菜单
Tightens inner-loop math, mixed precision, fused epilogues, and fast math intrinsics for GPU kernels. Use when kernels are compute-bound, epilogue-bound, or dominated by elementwise ops around matmul-like cores.
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
| name | gpu-opt-tier3-compute-fusion |
| description | Tightens inner-loop math, mixed precision, fused epilogues, and fast math intrinsics for GPU kernels. Use when kernels are compute-bound, epilogue-bound, or dominated by elementwise ops around matmul-like cores. |
AutoKernel Tier 3: reduce instructions per useful flop and fuse memory round-trips into the same kernel pass.
tl.dot acc) to preserve accuracy.Fuse into the same kernel to avoid extra global traffic:
| Pattern | When it matters |
|---|---|
| Bias (row/col) | LLM linear layers, conv fc |
| Scale / per-channel scale | quantization, norm fusion |
| Activation (GELU, ReLU, clamp) | MLP blocks |
| Quantize to int8/fp8 | inference export |
Rule: if the epilogue is memory-bandwidth limited, fusion is high ROI; if already compute-saturated, fusion still saves launches but may need register tuning.
if in inner loops.| Intrinsic / pattern | Typical use | Caveat |
|---|---|---|
__expf, __exp10f | softmax, activations | lower precision vs libm |
rsqrtf, __frsqrt_rn | norm, attention | acceptable for ML norms with care |
__sincosf | rotary, trig | pair eval saves bandwidth |
__fdividef | non-critical divisions | faster, less IEEE |
When each matters: transcendentals dominate in softmax/attention/layer-norm; fdivide helps cheap color-space or scaling loops. Prefer intrinsics in hot paths, library precision on numerically sensitive reductions.
--use-fast_math (or equivalent) globally enables aggressive approximations; prefer scoped intrinsics when you need deterministic training behavior.tl.math / libdevice-style ops that match the desired IEEE vs fast semantics for the pass.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.