arXiv subjects

Lin Gan

Publications and source records attributed to Lin Gan.

At least 19 recordsLinked to original sources

arXiv2026

ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents

Purpose-built kernel agents can write high-performance GPU kernels, but existing systems generally lack cross-run knowledge accumulation and architectural evolution. Each run starts from the same initial knowledge, and later iterations largely optimize within an architecture fixed early in the search. We present ForgeKernel, a kernel-generation system that accumulates and reuses knowledge across runs while enabling continued evolution of kernel architectures. It analyzes the hardware behavior of a high-performance reference and distills validated architectural decisions into reusable knowledge of hardware mechanisms rather than imitating source code. It then seals a kernel that passes correctness and performance validation and uses it as the reference for the next optimization run, enabling architectural decisions to be revisited. Across 27 operators and 29 input configurations on NVIDIA Hopper and Blackwell GPUs, ForgeKernel matches or beats its expert reference on 25 configurations, with a geometric-mean performance gain of 10%. On Ascend 910C, it outperforms the per-operator public best on the public case set of all 8 level-4 operators in the CANN-Bench Season 1 final leaderboard, with a geometric-mean gain of 25.1%, while its full-suite submission ranks first on the A3 leaderboard. As of 2026-10-30, public submissions generated by ForgeKernel rank first on all three CANN-Bench 2.0 solution leaderboards (Ascend 910C, 950DT, and 950PR) and on the SOL-ExecBench FlashInfer-Bench track.

cs.DC↗

arXiv2026

JustGoal: With Only a Goal, the AI Baseline Is Above the Expert Kernels

A small number of GPU kernels often determines the efficiency of large language model training and inference. High-performance kernel development typically relies on expert knowledge through manual implementation, collaboration with general-purpose models, or specialized agents equipped with domain knowledge, evaluation methods, and search workflows. We investigate how much value this knowledge provides as foundation models become stronger. We introduce JustGoal, a goal-only baseline in which a general-purpose agent receives only the optimization objective and constraints, without kernel-specific infrastructure, knowledge bases, or predefined workflows, and builds its own evaluation loop. On NVIDIA B200 GPUs, JustGoal outperforms open-source libraries on 8 of 10 operators, achieving per-operator geometric mean speedups of 1.14x to 2.34x. It also surpasses the fastest publicly released specialized-agent kernel on 5 of 9 operators with available comparisons. In a controlled same-model comparison using Fable 5.1 on single shapes, the KDA-0.5 harness reaches 0.95x of JustGoal's performance, while single-mechanism ablations show no consistent benefit from domain knowledge or predefined workflows. These results indicate that specialized machinery should be justified by clear gains over a same-model, same-scope goal-only baseline and reevaluated whenever the foundation model changes.

cs.DC↗

arXiv2026

ForgeStencil: Automating Per-Case Stencil Specialization from Kernels to 100+ Real Applications

Industrial and scientific computing rests on a few core kernels, and the stencil is among the most widely used: weather and climate models, seismic imaging, fluid dynamics, and image processing all run on it. No single stencil implementation is fastest: the optimal kernel changes qualitatively with stencil shape, grid shape, precision, and host application. For two decades the field has answered with general methods (DSLs, code generators, autotuners), because specialized solutions were too expensive to build per case, so all reuse one human-authored recipe. That reuse costs performance; we call the cost the generality tax. This premise no longer holds: code-synthesis agents now build a correct, specialized solution per case at acceptable cost. ForgeStencil automates this. A Kernel Agent synthesizes CUDA and forges a per-configuration map of specialized operators, removing the tax case by case. On an A100 the map beats the strongest public baseline in 37 of 37 cases: geometric mean 2.35x against same-precision f32 baselines and 1.95x for fp16, each reported under its own precision. The same change reaches end-to-end application performance. A generic operator library is tuned once for its own general case and reused across applications, so its shapes, layouts, and launch boundaries are optimal for none of them: using it is the application-level form of the tax. An App Agent instead forges a specialized solution per application, locating hotspots, rewriting application structure, and validating and integrating each change. Across 100 real industrial and scientific codes the end-to-end median speedup is 1.41x against each application's own GPU baseline. To our knowledge this is the first demonstration that per-case synthesis carries from a kernel library to complete applications at this breadth, and evidence that reuse is no longer the default in a domain built on it for two decades.

