菜单
Classifies GPU kernel bottlenecks (memory, compute, latency) from profiling metrics and applies GEAK-style workload guidance: what to prefer, consider, or deprioritize. Use after NCU/rocprof or when choosing between fusion, tiling, and scheduling changes without a clear winner.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Optimize fused cross-entropy loss kernels in Triton for NVIDIA and AMD GPUs. Covers fused log-softmax + NLL, online log-sum-exp, and large vocabulary handling. Use when writing or optimizing cross-entropy, label-smoothed CE, or similar loss kernels.
Optimize FlashAttention-style fused attention kernels in Triton for NVIDIA and AMD GPUs. Covers online softmax, tiled QK/AV GEMM, causal masking, and memory-efficient attention. Use when writing or optimizing self-attention, cross-attention, or any QKV attention kernel.
Optimize dense matrix multiplication (GEMM) kernels in Triton for NVIDIA and AMD GPUs. Covers tiled blocking, L2 cache grouping, tensor core utilization, and dual-platform autotune. Use when writing or optimizing matmul, linear layers, or any GEMM-based kernel.
Optimize RMS Normalization kernels in Triton for NVIDIA and AMD GPUs. Covers row-parallel reduction, vectorized loads, warp shuffle patterns, and rsqrt usage. Use when writing or optimizing RMSNorm, LayerNorm, or similar normalization kernels.
Optimize Rotary Position Embedding (RoPE) kernels in Triton for NVIDIA and AMD GPUs. Covers interleaved/sequential layouts, sincos precomputation, and multi-row processing. Use when writing or optimizing RoPE, position embeddings, or rotary transformations.
Optimize fused softmax kernels in Triton for NVIDIA and AMD GPUs. Covers online max/sum reduction, multi-row processing, numerical stability, and warp-level patterns. Use when writing or optimizing softmax, log-softmax, or similar row-wise normalization kernels.
| name | gpu-kernel-bottleneck-diagnosis |
| description | Classifies GPU kernel bottlenecks (memory, compute, latency) from profiling metrics and applies GEAK-style workload guidance: what to prefer, consider, or deprioritize. Use after NCU/rocprof or when choosing between fusion, tiling, and scheduling changes without a clear winner. |
Integrates GEAK-style workload_guidance with the AutoKernel six-tier playbook: diagnosis picks which tier to emphasize next.
Use measured % peak memory bandwidth and % peak FP16/FP8 TFLOPS (or vendor counters):
IF dram_bw_pct is high AND tflops_pct is low
→ memory-bound → Tiers 1–2 first, then fusion (Tier 3)
ELIF tflops_pct is high AND dram_bw_pct is moderate/low
→ compute-bound → Tier 3 math/MMA, then Tier 1 occupancy/reg tuning
ELIF both percentages are low AND many short kernels / low occupancy
→ latency-bound → Tier 4 persistence/fusion, reduce launches
ELSE
→ mixed / unclear → gather second metrics (L2 hit, TC active, wave occupancy)
| Priority | Meaning | Example actions |
|---|---|---|
| Prefer First | Highest ROI for this diagnosis | Memory-bound → L2 tiling + vector loads |
| Consider | Secondary, validate with profile | Compute-bound → fast math after MMA layout fixed |
| Deprioritize | Low ROI or harmful until root cause shifts | Latency-bound → micro-optimizing div/sqrt in cold code |
| Bottleneck | Primary tiers |
|---|---|
| Memory | 1–2 (tile + hierarchy), 3 fusion |
| Compute | 1 occupancy, 3 math, 5 vendor MMA |
| Latency | 4 scheduling, 3 fusion, multi-GPU overlap |
dram_throughput_pct, sm_throughput_pct / MFMA, occupancy_pct, kernel count, duration) from the profiling skill.