Existing deep neural network (DNN) frameworks optimize the computation graph of a DNN by applying graph transformations manually designed by human experts. This approach misses possible graph optimizations and is difficult to scale, as new DNN operators are introduced on a regular basis.
We propose TASO, the first DNN computation graph optimizer that automatically generates graph substitutions. TASO takes as input a list of operator specifications and generates candidate substitutions using the given operators as basic building blocks. All generated substitutions are formally verified against the operator specifications using an automated theorem prover. To optimize a given DNN computation graph, TASO performs a cost-based backtracking search, applying the substitutions to find an optimized graph, which can be directly used by existing DNN frameworks.
Our evaluation on five real-world DNN architectures shows that TASO outperforms existing DNN frameworks by up to 2.8X, while requiring significantly less human effort. For example, TensorFlow currently contains approximately 53,000 lines of manual optimization rules, while the operator specifications needed by TASO are only 1,400 lines of code.
more »
« less
Marvel: A Data-Centric Approach for Mapping Deep Learning Operators on Spatial Accelerators
A spatial accelerator’s efficiency depends heavily on both its mapper and cost models to generate optimized mappings for various operators of DNN models. However, existing cost models lack a formal boundary over their input programs (operators) for accurate and tractable cost analysis of the mappings, and this results in adaptability challenges to the cost models for new operators. We consider the recently introduced Maestro Data-Centric (MDC) notation and its analytical cost model to address this challenge because any mapping expressed in the notation is precisely analyzable using the MDC’s cost model. In this article, we characterize the set of input operators and their mappings expressed in the MDC notation by introducing a set of conformability rules . The outcome of these rules is that any loop nest that is perfectly nested with affine tensor subscripts and without conditionals is conformable to the MDC notation. A majority of the primitive operators in deep learning are such loop nests. In addition, our rules enable us to automatically translate a mapping expressed in the loop nest form to MDC notation and use the MDC’s cost model to guide upstream mappers. Our conformability rules over the input operators result in a structured mapping space of the operators, which enables us to introduce a mapper based on our decoupled off-chip/on-chip approach to accelerate mapping space exploration. Our mapper decomposes the original higher-dimensional mapping space of operators into two lower-dimensional off-chip and on-chip subspaces and then optimizes the off-chip subspace followed by the on-chip subspace. We implemented our overall approach in a tool called Marvel , and a benefit of our approach is that it applies to any operator conformable with the MDC notation. We evaluated Marvel over major DNN operators and compared it with past optimizers.
more »
« less
- Award ID(s):
- 1909900
- NSF-PAR ID:
- 10348661
- Date Published:
- Journal Name:
- ACM Transactions on Architecture and Code Optimization
- Volume:
- 19
- Issue:
- 1
- ISSN:
- 1544-3566
- Page Range / eLocation ID:
- 1 to 26
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The increasing reliance on robust data-driven decision-making across many domains has made it necessary for data management systems to manage many thousands to millions of versions of datasets, acquired or constructed at various stages of analysis pipelines over time. Delta encoding is an effective and widely-used solution to compactly store a large number of datasets, that simultaneously exploits redundancies across them and keeps the average retrieval cost of reconstructing any dataset low. However, supporting any kind of rich retrieval or querying functionality, beyond single dataset checkout, is challenging in such storage engines. In this paper, we initiate a systematic study of this problem, and present DEX, a novel stand-alone delta-oriented execution engine, whose goal is to take advantage of the already computed deltas between the datasets for efficient query processing. In this work, we study how to execute checkout, intersection, union and t-threshold queries over record-based files; we show that processing of even these basic queries leads to many new and unexplored challenges and trade-offs. Starting from a query plan that confines query execution to a small set of deltas, we introduce new transformation rules based on the algebraic properties of the deltas, that allow us to explore the search space of alternative plans. For the case of checkout, we present a dynamic programming algorithm to efficiently select the optimal query plan under our cost model, while we design efficient heuristics to select effective plans that vastly outperform the base checkout-then-query approach for other queries. A key characteristic of our query execution methods is that the computational cost is primarily dependent on the size and the number of deltas in the expression (typically small), and not the input dataset versions (which can be very large). We have implemented DEX prototype on top of git, a widely used version control system. We present an extensive experimental evaluation on synthetic data with diverse characteristics, that shows that our methods perform exceedingly well compared to the baseline.more » « less