cs.DC↗

arXiv2026

Enhancing LLM Metacognition via Cognitive Pairwise Training

Reinforcement learning with verifiable rewards (RLVR) has become central to LLM reasoning, but its outcome-level rewards can make models more willing to give confident answers when evidence or reasoning is unreliable. Existing SFT or RL methods mainly teach LLMs to refuse or express uncertainty at the response level, which can overfit abstention behavior rather than improve reasoning reliability. To address this limitation, we propose Cognitive Pairwise Training (CPT), a cognitive mid-training alignment stage that turns pairwise comparisons over reasoning traces into a reusable alignment signal. By learning to distinguish trustworthy from flawed reasoning, CPT encourages the model to internalize a reasoning-quality discrimination boundary rather than memorize surface refusal patterns. Across five model scales and three model families, CPT improves the reasoning--metacognition trade-off. At 14B, CPT+RL outperforms the standard SFT+RL pipeline by +2.2 math-average points and +5.6 abstention-F1 points. Further analyses show that CPT improves trace quality and exhibits strong robustness and scalability across evaluation and training settings. Code and models are released at https://github.com/Tsinghua-dhy/CPT.

cs.LG↗

arXiv2026

Beyond Fast Contractions: Attenuation and Recovery of Matrix-Engine Speedups in High-Order Finite Elements

Modern processors increasingly provide matrix engines whose peak arithmetic throughput greatly exceeds conventional SIMD, but scientific applications rarely realize this advantage end to end. We examine this gap in SPECFEM3D's dominant stiffness operator on the Arm LX2 CPUs that power the flagship Lineshine supercomputer. Against a matched, high-performance SVE baseline on the same cores, SME's $4\times$ single-precision peak advantage falls to $2.2\times$ for isolated tensor contractions and $1.1\times$ for the complete operator. Our factorized diagnostic attributes the loss to pointwise computation, indirect field movement and synchronization, and irregular coefficient delivery. Explicit SIMD mitigates pointwise work, raising the full-operator speedup to $1.3\times$. Field-layout changes mitigate indirect movement and synchronization, while vector-blocked coefficient streaming reduces irregular-access costs; together they raise speedup to $1.6\times$ at high order. A contraction-free control bounds further contraction-only gains at $1.11$--$1.32\times$. Realizing matrix-engine performance therefore requires co-designing the entire operator path, not merely replacing its contraction kernel.

cs.DC↗

arXiv2026

StateFlow: Sequence Pipeline Parallelism for Long-Context Modeling with Linear Recurrence

Long-context training is increasingly important for large language models, and linear attention and state space models have become popular for improving long-context efficiency. However, efficiently parallelizing long-sequence training for recurrent and hybrid models remains challenging. We present StateFlow, a sequence pipeline parallelism system for models with linear recurrence. StateFlow partitions each sequence into chunks and schedules their execution while propagating boundary states and gradients across chunks, thereby reducing activation lifetimes and improving training throughput. StateFlow further uses profile-guided nonuniform chunking to balance recurrence and softmax attention computation in hybrid models, and overlaps state transitions that expose limited parallelism with surrounding computation. Applying StateFlow to models with up to 32B parameters and 256K context length, we achieve up to $2.22\times$ throughput improvements and $2.45\times$ memory reduction compared to conventional pipeline parallelism, enabling otherwise infeasible configurations.

cs.DC↗

arXiv2026

PEERS: A Parallel and Exact Effective Resistance Solver via Implicit Inversion and Augmented Symbolic Analysis

High-precision effective resistance computation is a cornerstone of Electronic Design Automation (EDA) sign-off, yet it remains a fundamental bottleneck in large-scale power grid analysis, spectral sparsification, and circuit reliability. Existing approaches face a prohibitive "precision-memory impasse": approximate methods lack the stringent accuracy required for high-stakes industrial sign-off, while exact methods either suffer from redundant query overheads or trigger $O(n^2)$ memory explosions. To resolve this, we propose PEERS, a Parallel and Exact Effective Resistance Solver powered by an implicit inverse computing model of the Cholesky factor. By integrating a state-inherited augmented depth-first search (DFS) with a dynamic query update mechanism, PEERS eliminates numerical redundancy and evaluates all-edge resistance queries in a single parallel sweep. We provide a rigorous Work-Span analysis, proving that for graphs satisfying an $O(n^α)$ separator theorem, PEERS achieves a theoretically optimal parallel span of $O(n^α)$ while strictly maintaining $O(nnz(L))$ space complexity. Numerical evaluations on industrial benchmarks demonstrate that PEERS achieves an average speedup of 83.3x over state-of-the-art parallel solvers under identical memory constraints. Notably, PEERS processes a 1-million-node industrial graph in just 18.8 seconds and scales to 17 million nodes in under an hour, providing the first computationally feasible path for exact all-edge resistance analysis in multi-million-gate designs.

