pith. sign in

arxiv: 2607.02335 · v1 · pith:D33O64A6new · submitted 2026-07-02 · 🧮 math.NA · cs.NA

Coalesced Matrix-Free Finite Elements in Cell-Wise Storage

Micha{\l} Wichrowski This is my paper

Pith reviewed 2026-07-03 07:35 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords matrix-free finite elementsconjugate gradientGPU computingcell-wise storagedirect stiffness summationhigh-order methodspreconditioners
0
0 comments X

The pith

A preconditioner with continuous image lets the full flexible conjugate gradient run exactly on unassembled cell-wise data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper keeps all field data in persistent cell-local vectors as the main representation for high-order finite elements. It proves that the flexible conjugate gradient iteration produces identical results on this unassembled form because a primal-dual pairing identity makes every Krylov scalar from local data match the assembled version. Inter-element communication is therefore needed only inside the preconditioner. The required direct stiffness summation occurs through a cascade of sequential face exchanges that avoids atomics, coloring, or indirect addressing. Cell-wise storage also frees memory layout choices, which the authors use to test blocked arrangements on GPUs.

Core claim

Given a preconditioner whose image is continuous, the entire flexible conjugate gradient iteration can be carried out exactly on the unassembled cell-wise representation: a simple primal-dual pairing identity shows that all Krylov scalars computed from local data coincide with those of the assembled solve, so inter-element communication is confined entirely to the preconditioner.

What carries the argument

primal-dual pairing identity that equates all Krylov scalars between cell-wise local data and the assembled global representation

If this is right

  • The direct stiffness summation is realized by a dimensionally-split cascade of one-to-one face exchanges that accumulates edge and vertex contributions automatically.
  • Unstructured macro-block interfaces and hanging nodes are handled by separate topological kernels without touching the main high-throughput sweeps.
  • Memory layout can be chosen independently of mesh topology, allowing benchmarks of blocked arrangements that trade coalescing against element contiguity.
  • Operator evaluation and the full solver achieve higher throughput than existing matrix-free codes on modern GPUs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same storage model and pairing argument could be tested on other Krylov methods that rely on inner products and matrix-vector products.
  • Decoupling layout from topology may allow new data arrangements on hardware with non-uniform memory access or on distributed systems where communication cost dominates.
  • The sequential face-exchange cascade might extend to other local summation tasks that currently require atomics or coloring.

Load-bearing premise

The preconditioner must produce continuous fields so that the pairing identity holds and local scalars remain identical to assembled ones.

What would settle it

Compute the flexible CG iteration scalars and final solution both from cell-wise unassembled data and from the standard assembled matrix for the same continuous preconditioner, then check whether any scalar or the solution vector differs.

Figures

Figures reproduced from arXiv: 2607.02335 by Micha{\l} Wichrowski.

Figure 1
Figure 1. Figure 1: Dimensionally-split DSS propagation in 2D, shown for a vertex DoF shared by four [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Unstructured DSS exchange at macro-vertices of irregular valence in 2D, here a three [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Volumetric (solid) and DSS (dashed) throughput of the mass operator in cell-wise [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: BP1 mass-operator throughput per unique DoF as a function of the number of unique [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: DSS-operator throughput on a curved (ball) geometry with 6 global refinements, 2 097 151 [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Shadow cells on a 2D h-adaptive mesh. Every coarse cell that touches the refinement interface is materialized in cell-wise storage as a shadow: the dormant parent of the fine children that lie across the interface (shaded gray). Cells in the interior of a refined patch, which border no coarse neighbour, need no shadow and are left unshaded. The red entries are the auxiliary degrees of freedom introduced so… view at source ↗
read the original abstract

