pipeline-gpu-kernel
关于
This skill restructures tiled GPU kernels to implement software pipelining, enabling the overlap of global memory loads with Tensor Core computations. It guides developers in choosing between LDG-register and cp.async methods based on compute/load ratios and managing shared memory against occupancy limits. Use it to optimize kernels by restructuring prologue, main loop, and epilogue, with verification at the SASS level.
快速安装
Claude Code
推荐npx 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/pipeline-gpu-kernel在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Pipeline GPU Kernel
Double-buffer tiled GPU kernel → tile N+1 global load overlaps tile N Tensor Core compute. Sequential load-sync-compute-sync K-loop → prologue/loop/epilogue. Pick LDG-reg vs cp.async (LDGSTS) by compute/load ratio. Verify smem under arch cliff. Confirm overlap in SASS.
Use When
analyze-kernel-bottleneckflags mem-bound kernel, low compute/load per tile- Warp interleave alone can't hide DRAM latency (~300 cyc GA104)
- Sequential load-sync-compute-sync K-loop → restructurable
- Skip → ratio >20:1 + 8+ warps active
In
- Required: CUDA kernel
.cuw/ tiled K-loop, separate load + compute phases - Required: GPU arch (e.g., GA104 / sm_86 → smem cliff + occupancy)
- Required: Tile sizes (BM, BN, BK) + dtype (FP16, FP32, INT8)
- Optional: Compute/load ratio per tile (from
analyze-kernel-bottleneck) - Optional: Baseline (non-pipelined perf at target size)
Do
Step 1: Verify Preconditions
K-loop has load + compute phases split by __syncthreads(). Calc doubled smem cost vs arch cliff.
- Locate K-loop. Structure: load A+B tiles global→smem,
__syncthreads(), compute (HMMA/IMMA/FFMA) on smem tiles,__syncthreads(). - Single-buffer smem:
smem_a_size = BM * BK * sizeof(T),smem_b_size = BK * BN * sizeof(T). - Double-buffer cost:
smem_doubled = smem_a_size * 2 + smem_b_size * 2. - Vs arch cliff. GA104 (sm_86): 100 KB max smem/SM, cliff 50 KB/block (>50 KB = 1 block/SM = 4 warps, 2x occupancy collapse).
Single buffer: smem_a[BM*BK] + smem_b[BK*BN] = 2 KB + 2 KB = 4 KB
Double buffer: smem_a[2][BM*BK] + smem_b[2][BK*BN] = 4 KB + 4 KB = 8 KB
8 KB << 50 KB cliff -> 2 blocks/SM -> 8 warps
- Loop count:
num_tiles = K / BK. Pipelining needsnum_tiles >= 2.
→ Smem budget table: single + double cost, doubled under cliff, ≥2 blocks/SM.
If err: doubled > cliff → halve BK or BM until smem_doubled <= 50 KB GA104. Or use reg-only prefetch (LDG variant), no smem doubling — stage in regs, write same single buffer after __syncthreads().
Step 2: Choose Variant
LDG-reg vs cp.async (LDGSTS) by compute/load ratio per tile.
- Ratio:
ratio = (2 * BM * BN * BK) / ((BM * BK + BK * BN) * sizeof(T))(GEMM-like: 2 FLOPs/MAD, bytes/tile). - Decide:
LDG-register (ratio >= 5 or CUDA < 11.0):
- LDG tile N+1 → regs (non-blocking global loads).
- Compute on
buf[N % 2](overlaps outstanding LDGs). __syncthreads(), STS regs →buf[(N+1) % 2],__syncthreads().- Simpler, no pipeline API dep.
- Reg pressure: ~
(BM * BK + BK * BN) / BLOCK_SIZEregs/thread for staging.
cp.async (LDGSTS) (ratio < 5, CUDA >= 11.0):
__pipeline_memcpy_asynctile N+1 →buf[(N+1) % 2](async, bypass reg file).__pipeline_commit()before compute.- Compute on
buf[N % 2]. __pipeline_wait_prior(0)+__syncthreads()after compute.- Better overlap, zero reg pressure for prefetch. Needs
#include <cuda_pipeline.h>.
- Thresholds (GA104 IGEMM 4096x4096x4096):
- <5:1 → cp.async (+35% on IGEMM).
- 5-20:1 → impl both, bench.
-
20:1 → likely no gain (warp interleave enough).
→ Variant + justification (ratio + arch).
If err: ambiguous (5-20:1) → impl both, bench. cp.async = safer default if CUDA supports.
