System and method for multi-color dilu preconditioner

Abstract

A system and method for preconditioning or smoothing (e.g., multi-color DILU preconditioning) for iterative solving of a system of equations. The method includes accessing a matrix comprising a plurality of coefficients of a system of equations and accessing coloring information corresponding to the matrix. The method further includes determining a diagonal matrix based on the matrix and the coloring information corresponding to the matrix. The determining of the diagonal matrix may be determined in parallel on a per color basis. The method may further include determining an updated solution to the system of equations where the updated solution is determined in parallel on a per color basis using the diagonal matrix.

Description

FIELD OF THE INVENTION

Embodiments of the present invention are generally related to solving linear systems of equations.

BACKGROUND OF THE INVENTION

As computer systems have advanced, processing power and speed have increased substantially. Computer systems have thus been able to solve increasing complicated problems. Iterative methods for solving large sparse linear systems of equations have become popular in many areas of scientific computing. Although direct solution methods have long been preferred because of their robustness and predictable behavior, the development of new efficient iterative solvers and the increased need for solving very large systems have caused iterative solvers to become the method of choice for solving sparse linear systems.

A wide variety of iterative algorithms exist to solve sparse linear systems of equations including stationary iterative methods (such as Jacobi, Gauss-Seidel, Successive Over-Relaxation (SOR)), Krylov subspace methods (such as Conjugate Gradient (CG), Bi-Conjugate Gradient (BiCG), Generalized Minimal Residual Method (GMRES)) and Algebraic MultiGrid (AMG) methods. Krylov subspace methods and AMG methods have been the most popular iterative methods to solve sparse linear systems arising from partial differential equations (PDEs) because of their robustness and efficiency. Unfortunately, these iterative algorithms can be sequential in nature. This sequential nature results from the dependencies between computations and thereby results in increased computation time as each computation is dependent upon the needs results from proceeding computations.

Thus, while iterative algorithms are desirable over direct solving methods because of their efficiency, the sequential nature of the computations limits performance and time saved.

SUMMARY OF THE INVENTION

Accordingly, what is needed is an efficient preconditioner or smoother that is parallel in nature thereby allowing efficient solving of systems of equations using iterative methods. Embodiments of the present invention implement a multi-color DILU preconditioner that is suitable for implementation on a parallel hardware architecture (e.g., GPU). Embodiments of the present invention are operable to use coloring to extract parallelism in a DILU smoother or preconditioner. Embodiments of the present invention are further operable to perform the multi-color DILU preconditioning in parallel thereby advantageously providing significantly enhanced performance over traditional ILU preconditioners which are difficult to parallelize. The parallelism of embodiments of the present invention advantageously allows faster completion of preconditioning or smoothing over sequential methods. Embodiments of the present invention further advantageously provide a multi-color DILU preconditioner that is strong, parallel, requires very low storage, and computationally inexpensive.

In one embodiment, the present invention is directed toward a method for preconditioning. The method includes accessing a matrix comprising a plurality of coefficients of a system of equations and accessing coloring information corresponding to the matrix. The method further includes determining a diagonal matrix based on the matrix and the coloring information corresponding to the matrix. The determining of the diagonal matrix may be determined in parallel on a per color basis. The determining of the diagonal matrix may be performed on a parallel hardware architecture (e.g., a graphics processing unit (GPU)). The method may further include accessing an initial solution and determining a preconditioning matrix based on the diagonal matrix (e.g., an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner). The method may further include determining an updated solution to the system of equations. In one embodiment, the updated solution is determined in parallel on a per color basis.

In one embodiment, the present invention is implemented as a system for preconditioning. The system includes a matrix access module operable to access a matrix corresponding to a system of equations and a coloring access module operable to access color information corresponding to the matrix. In one embodiment, the matrix corresponds to a plurality of coefficients of the system of equations. The system further includes a diagonal matrix determination module operable for determining a diagonal matrix based on the color information corresponding to the matrix in parallel on a per color basis. In one embodiment, the diagonal matrix is used to construct a preconditioner matrix that comprises the product of lower and upper triangular matrices. The diagonal matrix may be determined by a graphical processing unit (GPU).

The system may further include a solution access module operable for accessing an initial solution to the system of equations and an updated solution determination module operable for determining an updated solution based on the matrix, the initial solution, and the color information. In one embodiment, the determining of the updated solution is determined in parallel based on a per color basis. In one exemplary embodiment, the updated solution module is operable to determine the updated solution based on a lower triangular solving phase and an upper triangular solving phase. The determining of the updated solution may be determined by a GPU.

In another embodiment, the present invention is directed to a method for solving a system of equations. The method includes accessing a matrix comprising coefficients of a system of equations and accessing coloring information corresponding to the matrix. The method further includes determining a diagonal matrix based on the coloring information and the matrix. In one embodiment, the determining of the diagonal matrix is computed in parallel (e.g., by a GPU) for each respective color of the coloring information. The method further includes determining an updated solution to the system of equations. In one embodiment, the determining of the updated solution is based on an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner performed by a graphics processing unit (GPU) and the updated solution is computed in parallel on a per color basis. The determining of the updated solution may be based on an initial solution, the matrix, and the diagonal matrix. In one exemplary embodiment, the updated solution is determined based on a lower triangular solving phase and an upper triangular solving phase. In another embodiment, the updated solution is determined based on a common sparse matrix vector multiplication function. The determining of the updated solution may be a portion of an algebraic multigrid (AMG) iterative method.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the present invention are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which like reference numerals refer to similar elements.

