PaperlayerView exact revision on arXiv

Source-linked AI summary

TileCas: Cascading Kernel Development Across Programming Abstractions

Xinyu Xiao, Zihe Song, Dezhi Ran, Wei Yang, Tao Xie

arXiv:2610.14740v1cs.SE

TL;DR

Coding agents repeatedly generate, evaluate, and refine code, raising the question of whether human-oriented programming abstractions are effective interfaces for them. TileCas coordinates kernel refinement across abstraction levels by connecting feedback to code, selecting candidates and levels, and preserving search histories. Across the reported evaluations, TileCas improves AKO’s overall mean speedup over the strongest baseline and improves KDA’s mean speedup as well.

  • Problem

    The paper asks whether programming abstractions designed for human programmers are effective interfaces for coding agents, whose repeated generation, evaluation, and refinement is shaped by the search space those abstractions provide.

    Source: §1 · p. 2

  • Method

    TileCas is a pluggable method that coordinates cross-level kernel refinement by linking feedback to code, selecting a source candidate and target level, and preserving candidate histories.

    Source: §1 · p. 3, §3.1 · p. 5

  • Results

    29.0%: AKO+TileCas improves equal-weight overall mean speedup over the strongest baseline by overall mean; TileCas also improves KDA’s same metric by 69.2% over KDA.

    Source: §1 · p. 3, §7 · p. 19

  • Takeaways & Limitations

    The results support treating abstraction selection as an explicit part of agent-driven kernel development, guided by execution evidence and prior exploration history.

    Source: §7 · p. 20

    Evaluation covers two editable abstraction levels, 50 KernelBench Level 3 workloads, two accelerator platforms, and two coding models; generalization to additional levels and compiler representations remains unevaluated.

    Source: §6 · p. 18

Abstract

from arXiv · show

Programming abstractions have traditionally been designed for humans, but code generation is increasingly performed by models and autonomous agents. We investigate abstractions for machine-driven programming through GPU and AI-accelerator kernel generation. High-level domain-specific languages (DSLs) simplify search by delegating implementation details to compilers, while lower-level representations provide finer-grained control at the cost of a larger and more complex search space. In this paper, we present TileCas, a unified, pluggable interface that enables coding agents to develop kernels across multiple editable abstraction levels. Rather than committing to a single representation, TileCas allows an agent to transfer and refine candidates across levels through Cross-Level Diagnosis, which maps correctness and performance feedback to actionable controls at different abstraction levels; Abstraction Scheduling, which selects both the candidate and representation from which subsequent refinement should proceed; and a Heterogeneous Search Tree, which preserves and revisits alternative search trajectories spanning multiple representations. We integrate TileCas with two code agents AKO and KDA, and evaluate both integrations on 50 end-to-end KernelBench Level 3 workloads across two accelerator platforms and two language models. Giving equal weight to the four accelerator-model configurations, AKO with TileCas achieves an arithmetic-mean speedup of $4.84\times$, compared with $3.75\times$ for the strongest evaluated baseline without TileCas, an improvement of 29.0%. Integrating the same interface with KDA improves its mean speedup by 69.2%, with gains in every configuration. These results show that allowing coding agents to navigate across programming abstractions can improve machine-driven software development and can be reused across different base workflows.

1 Introduction

TileCas addresses the question of whether human-oriented abstractions best serve coding agents by coordinating kernel refinement across abstraction levels. Across four accelerator–model configurations, it improves AKO’s overall mean speedup over the strongest baseline and also improves KDA’s mean speedup.

1 Introduction

  • Coding agents repeatedly generate, evaluate, and refine programs, making the abstraction choice consequential because it shapes the search space they navigate.The paper asks whether abstractions designed for human programmers are also effective interfaces for coding agents.Source: §1 · p. 2
  • High-level DSLs simplify exploration by delegating implementation details to compilers, while low-level code exposes more controls at the cost of a larger, more complex search space.High-level search can progress quickly, while lower-level refinement can open further optimization opportunities after transferring a promising candidate.Source: §1 · p. 2
  • Cross-level programming must coordinate what to change, where to make the change, and which prior candidate to continue from while keeping observations tied to their context.Hardware evidence alone may not reveal the programming choices responsible for performance behavior.Source: §1 · pp. 2–3
  • TileCas coordinates cross-level refinement through diagnosis linking feedback to code, scheduling choosing a source and target level, and a search tree preserving candidate histories.The method is designed as a pluggable extension to existing agentic kernel workflows.Source: §1 · pp. 3–4
  • 29.0%: AKO+TileCas improves equal-weight overall mean speedup over the strongest baseline by overall mean across four accelerator–model configurations.Integrating TileCas with KDA improves its arithmetic mean speedup by 69.2%.Source: §1 · p. 3

2 Motivation and Problem Definition

The motivation contrasts the speed of high-level exploration with the additional controls and search costs of low-level development. The problem is to choose a source candidate and abstraction level using code, feedback, and history, while improving the final correct implementation under a specified execution protocol.