We present a GPU-oriented formulation of continuous high-order finite elements in which the redundant, cell-wise (element-local) vector is the persistent primary representation of all field data, rather than a transient stage of matrix-free operator evaluation. We prove that, given a preconditioner whose image is continuous, the entire flexible conjugate gradient iteration can be carried out exactly on this unassembled representation: a simple primal-dual pairing identity shows that all Krylov scalars computed from local data coincide with those of the assembled solve, so inter-element communication is confined entirely to the preconditioner. The required direct stiffness summation (DSS) is then realized without indirect gather-scatter, atomics, or coloring, by a dimensionally-split cascade of one-to-one face exchanges that provably accumulates edge and vertex contributions as a byproduct of sequential axis passes; unstructured macro-block interfaces and $h$-adaptive hanging nodes are handled by disjoint topological kernels and a shadow-cell wrapper that leaves the high-throughput sweeps untouched. The cell-wise storage decouples the memory layout from the mesh topology, and we exploit this freedom to benchmark blocked layouts that trade memory coalescing against element contiguity. Numerical experiments on modern GPUs demonstrate that the resulting operator evaluation and solver outperform state-of-the-art matrix-free implementations, signifficantly exceeding throughput of existing implementations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The manuscript presents a GPU-oriented formulation of continuous high-order finite elements that uses the redundant cell-wise (unassembled) vector as the persistent primary data representation. It proves that, given a preconditioner whose image lies in the continuous finite-element space, the entire flexible conjugate gradient iteration can be performed exactly on this unassembled representation via a primal-dual pairing identity that makes all Krylov scalars computed from local data identical to the assembled case. Inter-element communication is thereby confined to the preconditioner; direct stiffness summation is realized by a dimensionally-split cascade of one-to-one face exchanges that accumulates edge and vertex contributions without atomics, coloring, or indirect gather-scatter. The approach handles unstructured macro-block interfaces and h-adaptive hanging nodes via disjoint topological kernels and a shadow-cell wrapper, decouples memory layout from mesh topology to enable blocked coalesced layouts, and reports benchmark results showing higher throughput than existing matrix-free implementations.

Significance. If the central identity and implementation details hold, the work offers a mathematically grounded route to reduced communication and topology-independent memory optimization in high-order matrix-free solvers on GPUs. The explicit continuity hypothesis on the preconditioner and the conditional character of the proof are strengths; the dimensionally-split DSS and hanging-node handling are presented as preserving the identity while preserving high-throughput sweeps.

minor comments (1)
  1. [Abstract] Abstract: 'signifficantly' is a typographical error and should read 'significantly'.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of the manuscript, the detailed summary of its contributions, and the recommendation of minor revision. No specific major comments are listed in the report.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The central claim is a conditional mathematical identity: given a preconditioner whose image lies in the continuous FE space, a primal-dual pairing makes all Krylov scalars computed from cell-wise data identical to the assembled case. This is presented as an explicit hypothesis plus a standard pairing identity rather than a fitted parameter, self-definition, or self-citation chain. No equations reduce the claimed equivalence to a tautology by construction, no ansatz is smuggled via prior work, and the DSS implementation details are described as preserving the identity without redefining it. The derivation is therefore self-contained against external mathematical benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5758 in / 1092 out tokens · 24766 ms · 2026-07-03T07:35:58.952458+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

107 extracted references · 14 canonical work pages · 1 internal anchor

  1. [1]

    arXiv preprint , year=

    Coalesced Matrix-Free Finite Elements in Cell-Wise Storage , author=. arXiv preprint , year=

  2. [2]

    arXiv preprint , year =

    Coalesced Matrix-Free Geometric Multigrid on Persistent Cell-Wise Storage , author =. arXiv preprint , year =

  3. [3]

    and Falk, Richard S

    Arnold, Douglas N. and Falk, Richard S. and Winther, R. , TITLE =. Math. Comput. , VOLUME =. 1997 , NUMBER =

    1997

  4. [4]

    and Falk, Richard S

    Arnold, Douglas N. and Falk, Richard S. and Winther, Ragnar , TITLE =. Numer. Math. , FJOURNAL =

  5. [5]

