analyze-kernel-bottleneck
À propos
Cette Compétence Claude analyse les noyaux GPU pour les classer comme limités par le calcul, par la mémoire ou par la latence, en utilisant l'analyse roofline, les calculs d'occupation et l'inspection des instructions SASS. Elle fournit une matrice de décision pour recommander des stratégies d'optimisation spécifiques comme cp.async ou le tiling. Utilisez-la pour le profilage avancé des performances des noyaux CUDA et pour guider des optimisations GPU ciblées.
Installation rapide
Claude Code
Recommandénpx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/analyze-kernel-bottleneckCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Analyze Kernel Bottleneck
Identify GPU kernel = compute-bound, memory-bound, latency-bound. Baseline perf → roofline classify → occupancy + compute/load ratio/tile → SASS instr mix + stall codes → smem cliff → decision matrix → right opt strategy.
Use When
- Pre-opt any CUDA kernel → baseline + classify
- After 1st working ver → ID opt path
- Underperforms vs theoretical peak
- Deciding cp.async vs larger tiles vs algorithmic restructure
In
- Required: Compiled kernel (
.cubinor.cu+ build cmd) - Required: Bench harness launching via CUDA event timing
- Required: Problem dims (M, N, K for GEMM; seq_len, heads, head_dim for attention)
- Optional: Target GPU arch (default: GA104 / sm_86 / RTX 3070 Ti)
- Optional: Expected peak util % for compare
- Optional: Prior profiling data (Nsight Compute)
Do
Step 1: Baseline Perf
Run kernel w/ CUDA events (BenchTimer), record ms. Calc effective throughput:
- Compile if not built:
nvcc --cubin -arch=sm_86 -O2 -o kernel.sm_86.cubin kernel.cu nvcc -arch=sm_86 -O2 -o bench bench.cu -lcuda -I../../phase2/common - Run representative sizes, warmup pre-measurement:
./bench 4096 4096 4096 - Record kernel ms from CUDA events (not wall-clock).
- Calc effective GFLOPS + BW:
- GEMM:
effective_gflops = (2 * M * N * K) / (time_ms / 1000) / 1e9 - BW-limited:
effective_bw = total_bytes / (time_ms / 1000) / 1e9 - Flash Attention:
effective_gflops = (4 * batch * heads * seq_len^2 * head_dim) / (time_ms / 1000) / 1e9
- GEMM:
→ Baseline: kernel ms, effective GFLOPS, effective BW.
If err: Check launches no err (CHECK_CU). Warmup pre-measurement. Dims large enough saturate GPU (small → launch overhead bottleneck).
Step 2: Roofline Classify
Arithmetic intensity vs machine balance → classify:
- Calc AI:
AI = FLOPs / bytes_loaded_from_global_memory. Count only unique bytes from DRAM (not shared mem or register reuse). - Lookup balance:
balance = peak_compute / peak_bandwidth. - Classify:
AI < balance→ memory-bound.AI > balance→ compute-bound.
GA104 (RTX 3070 Ti) Reference:
| Resource | Peak | Unit |
|---|---|---|
| FP32 FFMA | 21.7 | TFLOPS |
| FP16 Tensor Core (HMMA) | 174 | TFLOPS |
| INT8 Tensor Core (IMMA) | 696 | TOPS |
| DRAM Bandwidth | 608 | GB/s |
| L2 Cache | 4 | MB |
| SMs | 48 |
Derived Balance Points:
| Precision | Balance Point (FLOP/byte) |
|---|---|
| FP32 FFMA | 21700 / 608 = 35.7 |
| FP16 TC | 174000 / 608 = 286.2 |
| INT8 TC | 696000 / 608 = 1144.7 |
- Compute attained:
attained = effective_throughput / peak_throughput. Memory-bound → compare effective BW to 608 GB/s. Compute-bound → compare effective GFLOPS to relevant peak.
→ Classification: compute-bound, memory-bound, latency-bound (low occupancy → neither saturated) + numerical justification.
If err: Recheck byte counting. Watch redundant re-reads (e.g., 9x in direct conv2d no im2col). Neither saturated → latency-bound (Step 3).
