SugarBLAS.jl

Syntactic sugar for BLAS polynomials
Author lopezm94
Sub Category Linear Algebra
Category Mathematics
Github
Popularity
30 Stars
Updated Last
1 Year Ago
Started In
August 2016

SugarBLAS

codecov Build Status

BLAS functions are unaesthetic and annoying without good knowledge of the positional arguments. This package provides macros for BLAS functions representing polynomials. The main macro of the package is @blas! for most of the use cases: copy!, scale! and axpy!. Non mutable versions of this operator are already very easy to write so they are not included.

The macros will output a function from BASE module, this allows defining new behavior for custom types. Note that the output won't necessarily belong to the julia BLAS API, e.g. copy! is used instead of BASE.LinAlg.BLAS.blascopy! for better performance.

For now the package supports the most common BLAS functions from the internal API. The access for these functions is private since the official API is private aswell and may change in the future.

This documentation offers great examples but is by no means super extensive, for more examples check the test folder of the repository.

Installing

To install the package, use the following command inside Julia's REPL:

Pkg.add("SugarBLAS")

Usage

@blas! matches the expression and decides which function to call. As long as it is correctly parenthesized putting more variables won't be an issue.

julia> macroexpand(SugarBLAS, :(@blas! Y = (a*b +c)*(X*Z) + Y))
:(Base.LinAlg.axpy!(a * b + c,X * Z,Y))

julia> macroexpand(SugarBLAS, :(@blas! X = (a+c)*X))
:(scale!(a + c,X))

When doing this just imagine the BLAS expression.

Y = a*X + Y
->
a := (a*b +c); X := (X*Z)
->
Y = (a*b +c)*(X*Z) + Y

Updating operators

Both *= and += are supported. *= can only be used for scaling given that is pretty unambigous.

julia> macroexpand(SugarBLAS, :(@blas! Y += X)) == macroexpand(SugarBLAS, :(@blas! Y = Y + X))
true

Commutativity

+ is assumed as the only commutative operator, it is important to note here that * is not treated as commutative and therefore some expressions will lead to errors.

julia> a = 2.3;

julia> X = rand(10,10);

julia> Y = rand(10,10);

julia> @blas! Y += X*a
ERROR: MethodError: `axpy!` has no method matching axpy!(::Array{Float64,2}, ::Float64, ::Array{Float64,2})

The package assumes types by its position in the multiplication, this doesn't happen with addition and that's why it conserves its property.

julia> macroexpand(SugarBLAS, :(@blas! Y = X + Y)) == macroexpand(SugarBLAS, :(@blas! Y = Y + X))
true

Macros

blas!

Internal API

blas!

Macro for most of the use cases: copy!, scale! and axpy!

scale!

Scale an array X by a scalar a overwriting X in-place.

Polynomials

  • X *= a
  • X = a*X

Example

julia> macroexpand(SugarBLAS, :(@blas! X *= a))
:(scale!(a,X))

axpy!

Overwrite Y with a*X + Y. Return Y.

Polynomials

  • Y += X
  • Y += a*X

Example

julia> macroexpand(SugarBLAS, :(@blas! Y += X))
:(Base.LinAlg.axpy!(1.0,X,Y))

julia> macroexpand(SugarBLAS, :(@blas! Y += a*X))
:(Base.LinAlg.axpy!(a,X,Y))

copy!

Copy all elements from collection Y to array X. Return X.

Polynomials

  • X = Y

Example

julia> macroexpand(SugarBLAS, :(@blas! X = Y))
:(copy!(X,Y))

Internal API

Macro for most of the functions available in the JuliaLang internal BLAS API.

scale!

Scale an array X by a scalar a overwriting X in-place.

Polynomials

  • X *= a
  • X = a*X

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@scale! X *= a))
:(scale!(a,X))

axpy!

Overwrite Y with a*X + Y. Return Y.