FIG. 1 shows a computer system in accordance with one embodiment of the present invention.

FIG. 2 shows a flowchart of an exemplary electronic component controlled process for solving a system of equations in accordance with one embodiment of the present invention.

FIG. 3 shows a flowchart of an exemplary electronic component controlled process for setup of a preconditioner in accordance with one embodiment of the present invention.

FIG. 4 shows a block diagram of a colored matrix in accordance with one embodiment of the present invention.

FIG. 5 shows a flowchart of an exemplary electronic component controlled process for updating a solution based on the preconditioner in accordance with one embodiment of the present invention.

FIG. 6 shows a block diagram of exemplary components of a system in accordance with one embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

Reference will now be made in detail to the preferred embodiments of the present invention, examples of which are illustrated in the accompanying drawings. While the invention will be described in conjunction with the preferred embodiments, it will be understood that they are not intended to limit the invention to these embodiments. On the contrary, the invention is intended to cover alternatives, modifications and equivalents, which may be included within the spirit and scope of the invention as defined by the appended claims. Furthermore, in the following detailed description of embodiments of the present invention, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be recognized by one of ordinary skill in the art that the present invention may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail as not to unnecessarily obscure aspects of the embodiments of the present invention.

Notation and Nomenclature:

Some portions of the detailed descriptions, which follow, are presented in terms of procedures, steps, logic blocks, processing, and other symbolic representations of operations on data bits within a computer memory. These descriptions and representations are the means used by those skilled in the data processing arts to most effectively convey the substance of their work to others skilled in the art. A procedure, computer executed step, logic block, process, etc., is here, and generally, conceived to be a self-consistent sequence of steps or instructions leading to a desired result. The steps are those requiring physical manipulations of physical quantities. Usually, though not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared, and otherwise manipulated in a computer system. It has proven convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers, or the like.

It should be borne in mind, however, that all of these and similar terms are to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantities. Unless specifically stated otherwise as apparent from the following discussions, it is appreciated that throughout the present invention, discussions utilizing terms such as “processing” or “accessing” or “executing” or “storing” or “rendering” or the like, refer to the action and processes of an integrated circuit (e.g., computing system 100 of FIG. 1), or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system memories or registers or other such information storage, transmission or display devices.

Computer System Environment

FIG. 1 shows an exemplary computer system 100 in accordance with one embodiment of the present invention. Computer system 100 depicts the components of a generic computer system in accordance with embodiments of the present invention providing the execution platform for certain hardware-based and software-based functionality. In general, computer system 100 comprises at least one CPU 101, a system memory 115, and at least one graphics processor unit (GPU) 110. The CPU 101 can be coupled to the system memory 115 via a bridge component/memory controller (not shown) or can be directly coupled to the system memory 115 via a memory controller (not shown) internal to the CPU 101. The GPU 110 may be coupled to a display 112. One or more additional GPUs can optionally be coupled to system 100 to further increase its computational power. The GPU(s) 110 is coupled to the CPU 101 and the system memory 115. The GPU 110 can be implemented as a discrete component, a discrete graphics card designed to couple to the computer system 100 via a connector (e.g., AGP slot, PCI-Express slot, etc.), a discrete integrated circuit die (e.g., mounted directly on a motherboard), or as an integrated GPU included within the integrated circuit die of a computer system chipset component (not shown). Additionally, a local graphics memory 114 can be included for the GPU 110 for high bandwidth graphics data storage.

The CPU 101 and the GPU 110 can also be integrated into a single integrated circuit die and the CPU and GPU may share various resources, such as instruction logic, buffers, functional units and so on, or separate resources may be provided for graphics and general-purpose operations. The GPU may further be integrated into a core logic component. Accordingly, any or all the circuits and/or functionality described herein as being associated with the GPU 110 can also be implemented in, and performed by, a suitably equipped CPU 101. Additionally, while embodiments herein may make reference to a GPU, it should be noted that the described circuits and/or functionality can also be implemented and other types of processors (e.g., general purpose or other special-purpose coprocessors) or within a CPU.

In one exemplary embodiment, GPU 110 is operable for General-purpose computing on graphics processing units (GPGPU) computing. GPU 110 may execute Compute Unified Device Architecture (CUDA) programs and Open Computing Language (OpenCL) programs. GPU 110 may thus be used for a variety of computing applications including simulations of molecular dynamics, computational fluid dynamics, reservoir simulations, and finite structural problems. It is appreciated that the parallel architecture of GPU 110 may have significant performance advantages over CPU 101.

System 100 can be implemented as, for example, a desktop computer system or server computer system having a powerful general-purpose CPU 101 coupled to a dedicated graphics rendering GPU 110. In such an embodiment, components can be included that add peripheral buses, specialized audio/video components, IO devices, and the like. Similarly, system 100 can be implemented as a handheld device (e.g., cellphone, etc.), direct broadcast satellite (DBS)/terrestrial set-top box or a set-top video game console device such as, for example, the Xbox®, available from Microsoft Corporation of Redmond, Wash., or the PlayStation3®, available from Sony Computer Entertainment Corporation of Tokyo, Japan. System 100 can also be implemented as a “system on a chip”, where the electronics (e.g., the components 101, 115, 110, 114, and the like) of a computing device are wholly contained within a single integrated circuit die. Examples include a hand-held instrument with a display, a car navigation system, a portable entertainment system, and the like.

