kernelsight / README.md
williamhtan's picture
Add README.md
7ac4192 verified
|
Raw
History Blame Contribute Delete
13.9 kB
---
license: apache-2.0
pretty_name: KernelSight v4 Per-Timestep GPU Workload Traces
size_categories:
- 1K<n<10K
task_categories:
- other
tags:
- gpu
- cuda
- profiling
- performance
- kernel
- hopper
- h100
- cutlass
- kernelbench
- time-series
- sequence-labeling
- systems
viewer: false
---
# KernelSight v4
**Per-timestep workload labels for GPU execution traces.**
KernelSight pairs every GPU workload trace with a dense, per-timestep workload
labeling. Each snapshot is a `[24, 512]` counter image — 24 hardware-counter
channels sampled across 512 equal-width time bins — paired with per-bin labels
drawn from a two-level hierarchy of **12 coarse (L1)** and **73 fine (L2)**
workload classes. The goal is to label *what a kernel is doing at each instant*
(matmul, attention, reduction, memory movement, …) rather than the single
coarse bottleneck label a profiler assigns per launch.
All counters are collected on a single **NVIDIA H100 80GB HBM3** (Hopper,
`sm_90a`). No data augmentation is applied: each snapshot is the trace as
measured, and class imbalance is handled at training time rather than by
resampling.
| | |
|---|---|
| **Snapshots** | 1,444 |
| **Tensor shape** | `[24, 512]` (24 channels × 512 time bins) |
| **Label vocab** | 12 L1 classes · 73 L2 classes |
| **Labeled segments** | 45,860 |
| **Overlap ground truth** | 472 snapshots (`has_overlap=1`) |
| **Standard split** | train 1,124 / val 160 / test 160 |
| **Hardware** | NVIDIA H100 80GB HBM3 (`sm_90a`) |
| **On-disk size** | ~68 MB (4,332 `.npz` files) |
| **License** | Apache-2.0 |
---
## Quick start
The dataset ships as raw NumPy `.npz` files in a fixed directory layout, indexed
by JSON split files. It is **not** loadable through `datasets.load_dataset()`;
download the folder and read the `.npz` files directly with NumPy.
```python
from huggingface_hub import snapshot_download
import numpy as np, json, os
root = snapshot_download(repo_id="williamhtan/kernelsight", repo_type="dataset")
# Load a split (paths inside are relative to `root`)
split = json.load(open(os.path.join(root, "splits/train.json")))
rec = split["traces"][0]
# Input tensor
t = np.load(os.path.join(root, rec["path"]), allow_pickle=True)
X = t["data"] # (24, 512) float32
# Labels live alongside the input
lpath = rec["path"].replace("/input/", "/labels/").replace("tensor_input.npz", "labels.npz")
l = np.load(os.path.join(root, lpath), allow_pickle=True)
y_l1 = l["workload_l1"] # (512,) int32, -1 where unlabeled
y_l2 = l["workload_l2"] # (512,) int32
mask = l["mask_labeled"] # (512,) uint8
y_mh = l["workload_l1_multihot"] # (512, 12) uint8, multi-label overlap track
```
A reference PyTorch `Dataset` is bundled at `tools/sass_dataloader_stub.py`, and
the label vocabularies live in `tools/workload_taxonomy.py` (single source of
truth). Full schema documentation is in [`data_info.md`](data_info.md).
---
## Directory layout
```
kernelsight_dataset_v4/
├── README.md # this card
├── MANIFEST_v4.md # release notes / per-class counts
├── data_info.md # full on-disk schema reference
├── splits/ # 7 split JSONs (see Splits)
├── tools/
│ ├── workload_taxonomy.py # 12 L1 + 73 L2 + 8 flags + 5 spatial (source of truth)
│ └── sass_dataloader_stub.py
└── kernels/<motif>/_out/<variant>/
├── input/tensor_input.npz # [24, 512] profiler heatmap + metadata
├── labels/labels.npz # per-bin + per-segment L1/L2 labels, vocabs
└── fingerprint/fingerprint.npz # 32-D instruction-mix fingerprint
```
Each `<variant>` is one snapshot (e.g. parameter-swept geometry like
`cutlass_gemm/_out/m8192_n1024_k4096_.../`). Split JSON `path` fields are
relative to the dataset root and point at `.../input/tensor_input.npz`.
