Writing study: FlashSketch: Sketch-Kernel Co-Design for Fast Sparse Sketching on GPUs

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Rajat Vadiraj Dwaraknath, Sungyoon Kim, Mert Pilanci · Official ICML 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: Efficiency, inference & systems (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

51 physical pages

Before references / appendix

≈ 8.27 page-equivalents

Main figures / tables

4 / 1

Attached appendix figures / tables

130 / 0

Attached appendix span

≈ 40.96 page-equivalents

Before references / appendix: ≈ 16.22%References / intervening material: ≈ 3.47%Attached appendix: ≈ 80.31%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Small-cap method names suppressed section4 and Appendix C, breaking subsequent sequential detection. Restore main sections4-8, Appendices C-G and their wrapped continuations from physical-page text.

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	≈ 1.42 pp.	2.78%
2. Related Work	2	≈ 0.55 pp.	1.08%
3. Background	3	≈ 1.2 pp.	2.35%
4. BLOCKPERM-SJLT: A Block-Permuted Sparse JL Transform	4	≈ 0.75 pp.	1.47%
5. FLASHSKETCH: CUDA Kernel	5	≈ 0.95 pp.	1.86%
6. Theory: Permutation-Coupled Localization	6	≈ 1 pp.	1.96%
7. Experiments	7	≈ 1.5 pp.	2.94%
8. Limitations and Future Directions	8	≈ 0.39 pp.	0.76%
Acknowledgments	9	≈ 0.08 pp.	0.16%
Impact Statement	9	≈ 0.05 pp.	0.1%
References	9	≈ 1.78 pp.	3.49%
A. Additional Theory and Proofs	11	≈ 5.26 pp.	10.31%
B. Additional Details on FLASHSKETCH	16	≈ 0.31 pp.	0.61%
C. FLASHBLOCKROW: A Fast but Fragile Alternative	16	≈ 0.63 pp.	1.24%
D. Permutation Wiring and On-The-Fly Generation	17	≈ 0.8 pp.	1.57%
E. Additional Details on GraSS	18	≈ 0.72 pp.	1.41%
F. Additional Experiments and Ablations for RandNLA Tasks	18	≈ 17.28 pp.	33.88%
G. Additional Experiments and Ablations (RTX A6000)	36	≈ 15.96 pp.	31.29%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 5
• Figure 3 · p. 7
• Table 1 · p. 8
• Figure 4 · p. 8

