Showing 1–14 of 14 results for author: Merrill, D

Stream-K: Work-centric Parallel Decomposition for Dense Matrix-Matrix Multiplication on the GPU

Authors: Muhammad Osama, Duane Merrill, Cris Cecka, Michael Garland, John D. Owens

Abstract: We introduce Stream-K, a work-centric parallelization of matrix multiplication (GEMM) and related computations in dense linear algebra. Whereas contemporary decompositions are primarily tile-based, our method operates by partitioning an even share of the aggregate inner loop iterations among physical processing elements. This provides a near-perfect utilization of computing resources, regardless o… ▽ More We introduce Stream-K, a work-centric parallelization of matrix multiplication (GEMM) and related computations in dense linear algebra. Whereas contemporary decompositions are primarily tile-based, our method operates by partitioning an even share of the aggregate inner loop iterations among physical processing elements. This provides a near-perfect utilization of computing resources, regardless of how efficiently the output tiling for any given problem quantizes across the underlying processing elements. On GPU processors, our Stream-K parallelization of GEMM produces a peak speedup of up to 14$\times$ and 6.7$\times$, and an average performance response that is both higher and more consistent across 32,824 GEMM problem geometries than state-of-the-art math libraries such as CUTLASS and cuBLAS. Furthermore, we achieve this performance from a single tile size configuration per floating-point precision, whereas today's math libraries employ complex kernel-selection heuristics to select from a large ensemble of kernel variants. △ Less

Submitted 9 January, 2023; originally announced January 2023.

Comments: This work previously appeared in the author's PhD dissertation, available at arXiv:2212.08964

Onesweep: A Faster Least Significant Digit Radix Sort for GPUs

Authors: Andy Adinets, Duane Merrill

Abstract: We present Onesweep, a least-significant digit (LSD) radix sorting algorithm for large GPU sorting problems residing in global memory. Our parallel algorithm employs a method of single-pass prefix sum that only requires ~2n global read/write operations for each digit-binning iteration. This exhibits a significant reduction in last-level memory traffic versus contemporary GPU radix sorting implemen… ▽ More We present Onesweep, a least-significant digit (LSD) radix sorting algorithm for large GPU sorting problems residing in global memory. Our parallel algorithm employs a method of single-pass prefix sum that only requires ~2n global read/write operations for each digit-binning iteration. This exhibits a significant reduction in last-level memory traffic versus contemporary GPU radix sorting implementations, where each iteration of digit binning requires two passes through the dataset totaling ~3n global memory operations. On the NVIDIA A100 GPU, our approach achieves 29.4 GKey/s when sorting 256M random 32-bit keys. Compared to CUB, the current state-of-the-art GPU LSD radix sort, our approach provides a speedup of ~1.5x. For 32-bit keys with varied distributions, our approach provides more consistent performance compared to HRS, the current state-of-the-art GPU MSD radix sort, and outperforms it in almost all cases. △ Less

Submitted 3 June, 2022; originally announced June 2022.

Comments: 12 pages, 11 figures, 2 tables

ACM Class: F.2.2; D.1.3

Wavelet sets for crystallographic groups

Authors: Kathy D. Merrill

Abstract: Single wavelet sets, and thus single wavelets, are shown to exist for the actions of all crystallographic groups on $\mathbb R^2$ under all integer dilations. Examples of such sets satisfying the additional requirement that they are finite unions of convex sets are presented for each of the groups under dilation by two. Single wavelet sets, and thus single wavelets, are shown to exist for the actions of all crystallographic groups on $\mathbb R^2$ under all integer dilations. Examples of such sets satisfying the additional requirement that they are finite unions of convex sets are presented for each of the groups under dilation by two. △ Less

Submitted 4 May, 2020; originally announced May 2020.

Comments: 15 pages; 9 figures

MSC Class: 42C15; 42C40

Decomposing the wavelet representation for shifts by wallpaper groups

Authors: Lawrence W. Baggett, Kathy D. Merrill, Judith A. Packer, Keith F. Taylor

Abstract: The wavelet group and wavelet representation associated with shifts coming from a two dimensional crystal symmetry group $Γ$ and dilations by powers of 3, are defined and studied. The main result is an explicit decomposition of the $3Γ-$wavelet representation into irreducible representations of the wavelet group. The wavelet group and wavelet representation associated with shifts coming from a two dimensional crystal symmetry group $Γ$ and dilations by powers of 3, are defined and studied. The main result is an explicit decomposition of the $3Γ-$wavelet representation into irreducible representations of the wavelet group. △ Less

Submitted 7 February, 2020; originally announced February 2020.