Step 3: Restructure K-Loop
Sequential load-sync-compute-sync → prologue/loop/epilogue.
-
Three sections:
- Prologue: load tile 0 →
buf[0], sync, enter loop. - Main loop: tiles 1 to
num_tiles - 1, overlap load N+1 w/ compute N. - Epilogue: compute last tile (loaded by final main iter).
- Prologue: load tile 0 →
-
LDG-register:
// === LDG-register variant ===
// Prologue: load tile 0 into buf[0]
cooperative_load_tile(smem_a[0], smem_b[0], global_a, global_b, /*k_offset=*/0);
__syncthreads();
for (int tile = 0; tile < num_tiles - 1; tile++) {
int cur_buf = tile & 1;
int next_buf = 1 - cur_buf;
// Phase 1: LDG next tile into registers (non-blocking)
float reg_a[ELEMS_PER_THREAD_A], reg_b[ELEMS_PER_THREAD_B];
prefetch_tile_to_registers(reg_a, reg_b, global_a, global_b,
(tile + 1) * BK);
// Phase 2: Compute on current buffer (overlaps with LDG flight)
tensor_core_mma(smem_a[cur_buf], smem_b[cur_buf], acc);
// Phase 3: Drain registers into next buffer
__syncthreads();
store_registers_to_smem(smem_a[next_buf], smem_b[next_buf],
reg_a, reg_b);
__syncthreads();
}
// Epilogue: compute last tile
tensor_core_mma(smem_a[(num_tiles - 1) & 1], smem_b[(num_tiles - 1) & 1], acc);
- cp.async:
// === cp.async variant ===
#include <cuda_pipeline.h>
// Prologue: async load tile 0 into buf[0]
cpasync_load_tile(smem_a[0], smem_b[0], global_a, global_b, /*k_offset=*/0);
__pipeline_commit();
__pipeline_wait_prior(0);
__syncthreads();
for (int tile = 0; tile < num_tiles - 1; tile++) {
int cur_buf = tile & 1;
int next_buf = 1 - cur_buf;
// Phase 1: cp.async next tile into next buffer (async, direct to smem)
cpasync_load_tile(smem_a[next_buf], smem_b[next_buf],
global_a, global_b, (tile + 1) * BK);
__pipeline_commit();
// Phase 2: Compute on current buffer (overlaps with LDGSTS in flight)
tensor_core_mma(smem_a[cur_buf], smem_b[cur_buf], acc);
// Phase 3: Wait for async copies to complete
__pipeline_wait_prior(0);
__syncthreads();
}
// Epilogue: compute last tile
tensor_core_mma(smem_a[(num_tiles - 1) & 1], smem_b[(num_tiles - 1) & 1], acc);
- Loop count: main loop runs
num_tiles - 1iters. Epilogue computes tile from last iter.
→ Restructured K-loop w/ clear prologue, loop, epilogue.
If err: most common bug → off-by-one buf index or skipped epilogue. Verify: prologue → buf[0], first iter compute buf[0] + load buf[1], second compute buf[1] + load buf[0], etc. Epilogue → buf[(num_tiles - 1) & 1].
Step 4: Implement Double-Buffer
Declare double-buffered smem + load fns.
- Single → double:
// Before (single buffer)
__shared__ half smem_a[BM * BK];
__shared__ half smem_b[BK * BN];
// After (double buffer)
__shared__ half smem_a[2][BM * BK];
__shared__ half smem_b[2][BK * BN];
- cp.async load fn (pipeline API):
__device__ void cpasync_load_tile(half* dst_a, half* dst_b,
const half* src_a, const half* src_b,
int k_offset) {
// Each thread copies its portion (16 bytes = 8 half values per cp.async)
int tid = threadIdx.x;
int bytes_per_thread = 16; // cp.async.cg supports 4, 8, or 16 bytes
// A tile: BM * BK elements, distributed across BLOCK_SIZE threads
int elems_a = BM * BK / BLOCK_SIZE;
for (int i = 0; i < elems_a; i += 8) {
int idx = tid * elems_a + i;
__pipeline_memcpy_async(dst_a + idx,
src_a + k_offset * BM + idx,
bytes_per_thread);
}
// B tile: BK * BN elements, distributed similarly
int elems_b = BK * BN / BLOCK_SIZE;
for (int i = 0; i < elems_b; i += 8) {
int idx = tid * elems_b + i;
__pipeline_memcpy_async(dst_b + idx,
src_b + k_offset * BN + idx,
bytes_per_thread);
}
}
- LDG variant: reg staging arrays + store fns:
// Declare register staging (size = elements per thread)
half reg_a[BM * BK / BLOCK_SIZE];
half reg_b[BK * BN / BLOCK_SIZE];
// Prefetch: LDG from global to registers (non-blocking, issued early)
for (int i = 0; i < BM * BK / BLOCK_SIZE; i++) {
int idx = threadIdx.x * (BM * BK / BLOCK_SIZE) + i;
reg_a[i] = global_a[k_offset * BM + idx];
}
// ... similarly for reg_b
// Store: STS from registers to shared memory (after __syncthreads)
for (int i = 0; i < BM * BK / BLOCK_SIZE; i++) {
int idx = threadIdx.x * (BM * BK / BLOCK_SIZE) + i;
smem_a[next_buf][idx] = reg_a[i];
}
- Keep
__launch_bounds__(BLOCK_SIZE)→ accurate occupancy info to compiler. - Compile:
nvcc --cubin -arch=sm_86 -O2 -o kernel.sm_86.cubin kernel.cu.