Step 3: Occupancy
Active warps/SM from launch config + resource usage:
- Extract resource usage:
nvcc --cubin -arch=sm_86 -O2 --resource-usage -o kernel.sm_86.cubin kernel.cu 2>&1 | grep -E 'registers|smem' - Launch config:
warps_per_block = threads_per_block / 32. - Blocks/SM per limiting factor:
- Register:
floor(65536 / (registers_per_thread * threads_per_block)) - Smem:
floor(available_smem_per_SM / smem_per_block)→ see Step 6 cliff - Warp:
floor(48 / warps_per_block)(GA104 max: 48 warps/SM) - Block: 16 blocks/SM max GA104
- Register:
- Actual blocks/SM =
min(register_limit, smem_limit, warp_limit, block_limit). - Active warps/SM =
blocks_per_SM * warps_per_block. - Key threshold: 8 warps/SM enough latency hiding GA104. <8 = structural → latency-bound.
→ Occupancy table: blocks/SM, active warps/SM, limiting factor (registers, smem, warps).
If err: Check cuFuncSetAttribute for dynamic smem. Verify --resource-usage matches actual launch config. High register → --maxrregcount=N (trade spills for occupancy).
Step 4: Compute/Load Ratio/Tile
Count compute instrs + load bytes/K-tile from SASS (not src):
- Disassemble:
cuobjdump -sass kernel.sm_86.cubin > kernel.sass - Count compute/tile (inner K-tile loop):
grep -c 'HMMA' kernel.sass→ FP16 TC opsgrep -c 'IMMA' kernel.sass→ INT8 TC opsgrep -c 'FFMA' kernel.sass→ FP32 FMA
- Count global loads/tile:
grep -c 'LDG' kernel.sass→ global mem loads- Multiply bytes/load (typically 16 bytes for LDG.128)
- Ratio:
compute_ops / load_opsper tile. - Classify (cp.async threshold, gpu_reflections.md Insight 2):
- High (>20:1): cp.async net-neg; warp interleaving already hides DRAM latency. Focus algorithmic. Ref: Flash Attention 64 HMMA/tile = high, cp.async -5%.
- Medium (5-20:1): cp.async may help, benchmark both paths.
- Low (<5:1): cp.async strongly beneficial; loads dominate, async copy hides latency. Ref: IGEMM 8 IMMA/tile = low, cp.async +35%.
→ Compute/load ratio + classification (high/medium/low) + cp.async rec.
If err: Count from SASS not src — compiler may fuse, eliminate, reorder. Inner loop only (K-tile iter) not entire kernel.
Step 5: SASS Instr Inspect
Full SASS instr mix + stall codes:
- Disassemble (if not Step 4):
cuobjdump -sass kernel.sm_86.cubin > kernel.sass - Count instr types:
grep -c 'HMMA.16816' kernel.sass # FP16 Tensor Core grep -c 'IMMA.16816' kernel.sass # INT8 Tensor Core grep -c 'FFMA' kernel.sass # FP32 fused multiply-add grep -c 'LDGSTS' kernel.sass # cp.async (global->shared) grep -c 'LDG' kernel.sass # Global load grep -c 'STS' kernel.sass # Shared store grep -c 'LDS' kernel.sass # Shared load grep -c 'BAR.SYNC' kernel.sass # Barrier synchronization grep -c 'SHFL' kernel.sass # Warp shuffle (reductions) grep -c 'MUFU' kernel.sass # Special function unit - Stall codes critical instrs:
grep 'HMMA' kernel.sass | head -5 # Expect S08 minimum (hardware constraint) grep 'IMMA' kernel.sass | head -5 # Compiler emits S04, reducible to S02 via CuAssembler grep 'FFMA' kernel.sass | head -5 # Check for S04 (reducible to S01 on independent FFMAs) - ID opt targets:
- HMMA S08: hardware min Ampere, no reduce. Focus elsewhere.
- IMMA S04: compiler conservative. CuAssembler → S02 (15-20% gain).
- FFMA S04: independent → S01 via CuAssembler.
- Excessive BAR.SYNC: over-sync between pipeline stages.
→ Instr count table + stall code summary + ID'd opt targets.
If err: cuobjdump arch matches kernel compile target (both sm_86). SASS out empty → cubin corrupt → recompile.
Step 6: Smem Cliff
Smem usage crosses arch-specific occupancy cliff?
- Read smem/block from
--resource-usage(Step 3) orcuobjdump --res-usage kernel.sm_86.cubin. - Vs cliff:
- GA104 (sm_86): 100 KB max smem/SM. Cliff at 50 KB/block.