    Multigrid Methods , author =

  6. [6]

    p4est: Scalable algorithms for parallel adaptive mesh refinement on forests of octrees , author=. SIAM J. Sci. Comput. , volume=. 2011 , publisher=

    2011

  7. [7]

    ACM Trans

    Algorithms and data structures for massively parallel generic adaptive finite element codes , author=. ACM Trans. Math. Softw. , volume=. 2012 , publisher=

    2012

  8. [8]

    Daniel Arndt and Wolfgang Bangerth and Denis Davydov and Timo Heister and Luca Heltai and Martin Kronbichler and Matthias Maier and Jean-Paul Pelteret and Bruno Turcksin and David Wells , journal =. The. 2021 , DOI =

    2021

  9. [9]

    Daniel Arndt and Wolfgang Bangerth and Maximilian Bergbauer and Bruno Blais and Marc Fehling and Rene Gassmöller and Timo Heister and Luca Heltai and Martin Kronbichler and Matthias Maier and Peter Munch and Sam Scheuerman and Bruno Turcksin and Siarhei Uzunbajakau and David Wells and Michał Wichrowski , journal =. The. doi:doi:10.1515/jnma-2025-0115 , year =

    work page doi:10.1515/jnma-2025-0115 2025

  10. [10]

    and Farrell, Patrick E

    Brubeck, Pablo D. and Farrell, Patrick E. , title =. SIAM J. Sci. Comput. , volume =. 2022 , doi =

    2022

  11. [11]

    The Finite Element Method for Elliptic Problems , author =

  12. [12]

    Multi-grid Methods and Applications , author =

  13. [13]

    and Kanschat, G

    Janssen, B. and Kanschat, G. , title =. SIAM J. Sci. Comput. , year = 2011, volume = 33, number = 4, pages =

    2011

  14. [14]

    , title =

    Kanschat, G. , title =. J. Comput. Appl. Math. , year = 2008, volume = 218, pages =

    2008

  15. [15]

    and Mao, Y

    Kanschat, G. and Mao, Y. , title =. J. Numer. Math. , year = 2015, volume = 23, number = 1, pages =

    2015

  16. [16]

    2012 , doi =

    Kronbichler, Martin and Kormann, Katharina , journal =. 2012 , doi =

    2012

  17. [17]

    ACM Trans

    Kronbichler, Martin and Kormann, Katharina , title =. ACM Trans. Math. Softw. , pages =. 2019 , volume =

    2019

  18. [18]

    Enhancing data locality of the conjugate gradient method for high-order matrix-free finite-element implementations , author =. Int. J. High-Performance Computing Appl. , year =

  19. [19]

    Lucero Lorca, J. P. and G. Kanschat , title =. Electron. Trans. Numer. Anal. , volume =. 2021 , pages =

    2021

  20. [20]

    Lynch, R. E. and Rice, J. R. and Thomas, D. H. , title =. Numer. Math. , volume = 6, pages =

  21. [21]

    ACM Trans

    Multidimensional intratile parallelization for memory-starved stencil computations , author=. ACM Trans. Parallel Comput. , volume=. 2017 , doi=

    2017

  22. [22]

    Markus Melenk and Klaus Gerdes and Christoph Schwab

    J. Markus Melenk and Klaus Gerdes and Christoph Schwab. Fully discrete hp-finite elements: fast quadrature. Comp. Meth. Appl. Mech. Engrg. 2001

    2001

  23. [23]

    , journal =

    Orszag, Steven A. , journal =. Spectral methods for problems in complex geometries , volume =

  24. [24]

    Trojak and R

    W. Trojak and R. Watson and F.D. Witherden , title =. doi:10.1016/j.cpc.2021.108235 , year =

    work page doi:10.1016/j.cpc.2021.108235 2021

  25. [25]

    Proceedings of the 7th IEEE International Conference on eScience , year =

    Kormann, Katharina and Kronbichler, Martin , title =. Proceedings of the 7th IEEE International Conference on eScience , year =