Exemplary Systems and Methods for Multi-Color DILU Preconditioner

Embodiments of the present invention implement a multi-color DILU preconditioner that is suitable for implementation on a parallel hardware architecture (e.g., GPU). Embodiments of the present invention are operable to use coloring to extract parallelism in a DILU smoother or preconditioner. Embodiments of the present invention are further operable to perform the multi-color DILU preconditioning in parallel thereby advantageously providing significantly enhanced performance over traditional ILU preconditioners which are difficult to parallelize. The parallelism of embodiments of the present invention advantageously allows faster completion of preconditioning or smoothing over sequential methods. Embodiments of the present invention further advantageously provide a multi-color DILU preconditioner that is strong, parallel, requires very low storage, and computationally cheap. The low-storage of embodiments of the present invention advantageously allows for the solution of larger systems of equations on a system with finite amount of memory.

Embodiments of the present invention further provide an efficient preconditioner or smoother operable for use with a variety of iterative methods (e.g., iterative methods that utilize preconditioners or smoothers). Embodiments of the present invention are operable for use with a variety of simulations including molecular dynamics, computational fluid dynamics, reservoir simulations and finite structural problems. For example, in a fluid flow simulation, the system of equations may be used to predict how a fluid behaves as the fluid flows over an object.

Preconditioning

The efficiency and robustness of iterative methods (e.g., Krylov subspace iterative methods) can be improved by using a preconditioner. Preconditioning is a way to transform the original linear system into one which has the same solution, but which is likely to be easier to solve with an iterative solver. In general, the reliability of iterative techniques, when dealing with various applications, depends much more on the quality of the preconditioner than on the particular iterative method used. Popular preconditioners include stationary iterative methods such as Jacobi and Gauss-Seidel and incomplete (LU) factorization methods (hereinafter ILU preconditioners). It is noted that LU and ILU factorization factorizes a matrix as the product of a lower triangular matrix and an upper triangular matrix.

In the context of algebraic multigrid (AMG) methods, the preconditioner may be used as a smoother, whose objective is to damp the oscillatory components of the error. It is appreciated that good preconditioners are generally good smoothers in the context of AMG methods. It is further appreciated that an efficient preconditioner or smoother is one that requires low storage, facilitates reduction of the number of iterations of the iterative solver (e.g., Krylov subspace method or AMG method) and is computational cheap.

ILU(0) Preconditioner

Embodiments of the present invention are described herein with reference to exemplary sparse linear system Ax=b. It is noted the exemplary sparse linear system Ax=b is used for explanatory purposes. It is further noted that embodiments of the present invention are operable for use with linear systems of equations (e.g., sparse linear system of equations) and not intended to be limited. For example, in the smoothing step of an AMG algorithm, the solution is updated using the preconditioning matrix M as follows:x_k+1=x_k+M⁻¹(b−Ax_k),where x_k+1 is the updated solution, x_k is the initial or previously updated solution, M is the preconditioning matrix, b is a vector, and A is a matrix.

For the preconditioner to be computationally cheap, the preconditioning matrix M needs to be easily invertible. In an ILU preconditioner, M=LŪ where L is a sparse lower triangular matrix and Ū is a sparse upper triangular matrix such that the residual matrix R=LŪ−A satisfies certain constraints (e.g., having zero entries in some locations). The ILU factorization technique with no fill-in, denoted by ILU(0) is probably the most popular ILU preconditioner. In the ILU(0) preconditioner, M=LŪ, where L is a sparse lower triangular matrix with the same non-zero structure as the lower part of A, Ū is a sparse upper triangular matrix with the same non-zero structure as the upper part of A, and the entries in the matrix M satisfy m_ij=a_ij if a_ij≠0.

DILU Preconditioner

It is noted that that the amount of memory required to store the preconditioning matrix M of the ILU(0) preconditioner is the same amount of memory required to store matrix A. To further reduce the memory requirements, a variant of the ILU(0) preconditioner has been proposed, often referred to as a DILU preconditioner. The DILU preconditioner is a form of ILU preconditioners, and hence much better numerically than naïve smoothers (e.g., Jacobi or Gauss-Seidel). In a DILU preconditioner, the preconditioning matrix M has the following form:M=(E+L)E⁻¹(E+U),where L and U are the strict lower and upper triangular parts of A, and E is a diagonal matrix that is constructed such that diag(M)=diag(A).

It can be shown that the DILU preconditioner is equivalent to the ILU(0) preconditioner when the product of the strict-lower part and the strict-upper part of A consists only of diagonal elements and fill-in elements, which is true for example for the standard 5-point difference approximations to second order partial differential operators. The main advantage of the DILU preconditioner over the ILU(0) is that it requires only one extra diagonal of storage (to store that diagonal matrix E).