Polynomials

  • Y ±= X
  • Y ±= a*X

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@axpy! Y += X))
:(Base.LinAlg.axpy!(1.0,X,Y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@axpy! Y += a*X))
:(Base.LinAlg.axpy!(a,X,Y))

copy!

Copy all elements from collection Y to array X. Return X.

Polynomials

  • X = Y

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@copy! X = Y))
:(copy!(X,Y))

ger!

Rank-1 update of the matrix A with vectors x and y as alpha*x*y' + A.

Polynomials

  • A ±= alpha*x*y'

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@ger! A -= alpha*x*y'))
:(Base.LinAlg.BLAS.ger!(-alpha,x,y,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@ger! A += alpha*x*y'))
:(Base.LinAlg.BLAS.ger!(alpha,x,y,A))

syr!

Rank-1 update of the symmetric matrix A with vector x as alpha*x*(x)ᵀ + A. When left side has A['U'] the upper triangle of A is updated ('L' for lower triangle). Return A.

Polynomials

  • A[uplo] ±= alpha*x*(x)ᵀ

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syr! A['U'] -= alpha*x*(x)ᵀ))
:(Base.LinAlg.BLAS.syr!('U',-alpha,x,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syr! A['L'] += alpha*x*(x)ᵀ))
:(Base.LinAlg.BLAS.syr!('L',alpha,x,A))

syrk

Return either the upper triangle or the lower triangle, depending on ('U' or 'L'), of alpha*A*(A)ᵀ or alpha*(A)ᵀ*A.

Polynomials

  • alpha*A*(A)ᵀ uplo=ul
  • alpha*(A)ᵀ*A uplo=ul

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syrk alpha*A*(A)ᵀ uplo='U'))
:(Base.LinAlg.BLAS.syrk('U','N',alpha,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syrk alpha*(A)ᵀ*A uplo='L'))
:(Base.LinAlg.BLAS.syrk('L','T',alpha,A))

syrk!

Rank-k update of the symmetric matrix C as alpha*A*(A)ᵀ + beta*C or alpha*(A)ᵀ*A + beta*C. When the left hand side isC['U'] the upper triangle of C is updated ('L' for lower triangle). Return C.

Polynomials

  • C[uplo] ±= alpha*A*(A)ᵀ
  • C[uplo] = beta*C ± alpha*(A)ᵀ*A

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syrk! C['U'] -= alpha*A*(A)ᵀ))
:(Base.LinAlg.BLAS.syrk!('U','N',-alpha,A,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syrk! C['L'] = beta*C - alpha*(A)ᵀ*A))
:(Base.LinAlg.BLAS.syrk!('L','T',-alpha,A,beta,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syrk! C['U'] += alpha*A*(A)ᵀ))
:(Base.LinAlg.BLAS.syrk!('U','N',alpha,A,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@syrk! C['L'] = alpha*(A)ᵀ*A + beta*C))
:(Base.LinAlg.BLAS.syrk!('L','T',alpha,A,beta,C))

her!

Methods for complex arrays only. Rank-1 update of the Hermitian matrix A with vector x as alpha*x*x' + A. Whenthe left hand side is A['U'] the upper triangle of A is updated ('L' for lower triangle). Return A.

Polynomials

  • A[uplo] ±= alpha*x*x'

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@her! A['U'] -= alpha*x*x'))
:(Base.LinAlg.BLAS.her!('U',-alpha,x,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@her! A['L'] = A - alpha*x*x'))
:(Base.LinAlg.BLAS.her!('L',-alpha,x,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@her! A['U'] += alpha*x*x'))
:(Base.LinAlg.BLAS.her!('U',alpha,x,A))

herk

Methods for complex arrays only. Returns either the upper triangle or the lower triangle, according to uplo ('U' or 'L'), of alphaAA' or alpha*A'*A, according to trans ('N' or 'T').