  26. [26]

    and Arndt, D

    Witte, J. and Arndt, D. and Kanschat, G. , title =. Comput. Meth. Appl. Math. , number =. 2021 , pages =

    2021

  27. [27]

    doi:10.3390/axioms7030063 , year =

    Michael Dumbser and Francesco Fambri and Maurizio Tavelli and Michael Bader and Tobias Weinzierl , title =. doi:10.3390/axioms7030063 , year =

    work page doi:10.3390/axioms7030063

  28. [28]

    Textbook efficiency: massively parallel matrix-free multigrid for the

    Kohl, Nils and R\"ude, Ulrich , journal=. Textbook efficiency: massively parallel matrix-free multigrid for the. 2022 , doi=

    2022

  29. [29]

    ACM Trans

    Multigrid for matrix-free high-order finite element computations on graphics processors , author=. ACM Trans. Parallel Comput. , volume=. 2019 , doi=

    2019

  30. [30]

    An overlapping

    Fischer, Paul F and Miller, Neil I and Tufo, Henry M , journal =. An overlapping. 2000 , publisher=

    2000

  31. [31]

    A massively parallel nonoverlapping additive Schwarz method for discontinuous

    Dryja, Maksymilian and Krzy. A massively parallel nonoverlapping additive Schwarz method for discontinuous. Numer. Math. , volume=. 2016 , publisher=

    2016

  32. [32]

    From h to p efficiently: Implementing finite and spectral/hp element methods to achieve optimal performance for low-and high-order discretisations , author=. J. Comput. Phys. , volume=

  33. [33]

    Implementation of the

    Bauer, Petr and Klement, Vladimir and Oberhuber, Tom. Implementation of the. Computer Phys. Comm. , volume=

  34. [34]

    High-performance Implementation of Matrix-free High-order Discontinuous Galerkin Methods

    M. High-performance implementation of matrix-free high-order discontinuous. arXiv preprint arXiv:1711.10885 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    Textbook multigrid efficiency for fluid simulations , author=. Annu. Rev. Fluid Mech. , volume=. 2003 , doi=

    2003

  36. [36]

    A matrix-free high-order discontinuous

    Fehn, Niklas and Wall, Wolfgang A and Kronbichler, Martin , journal=. A matrix-free high-order discontinuous. 2019 , publisher=

    2019

  37. [37]

    A robust multigrid method for discontinuous

    Hong, Qingguo and Kraus, Johannes and Xu, Jinchao and Zikatanov, Ludmil , journal=. A robust multigrid method for discontinuous. 2016 , publisher=

    2016

  38. [38]

    A study of vectorization for matrix-free finite element methods , author=. Int. J. High Perf. Comput. Appl. , volume=. 2020 , publisher=

    2020

  39. [39]

    doi:10.1145/3126908.3126963 , year =

    Ken Raffenetti and Abdelhalim Amer and Lena Oden and Charles Archer and Wesley Bland and Hajime Fujita and Yanfei Guo and Tomislav Janjusic and Dmitry Durnov and Michael Blocksome and Min Si and Sangmin Seo and Akhil Langer and Gengbin Zheng and Masamichi Takagi and Paul Coffman and Jithin Jose and Sayantan Sur and Alexander Sannikov and Sergey Oblomov an...

    work page doi:10.1145/3126908.3126963

  40. [40]

    Peter Munch and Martin Kronbichler , title =. Int. J. High Perf. Comput. Appl. , year = 2023, note =

    2023

  41. [41]

    Wall and Martin Kronbichler , title =

    Niklas Fehn and Peter Munch and Wolfgang A. Wall and Martin Kronbichler , title =. doi:10.1016/j.jcp.2020.109538 , year =

    work page doi:10.1016/j.jcp.2020.109538 2020

  42. [42]

    The International Journal of High Performance Computing Applications , volume=

    Enhancing data locality of the conjugate gradient method for high-order matrix-free finite-element implementations , author=. The International Journal of High Performance Computing Applications , volume=. 2023 , publisher=

    2023

  43. [43]