With the DILU preconditioner, the matrix E can be obtained recursively by realizing that:diag(A)=diag(M)=diag((E+L)E⁻¹(E+U))=diag(E+LE⁻¹U)

so that E can be obtained as:E₁₁=A₁₁ E₂₂=A₂₂−L₂₁E₁₁⁻¹U₁₂ E₃₃=A₃₃−L₃₁E₁₁⁻¹U₁₃−L₃₂E₂₂⁻¹U₂₃ E₄₄=A₄₄−L₄₁E₁₁⁻¹U₁₄−L₄₂E₂₂⁻¹U₂₄−L₄₃E₃₃⁻¹U₃₄

It is appreciated that the above equations reflect the sequential nature of conventional solutions for determining the diagonal matrix E.

In the smoothing step, since M=(E+L)E⁻¹(E+U) (a LU decomposition of M), the productM⁻¹(b−Ax_k)

can be obtained by successively solving a lower-triangular system followed by an upper triangular system. In other words, one needs to computez=M⁻¹(b−Ax_k)=(E+U)⁻¹E(E+L)⁻¹(b−Ax_k)

which can be obtained by solving

1) Lower triangular solve(E+L)y=b−Ax_k

2) Upper triangular solve(E+U)z=Ey

With reference to FIGS. 2, 3, and 5, flowcharts 200, 300, and 500 illustrate example functions used by various embodiments of the present invention. Although specific function blocks (“blocks”) are disclosed in flowcharts 200, 300, and 500, such steps are examples. That is, embodiments are well suited to performing various other blocks or variations of the blocks recited in flowcharts 200, 300, and 500. It is appreciated that the blocks in flowcharts 200, 300, and 500 may be performed in an order different than presented, and that not all of the blocks in flowcharts 200, 300, and 500 may be performed. Processes 200, 300, and 500 may be used with or as part of a variety of iterative solving methods including, but not limited to, Krylov subspace methods (e.g., Conjugate Gradient (CG), Bi-Conjugate Gradient (BiCG), Generalized Minimal Residual Method (GMRES)) and Algebraic MultiGrid (AMG) methods. Portions of flowcharts 200, 300, and 500 may be performed or executed with a parallel hardware architecture (e.g., a graphics processing unit (GPU)).

FIG. 2 shows a flowchart of an exemplary electronic component controlled process for solving a system of equations in accordance with one embodiment of the present invention. In one embodiment, flowchart 200 depicts a process for using numerical analysis to iteratively solve a system of equations. Portions of Flowchart 200 may be executed by a GPU (e.g., GPU 110). In one exemplary embodiment, embodiments of the present invention are used as a preconditioner to achieve better approximations of updated solutions for numerical analysis.

At block 202, data corresponding to a system of equations is accessed. The data corresponding to the system of equations may be operable for use in solving the system of equations (e.g., a matrix A of the values of coefficients corresponding to the system of equations, a vector x corresponding the unknowns of the system of equations, and the right hand vector b for the linear system of equations Ax=b)

At block 204, a first preconditioner is selected. The first preconditioner may be a multi-color DILU preconditioner, as described herein, or may be any of a variety of preconditioners or smoothers selected based on the iterative method being used.

At block 206, a first updated solution is determined based on the first preconditioner. Block 206 may be performed multiple times thereby using the first preconditioner to determine a plurality of updated solutions, each of which are better approximations or those having less error. Block 212 may be performed based on the updated solution converging to a value within a tolerance.

At block 208, a second preconditioner is selected. The second preconditioner may be a multi-color DILU preconditioner, as described herein, or may be any of a variety of preconditioners or smoothers selected based on the iterative method being used.

At block 210, a second updated solution is determined based on the second preconditioner. Block 210 may be performed multiple times thereby using the updated solution from the first preconditioner to determine further updated solutions which are better approximations or those having less error. Block 204 may be performed for an iterative method that utilizes both the first preconditioner and second preconditioner more than once. Embodiments of the present invention are operable for use with iterative methods that iterate between two or more preconditioners or smoothers.

At block 212, a solution is output. In one embodiment, the solution may be a better approximate solution or updated solution which may be used with other portions of an iterative method (e.g., additional preconditions in process 200). In another embodiment, the solution output may be a final solution to the iterative method. The solution may be output based on a convergence to a solution that falls within some tolerance.

FIG. 3 shows a flowchart of an exemplary electronic component controlled process for setup of a preconditioner in accordance with one embodiment of the present invention. In one embodiment, flowchart 300 depicts a setup phase of Multi Color DILU preconditioner.

Unfortunately, with the DILU preconditioner, the setup phase including the computation of the E matrix is very sequential in nature. Embodiments of the present invention are operable to extract parallelism during the setup phase (e.g. computation of the diagonal matrix E) by using coloring to implicitly renumber the rows (e.g., the order of processing the rows) of the system of equations (e.g., matrix A).

At block 302, a matrix is accessed. In one embodiment, a matrix of the coefficients of the system of equations is accessed (e.g., matrix A for the system Ax=b).