Comments: 19 pages

MSC Class: 42C40; 43A65; 42C05

Generalized Integrated Gradients: A practical method for explaining diverse ensembles

Authors: John Merrill, Geoff Ward, Sean Kamkar, Jay Budzik, Douglas Merrill

Abstract: We introduce Generalized Integrated Gradients (GIG), a formal extension of the Integrated Gradients (IG) (Sundararajan et al., 2017) method for attributing credit to the input variables of a predictive model. GIG improves IG by explaining a broader variety of functions that arise from practical applications of ML in domains like financial services. GIG is constructed to overcome limitations of Sha… ▽ More We introduce Generalized Integrated Gradients (GIG), a formal extension of the Integrated Gradients (IG) (Sundararajan et al., 2017) method for attributing credit to the input variables of a predictive model. GIG improves IG by explaining a broader variety of functions that arise from practical applications of ML in domains like financial services. GIG is constructed to overcome limitations of Shapley (1953) and Aumann-Shapley (1974), and has desirable properties when compared to other approaches. We prove GIG is the only correct method, under a small set of reasonable axioms, for providing explanations for mixed-type models or games. We describe the implementation, and present results of experiments on several datasets and systems of models. △ Less

Submitted 6 September, 2019; v1 submitted 4 September, 2019; originally announced September 2019.

Comments: 38 pages, submitted to JMLR 9/3/2019

ACM Class: I.2.6; H.1.2; K.4

Quantum transport properties of industrial $^{28}$Si/$^{28}$SiO$_2$

Authors: D. Sabbagh, N. Thomas, J. Torres, R. Pillarisetty, P. Amin, H. C. George, K. Singh, A. Budrevich, M. Robinson, D. Merrill, L. Ross, J. Roberts, L. Lampert, L. Massa, S. Amitonov, J. Boter, G. Droulers, H. G. J. Eenink, M. van Hezel, D. Donelson, M. Veldhorst, L. M. K. Vandersypen, J. S. Clarke, G. Scappucci

Abstract: We investigate the structural and quantum transport properties of isotopically enriched $^{28}$Si/$^{28}$SiO$_2$ stacks deposited on 300 mm Si wafers in an industrial CMOS fab. Highly uniform films are obtained with an isotopic purity greater than 99.92\%. Hall-bar transistors with an equivalent oxide thickness of 17 nm are fabricated in an academic cleanroom. A critical density for conduction of… ▽ More We investigate the structural and quantum transport properties of isotopically enriched $^{28}$Si/$^{28}$SiO$_2$ stacks deposited on 300 mm Si wafers in an industrial CMOS fab. Highly uniform films are obtained with an isotopic purity greater than 99.92\%. Hall-bar transistors with an equivalent oxide thickness of 17 nm are fabricated in an academic cleanroom. A critical density for conduction of $1.75\times10^{11}$ cm$^{-2}$ and a peak mobility of 9800 cm$^2$/Vs are measured at a temperature of 1.7 K. The $^{28}$Si/$^{28}$SiO$_2$ interface is characterized by a roughness of $Δ=0.4$ nm and a correlation length of $Λ=3.4$ nm. An upper bound for valley splitting energy of 480 $μ$eV is estimated at an effective electric field of 9.5 MV/m. These results support the use of wafer-scale $^{28}$Si/$^{28}$SiO$_2$ as a promising material platform to manufacture industrial spin qubits. △ Less

Submitted 21 January, 2019; v1 submitted 15 October, 2018; originally announced October 2018.

Comments: 5 pages, 3 figures

Journal ref: Phys. Rev. Applied 12, 014013 (2019)

Negative Learning Rates and P-Learning

Authors: Devon Merrill

Abstract: We present a method of training a differentiable function approximator for a regression task using negative examples. We effect this training using negative learning rates. We also show how this method can be used to perform direct policy learning in a reinforcement learning setting. We present a method of training a differentiable function approximator for a regression task using negative examples. We effect this training using negative learning rates. We also show how this method can be used to perform direct policy learning in a reinforcement learning setting. △ Less

Submitted 17 June, 2018; v1 submitted 27 March, 2016; originally announced March 2016.