-
Embedding properties of network realizations of dissipative reduced order models Jörn Zimmerling, Mikhail Zaslavsky,Rob Remis, Shasri Moskow, Alexander Mamonov, Murthy Guddati, Vladimir Druskin, and Liliana Borcea Mathematical Sciences Department, Worcester Polytechnic Institute https://www.wpi.edu/people/vdruskin Abstract Realizations of reduced order models of passive SISO or MIMO LTI problems can be transformed to tridiagonal and block-tridiagonal forms, respectively, via dierent modi cations of the Lanczos algorithm. Generally, such realizations can be interpreted as ladder resistor-capacitor-inductor (RCL) networks. They gave rise to network syntheses in the rst half of the 20th century that was at the base of modern electronics design and consecutively to MOR that tremendously impacted many areas of engineering (electrical, mechanical, aerospace, etc.) by enabling e cient compression of the underlining dynamical systems. In his seminal 1950s works Krein realized that in addition to their compressing properties, network realizations can be used to embed the data back into the state space of the underlying continuum problems. In more recent works of the authors Krein's ideas gave rise to so-called nite-dierence Gaussian quadrature rules (FDGQR), allowing to approximately map the ROM state-space representation to its full order continuum counterpart on a judicially chosen grid. Thus, the state variables can be accessed directly from the transfer function without solving the full problem and even explicit knowledge of the PDE coe cients in the interior, i.e., the FDGQR directly learns" the problem from its transfer function. This embedding property found applications in PDE solvers, inverse problems and unsupervised machine learning. Here we show a generalization of this approach to dissipative PDE problems, e.g., electromagnetic and acoustic wave propagation in lossy dispersive media. Potential applications include solution of inverse scattering problems in dispersive media, such as seismic exploration, radars and sonars. To x the idea, we consider a passive irreducible SISO ROM fn(s) = Xn j=1 yi s + σj , (62) assuming that all complex terms in (62) come in conjugate pairs. We will seek ladder realization of (62) as rjuj + vj − vj−1 = −shˆjuj , uj+1 − uj + ˆrj vj = −shj vj , (63) for j = 0, . . . , n with boundary conditions un+1 = 0, v1 = −1, and 4n real parameters hi, hˆi, ri and rˆi, i = 1, . . . , n, that can be considered, respectively, as the equivalent discrete inductances, capacitors and also primary and dual conductors. Alternatively, they can be viewed as respectively masses, spring stiness, primary and dual dampers of a mechanical string. Reordering variables would bring (63) into tridiagonal form, so from the spectral measure given by (62 ) the coe cients of (63) can be obtained via a non-symmetric Lanczos algorithm written in J-symmetric form and fn(s) can be equivalently computed as fn(s) = u1. The cases considered in the original FDGQR correspond to either (i) real y, θ or (ii) real y and imaginary θ. Both cases are covered by the Stieltjes theorem, that yields in case (i) real positive h, hˆ and trivial r, rˆ, and in case (ii) real positive h,r and trivial hˆ,rˆ. This result allowed us a simple interpretation of (62) as the staggered nite-dierence approximation of the underlying PDE problem [2]. For PDEs in more than one variables (including topologically rich data-manifolds), a nite-dierence interpretation is obtained via a MIMO extensions in block form, e.g., [4, 3]. The main di culty of extending this approach to general passive problems is that the Stieltjes theory is no longer applicable. Moreover, the tridiagonal realization of a passive ROM transfer function (62) via the ladder network (63) cannot always be obtained in port-Hamiltonian form, i.e., the equivalent primary and dual conductors may change sign [1]. 