cs.AR↗

arXiv2026

High-Order Spectral Element Methods for Wave Propagation on ARM Multicore CPU with SME: Optimizations and Implications

Wave propagation based on the spectral element method (SEM) is a representative HPC workload, but existing SEM implementations are not well matched to emerging ARM multicore CPUs with Scalable Matrix Extension (SME). We present an SME-enabled optimization of \textsc{SPECFEM3D} on the emerging LX2 processor that combines an SME-aware batched small-matrix kernel for SEM tensor-product operators, a memory-aware hybrid MPI+OpenMP execution scheme for limited-HBM systems, and a dispersion-based iso-accuracy study of the $(h,p)$ tradeoff. At fixed polynomial order, the optimized implementation improves full-application performance by 4--6$\times$ over the original code and delivers clear gains over optimized non-SME CPU baselines. Beyond these implementation-level gains, our results suggest that SME shifts the performance-favorable operating point toward higher polynomial orders along the dispersion-based iso-accuracy frontier, further reducing time-to-solution and working-set size. These results indicate that SME affects not only kernel efficiency, but also the practical discretization tradeoff for SEM on modern ARM multicore platforms.

cs.DC↗

arXiv2025

FastMPS: Revisit Data Parallel in Large-scale Matrix Product State Sampling

Matrix Product State (MPS) is a versatile tensor network representation widely applied in quantum physics, quantum chemistry, and machine learning, etc. MPS sampling serves as a critical fundamental operation in these fields. As the problems become more complex, the scale of MPS is rapidly increasing. Traditional data parallelism is limited by memory and heavy I/O in large-scale MPS. Model parallelism that can handle large-scale MPS imposes rigid process bindings and lacks scalability. This work proposes Fast-MPS, a multi-level parallel framework for scalable MPS sampling. Our design combines data parallelism across samples with tensor parallelism along bond dimensions. We eliminate memory and I/O pressure through compression and overlapping, and revive data parallel in large-scale MPS sampling. We evaluate our approach on Gaussian Boson Sampling, a representative and demanding application. Fast-MPS achieves over 10x speedup compared to existing simulators, scales to thousands of processes, and enables simulations with 8,176 sites and bond dimension chi = 10^4, significantly outperforming the state of the art. Fast-MPS has demonstrated great potential in high-performance tensor network applications.

cs.DC↗

arXiv2025

New record in optical gain and room-temperature nanolasers in multiple wavelengths in 2D ErOCl single crystals

Erbium-based materials have long been recognized for their important telecom-band applications, yet their widespread adoption in integrated optoelectronics has been hindered by two fundamental limitations: the difficulty in achieving high erbium density without concentration quenching which leads to small optical gain in doped materials, and the difficulty in fabricating a practical device with single crystal nanowires that demonstrated high optical gain previously1,2. Here, we overcome these limitations by synthesizing 2D single crystal ErOCl that has an Er density of 1.75*1022 cm-3. The high-quality single crystal material significantly reduces the density-related quenching effect that dominates in randomly doped materials with high Er concentration. This results in a record optical gain coefficient over 1500 dB/cm at 1536 nm band, at least larger by an order of magnitude than the previous gain record in Er materials. Leveraging this exceptional gain medium, we demonstrate room-temperature continuous-wave lasing operation by integrating with a photonic crystal microcavity, achieving a record-low threshold of 7 μW with the most compact size of any Er-based lasers. Furthermore, the unique Stark splitting characteristics of ErOCl provide optical gain in three wavelength bands and lead to lasing in these wavelengths by engineering the cavity. This is the first time that optical gain has been shown in three different wavelength bands in Er materials, together with the smallest size of laser cavity, could have many important applications in on-chip sensing and optical communication.

physics.app-ph↗

arXiv2025

Robust quantum computational advantage with programmable 3050-photon Gaussian boson sampling

The creation of large-scale, high-fidelity quantum computers is not only a fundamental scientific endeavour in itself, but also provides increasingly robust proofs of quantum computational advantage (QCA) in the presence of unavoidable noise and the dynamic competition with classical algorithm improvements. To overcome the biggest challenge of photon-based QCA experiments, photon loss, we report new Gaussian boson sampling (GBS) experiments with 1024 high-efficiency squeezed states injected into a hybrid spatial-temporal encoded, 8176-mode, programmable photonic quantum processor, Jiuzhang 4.0, which produces up to 3050 photon detection events. Our experimental results outperform all classical spoofing algorithms, particularly the matrix product state (MPS) method, which was recently proposed to utilise photon loss to reduce the classical simulation complexity of GBS. Using the state-of-the-art MPS algorithm on the most powerful supercomputer EI Capitan, it would take > $10^{42}$ years to construct the required tensor network for simulation, while our Jiuzhang 4.0 quantum computer takes 25.6 $μ$s to produce a sample. This work establishes a new frontier of QCA and paves the way to fault-tolerant photonic quantum computing hardware.