Attached appendix

• Figure 5 · p. 19
• Figure 6 · p. 19
• Figure 7 · p. 19
• Figure 8 · p. 19
• Figure 9 · p. 20
• Figure 10 · p. 20
• Figure 11 · p. 20
• Figure 12 · p. 20
• Figure 13 · p. 21
• Figure 14 · p. 21
• Figure 15 · p. 21
• Figure 16 · p. 21
• Figure 17 · p. 22
• Figure 18 · p. 22
• Figure 19 · p. 22
• Figure 20 · p. 22
• Figure 21 · p. 23
• Figure 22 · p. 23
• Figure 23 · p. 24
• Figure 24 · p. 24
• Figure 25 · p. 24
• Figure 26 · p. 24
• Figure 27 · p. 25
• Figure 28 · p. 25
• Figure 29 · p. 25
• Figure 30 · p. 25
• Figure 31 · p. 26
• Figure 32 · p. 26
• Figure 33 · p. 26
• Figure 34 · p. 26
• Figure 35 · p. 27
• Figure 36 · p. 27
• Figure 37 · p. 27
• Figure 38 · p. 27
• Figure 39 · p. 28
• Figure 40 · p. 28
• Figure 41 · p. 28
• Figure 42 · p. 28
• Figure 43 · p. 29
• Figure 44 · p. 29
• Figure 45 · p. 29
• Figure 46 · p. 29
• Figure 47 · p. 30
• Figure 48 · p. 30
• Figure 49 · p. 30
• Figure 50 · p. 30
• Figure 51 · p. 31
• Figure 52 · p. 31
• Figure 53 · p. 31
• Figure 54 · p. 31
• Figure 55 · p. 32
• Figure 56 · p. 32
• Figure 57 · p. 32
• Figure 58 · p. 32
• Figure 59 · p. 33
• Figure 60 · p. 33
• Figure 61 · p. 33
• Figure 62 · p. 33
• Figure 63 · p. 34
• Figure 64 · p. 34
• Figure 65 · p. 34
• Figure 66 · p. 34
• Figure 67 · p. 35
• Figure 68 · p. 35
• Figure 69 · p. 35
• Figure 70 · p. 35
• Figure 71 · p. 36
• Figure 72 · p. 36
• Figure 73 · p. 36
• Figure 74 · p. 36
• Figure 75 · p. 37
• Figure 76 · p. 37
• Figure 77 · p. 37
• Figure 78 · p. 37
• Figure 79 · p. 38
• Figure 80 · p. 38
• Figure 81 · p. 38
• Figure 82 · p. 38
• Figure 83 · p. 39
• Figure 84 · p. 39
• Figure 85 · p. 39
• Figure 86 · p. 39
• Figure 87 · p. 40
• Figure 88 · p. 40
• Figure 89 · p. 40
• Figure 90 · p. 40
• Figure 91 · p. 41
• Figure 92 · p. 41
• Figure 93 · p. 41
• Figure 94 · p. 41
• Figure 95 · p. 42
• Figure 96 · p. 42
• Figure 97 · p. 42
• Figure 98 · p. 42
• Figure 99 · p. 43
• Figure 100 · p. 43
• Figure 101 · p. 43
• Figure 102 · p. 43
• Figure 103 · p. 44
• Figure 104 · p. 44
• Figure 105 · p. 44
• Figure 106 · p. 44
• Figure 107 · p. 45
• Figure 108 · p. 45
• Figure 109 · p. 45
• Figure 110 · p. 45
• Figure 111 · p. 46
• Figure 112 · p. 46
• Figure 113 · p. 46
• Figure 114 · p. 46
• Figure 115 · p. 47
• Figure 116 · p. 47
• Figure 117 · p. 47
• Figure 118 · p. 47
• Figure 119 · p. 48
• Figure 120 · p. 48
• Figure 121 · p. 48
• Figure 122 · p. 48
• Figure 123 · p. 49
• Figure 124 · p. 49
• Figure 125 · p. 49
• Figure 126 · p. 49
• Figure 127 · p. 50
• Figure 128 · p. 50
• Figure 129 · p. 50
• Figure 130 · p. 50
• Figure 131 · p. 51
• Figure 132 · p. 51
• Figure 133 · p. 51
• Figure 134 · p. 51

Argument and organization

1. 1 Introduction; 2 Related Work; 3 Background

   Sets up the conflict between sparse random mixing and GPU-local computation.

2. 4 BLOCKPERM-SJLT; 5 FLASHSKETCH; 6 Theoretical Analysis

   Introduces the sketch distribution, then its kernel, then conditional embedding arguments.

3. 7 Experiments; 8 Limitations

   Tests approximation-runtime tradeoffs before identifying occupancy, mixing and implementation limits.

Problem, claim and contribution

The problem is co-design of the random transform and its implementation, not just accelerating an unchanged sparse matrix product.

Introduction, p.1; Section 4, p.4 · Main paper

Argument and section order

The mixing parameter connects block wiring, shared-memory implementation and neighborhood-coherence analysis before the experimental Pareto comparisons.

Sections 4-6, pp.4-6 · Main paper

Figures, captions and visual explanation

Colored block permutations and a neighboring kernel listing explain the mechanism; the benchmark page uses consistent method colors and marker shapes for approximation-runtime tradeoffs.

Figure 2 and Algorithm 1, p.5; Figure 4, p.8 · Main paper

Tables, comparisons and evidence

GraSS attribution quality is paired with projection-step timing; this timing should not be paraphrased as the speedup of the entire attribution pipeline.

Section 7 and Figure 4, p.8 · Main paper

Appendix and reproducibility

Implementation and theoretical caveats are concrete: split-column fallback reintroduces global atomics, and independently sampled permutations in the analysis differ from distinct structured permutations in the kernel.

Appendix discussion, p.16 · Main paper

Clarity and writing conventions

The limitations specify when occupancy or mixing harms performance and distinguish empirically useful parameter choices from guarantees.

Section 8, pp.8-9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Tie an approximation knob to both its mathematical effect and its hardware cost, and label the precise timed subroutine beside downstream quality.

Boundary: The theory uses stated smoothing and independence assumptions; it is not an unconditional guarantee for every implemented permutation family, and projection-time gains are not whole-pipeline gains.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan of the main argument through physical p.9, with focused physical-page text inspection on 1, 3, 4, 5, 6, 7, 8, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 16, 17, 19. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

5, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 51 pages