Polynomials

  • alpha*A*A' uplo=ul
  • alpha*A'*A uplo=ul

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk alpha*A*A' uplo='U'))
:(Base.LinAlg.BLAS.herk('U','N',alpha,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk alpha*A'*A uplo='U'))
:(Base.LinAlg.BLAS.herk('U','T',alpha,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk alpha*A*A' uplo='L'))
:(Base.LinAlg.BLAS.herk('L','N',alpha,A))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk alpha*A'*A uplo='L'))
:(Base.LinAlg.BLAS.herk('L','T',alpha,A))

herk!

Methods for complex arrays only. Rank-k update of the Hermitian matrix C as alpha*A*A' + beta*C or alpha*A'*A + beta*C. When the left hand side is C['U'] the upper triangle of C is updated ('L' for lower triangle). Return C.

Polynomials

  • C[uplo] ±= alpha*A*A'
  • C[uplo] = beta*C ± alpha*A'*A

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk! C['L'] -= alpha*A'*A))
:(Base.LinAlg.BLAS.herk!('L','T',-alpha,A,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk! C['U'] = C - alpha*A*A'))
:(Base.LinAlg.BLAS.herk!('U','N',-alpha,A,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk! C['L'] = beta*C - alpha*A'*A))
:(Base.LinAlg.BLAS.herk!('L','T',-alpha,A,beta,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk! C['U'] += alpha*A*A'))
:(Base.LinAlg.BLAS.herk!('U','N',alpha,A,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@herk! C['L'] = alpha*A'*A + beta*C))
:(Base.LinAlg.BLAS.herk!('L','T',alpha,A,beta,C))

gbmv

Return alpha*A*x or alpha*A'*x. The matrix A is a general band matrix of dimension m by size(A,2) with kl sub-diagonals and ku super-diagonals.

Polynomials

  • alpha*A[kl:ku,h=m]*x
  • alpha*A[h=m,kl:ku]'*x

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gbmv alpha*A[0:ku,h=2]*x))
:(Base.LinAlg.BLAS.gbmv('N',2,0,ku,alpha,A,x))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gbmv alpha*A[h=m,-kl:ku]*x))
:(Base.LinAlg.BLAS.gbmv('N',m,kl,ku,alpha,A,x))

gbmv!

Update vector y as alpha*A*x + beta*y or alpha*A'*x + beta*y. The matrix A is a general band matrix of dimension m by size(A,2) with kl sub-diagonals and ku super-diagonals. Return the updated y.

Polynomials

  • y ±= alpha*A[kl:ku,h=m]*x
  • y = beta*y ± alpha*A[h=m,kl:ku]'*x

Example

@test macroexpand(SugarBLAS, :(SugarBLAS.@gbmv! y -= alpha*A[h=m,-kl:ku]*x))
:(Base.LinAlg.BLAS.gbmv!('N',m,kl,ku,-alpha,A,x,1.0,y))

@test macroexpand(SugarBLAS, :(SugarBLAS.@gbmv! y = beta*y - alpha*A[h=2, 0:ku]'*x))
:(Base.LinAlg.BLAS.gbmv!('T',2,0,ku,-alpha,A,x,beta,y))

@test macroexpand(SugarBLAS, :(SugarBLAS.@gbmv! y = alpha*A[0:ku,h=2]*x + y))
:(Base.LinAlg.BLAS.gbmv!('N',2,0,ku,alpha,A,x,1.0,y))

@test macroexpand(SugarBLAS, :(SugarBLAS.@gbmv! y += alpha*A[h=m,-kl:ku]*x))
:(Base.LinAlg.BLAS.gbmv!('N',m,kl,ku,alpha,A,x,1.0,y))

@test macroexpand(SugarBLAS, :(SugarBLAS.@gbmv! y = alpha*A[kl:ku, h=m]'*x + beta*y))
:(Base.LinAlg.BLAS.gbmv!('T',m,-kl,ku,alpha,A,x,beta,y))

sbmv

Return alpha*A*x where A is a symmetric band matrix of order size(A,2) with k super-diagonals stored in the argument A.