Comments: Embarrassingly poor manuscript with many errors

Simple wavelet sets in R^n

Authors: Kathy D. Merrill

Abstract: Wavelet sets that are finite unions of convex sets are constructed in $\mathbb R^n$, $n\geq 2$, for dilation by any expansive matrix that has a power equal to a scalar times the identity and also has all singular values greater than $\sqrt n$. In particular, we produce simple wavelet sets in any dimension for dilation by any real scalar greater than 1. Wavelet sets that are finite unions of convex sets are constructed in $\mathbb R^n$, $n\geq 2$, for dilation by any expansive matrix that has a power equal to a scalar times the identity and also has all singular values greater than $\sqrt n$. In particular, we produce simple wavelet sets in any dimension for dilation by any real scalar greater than 1. △ Less

Submitted 10 November, 2013; v1 submitted 19 April, 2013; originally announced April 2013.

Comments: To appear in Journal of Geometric Analysis; 13 pages, 10 figures

MSC Class: 42C40; 52C22

Journal ref: Journal of Geometric Analysis, Volume 25, Issue 2 (2015) pp 1295-1305

Probability measures on solenoids corresponding to fractal wavelets

Authors: Lawrence W. Baggett, Kathy D. Merrill, Judith A. Packer, Arlan B. Ramsay

Abstract: The measure on generalized solenoids constructed using filters by Dutkay and Jorgensen is analyzed further by writing the solenoid as the product of a torus and a Cantor set. Using this decomposition, key differences are revealed between solenoid measures associated with classical filters in $\mathbb R^d$ and those associated with filters on inflated fractal sets. In particular, it is shown that t… ▽ More The measure on generalized solenoids constructed using filters by Dutkay and Jorgensen is analyzed further by writing the solenoid as the product of a torus and a Cantor set. Using this decomposition, key differences are revealed between solenoid measures associated with classical filters in $\mathbb R^d$ and those associated with filters on inflated fractal sets. In particular, it is shown that the classical case produces atomic fiber measures, and as a result supports both suitably defined solenoid MSF wavelets and systems of imprimitivity for the corresponding wavelet representation of the generalized Baumslag-Solitar group. In contrast, the fiber measures for filters on inflated fractal spaces cannot be atomic, and thus can support neither MSF wavelets nor systems of imprimitivity. △ Less

Submitted 2 November, 2010; originally announced November 2010.

Comments: 27 pages

MSC Class: 42C40; 22D25

Classification of Generalized Multiresolution Analyses

Authors: Lawrence W. Baggett, Veronika Furst, Kathy D. Merrill, Judith A. Packer

Abstract: We discuss how generalized multiresolution analyses (GMRAs), both classical and those defined on abstract Hilbert spaces, can be classified by their multiplicity functions $m$ and matrix-valued filter functions $H$. Given a natural number valued function $m$ and a system of functions encoded in a matrix $H$ satisfying certain conditions, a construction procedure is described that produces an abs… ▽ More We discuss how generalized multiresolution analyses (GMRAs), both classical and those defined on abstract Hilbert spaces, can be classified by their multiplicity functions $m$ and matrix-valued filter functions $H$. Given a natural number valued function $m$ and a system of functions encoded in a matrix $H$ satisfying certain conditions, a construction procedure is described that produces an abstract GMRA with multiplicity function $m $ and filter system $H$. An equivalence relation on GMRAs is defined and described in terms of their associated pairs $(m,H)$. This classification system is applied to classical examples in $L^2 (\mathbb R^d)$ as well as to previously studied abstract examples. △ Less

Submitted 28 October, 2009; originally announced October 2009.

Comments: 18 pages including bibliograph

MSC Class: 42C40; 47D03

Generalized low-pass filters and multiresolution analyses

Authors: Lawrence W. Baggett, Veronika Furst, Kathy D. Merrill, Judith A. Packer

Abstract: We study generalized filters that are associated to multiplicity functions and homomorphisms of the dual of an abelian group. These notions are based on the structure of generalized multiresolution analyses. We investigate when the Ruelle operator corresponding to such a filter is a pure isometry, and then use that characterization to study the problem of when a collection of closed subspaces, w… ▽ More We study generalized filters that are associated to multiplicity functions and homomorphisms of the dual of an abelian group. These notions are based on the structure of generalized multiresolution analyses. We investigate when the Ruelle operator corresponding to such a filter is a pure isometry, and then use that characterization to study the problem of when a collection of closed subspaces, which satisfies all the conditions of a GMRA except the trivial intersection condition, must in fact have a trivial intersection. In this context, we obtain a generalization of a theorem of Bownik and Rzeszotnik. △ Less

Submitted 10 December, 2008; originally announced December 2008.