→ Compilable kernel, double-buffered smem, chosen load mech. Cubin gen no errors.
If err: pipeline API fail → #include <cuda_pipeline.h> + CUDA >= 11.0. Reg spills (nvcc --resource-usage) → shrink reg staging via larger BLOCK_SIZE or smaller BK.
Step 5: Verify Correctness
Pipelined kernel vs CPU ref → identical numerical out.
- Compile bench:
nvcc -arch=sm_86 -O2 -o bench bench.cu -lcuda -I../../phase2/common. - Small problem first (512x512x512) → catch index bugs before scale.
- Tolerance per dtype:
- INT8 Tensor Core (IMMA):
abs=0.5, rel=0.1 - FP16 Tensor Core (HMMA):
abs=1e-2, rel=1e-2 - FP32 scalar (FFMA):
abs=1e-3, rel=1e-3
- INT8 Tensor Core (IMMA):
- Pipelining doesn't change arithmetic — reorders loads. Fail → bug in buf index, not compute.
- Test target size (e.g., 4096x4096x4096) → verify boundary handling.
→ PASS at small + target sizes, error bounds = non-pipelined baseline.
If err: buf index bug = top suspect. Verify: compute reads buf[tile & 1], loads write buf[1 - (tile & 1)]. Epilogue uses (num_tiles - 1) & 1, not num_tiles & 1. cp.async → __pipeline_wait_prior(0) before __syncthreads(), else compute reads partial.
Step 6: Benchmark + Compare
Pipelined vs non-pipelined baseline at target size.
- Run baseline → record GFLOPS or bandwidth.
- Run each pipelined variant → same metric.
- Speedup:
speedup = pipelined_metric / baseline_metric. - Expected gains by ratio (GA104):
- Low (<5:1): +15-35% from cp.async (IGEMM: LDG +18%, cp.async +35% at 4096x4096x4096).
- Med (5-20:1): +5-15%.
- High (>20:1): 0-5% or regress.
- Both impl → pick faster for prod.
| Variant | GFLOPS | Speedup vs Baseline |
|------------------|--------|---------------------|
| Baseline | XXX | 1.00x |
| LDG-register | XXX | X.XXx |
| cp.async (LDGSTS)| XXX | X.XXx |
→ Perf table showing improvement. Chosen variant → measurable speedup matching ratio prediction.
If err: regress → check 3: (1) SASS for unexpected overhead (extra BAR.SYNC, reg spills). (2) Smem didn't cross cliff — nvcc --resource-usage or cuobjdump -res-usage. (3) Enough tiles (K / BK >= 4) → amortize prologue/epilogue.
Step 7: Verify SASS Overlap
Inspect SASS → global loads + Tensor Core overlap in main loop.
- Disassemble:
cuobjdump -sass kernel.sm_86.cubin | grep -E 'IMMA|HMMA|LDGSTS|LDG|BAR'. - Main loop pattern:
LDGSTSorLDGbeforeHMMAorIMMA.- No
BAR.SYNCbetween loads + compute (must overlap in warp scheduler). BAR.SYNCafter compute → gates next iter's use of loaded data.
- Stall codes on HMMA/IMMA: S08 HMMA pipeline delay = expected. S01-S04 IMMA = normal. LDG/LDGSTS stalls low (S01) → scheduler switches to compute while loads in flight.