---
## Input tensor — `tensor_input.npz`
- `data` — `[24, 512]` float32: 24 counter channels × 512 time bins.
- `time_edges_ns` — `[513]` int64: bin-boundary timestamps (bins are equal-width
*per trace*, ~0.5 ms for fast matmul up to ~30 ms for long scatter; the window
is clipped to the kernel-active span).
- `counter_names` — `[24]`: channel names. `kernels`, `kernel_names`,
`kernel_function_index` — per-launch identity metadata.
Channels divide into five semantic groups. Each channel is divided by a fixed
physical-scale divisor (pipe/throughput ÷100, warp-stall ÷64, coalescing ÷8) so
values land in ~`[0, 1]` while *preserving* magnitude differences (no per-channel
min/max rescale).
| Rows | Group | Source | Channels |
|---|---|---|---|
| 0–6 | Pipe signature | CUPTI | `tensor_op_hmma`, `xu`, `fma`, `alu`, `lsu`, `cbu`, `tma` |
| 7–8 | Memory access | CUPTI/ncu | `hit: l2`, `atom: lts_atomic_input_pct` |
| 9–12 | Discriminators | ncu/NVBit | `short_scoreboard`, `barrier`, `pred_on_per_inst_ratio`, `gmem_coalesce_ratio` |
| 13–16 | System bandwidth | Nsight Systems | `SMs Active %`, `DRAM Read %`, `DRAM Write %`, `Tensor Active %` |
| 17–23 | SASS modality | NVBit | `compute_fma`, `compute_tensor`, `memory_global`, `memory_shared`, `memory_tma`, `control`, `misc` |
---
## Labels — `labels.npz`
**Per-bin arrays** (length `T = 512`):
| Key | dtype | Meaning |
|---|---|---|
| `workload_l1` | int32 | L1 class id per bin (`-1` if unlabeled) |
| `workload_l2` | int32 | L2 class id per bin (`-1` if unlabeled) |
| `workload_l1_multihot` | uint8 `[T,12]` | Multi-hot per-bin L1 (overlap track) |
| `workload_l2_multihot` | uint8 `[T,73]` | Multi-hot per-bin L2 |
| `multihot_n_active` | uint8 `[T]` | # active L1 classes per bin |
| `multihot_has_overlap` | uint8 `[]` | 1 if any bin asserts ≥2 classes |
| `segment_id` | int32 `[T]` | 0-based segment ordinal per bin (`-1` if none) |
| `mask_any_kernel` | uint8 `[T]` | 1 if a kernel interval overlaps this bin |
| `mask_labeled` | uint8 `[T]` | 1 if `workload_l1 >= 0` |
| `time_edges_ns` | int64 `[T+1]` | Bin boundaries |
**Per-segment arrays** (length `S`, varies by motif): `segment_starts`,
`segment_ends`, `segment_label_l1`, `segment_label_l2`, `segment_kernel_names`,
`segment_predecessor_l1/l2`, `segment_position`, `attribute_flags` `[S,8]`.
**Vocabularies** (carried in *every* `labels.npz`): `vocab_l1[12]`,
`vocab_l2[73]`, `attribute_flag_names[8]`, `spatial_state_vocab[5]`,
`l2_parent_l1[73]`.
The single-label fields are always a subset of the multi-hot tracks. On the
sequential corpus the multi-hot is effectively one-hot; genuine overlap comes
from 472 `cutlass_ws_overlap` snapshots whose producer (TMA load →
`memory_movement`) and consumer (WGMMA → `matmul`) phases co-occur, derived from
device `%globaltimer` markers (independent of the 24 counter channels — no label
leakage).
---
## Label taxonomy
**L1 (12):** `matmul`, `conv`, `activation`, `normalization`, `softmax`,
`pooling`, `reduction`, `attention`, `loss`, `elementwise`, `memory_movement`,
`other`.