2.1 Motivation

  • High-level search can reach a strong implementation early, while lower-level development exposes additional controls but requires managing more interacting choices.Exporting a promising high-level candidate lets low-level development build on prior progress; deciding when to switch remains difficult because a plateau may precede effective same-level edits.Source: §2.1 · p. 4

2.2 Problem Definition

  • The setting uses a task, target hardware, and base workflow to develop candidates at two editable levels: high and low.A candidate is a version of a complete implementation, possibly containing multiple kernels; correctness checks and measured performance determine its quality.Source: §2.2 · p. 4
  • The refinement state retains the current and editable historical candidates alongside modification, conversion, evaluation, and hardware-evidence records.A development decision specifies an action, source candidate, and target level; the objective is the best measured correct candidate under the workflow’s run budget, acceptance requirements, and stopping conditions.Source: §2.2 · pp. 4–5

3 TileCas

TileCas plugs into an agentic kernel workflow to guide refinement across representations. Its diagnosis, scheduling, and search-tree components connect execution evidence to code, select the next development plan, and retain candidate histories for later decisions.

3.1 Overview

  • TileCas integrates with the base workflow’s analysis/planning and evaluation/recording stages, while leaving its generation strategy and evaluation procedure intact.Integration requires access to candidate state and evaluation feedback, plus support for executing the selected plan.Source: §3.1 · pp. 5–6

3.2 Cross-Level Diagnosis

  • Cross-Level Diagnosis links measurements to the corresponding candidate code and connects higher-level calls to generated lower-level implementations for inspection.A recorded timing covers its matched kernel or call group as a whole; attributing it to an individual operation requires additional evidence.Source: §3.2 · pp. 6–7

3.3 Abstraction Scheduling

  • Abstraction Scheduling forms a testable modification hypothesis, compares feasible sources and controls, then selects an action, source candidate, and target level.Stay continues in the source representation, Lower converts to a supported lower level, and Rollback resumes a retained higher-level source; candidate correctness is checked after execution.Source: §3.3 · pp. 8–9

3.4 Heterogeneous Search Tree

  • The Heterogeneous Search Tree retains versioned candidates, their evaluation context, and relationships between edits and conversions so prior observations can be reassessed.It also supports compatible cross-branch code reuse: the combined candidate iter-32 passed correctness checks and ran in 4.330 ms, compared with 4.468 ms for iter-30.Source: §3.4 · pp. 9–10

4 Evaluation

Across 50 KernelBench Level 3 workloads, TileCas is evaluated with two base workflows, two accelerator platforms, and two language models. AKO+TileCas achieves the highest mean speedup in each configuration, while the evaluation also examines search efficiency, component contributions, and cross-level decisions.

4.1 Experimental Setup

  • The evaluation covers 50 KernelBench Level 3 workloads across four accelerator–model configurations, using five seeds, a 120-minute run budget, and majority correctness requiring at least three correct runs.AKO is compared with five baselines and TileCas; correctness and latency follow a common evaluation protocol, and endpoint scores are averaged across seeds and workloads.Source: §4.1 · pp. 10–11

4.2 RQ1: End-to-End Effectiveness

  • AKO+TileCas achieved the highest mean speedup in every accelerator–model configuration and an equal-weight overall mean of 4.836×, versus 3.748× for AKO-Random, a 29.0% relative improvement.Relative to the strongest AKO baseline within each configuration, mean-speedup gains ranged from 13.7% to 22.5%.Source: §4.2 · p. 12
  • KDA+TileCas improves mean speedup over KDA by 69.2% and increases mean speedup in all four accelerator–model configurations.The evaluation compares KDA directly with KDA+TileCas because KDA does not natively expose the High/Low interface.Source: §4.1 · p. 11, §4.2 · p. 12

4.3 RQ2: Search Efficiency

  • Although High reaches first correctness fastest at 6.81 minutes, TileCas reaches 4.806× at 90 minutes and 4.790× after 32 candidate evaluations.Search progress is reported against both elapsed time and candidate-evaluation budget.Source: §4.3 · pp. 13–14, Figure 6 · p. 13, Table 3 · p. 14
  • TileCas’s gains approach saturation: extending the search from 90 to 120 minutes raises mean speedup from 4.806× to 4.836×, about 0.6%.Doubling the candidate-evaluation budget from 32 to 64 raises mean speedup from 4.790× to 4.836×, about 1.0%.Source: §4.3 · p. 14

4.4 RQ3: Component Contributions

  • Removing Diagnosis, Scheduling, or the Heterogeneous Search Tree reduces mean speedup by 12.6%, 20.3%, or 14.4%, respectively, under similar search budgets.Removing Scheduling causes the largest reduction; the ablation losses are not additive because the components affect one another’s evidence, branches, and retained states.Source: §4.4 · p. 15, Table 4 · p. 15