- Count HMMA/IMMA per iter → should match non-pipelined (pipelining ≠ compute volume change).
# Full SASS pipeline verification
cuobjdump -sass kernel.sm_86.cubin | grep -E 'IMMA|HMMA|LDGSTS|LDG|BAR'
# Count compute instructions per loop
cuobjdump -sass kernel.sm_86.cubin | grep -c 'HMMA\|IMMA'
# Check for register spills
nvcc --resource-usage --cubin -arch=sm_86 -O2 kernel.cu 2>&1 | grep -i spill
→ SASS shows load-before-compute, no intervening barriers. Zero reg spills.
If err: compiler reordered loads after compute (overlap defeated) → (1) #pragma unroll 1 on main loop → no over-aggressive unroll. (2) Split load + compute into distinct inline fns → sequencing hint. (3) asm volatile("" ::: "memory") as compiler fence (last resort, may inhibit other opts).
Check
- Double-buffer smem under arch cliff (GA104: 50 KB/block)
- Both buffers alternate (
buf[tile & 1]) - Prologue → tile 0 in
buf[0] - Epilogue → compute
buf[(num_tiles - 1) & 1] - Correctness PASS vs CPU ref at small + target
- SASS confirms overlap (no
BAR.SYNCbetween LDGSTS/LDG + IMMA/HMMA) - Perf > non-pipelined baseline
- No reg spill (LDG variant) — check
nvcc --resource-usage
Traps
- Cross smem cliff via doubling — GA104 cliff 50 KB/block, not 64. Always calc
smem_doubledbefore impl. 28 KB single → 56 KB doubled crosses cliff, halves occupancy. +20% pipelining gain → -50% occupancy regress. - Skip epilogue compute — Last tile loaded in final iter needs own compute outside loop. Without → last BK cols of K silently dropped → incorrect results, may look like small numerical noise not obvious fail.
- Buf index off-by-one — Use
buf[tile & 1]for compute,buf[1 - (tile & 1)]for next load. Common err:buf[(tile + 1) & 1]for next = same asbuf[1 - (tile & 1)]only when 2 buffers — wrong if applied to compute index. - cp.async commit/wait order —
__pipeline_commit()BEFORE compute (seals async batch).__pipeline_wait_prior(0)AFTER compute (blocks until copies done). Swap → async becomes synchronous, kills overlap. - Missing __syncthreads — LDG variant:
__syncthreads()between compute + STS drain (compute finishes reading current buf before overwrite). Another after STS drain (all threads done writing before next iter reads). cp.async:__syncthreads()after__pipeline_wait_prior(0)→ all threads see completed copies. - Boundary in cp.async —
__pipeline_memcpy_asyncneeds valid + aligned src. Matrix edges where K not multiple of BK → last tile reads OOB. Fall back to scalar loads w/ bounds check for final, or pad inputs to BK multiple.
→
analyze-kernel-bottleneck— identify mem-bound, calc compute/load ratio for variant pick
GitHub 仓库
相关推荐技能
llamaguard
其他LlamaGuard是Meta推出的7-8B参数内容审核模型,专门用于过滤LLM的输入和输出内容。它能检测六大安全风险类别(暴力/仇恨、性内容、武器、违禁品、自残、犯罪计划),准确率达94-95%。开发者可通过HuggingFace、vLLM或Sagemaker快速部署,并能与NeMo Guardrails集成实现自动化安全防护。
cost-optimization
其他这个Claude Skill帮助开发者优化云成本,通过资源调整、标记策略和预留实例来降低AWS、Azure和GCP的开支。它适用于减少云支出、分析基础设施成本或实施成本治理策略的场景。关键功能包括提供成本可视化、资源规模调整指导和定价模型优化建议。
quantizing-models-bitsandbytes
其他这个Skill使用bitsandbytes库量化大语言模型,能在GPU内存有限时通过8位或4位量化减少50-75%内存占用,同时保持精度损失最小。它支持INT8、NF4、FP4等多种量化格式,可与HuggingFace Transformers无缝集成,适用于需要部署更大模型或加速推理的场景。还提供QLoRA训练和8位优化器支持,让开发者能轻松实现高效模型压缩。
dispatching-parallel-agents
其他该Skill用于并行处理3个以上无依赖关系的独立故障,可为每个问题域分派专属Claude代理同时执行调查修复。它通过并发处理多个独立问题显著提升故障排查效率,特别适用于测试文件、子系统等无共享状态的场景。