**L2 (73), nested under L1 parents:**
| L1 | L2 classes |
|---|---|
| matmul | `bmm`, `gemm`, `matvec` |
| conv | `conv1d_standard`, `conv2d_depthwise`, `conv2d_pointwise`, `conv2d_standard`, `conv3d_standard`, `convtranspose1d`, `convtranspose2d`, `convtranspose3d` |
| activation | `elu`, `gelu`, `hardsigmoid`, `hardswish`, `hardtanh`, `leaky_relu`, `mish`, `other`, `relu`, `selu`, `sigmoid`, `softplus`, `softsign`, `swish`, `tanh` |
| normalization | `batchnorm`, `frobeniusnorm`, `groupnorm`, `instancenorm`, `l1norm`, `l2norm`, `layernorm`, `rmsnorm` |
| softmax | `log_softmax`, `logsumexp`, `softmax` |
| pooling | `avg_pool`, `global_avg_pool`, `max_pool` |
| reduction | `argmax`, `argmin`, `cumprod`, `cumsum`, `max`, `mean`, `min`, `prod`, `sum` |
| attention | `scaled_dot_product` |
| loss | `cross_entropy`, `hinge`, `huber`, `kldiv`, `mse`, `triplet_margin` |
| elementwise | `add`, `bias_add`, `cast`, `clamp`, `div`, `mul`, `residual_add`, `scalar_multiplication`, `scaling`, `sub` |
| memory_movement | `copy`, `embedding`, `gather`, `scatter`, `transpose` |
| other | `dropout`, `misc` |
**Attribute flags (8, multi-label per segment):** `sparse`, `tma`, `cluster`,
`masked`, `persistent`, `vectorized_store`, `atomic_accum`, `ldgsts`.
**Spatial-state vocab (5, exposed for the model side):** `uniform`,
`wavefront_transition`, `tail_effect`, `load_imbalanced`, `hotspot`.
---
## Corpus composition
| Source | Motif | Snapshots | Notes |
|---|---|---|---|
| Microbench | `vector_add` | 20 | coalesced BW-bound elementwise |
| Microbench | `gather` | 17 | random-indexed memory movement |
| Microbench | `reduction` | 16 | tree + atomic reductions |
| Microbench | `scatter` | 31 | atomic histogram scatter |
| Microbench | `wgmma` | 1 | tensor-core GEMM baseline |
| KernelBench | `kernelbench` | 480 | PyTorch L1 + L2 ops (11 populated L1 classes) |
| CUTLASS | `cutlass_gemm` | 278 | ex48 TF32 warp-specialized GEMM |
| CUTLASS | `cutlass_fmha` | 85 | ex88 FlashAttention-3 |
| CUTLASS | `cutlass_ws_overlap` | 472 | ex48 + device `%globaltimer` markers |
| CUTLASS | `cutlass_fp8_gemm` | 14 | ex54 FP8 WS-GEMM |
| CUTLASS | `cutlass_sparse_gemm` | 18 | ex62 2:4 structured sparsity |
| CUTLASS | `cutlass_grouped_gemm` | 12 | ex57 grouped GEMM |
**Per-L1 distribution** (snapshots containing each class / labeled segments):
matmul 849 / 3,257 · activation 147 / 11,654 · reduction 125 / 7,155 · conv 98 /
6,418 · attention 92 / 295 · elementwise 86 / 2,887 · normalization 79 / 6,546 ·
pooling 62 / 2,019 · memory_movement 48 / 48 · loss 42 / 4,860 · softmax 28 /
721. The corpus is heavily imbalanced (matmul dominates bin count via long
CUTLASS traces), which motivates a class-balanced objective and macro-F1.
---
## Splits
Each split JSON is `{split, seed, n, traces, notes}`; every `traces[i]` records
`path`, `motif`, `n_kernels`, `n_unique_kernels`, `T`, `l1_labels`, `l2_labels`,
`dominant_l1`, `dominant_l2`.