4.5 RQ4: Cross-Level Search Behavior

  • TileCas reaches 4.836× mean speedup versus Random’s 3.748×, a 29.0% improvement, despite Random performing more Lower actions and Rollbacks per run.The difference holds across all four accelerator–model configurations; Random is frequency-matched but ignores diagnostic evidence when choosing when to transition.Source: §4.5 · pp. 15–16, Table 5 · p. 16
  • In matched 10-evaluation continuations, TileCas-selected actions beat Stay in 42 of 50 checkpoints (84.0%); Lower wins at 23 of 25 High checkpoints and Rollback at 19 of 25 Low checkpoints.Lower has an immediate productive effect in 84.4% of eligible windows; longer Rollback windows are productive in 79.8%, but that statistic includes subsequent modifications and is descriptive, not an isolated causal effect.Source: §4.5 · pp. 16–17, Figure 7 · p. 17

5 Related Work

Prior work spans agentic kernel workflows, agent-facing abstractions, and structured multilevel search; TileCas adds feedback-guided selection of both a retained candidate and editable abstraction level. Its evaluation uses KernelBench Level 3 as a common workload contract.

Agentic kernel generation and optimization

  • TileCas adds joint selection of a retained candidate and editable level to agentic kernel workflows, which otherwise provide the development-and-evaluation loop.The integrations preserve that loop and use available evidence and search history to select the programming level.Source: §5 · pp. 17–18

Agent–compiler cooperation and agent-oriented abstractions

  • Unlike ACCLAIM [39], which coordinates LLM rewrites with compiler components for C on CPUs, TileCas associates measured kernel costs with code across abstraction levels.This diagnosis links kernel costs to DSL operations and generated low-level implementations, helping determine whether to edit at the DSL or low level.Source: §5 · p. 18
  • Ave [38] and CAKE improve feedback or controls within designed abstractions, whereas TileCas addresses when to stay at one level or resume elsewhere.TileCas preserves alternative branches across existing High and Low representations without requiring a new IR.Source: §5 · p. 18

Programming abstractions and multi-level search

  • Earlier structured and multilevel search work shows the value of staged refinement and alternative implementations; TileCas studies how runtime evidence and candidate history guide the next programming context.This context includes feedback-guided implementation search such as SPIRAL, and autotuning and equality-saturation approaches [7] [65] [71].Source: §5 · p. 18

6 Limitations and Threats to Validity

The evaluation and diagnosis have bounded scope: they depend on selected workloads, two editable levels, available profiling evidence, and tested intervention settings. Portability beyond AKO and KDA integrations requires additional workflow and backend support.

Evaluation scope

  • The evaluation covers two editable abstraction levels, 50 KernelBench Level 3 workloads, two accelerator platforms, and two coding models.Generalization to additional levels and compiler representations remains unevaluated; selected workloads also limit conclusions about dynamic shapes, communication, and multi-device execution.Source: §6 · p. 18

Diagnosis and scheduling

  • Diagnosis depends on available evidence and model interpretation; kernel or call-group timings provide limited evidence about individual memory accesses or synchronization operations.Misinterpretation can lead to ineffective modifications or unnecessary transitions that consume search budget, and edits or conversions can change the relevance of earlier observations.Source: §6 · p. 19

Measurement and variability

  • Correctness and performance depend on tested inputs and run-to-run variability; generated-input checks at tolerance 10^-3 leave behavior on other inputs unverified.Warm-up and median timing reduce noise, but device state and model sampling can still affect results; the study uses five independent seeds and workload-level bootstrap intervals.Source: §6 · p. 19

Intervention scope

  • Matched interventions test selected cross-level actions against Stay at 50 switching checkpoints under a 10-evaluation budget, not Stay decisions elsewhere.Longer continuations or different modifications may change which action performs best; longer Rollback observation windows combine restoration and subsequent edits, so gains cannot be attributed to restoration alone.Source: §6 · p. 19

Portability

  • Portability has been evaluated through AKO and KDA; other workflows need candidate-state and feedback interfaces, plan execution, and candidate restoration.New backends also need code conversion and source correspondence, profiler availability may limit diagnosis evidence, and adaptation engineering effort was not quantified.Source: §6 · p. 19

7 Conclusion

TileCas coordinates what to modify, where to modify it, and which candidate to continue from across abstraction levels. Results with AKO and KDA support making abstraction selection explicit in agent-driven kernel development.

7 Conclusion

  • High-level representations simplify structural exploration, while native code exposes finer hardware control; restricting development to one level can limit progress.Source: §7 · p. 19
  • 29.0% overall arithmetic-mean speedup improvement for AKO over the strongest evaluated baseline, and 69.2% improvement for KDA over KDA without TileCas.The evaluation covers 50 KernelBench Level 3 workloads across two accelerator platforms and two language models.Source: §7 · p. 19
  • TileCas addresses cross-level development decisions with Cross-Level Diagnosis, Abstraction Scheduling, and a Heterogeneous Search Tree.Together, these components relate execution feedback to candidate code, guide the next modification and its implementation level, and preserve candidate histories.Source: §7 · p. 19
  • The findings support treating abstraction selection as an explicit part of agent-driven kernel development, guided by execution evidence and prior exploration history.Source: §7 · p. 20

Related papers

Loading 2610.14740v1…