LinearAlgebra¶

Usage

use LinearAlgebra;

Submodules

Sparse

A high-level interface to linear algebra operations and procedures.

Compiling with Linear Algebra¶

Programs using the LinearAlgebra module can be built with no additional dependencies if they do not use any procedures that rely on BLAS or LAPACK. Procedure dependencies are specified in procedure documentation below.

If a program calls a procedure that depends on BLAS or LAPACK, the headers and library will need to be available during compilation, typically through compiler flags and/or environment variables.

Some procedures have implementations both with and without dependencies. By default, the implementation with dependencies will be selected. Users can explicitly opt out of using the BLAS and LAPACK dependent implementations by setting the blasImpl and lapackImpl flags to none.

Building programs with no dependencies

// example1.chpl
var A = Matrix([0.0, 1.0, 1.0],
               [1.0, 0.0, 1.0],
               [1.0, 1.0, 0.0]);
var I = eye(3,3);
var B = A + I;

The program above has no dependencies and can therefore be compiled without the BLAS or LAPACK headers and libraries available:

chpl example1.chpl

If this program had used a procedure with a dependency such as cholesky (depends on LAPACK), compilation without LAPACK headers and libraries available would result in a compilation error.

Building programs with dependencies

// example2.chpl
var A = Matrix([2.0, 1.0],
               [1.0, 2.0]);
var (eigenvalues, eigenvectors) = eig(A, right=true);

The program above uses eig, which depends on LAPACK. Compilation without LAPACK headers and libraries available would result in a compilation error.

Following the instructions from the LAPACK module documentation, the above program could be compiled if LAPACK is available on the system and specified with the following compilation flags:

# Building with LAPACK dependency
chpl -I$PATH_TO_LAPACKE_INCLUDE_DIR \
     -L$PATH_TO_LIBGFORTRAN -lgfortran \
     -L$PATH_TO_LAPACK_BINARIES -llapacke -llapack -lrefblas \
     example2.chpl

Building programs with optional dependencies

// example3.chpl
var A = Matrix(3,5);
A = 2;
var AA = A.dot(A.T);

The program above uses dot, which has two available implementations: one that depends on BLAS and one that is written in Chapel. The program will default to using the more performant BLAS implementation of matrix-matrix multiplication.

Following the instructions from the BLAS module documentation, the above program could be compiled if BLAS is available on the system and specified with the following compilation flags:

# Building with BLAS dependency
chpl -I$PATH_TO_CBLAS_DIR \
     -L$PATH_TO_BLAS_LIBS -lblas \
     example3.chpl

Note

Users can set environment variables like LDFLAGS for -L arguments and CFLAGS for -I arguments, to avoid throwing these flags every time.

Additionally, the required linker flags (-l) may vary depending on the BLAS and LAPACK implementations being used.

To opt out of using the BLAS implementation, users can add the --set blasImpl=none flag, so that BLAS is no longer a dependency:

# Building with BLAS dependency explicitly disabled
chpl --set blasImpl=none example3.chpl

Similarly, users can opt out of of LAPACK implementations with the --set lapackImpl=none flag. Setting both flags to none will always choose the Chapel implementation when available, and will emit a compiler error when no native implementation is available:

# Building with all dependencies explicitly disabled
chpl --set lapackImpl=none --set blasImpl=none example3.chpl

See the documentation of BLAS or LAPACK for more details on these flags.

Linear Algebra Interface¶

Matrix and Vector representation

Matrices and vectors are represented as Chapel arrays, which means the convenience constructors provided in this module are not necessary to use this module's functions. This also means that matrices (2D arrays) and vectors (1D arrays) will work with any other Chapel library that works with arrays.

Note

This documentation uses the terms matrix to refer to 2D arrays, and vector to refer to 1D arrays.

Domain offsets

All functions that return arrays will inherit their domains from the input array if possible. Otherwise they will return arrays with 1-based indices.

Row vs Column vectors

Row and column vectors are both represented as 1D arrays and are indistinguishable in Chapel. In the dot function, matrix-vector multiplication assumes a column vector, vector-matrix multiplication assumes a row vector, vector-vector multiplication is always treated as an inner-product, as the function name implies. An outer product can be computed with the outer function.