    SIAM Journal on Scientific Computing , volume=

    Smoothers with Localized Residual Computations for Geometric Multigrid Methods for Higher-Order Finite Elements , author=. SIAM Journal on Scientific Computing , volume=. 2025 , publisher=

    2025

  44. [44]

    A matrix-free multilevel preconditioner for the generalized

    Wichrowski, Micha. A matrix-free multilevel preconditioner for the generalized. Journal of Computational Science , volume=. 2022 , publisher=

    2022

  45. [45]

    An implementation of tensor product patch smoothers on

    Cui, Cu and Grosse-Bley, Paul and Kanschat, Guido and Strzodka, Robert , journal=. An implementation of tensor product patch smoothers on. 2025 , publisher=

    2025

  46. [46]

    Computational Methods in Applied Mathematics , number=

    Tensor-product vertex patch smoothers for biharmonic problems , author=. Computational Methods in Applied Mathematics , number=. 2025 , publisher=

    2025

  47. [47]

    Matrix-free multigrid block-preconditioners for higher order discontinuous

    Bastian, Peter and M. Matrix-free multigrid block-preconditioners for higher order discontinuous. Journal of Computational Physics , volume=. 2019 , publisher=

    2019

  48. [48]

    Numerische Mathematik , volume=

    Additive Schwarz methods for the p-version finite element method , author=. Numerische Mathematik , volume=. 1993 , publisher=

    1993

  49. [49]

    Numerical methods for partial differential equations , pages=

    Spectral methods for problems in complex geometrics , author=. Numerical methods for partial differential equations , pages=. 1979 , publisher=

    1979

  50. [50]

    Journal of Computational Physics , volume=

    Scaling to the stars--a linearly scaling elliptic solver for p-multigrid , author=. Journal of Computational Physics , volume=. 2019 , publisher=

    2019

  51. [51]

    Approximate tensor-product preconditioners for very high order discontinuous

    Pazner, Will and Persson, Per-Olof , journal=. Approximate tensor-product preconditioners for very high order discontinuous. 2018 , publisher=

    2018

  52. [52]

    2016 , publisher=

    Remacle, J-F and Gandham, Rajesh and Warburton, Tim , journal=. 2016 , publisher=

    2016

  53. [53]

    Efficient low-order refined preconditioners for high-order matrix-free continuous and discontinuous

    Pazner, Will , journal=. Efficient low-order refined preconditioners for high-order matrix-free continuous and discontinuous. 2020 , publisher=

    2020

  54. [54]

    SIAM Journal on Scientific Computing , volume=

    Scalable low-order finite element preconditioners for high-order spectral element Poisson solvers , author=. SIAM Journal on Scientific Computing , volume=. 2019 , publisher=

    2019

  55. [55]

    An hp multigrid approach for tensor-product space-time finite element discretizations of the

    Margenberg, Nils and Bause, Markus and Munch, Peter , journal=. An hp multigrid approach for tensor-product space-time finite element discretizations of the

  56. [56]

    An efficient smoother for the

    Braess, Dietrich and Sarazin, Regina , journal=. An efficient smoother for the. 1997 , publisher=

    1997

  57. [57]

    Matrix-free monolithic multigrid methods for

    Jodlbauer, Daniel and Langer, Ulrich and Wick, Thomas and Zulehner, Walter , journal=. Matrix-free monolithic multigrid methods for. 2024 , publisher=

    2024

  58. [58]

    Computing , volume=

    A class of smoothers for saddle point problems , author=. Computing , volume=. 2000 , publisher=

    2000

  59. [59]

    Monolithic multigrid preconditioners for high-order discretizations of

    Voronin, Alexey and Harper, Graham and MacLachlan, Scott and Olson, Luke N and Tuminaro, Raymond S , journal=. Monolithic multigrid preconditioners for high-order discretizations of. 2025 , publisher=

    2025

  60. [60]

    An energy-efficient

    Anselmann, Mathias and Bause, Markus and Margenberg, Nils and Shamko, Pavel , journal=. An energy-efficient. 2024 , publisher=

