumps don't include one. Do you want .ncu-rep for sm_86 too? (I can run it on the physical 3090.)

Scope I can cover: the 12 kernels that already have sm_89 dumps — basics 01–08, warp_collectives 09–10, math_and_spills 11–12 — one dump per PR per your workflow. The tensor_cores chapters (13–25) target Blackwell features (fp8 QMMA, fp4, sparse MMA, stmatrix) that Ampere can't compile, so those are out of scope for a 3090.

I'll wait for your go-ahead on flags/layout before opening any PR.