quant-ph↗

arXiv2025

MMStencil: Optimizing High-order Stencils on Multicore CPU using Matrix Unit

Matrix-accelerated stencil computation is a hot research topic, yet its application to three-dimensional (3D) high-order stencils and HPC remains underexplored. With the emergence of matrix units on multicore CPUs, we analyze matrix-based acceleration strategies and tailor an optimal approach for 3D high-order stencils. We introduce algorithmic optimizations based on SIMD and matrix units to address strided memory accesses, alignment conflicts, and redundant accesses. We propose memory optimizations to boost on-package memory efficiency, and a novel multi-thread parallelism paradigm to overcome data-sharing challenges caused by the absence of shared data caches. MMStencil sustains consistently high hardware utilization across diverse stencil shapes and dimensions. Our DMA-based inter-NUMA communication further mitigates NUMA effects and MPI limitations in hybrid parallelism. Combining all the innovations, MMStencil outperforms state-of-the-art libraries on Nvidia A100 GPGPU by up to 2.1x. Moreover, the performance improvements translate directly to real-world HPC applications and enable RTM applications to yield 1.8x speedup versus a highly optimized industrial Nvidia A100 GPGPU version.

cs.DC↗

arXiv2025

GenTT: Generate Vectorized Codes for General Tensor Permutation

Tensor permutation is a fundamental operation widely applied in AI, tensor networks, and related fields. However, it is extremely complex, and different shapes and permutation maps can make a huge difference. SIMD permutation began to be studied in 2006, but the best method at that time was to split complex permutations into multiple simple permutations to do SIMD, which might increase the complexity for very complex permutations. Subsequently, as tensor contraction gained significant attention, researchers explored structured permutations associated with tensor contraction. Progress on general permutations has been limited, and with increasing SIMD bit widths, achieving efficient performance for these permutations has become increasingly challenging. We propose a SIMD permutation toolkit, \system, that generates optimized permutation code for arbitrary instruction sets, bit widths, tensor shapes, and permutation patterns, while maintaining low complexity. In our experiments, \system is able to achieve up to $38\times$ speedup for special cases and $5\times$ for general gases compared to Numpy.

cs.DS↗

arXiv2025

SW-TNC : Reaching the Most Complex Random Quantum Circuit via Tensor Network Contraction

Classical simulation is essential in quantum algorithm development and quantum device verification. With the increasing complexity and diversity of quantum circuit structures, existing classical simulation algorithms need to be improved and extended. In this work, we propose novel strategies for tensor network contraction based simulator on Sunway architecture. Our approach addresses three main aspects: complexity, computational paradigms and fine-grained optimization. Data reuse schemes are designed to reduce floating-point operations, and memory organization techniques are employed to eliminate slicing overhead while maintaining parallelism. Step fusion strategy is extended by multi-core cooperation to improve the data locality and computation intensity. Fine-grained optimizations, such as in-kernel vectorized permutations, and split-K operators, are developed as well to address the challenges in new hotspot distribution and topological structure. These innovations can accelerate the simulation of the Zuchongzhi-60-24 by more than 10 times, using more than 1024 Sunway nodes (399,360 cores). Our work demonstrates the potential for enabling efficient classical simulation of increasingly complex quantum circuits.

cs.DC↗

arXiv2025

Reconfigurable nonlinear optical computing device for retina-inspired computing

Optical neural networks are at the forefront of computational innovation, utilizing photons as the primary carriers of information and employing optical components for computation. However, the fundamental nonlinear optical device in the neural networks is barely satisfied because of its high energy threshold and poor reconfigurability. This paper proposes and demonstrates an optical sigmoid-type nonlinear computation mode of Vertical-Cavity Surface-Emitting Lasers (VCSELs) biased beneath the threshold. The device is programmable by simply adjusting the injection current. The device exhibits sigmoid-type nonlinear performance at a low input optical power ranging from merely 3-250 μW. The tuning sensitivity of the device to the programming current density can be as large as 15 μW*mm2/mA. Deep neural network architecture based on such device has been proposed and demonstrated by simulation on recognizing hand-writing digital dataset, and a 97.3% accuracy has been achieved. A step further, the nonlinear reconfigurability is found to be highly useful to enhance the adaptability of the networks, which is demonstrated by significantly improving the recognition accuracy by 41.76%, 19.2%, and 25.89% of low-contrast hand-writing digital images under high exposure, low exposure, and high random noise respectively.