At block 304, a coloring or color information of the matrix is accessed. The coloring of the rows of a matrix (e.g., matrix A) allows the extraction of parallelism in its processing. The coloring is constructed such that different colors are independent and rows of the same color can be computed in parallel. In one embodiment, the coloring may assign a color to each unknown. Since there is a one-to-one relation between the unknowns (e.g., in vector x), the rows, and the columns of matrix A, this means that each row and each column of matrix A is assigned a particular color. The coloring may be used to renumber the rows and columns of the matrix, e.g., typically according to ascending color. For example, the processing order may indicate that first all rows of color 0 are processed then all rows of color 1, etc. and columns are processed in an order such that only columns of a color smaller than the current row color are processed. In this sense a coloring corresponds to an implicit renumbering that can be used as the processing order of the rows and columns of the matrix (e.g., which is not necessarily constructed or stored in memory explicitly). The coloring can be used to reorder the matrix according to this numbering (e.g., to store a reordered matrix). In one embodiment, the coloring information or other coloring information which corresponds to a renumbering of the coefficients of the corresponding systems of equations (e.g., matrix A) facilitates parallel computation of a diagonal matrix (e.g., matrix E). Embodiments of the present invention are operable for accessing coloring information that is determined by an external module.

The color information may include respective colors corresponding to unknowns that can be computed independently. For example, in a 4×4 matrix, rows 1 and 2 may correspond to the color red while rows 3 and 4 may correspond to the color green indicating that rows 1 and 2 can be computed in parallel and rows 3 and 4 can be computed in parallel.

At block 306, a coloring of the matrix is determined. In one exemplary embodiment, a coloring of matrix for the coefficients of the systems of equations is determined. The coloring of the matrix may be determined based on well known coloring methods.

At block 308, a diagonal matrix is determined. The diagonal matrix has non-zero values only on the diagonal of the matrix. In one embodiment, the coloring of the matrix and the matrix (e.g., matrix A) are used to determine a diagonal matrix (e.g., matrix E). Embodiments of the present invention are operable to determine a diagonal matrix (e.g., matrix E) in parallel on a per color basis. The diagonal matrix may be determined via a parallel hardware architecture (e.g., GPU).

FIG. 4 shows a block diagram of a colored matrix in accordance with one embodiment of the present invention. FIG. 4 depicts a matrix (e.g., matrix A) of the coefficients of a linear system with rows and columns reordered based on coloring the unknowns and thereby for parallelism. Matrix 400 includes sparse matrices 402 (e.g., G) and 404 (e.g., F) and diagonal matrices 406 (e.g., D₁-D₆).

In one embodiment, the amount of parallelism is now of order m/p, where m is the number of rows in A, and p is the number of colors. For example, for a 4×4 matrix, where unknowns 1 and 2 are colored red and unknowns 3 and 4 are colored black the available parallelism is twofold in each color. An exemplary setup phase of the DILU preconditioner would be:E₁₁=A₁₁ E₂₂=A₂₂ E₃₃=A₃₃−L₃₁E₁₁⁻¹U₁₃−L₃₂E₂₂⁻¹U₂₃ E₄₄=A₄₄−L₄₁E₁₁⁻¹U₁₄−L₄₂E₂₂⁻¹U₂₄

Elements E_1,1 and E_2,2 can be computed in parallel in the first step and elements E_3,3 and E_2,2 can be processed in parallel in the second step. It is appreciated that this can be generalized for multiple colors and multiple unknowns per color. In one embodiment, the matrix is not reordered corresponding to colors, but the above parallel processing still applies according to the implicit renumbering (e.g., the processing order of the matrix performed based on the color and thus with increased parallelism).

In one embodiment, the setup phase is performed by the pseudo code of Table I. The pseudo-algorithms listed herein describe how the “construction” of a preconditioner that can be implemented efficiently on massively parallel devices. It is noted that most research on parallel preconditioners focuses on their application, not their construction. It is appreciated that the pseudo code of Table I is operable for use with a matrix A having entries that are scalar entries or small block entries and is not intended to be limited as such.

TABLE I
Exemplary Setup Phase Pseudo Code
setupMulticolorDILU(A, Einv, sorted_rows_by_color, row_colors, n,
num_colors) {
/* Inputs: Matrix A
           Array sorted_rows_by_color (sorted list of
           rows by color)
           Array row_colors (list storing the color of each
           row)
           Integer n (size of A)
           Integer num_colors (number of colors used)
Output:    Diagonal matrix Einv (stores the inverse of
           the diagonal matrix E) */
for (c=0; c<num_colors; c++) {
           for (every row i of matrix A with color c) do in parallel {
           Einv_temp = 0;
           for (int j=0;j<n;j++) {
           if (j==i)
           Einv_temp += A[i][i];
           else if (A[i][j] ≠ 0 && row_color[j] < c)
           Einv_temp −= A[i][j]*Einv[j][j]*A[j][i]
           }
           Einv_temp = inverse(Einv_temp);
           Einv[i] = Einv_temp;
           }
}
}

FIG. 5 shows a flowchart of an exemplary electronic component controlled process for updating a solution based on the preconditioner in accordance with one embodiment of the present invention. In one embodiment, flowchart 500 depicts a process for determining an updated solution using a multi-color DILU preconditioner (e.g., a solving portion) based on a diagonal matrix (e.g., matrix E based on A, color information corresponding to A, and vector b for a linear system of equations Ax=b).

At block 502, a preconditioning matrix is determined. In one embodiment, the preconditioning matrix (e.g., matrix M) is determined based on a diagonal matrix (e.g., matrix E) and the matrix of coefficients of systems of equations (e.g., matrix A). The preconditioning matrix may be an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner.

At block 504, an initial solution is accessed. The initial solution may be an estimate or guess solution (e.g., x), predetermined value, or fixed value (e.g., zero) for the system of equations. The initial solution may depend on the iterative method being used.