100 Embedding of the Stieltjes problems, e.g., the case (i) was done by mapping h and hˆ into values of acoustic (or electromagnetic) impedance at grid cells, that required a special coordinate stretching (known as travel time coordinate transform) for continuous problems. Likewise, to circumvent possible non-positivity of conductors for the non-Stieltjes case, we introduce an additional complex s-dependent coordinate stretching, vanishing as s → ∞ [1]. This stretching applied in the discrete setting induces a diagonal factorization, removes oscillating coe cients, and leads to an accurate embedding for moderate variations of the coe cients of the continuum problems, i.e., it maps discrete coe cients onto the values of their continuum counterparts. Not only does this embedding yields an approximate linear algebraic algorithm for the solution of the inverse problems for dissipative PDEs, it also leads to new insight into the properties of their ROM realizations. We will also discuss another approach to embedding, based on Krein-Nudelman theory [5], that results in special data-driven adaptive grids. References [1] Borcea, Liliana and Druskin, Vladimir and Zimmerling, Jörn, A reduced order model approach to inverse scattering in lossy layered media, Journal of Scienti c Computing, V. 89, N1, pp. 136,2021 [2] Druskin, Vladimir and Knizhnerman, Leonid, Gaussian spectral rules for the three-point second dierences: I. A two-point positive de nite problem in a semi-in nite domain, SIAM Journal on Numerical Analysis, V. 37, N 2, pp.403422, 1999 [3] Druskin, Vladimir and Mamonov, Alexander V and Zaslavsky, Mikhail, Distance preserving model order reduction of graph-Laplacians and cluster analysis, Druskin, Vladimir and Mamonov, Alexander V and Zaslavsky, Mikhail, Journal of Scienti c Computing, V. 90, N 1, pp 130, 2022 [4] Druskin, Vladimir and Moskow, Shari and Zaslavsky, Mikhail LippmannSchwingerLanczos algorithm for inverse scattering problems, Inverse Problems, V. 37, N. 7, 2021, [5] Mark Adolfovich Nudelman The Krein String and Characteristic Functions of Maximal Dissipative Operators, Journal of Mathematical Sciences, 2004, V 124, pp 49184934 Go back to Plenary Speakers Go back to Speakers Go backmore » « less
-
This work proposes a new computational framework for learning a structured generative model for real-world datasets. In particular, we propose to learn a Closed-loop Transcriptionbetween a multi-class, multi-dimensional data distribution and a Linear discriminative representation (CTRL) in the feature space that consists of multiple independent multi-dimensional linear subspaces. In particular, we argue that the optimal encoding and decoding mappings sought can be formulated as a two-player minimax game between the encoder and decoderfor the learned representation. A natural utility function for this game is the so-called rate reduction, a simple information-theoretic measure for distances between mixtures of subspace-like Gaussians in the feature space. Our formulation draws inspiration from closed-loop error feedback from control systems and avoids expensive evaluating and minimizing of approximated distances between arbitrary distributions in either the data space or the feature space. To a large extent, this new formulation unifies the concepts and benefits of Auto-Encoding and GAN and naturally extends them to the settings of learning a both discriminative and generative representation for multi-class and multi-dimensional real-world data. Our extensive experiments on many benchmark imagery datasets demonstrate tremendous potential of this new closed-loop formulation: under fair comparison, visual quality of the learned decoder and classification performance of the encoder is competitive and arguably better than existing methods based on GAN, VAE, or a combination of both. Unlike existing generative models, the so-learned features of the multiple classes are structured instead of hidden: different classes are explicitly mapped onto corresponding independent principal subspaces in the feature space, and diverse visual attributes within each class are modeled by the independent principal components within each subspace.more » « less
-
Approximate computing is a promising way to improve the power efficiency of deep learning. While recent work proposes new arithmetic circuits (adders and multipliers) that consume substantially less power at the cost of computation errors, these approximate circuits decrease the end-to-end accuracy of common models. We present AutoApprox, a framework to automatically generate approximate low-power deep learning accelerators without any accuracy loss. AutoApprox generates a wide range of approximate ASIC accelerators with a TPUv3 systolic-array template. AutoApprox uses a learned router to assign each DNN layer to an approximate systolic array from a bank of arrays with varying approximation levels. By tailoring this routing for a specific neural network architecture, we discover circuit designs without the accuracy penalty from prior methods. Moreover, AutoApprox optimizes for the end-to-end performance, power and area of the the whole chip and PE mapping rather than simply measuring the performance of the arithmetic units in iso-lation. To our knowledge, our work is the first to demonstrate the effectiveness of custom-tailored approximate circuits in delivering significant chip-level energy savings with zero accuracy loss on a large-scale dataset such as ImageNet. AutoApprox synthesizes a novel approximate accelerator based on the TPU that reduces end-to-end power consumption by 3.2% and area by 5.2% at a sub-10nm process with no degradation in ImageNet validation top-1 and top-5 accuracy.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1145/3485137