physics.optics↗

arXiv2024

D-band MUTC Photodiode Module for Ultra-Wideband 160 Gbps Photonics-Assisted Fiber-THz Integrated Communication System

Current wireless communication systems are increasingly constrained by insufficient bandwidth and limited power output, impeding the achievement of ultra-high-speed data transmission. The terahertz (THz) range offers greater bandwidth, but it also imposes higher requirements on broadband and high-power devices. In this work, we present a modified uni-traveling-carrier photodiode (MUTC-PD) module with WR-6 waveguide output for photonics-assisted fiber-THz integrated wireless communications. Through the optimization of the epitaxial structure and high-impedance coplanar waveguide (CPW), the fabricated 6-um-diameter MUTC-PD achieves a high output power of -0.96 dBm at 150 GHz and ultra-flat frequency response at D-band. The MUTC-PD is subsequently packaged into a compact WR-6 module, incorporating planar-circuit-based RF-choke, DC-block and probe. The packaged PD module demonstrates high saturation power and flat frequency responses with minimal power roll-off of only 2 dB over 110-170 GHz. By incorporating the PD module into a fiber-THz integrated communication system, high data rates of up to 160 Gbps with 16 quadrature amplitude modulation (QAM) and a maximum symbol transmission rate of 60 Gbaud with QPSK modulation are successfully secured. The demonstration verifies the potential of the PD module for ultra-broadband and ultra-high-speed THz communications, setting a foundation for future research in high-speed data transmission.

physics.optics↗

arXiv2024

Ultra-High-Efficiency Dual-Band Thin-Film Lithium Niobate Modulator Incorporating Low-k Underfill with 220 GHz Extrapolated Bandwidth for 390 Gbit/s PAM8 Transmission

High-performance electro-optic modulators play a critical role in modern telecommunication networks and intra-datacenter interconnects. Low driving voltage, large electro-optic bandwidth, compact device size, and multi-band operation ability are essential for various application scenarios, especially energy-efficient high-speed data transmission. However, it is challenging to meet all these requirements simultaneously. Here, we demonstrate a high-performance dual-band thin-film lithium niobate electro-optic modulator with low-k underfill to achieve overall performance improvement. The low-k material helps reduce the RF loss of the modulator and achieve perfect velocity matching with narrow electrode gap to overcome the voltage-bandwidth limitation, extending electro-optic bandwidth and enhancing modulation efficiency simultaneously. The fabricated 7-mm-long modulator exhibits a low half-wave voltage of 1.9 V at C-band and 1.54 V at O-band, featuring a low half-wave voltage-length product of 1.33 V*cm and 1.08 V*cm, respectively. Meanwhile, the novel design yields an ultra-wide extrapolated 3 dB bandwidth of 220 GHz (218 GHz) in the C-band (O-band). High-speed data transmission in both C- and O-bands using the same device has been demonstrated for the first time by PAM8 with data rates up to 390 Gbit/s, corresponding to a record-low energy consumption of 0.69 fJ/bit for next-generation cost-effective ultra-high-speed optical communications.

physics.optics↗

arXiv2024

A Survey of Foundation Models for Music Understanding

Music is essential in daily life, fulfilling emotional and entertainment needs, and connecting us personally, socially, and culturally. A better understanding of music can enhance our emotions, cognitive skills, and cultural connections. The rapid advancement of artificial intelligence (AI) has introduced new ways to analyze music, aiming to replicate human understanding of music and provide related services. While the traditional models focused on audio features and simple tasks, the recent development of large language models (LLMs) and foundation models (FMs), which excel in various fields by integrating semantic information and demonstrating strong reasoning abilities, could capture complex musical features and patterns, integrate music with language and incorporate rich musical, emotional and psychological knowledge. Therefore, they have the potential in handling complex music understanding tasks from a semantic perspective, producing outputs closer to human perception. This work, to our best knowledge, is one of the early reviews of the intersection of AI techniques and music understanding. We investigated, analyzed, and tested recent large-scale music foundation models in respect of their music comprehension abilities. We also discussed their limitations and proposed possible future directions, offering insights for researchers in this field.

cs.SD↗