- Confirmed: 48 KB/block → 2 blocks/SM (good), 56 KB/block → 1 block/SM (2x regression).
- Above cliff (smem >50 KB/block):
- Blocks/SM drops to 1, active warps drop to warps_per_block (typically 4).
- 2x regression from exposed DRAM stalls.
- Double-buffering impact: Doubles smem. 30 KB current → 60 KB double-buf → crosses cliff. Eval async benefit vs occupancy loss.
- Record smem/block, blocks/SM, cliff crossed?
→ Smem/block + blocks/SM + explicit statement cliff crossed.
If err: Above cliff + occupancy bottleneck → change strategy: reduce tile → smem <50 KB, or accept 1 block/SM + compensate higher compute/load ratio (more register reuse, longer K-tiles).
Step 7: Decision Matrix
Synthesize Steps 2-6 → opt strategy:
| Condition | Strategy |
|---|---|
| Memory-bound + low compute/load (<5:1) + smem under cliff | SW pipelining cp.async (LDGSTS). Overlap global loads w/ compute. |
| Memory-bound + high compute/load (>20:1) + 8+ warps | Warp interleaving already hides. Focus algorithmic: implicit GEMM, split-Q, im2col. |
| Compute-bound + FFMA-heavy | CuAssembler stall tighten: S04 → S01 on independent FFMAs. |
| Compute-bound + HMMA-heavy | S08 hardware min, no reduce. Increase tile reuse (larger M/N, longer K-loop). |
| Compute-bound + IMMA-heavy | CuAssembler: S04 → S02 on IMMA (compiler conservative). |
| Latency-bound (low occupancy) | Reduce smem/registers → more blocks/SM. >8 warps/SM. |
| Smem above cliff | Reduce tile or restructure → smem/block <50 KB (GA104). |
- Rank strategies by expected gain, via compute/load + occupancy data.
- Estimate gain range per strategy, how far from relevant ceiling.
- Flag conflicts: cp.async doubles smem (may cross cliff), larger tiles → register pressure (may reduce occupancy).
→ Ranked list recommended opts + predicted gain + conflicts.
If err: No clear winner → micro-benchmarks isolate each (cp.async alone, reduced tile alone) → measure actual pre-combine.
Step 8: Doc Findings
Structured bottleneck report:
- Baseline: kernel ms, effective GFLOPS + BW, problem dims.
- Roofline: AI, classification, attained fraction.
- Occupancy: blocks/SM, active warps/SM, limiting factor.
- Compute/load: ratio, classification, cp.async rec.
- SASS summary: instr counts, stall findings, CuAssembler targets.
- Smem cliff: smem/block, blocks/SM, status.
- Rec: ranked opt strategies + gain estimates.
## Bottleneck Analysis Report: [kernel_name]
### Baseline
- Problem: [dimensions]
- Kernel time: [X] ms
- Effective GFLOPS: [Y] | Effective BW: [Z] GB/s
### Roofline Classification
- Arithmetic intensity: [AI] FLOP/byte
- Balance point: [BP] FLOP/byte ([precision])
- Classification: **[compute|memory|latency]-bound**
- Attained fraction: [X]% of peak
### Occupancy
| Resource | Per Block | Limit/SM | Blocks/SM |
|----------|-----------|----------|-----------|
| Registers | [N]/thread | 65536 | [B] |
| Shared mem | [X] KB | 100 KB (cliff: 50 KB) | [B] |
| Warps | [W] | 48 | [B] |
| **Limiting** | | | **[min(B)]** |
- Active warps/SM: [W] ([sufficient|insufficient] for latency hiding)
### Compute/Load Ratio
- Compute ops/tile: [N] [HMMA|IMMA|FFMA]
- Load bytes/tile: [N] bytes ([N] LDG x [N] bytes)
- Ratio: [X]:1 — **[high|medium|low]**
- cp.async recommendation: [beneficial|neutral|detrimental]
### SASS Instruction Mix
| Instruction | Count | Notes |
|-------------|-------|-------|
| HMMA.16816 | [N] | Stall: S08 (hardware min) |
| IMMA.16816 | [N] | Stall: S04 (reducible to S02) |
| FFMA | [N] | Stall: S04 (reducible to S01) |
| LDG | [N] | |
| LDGSTS | [N] | cp.async |
| BAR.SYNC | [N] | |
### Smem Cliff
- Smem/block: [X] KB — [under|over] 50 KB cliff
- Blocks/SM: [B] — [no occupancy loss|occupancy halved]
### Recommended Optimizations (ranked)
1. [Strategy] — estimated [X-Y]% gain
2. [Strategy] — estimated [X-Y]% gain
3. [Strategy] — estimated [X-Y]% gain
→ Complete MD report consumable by kernel-optimizer agent or dev.