Comments: 20 pages including bibliography

MSC Class: 42C40; 47D03

Generalized multiresolution analyses with given multiplicity functions

Authors: Lawrence W. Baggett, Nadia S. Larsen, Kathy D. Merrill, Judith A. Packer, Iain Raeburn

Abstract: Generalized multiresolution analyses are increasing sequences of subspaces of a Hilbert space $\H$ that fail to be multiresolution analyses in the sense of wavelet theory because the core subspace does not have an orthonormal basis generated by a fixed scaling function. Previous authors have studied a multiplicity function $m$ which, loosely speaking, measures the failure of the GMRA to be an MR… ▽ More Generalized multiresolution analyses are increasing sequences of subspaces of a Hilbert space $\H$ that fail to be multiresolution analyses in the sense of wavelet theory because the core subspace does not have an orthonormal basis generated by a fixed scaling function. Previous authors have studied a multiplicity function $m$ which, loosely speaking, measures the failure of the GMRA to be an MRA. When the Hilbert space $\H$ is $L^2(\mathbb R^n)$, the possible multiplicity functions have been characterized by Baggett and Merrill. Here we start with a function $m$ satisfying a consistency condition which is known to be necessary, and build a GMRA in an abstract Hilbert space with multiplicity function $m$. △ Less

Submitted 10 October, 2007; originally announced October 2007.

Comments: 16 pages including bibliography

MSC Class: 42C40; 47D03

Construction of Parseval wavelets from redundant filter systems

Authors: L. W. Baggett, P. E. T. Jorgensen, K. D. Merrill, J. A. Packer

Abstract: We consider wavelets in L^2(R^d) which have generalized multiresolutions. This means that the initial resolution subspace V_0 in L^2(R^d) is not singly generated. As a result, the representation of the integer lattice Z^d restricted to V_0 has a nontrivial multiplicity function. We show how the corresponding analysis and synthesis for these wavelets can be understood in terms of unitary-matrix-v… ▽ More We consider wavelets in L^2(R^d) which have generalized multiresolutions. This means that the initial resolution subspace V_0 in L^2(R^d) is not singly generated. As a result, the representation of the integer lattice Z^d restricted to V_0 has a nontrivial multiplicity function. We show how the corresponding analysis and synthesis for these wavelets can be understood in terms of unitary-matrix-valued functions on a torus acting on a certain vector bundle. Specifically, we show how the wavelet functions on R^d can be constructed directly from the generalized wavelet filters. △ Less

Submitted 5 May, 2005; v1 submitted 14 May, 2004; originally announced May 2004.

Comments: 34 pages, AMS-LaTeX ("amsproc" document class) v2 changes minor typos in Sections 1 and 4, v3 adds a number of references on GMRA theory and wavelet multiplicity analysis; v4 adds material on pages 2, 3, 5 and 10, and two more references

MSC Class: Primary 54C40; 14E20; 42A65; Secondary 46E25; 20C20

An analogue of Bratteli-Jorgensen loop group actions for GMRA's

Authors: L. W. Baggett, P. E. T. Jorgensen, K. D. Merrill, J. A. Packer

Abstract: Several years ago, O. Bratelli and P. Jorgensen developed the concept of m-systems of filters for dilation by a positive integer N>1 on L^2(R). They constructed a loop group action on m-systems. By work of Mallat and Meyer, these m-systems are important in constructing multi-resolution analyses and wavelets associated to dilation by N and translation by Z on L^2(R). In this paper, we discuss an… ▽ More Several years ago, O. Bratelli and P. Jorgensen developed the concept of m-systems of filters for dilation by a positive integer N>1 on L^2(R). They constructed a loop group action on m-systems. By work of Mallat and Meyer, these m-systems are important in constructing multi-resolution analyses and wavelets associated to dilation by N and translation by Z on L^2(R). In this paper, we discuss an extension of this loop-group construction to generalized filter systems, which we will call ``M-systems,'' associated with generalized multiresolution analyses. In particular, we show that every multiplicity function has an associated generalized loop group which acts freely and transitively on the set of M-systems corresponding to the multiplicity function. The results of Bratteli and Jorgensen correspond to the case where the multiplicity function is identically equal to 1. △ Less

Submitted 13 August, 2003; originally announced August 2003.

Comments: 15 pages; AMS-LaTeX; submitted to proceedings of AMS Special Session on Wavelets, Frames, and Operator Theory held at Baltimore

MSC Class: 54C40; 14E20 (Primary) 46E25; 20C20 (Secondary

Journal ref: Wavelets, Frames, and Operator Theory (College Park, Maryland, January 15-21, 2003) (C. Heil, P.E.T. Jorgensen, and D. Larson, eds.), Contemp. Math., vol. 345, American Mathematical Society, Providence, RI, 2004, pp. 11-25