Polynomials

  • A[0:k,uplo]*xv
  • alpha*A[0:k,uplo]*x

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv A['U',0:k]*x))
:(Base.LinAlg.BLAS.sbmv('U',k,A,x))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv alpha*A[0:k,'L']*x))
:(Base.LinAlg.BLAS.sbmv('L',k,alpha,A,x))

sbmv!

Update vector y as alpha*A*x + beta*y where A is a a symmetric band matrix of order size(A,2) with k super-diagonals stored in the argument A. If A[...,'U'] is used multiplication is done with A's upper triangle, L is for the lower triangle. Return updated y.

Polynomials

  • y ±= alpha*A[0:k,uplo]*x
  • y = beta*y ± alpha*A[0:k,uplo]*x

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv! y -= alpha*A['U',0:k]*x))
:(Base.LinAlg.BLAS.sbmv!('U',k,-alpha,A,x,1.0,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv! y = beta*y - alpha*A[0:k,'U']*x))
:(Base.LinAlg.BLAS.sbmv!('U',k,-alpha,A,x,beta,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv! y = beta*y - alpha*A[0:k,'L']*x))
:(Base.LinAlg.BLAS.sbmv!('L',k,-alpha,A,x,beta,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv! y += alpha*A[0:k,'L']*x))
:(Base.LinAlg.BLAS.sbmv!('L',k,alpha,A,x,1.0,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@sbmv! y = alpha*A['L',0:k]*x + beta*y))
:(Base.LinAlg.BLAS.sbmv!('L',k,alpha,A,x,beta,y))

gemm

Return alpha*A*B, alpha*A'*B, alpha*A*B' or alpha*A'*B'.

Polynomials

  • A*B
  • A'*B
  • A*B'
  • A'*B'
  • alpha*A*B
  • alpha*A'*B
  • alpha*A*B'
  • alpha*A'*B'

Example

@test macroexpand(SugarBLAS, :(SugarBLAS.@gemm alpha*A*B))
:(Base.LinAlg.BLAS.gemm('N','N',alpha,A,B))

@test macroexpand(SugarBLAS, :(SugarBLAS.@gemm A*B'))
:(Base.LinAlg.BLAS.gemm('N','T',A,B))

gemm!

Update C as alpha*A*B + beta*C or the other three variants according to the combination of transposes of A and B. Return updated C.

Polynomials

  • C ±= alpha*A*B
  • C ±= alpha*A'*B
  • C ±= alpha*A*B'
  • C ±= alpha*A'*B'
  • C = beta*C ± alpha*A*B
  • C = beta*C ± alpha*A'*B
  • C = beta*C ± alpha*A*B'
  • C = beta*C ± alpha*A'*B'

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemm! C -= alpha*A*B))
:(Base.LinAlg.BLAS.gemm!('N','N',-alpha,A,B,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemm! C = beta*C - alpha*A*B))
:(Base.LinAlg.BLAS.gemm!('N','N',-alpha,A,B,beta,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemm! C += alpha*A*B))
:(Base.LinAlg.BLAS.gemm!('N','N',alpha,A,B,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemm! C = 3.4*C - alpha*A'*B'))
:(Base.LinAlg.BLAS.gemm!('T','T',-alpha,A,B,3.4,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemm! C = alpha*A'*B + beta*C))
:(Base.LinAlg.BLAS.gemm!('T','N',alpha,A,B,beta,C))

gemv

Return alpha*A*x or alpha*A'*x.

Polynomials

  • A*x
  • A'*x
  • alpha*A*x
  • alpha*A'*x

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv A'*x))
:(Base.LinAlg.BLAS.gemv('T',A,x))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv alpha*A*x))
:(Base.LinAlg.BLAS.gemv('N',alpha,A,x))

gemv!