If err: Re-run different sizes (1024, 2048, 4096, 8192) → confirm not size-specific. Small may appear latency-bound when real bottleneck at scale is BW.
Check
- Baseline via CUDA events (not wall-clock)
- Roofline classification (compute/memory/latency bound)
- Occupancy + limiting factor
- Compute/load ratio/tile from SASS
- SASS instr mix + stall codes documented
- Smem cliff vs arch threshold
- Decision matrix + strategy rec
- Findings in structured report
Traps
- Re-read multiply: Direct conv2d reads weight 9x no im2col → byte count inflated 9x. Use actual unique bytes from DRAM, not total load instrs, for AI.
- Confuse FP16 TC peak w/ FP32: FP16 TC peak 174 TFLOPS, FP32 FFMA 21.7 TFLOPS — 8x diff. Wrong peak → roofline meaningless.
- Using 64 KB cliff not 50 KB GA104: GA104 (sm_86) 100 KB max smem/SM. Cliff 100/2 = 50 KB/block, not 64 KB. Arch-specific; other GPUs differ.
- Ignore warp interleaving when eval cp.async: 8 warps long compute (high compute/load) already hide DRAM via warp sched. cp.async → smem pressure + barrier overhead no benefit (Flash Attention -5%).
- Count instrs from src not SASS: Compiler may fuse, eliminate dead, unroll differently, reorder. Always from
cuobjdump -sass. - No warmup iters: 1st launch → JIT compile overhead + cold cache. 2-5 warmup pre-measured run.
→
pipeline-gpu-kernel— impl SW pipelining cp.async when memory-bound + low compute/loadsimulate-cpu-architecture— complementary arch analysis CPU-side bottlenecks in host-device workflows
Dépôt GitHub
Compétences associées
llamaguard
AutreLlamaGuard est le modèle de Meta, doté de 7 à 8 milliards de paramètres, conçu pour modérer les entrées et sorties des LLM selon six catégories de sécurité comme la violence et les discours haineux. Il offre une précision de 94 à 95 % et peut être déployé avec vLLM, Hugging Face ou Amazon SageMaker. Utilisez cette compétence pour intégrer facilement le filtrage de contenu et des garde-fous de sécurité dans vos applications d'IA.
cost-optimization
AutreCette compétence de Claude aide les développeurs à optimiser les coûts du cloud grâce au redimensionnement des ressources, aux stratégies d'étiquetage et à l'analyse des dépenses. Elle fournit un cadre pour réduire les dépenses cloud et mettre en œuvre une gouvernance des coûts sur AWS, Azure et GCP. Utilisez-la lorsque vous devez analyser les coûts d'infrastructure, redimensionner les ressources ou respecter des contraintes budgétaires.
quantizing-models-bitsandbytes
AutreCette compétence quantifie les LLMs en précision 8 bits ou 4 bits à l'aide de bitsandbytes, permettant une réduction de 50 à 75 % de la mémoire utilisée avec une perte de précision minime. Elle est idéale pour exécuter des modèles plus volumineux sur une mémoire GPU limitée ou pour accélérer l'inférence, prenant en charge des formats comme INT8, NF4 et FP4. La compétence s'intègre à HuggingFace Transformers et permet l'entraînement QLoRA ainsi que l'utilisation d'optimiseurs en 8 bits.
dispatching-parallel-agents
AutreCette compétence Claude déploie plusieurs agents pour enquêter et résoudre simultanément 3 problèmes indépendants ou plus. Elle est conçue pour des scénarios impliquant des défaillances non liées qui peuvent être résolues sans état partagé ni dépendances. La capacité fondamentale est la résolution de problèmes en parallèle, en assignant un agent par domaine problématique indépendant afin de maximiser l'efficacité.