    2024

  61. [61]

    A Geometric Multigrid Preconditioner for Discontinuous

    Wichrowski, Michal , journal=. A Geometric Multigrid Preconditioner for Discontinuous

  62. [62]

    arXiv preprint arXiv:2507.17053 , year=

    Matrix-Free Evaluation of High-Order Shifted Boundary Finite Element Operators , author=. arXiv preprint arXiv:2507.17053 , year=

    work page arXiv

  63. [63]

    A multigrid method for

    Cui, Cu and Kanschat, Guido , journal=. A multigrid method for

  64. [64]

    SIAM Journal on Scientific Computing , volume=

    High-performance matrix-free unfitted finite element operator evaluation , author=. SIAM Journal on Scientific Computing , volume=. 2025 , publisher=

    2025

  65. [65]

    Differencing of the diffusion equation in

    Kershaw, David S , journal=. Differencing of the diffusion equation in. 1981 , publisher=

    1981

  66. [66]

    SIAM Journal on Scientific Computing , volume=

    Flexible conjugate gradients , author=. SIAM Journal on Scientific Computing , volume=. 2000 , publisher=

    2000

  67. [67]

    2002 , publisher=

    High-Order Methods for Incompressible Fluid Flow , author=. 2002 , publisher=

    2002

  68. [68]

    Journal of Open Source Software , volume=

    libCEED: Fast algebra for high-order element-based discretizations , author=. Journal of Open Source Software , volume=

  69. [69]

    Parallel Computing , volume=

    NekRS, a GPU-accelerated spectral element Navier--Stokes solver , author=. Parallel Computing , volume=. 2022 , publisher=

    2022

  70. [70]

    International Journal for Numerical Methods in Engineering , volume=

    A method of finite element tearing and interconnecting and its parallel solution algorithm , author=. International Journal for Numerical Methods in Engineering , volume=. 1991 , publisher=

    1991

  71. [71]

    SIAM Journal on Scientific Computing , volume=

    A preconditioner for substructuring based on constrained energy minimization , author=. SIAM Journal on Scientific Computing , volume=. 2003 , publisher=

    2003

  72. [72]

    Computers & Mathematics with Applications , volume=

    MFEM: A modular finite element methods library , author=. Computers & Mathematics with Applications , volume=. 2021 , publisher=

    2021

  73. [73]

    2020 , institution=

    Nek5000 developments in support of industry and the NRC , author=. 2020 , institution=

    2020

  74. [74]

    and Fischer, Paul F

    Lottes, James W. and Fischer, Paul F. , title =

  75. [75]

    Bradbury, James and Frostig, Roy and Hawkins, Peter and Johnson, Matthew James and Leary, Chris and Maclaurin, Dougal and Necula, George and Paszke, Adam and Vander

  76. [76]

    Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages (MAPL) , pages=

    Triton: an intermediate language and compiler for tiled neural network computations , author=. Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages (MAPL) , pages=

  77. [77]

    arXiv preprint arXiv:2508.00441 , year=

    DGEMM without FP64 Arithmetic-Using FP64 Emulation and FP8 Tensor Cores with Ozaki Scheme , author=. arXiv preprint arXiv:2508.00441 , year=

    work page arXiv

  78. [78]

    arXiv preprint arXiv:2504.19442 , year=

    Triton-distributed: Programming overlapping kernels on distributed ai systems with the triton compiler , author=. arXiv preprint arXiv:2504.19442 , year=

    work page arXiv

  79. [79]

    Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1 , pages=

    Linear Layouts: Robust Code Generation of Efficient Tensor Computation Using F\_2 , author=. Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1 , pages=

  80. [80]

    arXiv preprint arXiv:2512.18134 , year=

    Optimal Software Pipelining and Warp Specialization for Tensor Core GPUs , author=. arXiv preprint arXiv:2512.18134 , year=

    work page arXiv

Showing first 80 references.