Update the vector y as alpha*A*x + beta*y or alpha*A'*x + beta*y. Return updated y.

Polynomials

  • y ±= alpha*A*x
  • y ±= alpha*A'*x
  • y = beta*y ± alpha*A*x
  • y = beta*y ± alpha*A'*x

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv! y -= alpha*A*x))
:(Base.LinAlg.BLAS.gemv!('N',-alpha,A,x,1.0,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv! y = beta*y - alpha*A*x))
:(Base.LinAlg.BLAS.gemv!('N',-alpha,A,x,beta,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv! y = beta*y - 1.5*A'*x))
:(Base.LinAlg.BLAS.gemv!('T',-1.5,A,x,beta,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv! y += alpha*A*x))
:(Base.LinAlg.BLAS.gemv!('N',alpha,A,x,1.0,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@gemv! y = alpha*A*x + beta*y))
:(Base.LinAlg.BLAS.gemv!('N',alpha,A,x,beta,y))

symm

Return alpha*A*B or alpha*B*A according to "symm". A is assumed to be symmetric. Only the uplo triangle of A is used ('L' for lower and 'U' for upper).

Polynomials

  • A["symm", uplo]*B
  • A*B["symm", uplo]
  • alpha*A["symm", uplo]*B
  • alpha*A*B["symm", uplo]

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symm alpha*A["symm", 'L']*B))
:(Base.LinAlg.BLAS.symm('L','L',alpha,A,B))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symm A*B["symm", 'U']))
:(Base.LinAlg.BLAS.symm('R','U',A,B))

symm!

Update C as alpha*A*B + beta*C or alpha*B*A + beta*C according to "symm". A is assumed to be symmetric. Only the uplo triangle of A is used ('L' for lower and 'U' for upper). Return updated C.

Polynomials

  • C = alpha*A["symm",uplo]*B
  • C = alpha*A*B["symm",uplo]
  • C = beta*C ± alpha*A["symm",uplo]*B
  • C = beta*C ± alpha*A*B["symm",uplo]

Example

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symm! C -= alpha*A["symm", 'L']*B))
:(Base.LinAlg.BLAS.symm!('L','L',-alpha,A,B,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symm! C = C - alpha*A["symm", 'U']*B))
:(Base.LinAlg.BLAS.symm!('L','U',-alpha,A,B,1.0,C))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symm! C = beta*C - alpha*A["symm", 'L']*B))
:(Base.LinAlg.BLAS.symm!('L','L',-alpha,A,B,beta,C))

symv

Return alpha*A*x. A is assumed to be symmetric. Only the uplo triangle of A is used ('L' for lower and 'U' for upper).

Polynomials

  • A[uplo]*x
  • alpha*A[uplo]*x
julia> macroexpand(SugarBLAS, :(SugarBLAS.@symv alpha*A['U']*x))
:(Base.LinAlg.BLAS.symv('U',alpha,A,x))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symv A['L']*x))
:(Base.LinAlg.BLAS.symv('L',A,x))

symv!

Update the vector y as alpha*A*x + beta*y. A is assumed to be symmetric. Only the uplo triangle of A is used ('L' for lower and 'U' for upper). Return updated y.

Polynomials

  • y ±= alpha*A[uplo]*x
  • y = beta*y ± alpha*A[uplo]*x
julia> macroexpand(SugarBLAS, :(SugarBLAS.@symv! y -= alpha*A['U']*x))
:(Base.LinAlg.BLAS.symv!('U',-alpha,A,x,1.0,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symv! y = y - alpha*A['L']*x))
(Base.LinAlg.BLAS.symv!('L',-alpha,A,x,1.0,y))

julia> macroexpand(SugarBLAS, :(SugarBLAS.@symv! y = beta*y + alpha*A['U']*x))
:(Base.LinAlg.BLAS.symv!('U',-alpha,A,x,beta,y))