At block 506, a vector is accessed (e.g., vector b). In one embodiment, the vector is a vector from the linear systems of equations (e.g., vector b of system Ax=b).

At block 508, an updated solution is determined. In one embodiment, the updated solution is a new approximate solution that is closer to the exact solution. The updated solution may be determined based on an initial solution, a vector (e.g., the right hand side vector b) a diagonal matrix, a matrix of the coefficients of the system of equations, and color information corresponding to the matrix of the coefficients of the system of equations. The updated solution may be determined based on parallel computation of unknowns of the same color for each color. For example, for a system of equations with a million unknowns and 3 colors, on average ⅓ of the unknowns may be solved for in parallel for each color.

The updated solution may be determined based on an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner. The updated solution may be a portion of an algebraic multigrid iterative method to solve a system of equations. In one embodiment, the updated solution is determined based on the equation:x_k+1=x_k+M(E)⁻¹(b−Ax_k)

Where x_k+1 is an updated solution (if k is the final iteration then x_k+1 is the final solution vector from this iterative solver), x_k is an initial solution (e.g., x₀) or previously updated solution (e.g., an intermediate solution vector, for iterations k=1, 2, . . . ), M is a preconditioning matrix, E is a diagonal matrix, A is a matrix of coefficients of the system Ax=b. In one embodiment, the goal of the preconditioner is to obtain an updated solution x_k+1 which is closer to the exact solution to x=A⁻¹b, where A⁻¹ is the inverse of A.

Embodiments of the present invention are operable, during the smoothing or solving phase, to utilize the coloring information (e.g., corresponding to the matrix of coefficients of the system of equation) to compute an updated solution in parallel on a per color basis. The updated solution may be determined via a parallel hardware architecture (e.g., a GPU). For example, the updated solution may be computed in parallel for each respective color of the coloring information. In one exemplary embodiment, each thread on a GPU is associated with respective row and the GPU can compute each row in parallel by executing each respective thread in parallel.

In one embodiment, the updated solution is determined based on a lower and upper triangular solving phase (e.g., of the smoothing phase). The lower triangular solving phase may be performed by the pseudo code of Table II.

TABLE II
Exemplary Lower Triangular Solving Pseudo Code
LowerSolveMulticolorDILU(
A, Einv, sorted_rows_by_color, row_colors, n, num_colors, x, b, y
) {
/* Inputs: Matrix A
           Diagonal matrix Einv (stores the inverse of the
           diagonal matrix E)
           Array x
           Array b
           Array row_colors (list storing the color of each
           row)
           Integer n (size of A)
           Integer num_colors (number of colors used)
Output:    Array y */
for (c=0; c<num_colors; c++) {
           for (every row i of matrix A with color c) do in parallel {
           temp = b[i];
           for (int j=0;j<n,j++) {
           if (A[i][j] ≠ 0) {
           temp −= A[i][j]*x[j];
           if (row_color[j] < c)
           temp −= A[i][j]*y[j];
           }
           }
           y[i] = Einv[i]*temp;
           }
}
}

The upper triangular solving phase may be performed by the pseudo code of Table III.

TABLE III
Exemplary Upper Triangular Solving Pseudo Code
UpperSolveMulticolorDILU(
A, Einv, sorted_rows_by_color, row_colors, n, num_colors, y, z
) {
/* Inputs: Matrix A
           Diagonal matrix Einv (stores the inverse of the
           diagonal matrix E)
           Array y
           Array row_colors (list storing the color of each
           row)
           Integer n (size of A)
           Integer num_colors (number of colors used)
Output:    Array z */
for (c=num_colors−1; c>0; c−−) {
           for (every row i of matrix A with color c) do in parallel {
           temp = 0.;
           for (int j=0;j<n;j++) {
           if (A[i][j] ≠ 0 && row_color[j] > c) {
           temp += A[i][j]*z[j];
           }
           }
           z[i] = y[i] − Einv[i]*temp;
           }
}
}

In another embodiment, the lower and upper triangular solve algorithms can be also expressed with a common sparse matrix vector multiplication function SpMV (scalar a, matrix A, vector x, scalar b, vector y) which performs the operation: y=a*Ax+b*y. The matrix A may be represented by color blocks A=(A_c,d)_c,d, with c,dε{0, . . . , num_colors−1}, where A_c,d contains all matrix elements A[i][j] for which row i is of color c and column j of color d. In the corresponding vector notation, v_c denotes all vector elements v[i] for which i is of color c. It is noted that with this notation the solve algorithms can be expressed in a more abstract fashion.

The lower triangular solving phase may be performed by the pseudo code of Table IV.

TABLE IV
Exemplary Lower Triangular Solving Pseudo Code
LowerSolveMulticolorDILU(A,Einv, n, num_colors, x, b, y) {
/* Inputs: Matrix A
           Diagonal matrix Einv (stores the inverse of the diagonal
           matrix E)
           Array y
           Array b
           Integer n (size of A)
           Integer num_colors (number of colors used)
Output:    Array y */
Auxiliary vector v = b;
SpMV(−1, A, x, 1, v);
for (c=0; c<num_colors; c++) {
           if(c>0) SpMV(−1, [Ac,0, ..., Ac,c−1], y, 0, vc);
           SpMV( 1, Einvc, vc, 0, yc);
}

The upper triangular solving phase may be performed by the pseudo code of Table V.

