# KernelSight Dataset v4.0

Camera-ready dataset for the CS231N project report.

## Overview

| | |
|---|---|
| **Snapshots** | 1,444 |
| **Tensor shape** | `[24, 512]` (24 channels, 512 time bins) |
| **Label vocab** | 12 L1 classes, 73 L2 classes |
| **Segments** | 45,860 labeled segments |
| **Overlap ground truth** | 472 snapshots (`has_overlap=1`) |
| **Splits** | train 1,124 / val 160 / test 160 |
| **CI** | 26,996 passed, 0 failed |

## Corpus composition

| Source | Motif | Snapshots | L1 classes | Notes |
|---|---|---|---|---|
| Microbenchmark | `vector_add` | 20 | elementwise | coalesced BW-bound |
| Microbenchmark | `gather` | 17 | memory_movement | random-indexed |
| Microbenchmark | `reduction` | 16 | reduction | tree + atomic |
| Microbenchmark | `scatter` | 31 | memory_movement | atomic histogram |
| Microbenchmark | `wgmma` | 1 | matmul | tensor-core GEMM |
| KernelBench | `kernelbench` | 480 | all 11 populated | PyTorch L1+L2 ops |
| CUTLASS | `cutlass_gemm` | 278 | matmul | ex48 TF32 WS-GEMM |
| CUTLASS | `cutlass_fmha` | 85 | attention | ex88 FA3 |
| CUTLASS | `cutlass_ws_overlap` | 472 | matmul (overlap) | ex48 + %globaltimer markers |
| CUTLASS | `cutlass_fp8_gemm` | 14 | matmul | ex54 FP8 WS-GEMM |
| CUTLASS | `cutlass_sparse_gemm` | 18 | matmul | ex62 2:4 structured sparsity |
| CUTLASS | `cutlass_grouped_gemm` | 12 | matmul | ex57 grouped GEMM |

## Per-L1 distribution (snapshots containing each class)

| L1 class | snapshots | 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 |

## Per-snapshot artifacts

Each snapshot directory contains:

```
<variant>/
  input/tensor_input.npz      24-channel [24, 512] profiler heatmap
  labels/labels.npz            L1+L2 workload labels, segments, multi-hot
  fingerprint/fingerprint.npz  32-D instruction-mix fingerprint
```

## Splits

| Split | n | Description |
|---|---|---|
| `train.json` | 1,124 | 80% L2-stratified training set |
| `val.json` | 160 | 10% validation |
| `test.json` | 160 | 10% test |
| `iid.json` | 433 | 30% random IID sample (overlapping tag) |
| `param_ood.json` | 956 | geometry-sweep variants (overlapping tag) |
| `composed.json` | 1,124 | multi-kernel traces (overlapping tag) |

## Tools included

- `workload_taxonomy.py` — L1/L2 vocab, anchor overrides (single source of truth)
- `sass_dataloader_stub.py` — PyTorch Dataset for loading tensor+labels

## Changes from v3.1

- Deleted `megakernel` (PoC, 1 snapshot) and `tiled_gemm_poc` (hand-written PoC, 590 snapshots)
- Added 3 new CUTLASS Hopper GEMM datapaths: FP8 (ex54), 2:4 sparse (ex62), grouped (ex57)
- Selective KernelBench expansion: activation, normalization, pooling, reduction, elementwise
- Microbenchmark geometry sweep: vector_add, gather, reduction, scatter
- CUTLASS geometry sweep: cutlass_gemm (264), cutlass_ws_overlap (456), cutlass_fmha (72)
- Corpus grew from 262 (v3.1) to 1,444 (v4.0); overlap ground truth from 29 to 472
- CI: 26,996 passed, 0 failed

## Collection environment

- GPU: NVIDIA H100 80GB HBM3 (sm_90a)
- Profilers: CUPTI Range Profiler (19-metric warp-stall set), NVBit region profiler, Nsight Systems
- CUTLASS: Hopper examples from the CUTLASS 3.x tree (build_ex48 / build_ex88)