Writing study: TileLang: Bridge Programmability and Performance in Modern Neural Kernels

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Lei Wang, Yu Cheng, Yining Shi, Zhiwen Mo, Zhengju Tang, Wenhao Xie, Tong Wu, Lingxiao Ma, Yuqing Xia, Jilong Xue, Fan Yang, Zhi Yang · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Inference, algorithms & systems efficiency (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

36 physical pages

Before references / appendix

≈ 10.36 page-equivalents

Main figures / tables

10 / 0

Attached appendix figures / tables

9 / 19

Attached appendix span

≈ 21.54 page-equivalents

Before references / appendix: ≈ 28.78%References / intervening material: ≈ 11.39%Attached appendix: ≈ 59.83%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: TileLang: Figure 10 is a real side-wrapped caption on physical page 10. Poppler split Figure, 10:, and the caption body into separate same-baseline lines. Visually checked page 10; restore the caption rather than treating this as an author numbering gap.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 1.17 pp.	3.25%
2 RELATED WORK	2	≈ 0.61 pp.	1.69%
3 PROGRAMMING MODEL	3	≈ 1.18 pp.	3.28%
4 SCHEDULING GUIDANCE AND AUTOMATION	4	≈ 3.27 pp.	9.08%
5 EVALUATION	7	≈ 3.03 pp.	8.42%
6 CONCLUSION	10	≈ 0.29 pp.	0.81%
7 REPRODUCIBILITY STATEMENT	11	≈ 0.14 pp.	0.39%
8 ACKNOWLEDGEMENT	11	≈ 0.12 pp.	0.33%
REFERENCES	11	≈ 4.1 pp.	11.39%
A EXTENDED DISCUSSION OF RELATED WORK	15	≈ 1.06 pp.	2.94%
B MLA ALGORITHM	16	≈ 0.28 pp.	0.78%
C PARTIAL LIST OF TILELANG PRIMITIVES	16	≈ 0.57 pp.	1.58%
D PLATFORM-SPECIFIC SCHEDULING	17	≈ 1.11 pp.	3.08%
E LAYOUT INFERENCE ALGORITHM	18	≈ 4.16 pp.	11.56%
F PIPELINE INFERENCE ALGORITHM	22	≈ 2.53 pp.	7.03%
G EFFECTIVENESS OF THE COST MODEL	25	≈ 0.42 pp.	1.17%
H TUNING TIME	25	≈ 0.76 pp.	2.11%
I MATMUL IMPLEMENTATION DIFFS: TVM VS. TILELANG VS. TRITON	26	≈ 0.38 pp.	1.06%
J COMPARISON OF FTG IR AND TVM IR	26	≈ 1.38 pp.	3.83%
K COMPARISON OF FLASHMLA IMPLEMENTATIONS ON DIFFERENT	28	≈ 0.61 pp.	1.69%
L COMPARISON WITH RECENT SYSTEMS	28	≈ 1.71 pp.	4.75%
M BASELINE COMMIT HASHES	30	≈ 0.31 pp.	0.86%
N OPERATOR SHAPES IN OUR BENCHMARK	30	≈ 2.38 pp.	6.61%
O KERNEL IMPLEMENTATIONS	33	≈ 3.89 pp.	10.81%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 4
• Figure 3 · p. 5
• Figure 4 · p. 5
• Figure 5 · p. 6
• Figure 6 · p. 7
• Figure 7 · p. 8
• Figure 8 · p. 9
• Figure 9 · p. 10
• Figure 10 · p. 10

Attached appendix

• Table 1 · p. 17
• Table 2 · p. 17
• Figure 11 · p. 18
• Table 3 · p. 25
• Table 4 · p. 25
• Table 5 · p. 26
• Table 6 · p. 26
• Figure 12 · p. 27
• Figure 13 · p. 27
• Figure 14 · p. 28
• Table 7 · p. 28
• Table 8 · p. 29
• Table 9 · p. 29
• Table 10 · p. 29
• Table 11 · p. 30
• Table 12 · p. 31
• Table 13 · p. 31
• Table 14 · p. 31
• Table 15 · p. 31
• Table 16 · p. 31
• Table 17 · p. 32
• Table 18 · p. 32
• Table 19 · p. 32
• Figure 15 · p. 33
• Figure 16 · p. 33
• Figure 17 · p. 34
• Figure 18 · p. 35
• Figure 19 · p. 36

Argument and organization

1. Programming model

   A running FlashMLA example connects explicit programmer control to compiler automation.

2. Scheduling and evaluation

   Recommendation/inference mechanisms precede cross-hardware performance, code-complexity comparisons, and ablations.

Problem, claim and contribution

The introduction presents performance and implementation complexity as a two-objective tension, then keeps both metrics visible throughout evaluation.

Pages 1–2, Figure 1. · Main paper

Argument and section order

A single MLA running example unifies programming primitives, graph representations, layouts, and pipeline scheduling instead of introducing unrelated examples.

Pages 4–7, sections 3.3–4.3. · Main paper

Figures, captions and visual explanation

Numbered callouts and red/blue dashed regions on executable-looking code distinguish recommended decisions from inferred details.

Page 4, Figure 2. · Main paper

Tables, comparisons and evidence

Small-multiple scatterplots repeat speedup versus lines-of-code axes across operators and separately group NVIDIA/AMD results; the caption tells readers which corner is desirable.

Page 9, Figure 8. · Main paper

Appendix and reproducibility

An appendix roadmap leads to pseudocode, hardware-specific code diffs, baseline commit hashes, benchmark shapes, and actual kernel implementations.

Page 15 roadmap; appendices E–O. · Main paper

Clarity and writing conventions

Explicit 'Goal and Scope' and 'Discussion' paragraphs translate long compiler algorithms into readable contracts.

Pages 18–25, appendices E–F. · Main paper

What to reuse—and what not to infer

Transferable lesson: Choose an executable running example and preserve its terminology across API explanation, intermediate representation, optimization mechanism, and empirical evaluation.

Boundary: Lines of code are a limited productivity proxy, and speedups depend on hardware, workload, baseline implementation, and tuning budget. Keep these qualifiers beside headline gains; dense code screenshots require larger zoomable versions.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–11, sections 1–8; references pages 11–14.

Appendix / supplement

Pages 15–36, appendices A–O: related work, primitive semantics, platform differences, algorithms, cost/tuning evaluation, baseline versions, shapes, and full kernel listings.

Visually inspected PDF pages

4, 9, 28

• ICLR 2026 final proceedings · 36 pages