TABLE V
Exemplary Upper Triangular Solving Pseudo Code
UpperSolveMulticolorDILU(A,Einv, n, num_colors, y, z) {
/* Inputs: Matrix A
           Diagonal matrix Einv (stores the inverse of the diagonal
           matrix E)
           Array y
           Integer n (size of A)
           Integer num_colors (number of colors used)
Output:    Array z */
Auxiliary vector v;
z = y;
for (c=num_colors−2; c>0; c−−) {
           SpMV(1, [Ac,c+1, ..., Ac,num_colors−1], z, 0, vc);
           SpMV(−1, Einvc, vc, 1, zc);
}

These formulations advantageously allow the use of highly optimized sparse matrix vector multiplication functions (SpMV) for certain matrix sizes, types or hardware architectures. In one embodiment, the SpMV functions can be formulated such that the use of the sub-matrices [A_c,0, . . . , A_c,c−1] and [A_c,c+1, . . . , A_{c,num_colors−1}] does not require their explicit construction, although typical SpMV implementations do not allow operations on a sub-matrix with an arbitrary selection of rows and columns.

In one embodiment, a commonly available four-vector SpMV for CSR (Compressed Sparse Row) and CSC (Compressed Sparse Column) formats could be used to implement the above algorithms. The rows and columns of Ā are sorted by ascending color to obtain a new representation of the matrix denoted by Ā (e.g., stored according to the implicit renumbering based on colors). If matrix Ā is stored as CSR (or CSC) with a values array, a row offset array (or column offset array for CSC) and a column index array (or row index array for CSC), then adding a row end array (or column end array for CSC) allows performance of the SpMV operations on sub-matrices [Ā_c,0, . . . , Ā_c,c−1] and [Ā_c,c+1, . . . , Ā_{c,num_colors−1}] as four-vector SpMVs (e.g., using the four mentioned arrays) without the need to construct these sub-matrices explicitly.

Block 508 may be performed multiple times, each time based on the previously updated solution determined at each execution of block 508. It is appreciated that the diagonal matrix (e.g., matrix E) may be determined once (e.g., via parallel computation) and used to determine multiple updated solutions. In one embodiment, the diagonal matrix and/or preconditioning matrix can be reused for solving a plurality of systems of equations where matrix of coefficients of the system of equations (e.g., matrix A) is the same while other vectors are different (e.g., for the a plurality of systems represented by Ax₁=b₁ and Ax₂=b₂, the A matrix is the same while the x and b vectors vary). As illustrated by flowchart 200, the updated solution may then be passed on to another portion of an iterative solver (e.g., another preconditioner or smoother). Block 508 may be performed multiple times until there is convergence to a solution or may be performed multiple times and used in conjunction with other preconditioners or smoothers as part of an iterative method.

In one embodiment, the preconditioning is part of an algebraic multigrid iterative solver and the preconditioning performed by embodiments of the present invention is operable both as a smoother and coarse grid solver.

FIG. 6 illustrates example components used by various embodiments of the present invention. Although specific components are disclosed in system 600, it should be appreciated that such components are examples. That is, embodiments of the present invention are well suited to having various other components or variations of the components recited in system 600. It is appreciated that the components in system 600 may operate with other components than those presented, and that not all of the components of system 600 may be required to achieve the goals of system 600.

FIG. 6 shows a block diagram of exemplary components in accordance with one embodiment of the present invention. System 600 depicts components of a preconditioning module of an iterative solver for solving systems of equations. Preconditioner module 602 includes setup module 604 and solving module 620. System 600 may be used with or as part of a variety of iterative solving methods including, but not limited to, Krylov subspace methods (e.g., Conjugate Gradient (CG), Bi-Conjugate Gradient (BiCG), Generalized Minimal Residual Method (GMRES)) and algebraic multigrid (AMG) methods.

Setup module 604 includes matrix access module 606, coloring access module 608, coloring determination module 610, and diagonal matrix determination module 612. In one embodiment, matrix access module 606 operable to access a matrix corresponding to a system of equations. Coloring access module 608 is operable to access color information corresponding to a matrix (e.g., matrix A). In one embodiment, the matrix (e.g., matrix A) corresponds to a plurality of coefficients of the system of equations. As described herein, the coloring information may include groups of rows or columns that can be used for parallel computations (e.g., the rows or columns are independent and can be computed independently). Optional coloring determination module 610 is operable to determine a coloring of a matrix (e.g., matrix A which corresponds to the coefficients of a system of equations) which can be used to determine a diagonal matrix in parallel on a per color basis.

Diagonal matrix determination module 612 is operable for determining a diagonal matrix based on the color information corresponding to a matrix (e.g., matrix A) in parallel on a per color basis. In one embodiment, the diagonal matrix is operable for computing an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner matrix. The diagonal matrix may be determined by a graphical processing unit (GPU).

Solving Module 620 includes matrix access module 622, initial solution access module 624, preconditioner matrix module 626, and updated solution determination module 630. Initial solution access module 624 is operable for accessing an initial solution to the system of equations. Matrix access module 622 is operable to access matrices to be used in solving a system of equations (e.g., matrices b and x for the system Ax=b). Initial solution access module 624 is operable to access an initial solution, as described herein. Preconditioner matrix module 626 is operable to determine a preconditioning matrix (e.g., matrix M), as described herein.