Domain maps

All of the functions in this module only support DefaultRectangular arrays (the default domain map), unless explicitly specified in the function's documentation. Other domain maps are supported through submodules, such LinearAlgebra.Sparse for the CS layout.

class LinearAlgebraError: Error¶

Base Error type for LinearAlgebra errors.

var info: string¶: Stores message to be emitted upon uncaught throw

proc Vector(length, type eltType = real)¶: Return a vector (1D array) over domain {1..length}

proc Vector(space: range, type eltType = real): Return a vector (1D array) over domain {space}

proc Vector(Dom: domain(1), type eltType = real): Return a vector (1D array) over domain Dom

proc Vector(A: [?Dom] ?Atype, type eltType = Atype): Return a vector (1D array) with domain and values of A

proc Vector(x: ?t, Scalars ...?n, type eltType)

Return a vector (1D array), given 2 or more numeric values

If type is omitted, it will be inferred from the first argument

proc Matrix(rows, type eltType = real)¶: Return a square matrix (2D array) over domain {1..rows, 1..rows}

proc Matrix(rows, cols, type eltType = real): Return a matrix (2D array) over domain {1..rows, 1..cols}

proc Matrix(space: range, type eltType = real): Return a square matrix (2D array) over domain {space, space}

proc Matrix(rowSpace: range, colSpace: range, type eltType = real): Return a matrix (2D array) over domain {rowSpace, colSpace}

proc Matrix(Dom: domain, type eltType = real): Return a matrix (2D array) over domain Dom

proc Matrix(A: [?Dom] ?Atype, type eltType = Atype)

Return a matrix (2D array) with domain and values of A.

A can be sparse (CS) or dense.

proc Matrix(const Arrays: ?t ...?n, type eltType)

Return a matrix (2D array), given 2 or more vectors, such that the vectors form the rows of the matrix. In other words, the vectors are concatenated such that the ith vector corresponds to the matrix slice: A[i, ..]

If type is omitted, it will be inferred from the first array.

For example:

var A = Matrix([1, 2, 3],
               [4, 5, 6],
               [7, 8, 9]);
/* Produces the 3x3 matrix of integers:
     1 2 3
     4 5 6
     7 8 9
 */

proc eye(m, type eltType = real)¶: Return a square identity matrix over domain {1..m, 1..m}

proc eye(m, n, type eltType = real): Return an identity matrix over domain {1..m, 1..n}

proc eye(Dom: domain(2), type eltType = real): Return an identity matrix over domain Dom

proc transpose(A: [?Dom] ?eltType)¶: Transpose vector, matrix, or domain.

Note

Since row vectors and columns vectors are indistinguishable, passing a vector to this function will return that vector unchanged

proc _array.T¶: Transpose vector or matrix

proc _array.plus(A: [?Adom] ?eltType)¶: Element-wise addition. Same as A + B.

proc _array.minus(A: [?Adom] ?eltType)¶: Element-wise subtraction. Same as A - B.

proc _array.times(A: [?Adom])¶: Element-wise multiplication. Same as A * B.

proc _array.elementDiv(A: [?Adom])¶: Element-wise division. Same as A / B.

proc dot(A: [?Adom] ?eltType, B: [?Bdom] eltType)¶: Generic matrix multiplication, A and B can be a matrix, vector, or scalar.

Note

When A is a vector and B is a matrix, this function implicitly computes dot(transpose(A), B), which may not be as efficient as passing A and B in the reverse order.

Note

Dense matrix-matrix and matrix-vector multiplication will utilize the BLAS module for improved performance, if available. Compile with --set blasImpl=none to opt out of the BLAS implementation.

proc _array.dot(A: [])¶: Compute the dot-product

Note

Dense matrix-matrix and matrix-vector multiplication will utilize the BLAS module for improved performance, if available. Compile with --set blasImpl=none to opt out of the BLAS implementation.

proc inner(A: [?Adom], B: [?Bdom])¶: Inner product of 2 vectors.

proc outer(A: [?Adom] ?eltType, B: [?Bdom] eltType)¶: Outer product of 2 vectors.

proc matPow(A: [], b)¶: Return the matrix A to the bth power, where b is a positive integral type.

Note

matPow will utilize the BLAS module for improved performance, if available. Compile with --set blasImpl=none to opt out of the BLAS implementation.

proc cross(A: [?Adom] ?eltType, B: [?Bdom] eltType)¶: Return cross-product of 3-element vectors A and B with domain of A.

proc diag(A: [?Adom] ?eltType, k = 0)¶: Return a Vector containing the diagonal elements of A if the argument A is of rank 2. Return a diagonal Matrix whose diagonal contains elements of A if argument A is of rank 1.

proc tril(A: [?D] ?eltType, k = 0)¶

Return lower triangular part of matrix, above the diagonal + k, where k = 0 includes the diagonal, and k = -1 does not include the diagonal. For example:

var A = Matrix(4, 4, eltType=int);
A = 1;

tril(A);
/* Returns:

