Datasets:

---

bcui2

/

NCU_report

Like 1

Summary

Add 110 TileLang-only Nsight Compute reports for all 45 operators under
NVIDIA_B200/<operator>/tilelang_<dtype>.ncu-rep.

The operator directory names and dtype suffixes match the existing Triton and
cuTile reports exactly, including fp8_e4m3fn and int8. No existing files are
replaced, and no other hardware folders are changed. A report contains the
operator's captured compute kernel(s); the filename identifies the backend and
dtype, not an individual generated kernel symbol.

Capture Environment

• NVIDIA B200; driver 595.58.03.
• Paper-era packages: PyTorch 2.10.0+cu130, Triton 3.6.0, cuda-tile 1.3.0,
  TileLang 0.1.11.
• Native CUDA compiler 13.0.88; Nsight Compute 2026.1.1.
• Each case uses the NCU-catalogued maximum input and recorded autotuned
  TileLang winning configuration, without a new autotune sweep.
• Existing TileBench profiling harness and cache methodology; non-operator
  allocation/cast/zero-fill helpers are excluded.
• Every captured compute kernel includes smsp__inst_executed.sum, kernel
  duration, and registers/thread. All 110 exports passed capture validation
  and SHA-256 verification against the local collection manifest.

Collection Limitations

97 reports use full collection, 5 use targeted collection, and 8 use reduced
core hardware metrics with kernel replay. The targeted reports are
linear_self_attention/fp32, bitonic_sort/fp16, bitonic_sort/fp32,
radix_sort/int32, and top_k_selection/fp32. The reduced reports are:

• batched_matmul: bf16, fp16.
• block_sparse_attention: fp16.
• matmul_fp32_fp16_fp8: fp16, fp8_e4m3fn.
• matmul_int8: int8.
• streamk_matmul: bf16, fp16 (both first-wave and full-tile kernels).

Application replay stalled on these paths; kernel replay successfully collected
the reduced metric set without changing operator code or winning configurations.
The underlying profiler versus kernel-synchronization cause is not established.
These are not full-section reports. Kernel replay with --cache-control none
does not repeat the harness's manual L2 flush before every replay pass, so its
cache-sensitive metrics are not identical in methodology to application replay.
NCU durations are profiling measurements, not replacements for benchmark CSV
latencies.