Menu
SkillsMP a collecté 8 skills depuis teddyjfpender/stwo-zig. Ouvrez un skill pour examiner sa source et ses détails.
Affichage de 8 skills collectés sur 8.
Isolate a Metal kernel and measure real GPU time, pipeline reflection (occupancy limits, execution width, threadgroup memory), and approximate achieved bandwidth via the generic runner; wrap whole commands in Metal System Trace. Use for GPU performance…
Langue du texte source : anglais
Implement and review CUDA prover kernels under exact field, transcript, AOT, ABI, memory-traffic, and occupancy contracts. Use after profiling identifies a dominant CUDA kernel or pass.
Langue du texte source : anglais
Design generic, process-resident CUDA proving architectures from ProofProgram and CudaPlan evidence. Use for runtime ownership, arenas, caching, streams, graphs, AOT modules, pass reduction, proof concurrency, and frontend/backend boundaries.
Langue du texte source : anglais
Attribute CUDA prover latency with product stage events, NVTX, Nsight Systems, and Nsight Compute while preserving an uninstrumented verified-request verdict. Use before changing CUDA schedules, memory passes, kernels, streams, graphs, allocation, or module…
Langue du texte source : anglais
Design and review high-performance Metal compute, render, and hybrid architectures for Apple GPUs. Use when writing or reviewing Metal host code or MSL; choosing storage modes, resource lifetimes, synchronization, command-buffer structure, bindings, pipeline…
Langue du texte source : anglais
Isolate Zig code into a scratch harness and profile it with in-process hardware counters (instructions, cycles, IPC, energy), stack sampling, codegen summaries, and ABBA A/B comparison via the stwo-perf statistics. Use for any CPU performance question —…
Langue du texte source : anglais
Capture, sanitize, and attach the agent session transcripts that produced an optimization — the default part of every stwo-perf submission flow, and the most valuable dataset the harness curates. Use from the moment work begins (capture as you go, not after)…
Langue du texte source : anglais
Match algorithm-design work to canonical problems, computational models, complexity and parameterized results, published algorithms, and maintained implementations before coding. Use before creating or replacing algorithms for optimization, decision, search,…
Langue du texte source : anglais