    1    0    0    0
    1    1    0    0
    1    1    1    0
    1    1    1    1
*/

tril(A, 1);
/* Returns:

    1    1    0    0
    1    1    1    0
    1    1    1    1
    1    1    1    1
*/

tril(A, -1);
/* Returns:

    0    0    0    0
    1    0    0    0
    1    1    0    0
    1    1    1    0
*/

proc triu(A: [?D] ?eltType, k = 0)¶

Return upper triangular part of matrix, above the diagonal + k, where k = 0 includes the diagonal, and k = 1 does not include the diagonal. For example:

var A = Matrix(4, 4, eltType=int);
A = 1;

triu(A);
/* Returns:

    1    1    1    1
    0    1    1    1
    0    0    1    1
    0    0    0    1
*/

triu(A, 1);
/* Returns:

    0    1    1    1
    0    0    1    1
    0    0    0    1
    0    0    0    0
*/

triu(A, -1);
/* Returns:

    1    1    1    1
    1    1    1    1
    0    1    1    1
    0    0    1    1
*/

proc isDiag(A: [?D] ?eltType)¶: Return true if matrix is diagonal

proc isHermitian(A: [?D])¶: Return true if matrix is Hermitian

proc isSymmetric(A: [?D]): bool¶: Return true if matrix is symmetric

proc isTril(A: [?D] ?eltType, k = 0): bool¶: Return true if matrix is lower triangular below the diagonal + k, where k = 0 does not include the diagonal, and k = 1 includes the diagonal

proc isTriu(A: [?D] ?eltType, k = 0): bool¶: Return true if matrix is upper triangular above the diagonal + k, where k = 0 does not include the diagonal, and k = -1 includes the diagonal

proc isSquare(A: [?D])¶: Return true if matrix is square

proc trace(A: [?D] ?eltType)¶: Return the trace (sum of diagonal elements) of A

proc cholesky(A: [] ?t, lower = true)¶: Perform a Cholesky factorization on matrix A. A must be square. Argument lower indicates whether to return the lower or upper triangular factor. Matrix A is not modified. Returns an array with the same shape as argument A with the lower or upper triangular Cholesky factorization of A.

Note

This procedure depends on the LAPACK module, and will generate a compiler error if lapackImpl is none.

proc eigvals(A: [] ?t)¶: Find the eigenvalues of matrix A. A must be square.

Note

This procedure depends on the LAPACK module, and will generate a compiler error if lapackImpl is none.

proc eig(A: [] ?t, param left = false, param right = false)¶

Find the eigenvalues and eigenvectors of matrix A. A must be square.

If left is true then the "left" eigenvectors are computed. The return value is a tuple with two elements: (eigenvalues, leftEigenvectors)
If right is true then the "right" eigenvectors are computed. The return value is a tuple with two elements: (eigenvalues, rightEigenvectors)
If left and right are both true then both eigenvectors are computed. The return value is a tuple with three elements: (eigenvalues, leftEigenvectors, rightEigenvectors)
If left and right are both false only the eigenvalues are computed, and returned as a single array.

Note

This procedure depends on the LAPACK module, and will generate a compiler error if lapackImpl is none.

proc svd(A: [?Adom] ?t) throws¶

Singular Value Decomposition.

Factorizes the m x n matrix A such that:

\[\mathbf{A} = \textbf{U} \cdot \Sigma \cdot \mathbf{V^H}\]

where

\(\mathbf{U}\) is an m x m unitary matrix,

\(\Sigma\) is a diagonal m x n matrix,

\(\mathbf{V}\) is an n x n unitary matrix, and \(\mathbf{V^H}\) is the Hermitian transpose.

This procedure returns a tuple of (U, s, Vh), where s is a vector containing the diagonal elements of \(\Sigma\), known as the singular values.

For example:

var A = Matrix([3, 2,  2],
               [2, 3, -2],
               eltType=real);
var (U, s, Vh) = svd(A);

LinearAlgebraError will be thrown if the SVD computation does not converge or an illegal argument, such as a matrix containing a NAN value, is given.

Note

A temporary copy of A will be created within this computation.

Note

This procedure depends on the LAPACK module, and will generate a compiler error if lapackImpl is none.

proc kron(A: [?ADom] ?eltType, B: [?BDom] eltType)¶: Return the Kronecker Product of matrix A and matrix B. If the size of A is x * y and of B is a * b then size of the resulting matrix will be (x * a) * (y * b)