Updated solution determination module 630 is operable for determining an updated solution based on the matrix (e.g., matrix A or matrix of coefficients), an initial solution (e.g., vector x₀), color information, and a preconditioning matrix (e.g., matrix M). The updated solution may be determined with use of a parallel hardware architecture (e.g., GPU). In one exemplary embodiment, the determining of the updated solution is determined in parallel on a per color basis.

In one embodiment, updated solution determination module 630 includes upper solve module 632 and lower solve module 634. Upper solve module 632 is operable to determine a portion of the updated solution based on an upper triangular solving phase (e.g., UpperSolveMulticolorDILU( )). Lower solve module 634 is operable to determine a portion of the updated solution based on a lower triangular solving phase (e.g., LowerSolveMulticolorDILU( ).

The foregoing descriptions of specific embodiments of the present invention have been presented for purposes of illustration and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed, and many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, to thereby enable others skilled in the art to best utilize the invention and various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the claims appended hereto and their equivalents.

Claims

1. A non-transitory computer-readable storage medium embodying instructions that, when executed by a processing device, cause the processing device to perform a method for preconditioning, said method comprising: accessing a matrix comprising a plurality of coefficients of a system of linear equations;accessing coloring information corresponding to said matrix; anddetermining a diagonal matrix of a preconditioning matrix of said matrix based on said matrix and said coloring information corresponding to said matrix, wherein said determining comprises: assigning a plurality of execution threads to a plurality of elements of a respective color in said diagonal matrix; andcomputing said plurality of elements of said respective color by executing said plurality of execution threads in said processing device.

2. The non-transitory computer-readable storage medium as described in claim 1, wherein the method further comprises: determining said preconditioning matrix based on said diagonal matrix.

3. The non-transitory computer-readable storage medium as described in claim 2 wherein said preconditioning matrix is an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner.

4. The non-transitory computer-readable storage medium as described in claim 1 wherein the method further comprises: accessing an initial solution.

5. The non-transitory computer-readable storage medium as described in claim 4 wherein the method further comprises: determining an updated solution to said system of equations, wherein said updated solution is determined in parallel on a per color basis.

6. The non-transitory computer-readable storage medium as described in claim herein said processing device comprises a graphics processing unit (GPU).

7. A system for preconditioning, said system comprising: a processor; andmemory coupled to said processor and comprising instructions that comprise:a matrix access module operable to access a matrix corresponding to a system of equations;a coloring access module operable to access color information corresponding to said matrix; anda diagonal matrix determination module operable for determining a diagonal matrix based on said color information corresponding to said matrix, wherein said determining comprises:assigning a plurality of execution threads to a plurality of elements of a respective color in said diagonal matrix; andcomputing said plurality of elements of said respective color by executing said plurality of execution threads in parallel in said processing device, wherein said diagonal matrix determination module is operable for computing an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner matrix.

8. The system as described in claim 7 wherein said diagonal matrix is determined by a graphical processing unit (GPU).

9. The system as described in claim 7 wherein said matrix corresponds to a plurality of coefficients of said system of equations.

10. The system as described in claim 7, wherein said instructions further comprise: a solution access module operable for accessing an initial solution to said system of equations.

11. The system as described in claim 10, wherein said instructions further comprise: an updated solution determination module operable for determining an updated solution based on said matrix, said initial solution, and said color information, wherein said determining said updated solution is determined in parallel based on a per color basis.

12. The system as described in claim 11 wherein said updated solution module is operable to determine said updated solution based on a lower triangular solving phase and an upper triangular solving phase.

13. The system as described in claim 11 wherein said determining said updated solution is determined by a GPU.

14. A non-transitory computer-readable storage medium embodying instructions that, when executed by a processing device, cause the processing device to perform a method for solving a system of equations, said method comprising: accessing a matrix comprising coefficients of a system of equations;accessing coloring information corresponding to said matrix, wherein each row of said matrix is assigned with a particular color;determining a diagonal matrix of a preconditioning matrix of said matrix, wherein said determining said diagonal matrix comprises: assigning a plurality of execution threads to a plurality of elements of a respective color in said diagonal matrix; andcomputing said plurality of elements of said respective color by executing said plurality of execution threads in parallel in said processing device; anddetermining an updated solution to said system of equations, wherein said determining said updated solution is based on an incomplete lower and upper triangular decomposition with only diagonal modified (DILU) preconditioner performed by a graphics processing unit (GPU) and said updated solution is computed in parallel on a per color basis.

15. The non-transitory computer-readable storage medium as described in claim 14 wherein said determining of said diagonal matrix is determined by said GPU.

16. The non-transitory computer-readable storage medium as described in claim 14 wherein said determining said updated solution is based on an initial solution, said matrix, and said diagonal matrix.

17. The non-transitory computer-readable storage medium as described in claim 14 wherein said updated solution is determined based on a lower triangular solving phase and an upper triangular solving phase.

18. The non-transitory computer-readable storage medium as described in claim 14 wherein said updated solution is determined based on a common sparse matrix vector multiplication function.

19. The non-transitory computer-readable storage medium as described in claim 14 wherein said determining said updated solution is a portion of an algebraic multigrid iterative process.