**Standard disjoint partition** (L2-stratified, trace-level leak-free):
| Split | n |
|---|---|
| `train.json` | 1,124 |
| `val.json` | 160 |
| `test.json` | 160 |
This measures *within-kernel generalization*: most test traces share kernel
identity with training and differ in geometry/precision/sweep parameters.
**Overlapping analysis tags** (views over the same corpus, not a partition):
| Tag | n | Selects |
|---|---|---|
| `iid.json` | 433 | random IID sample |
| `param_ood.json` | 956 | parameter-sweep variants (fixed op, unseen geometry) |
| `composed.json` | 1,124 | multi-kernel / multi-segment traces |
| `length_ood.json` | 0 | reserved (empty in v4) |
---
## Collection methodology
Workloads come from three sources — hand-written CUDA microbenchmarks isolating
canonical GPU behaviors, the [KernelBench](https://github.com/ScalingIntelligence/KernelBench)
Level-1 / Level-2 problem suite, and [CUTLASS](https://github.com/NVIDIA/cutlass)
Hopper examples (the single largest contributor, ~61% of the corpus) spanning six
warp-specialized datapaths (TF32, FP8, 2:4-sparse, grouped GEMM,
FlashAttention-3, and WS-GEMM with device markers).
Each workload is profiled by three complementary collectors and fused onto one
time grid:
1. **NVBit** — SASS-level dynamic binary instrumentation: per-PC instruction mix
and coalescing statistics.
2. **CUPTI Range Profiler** — replays each kernel for a 19-metric warp-stall
taxonomy (stall reasons, pipe utilizations, occupancy).
3. **Nsight Systems** — samples system throughput at ~10 kHz alongside the
CUDA/NVTX timeline; the only natively time-resolved source, so it defines the
time grid.
Labels come from NVTX markers + kernel boundaries, with kernel-name + SASS
pattern matching resolving the L2 class. The full collector fork and per-motif
`run.sh` reproduction harness are part of the KernelSight project (not bundled in
this dataset distribution, which ships the rendered tensors, labels, and splits).
---
## Changes from v3.1
- Dropped `megakernel` (1 PoC snapshot) and `tiled_gemm_poc` (590 hand-written
PoC snapshots).
- Added three CUTLASS Hopper datapaths: FP8 (ex54), 2:4 sparse (ex62), grouped
(ex57).
- Selective KernelBench expansion (activation, normalization, pooling, reduction,
elementwise) and geometry sweeps over microbenchmarks and CUTLASS GEMM/FMHA.
- Corpus 262 → 1,444 snapshots; overlap ground truth 29 → 472 snapshots.
- CI: 26,996 assertions passed, 0 failed.
See [`MANIFEST_v4.md`](MANIFEST_v4.md) for full release notes.
---
## Limitations
- Counters are from a **single H100** (`sm_90a`); cross-architecture transfer is
out of scope.
- Overlap timing is coarse: device-marker spans resolve producer/consumer
*envelopes* (≈ whole launch), so overlap is annotated at launch granularity.
- All 24 channels carry real signal, but many rows are legitimately zero where
the hardware is inactive for a given motif.
- The `spatial_state` vocab is exposed for the model side; per-bin spatial-state
derivation is not provided.
---
## License & provenance
Released under **Apache-2.0**. Derived workloads retain their upstream licenses:
- **KernelBench** problems — MIT (Scaling Intelligence Lab, Stanford University).
- **CUTLASS** examples — BSD-3-Clause (NVIDIA Corporation).
The profiler tooling builds on the Intra-Kernel Profiler (NVBit / CUPTI / Nsight
Systems). This release contains only derived, aggregated counter tensors and
labels — no third-party source code.
## Citation
```bibtex
@misc{tan2026kernelsight,
title = {KernelSight: Per-Timestep Workload Labeling of GPU Execution Traces},
author = {Tan, William},
year = {2026},
note = {CS231N project, Stanford University},
howpublished = {\url{https://huggingface.co/datasets/williamhtan/kernelsight}}
}
```