add special type for views of Band, macro to implement banded linear algebra #37

dlfivefifty · 2017-10-24T11:19:02Z

Now view(A, band(0)) returns a view of A, not A.data.

This enables faster implementations of, for example, Vector{Float64}(view(A, band(0))).

randy3k

In general, we need to check if the matrix type if BLAS compatible as well as the existence of the methods leadingdimension(::T) and pointer(::T).

randy3k · 2017-10-31T18:12:12Z

src/generic/linalg.jl

+        BandedMatrices.@banded_banded_linalg($Typ, $Typ)
+        # default is to use dense axpy!
+        BandedMatrices.banded_axpy!(a::Number, X::$Typ, Y::AbstractMatrix) =
+                BandedMatrices.banded_dense_axpy!(a, X, Y)


perhaps there is obvious, but why not using banded_generic_axpy if the interface has been implemented?

randy3k · 2017-10-31T18:18:35Z

src/generic/linalg.jl

+
+macro banded_banded_banded_linalg(Typ1, Typ2, Typ3)
+    ret = quote
+        BandedMatrices.positively_banded_matmatmul!(C::$Typ1{T}, tA::Char, tB::Char, A::$Typ2{T}, B::$Typ3{T}) where {T <: BlasFloat} =


We should only eval this when Type 1, 2, 3 are BLAS compatible. Also, unless bothleadingdimension(::T) and pointer(::T) exist, this function will dispatch gbmm and produce an error.

randy3k · 2017-10-31T18:20:10Z

src/generic/linalg.jl

+        Base.:*(A::$Typ{U}, b::StridedVector{V}) where {U, V} =
+            Base.LinAlg.A_mul_B!(Vector{promote_type(U, V)}(size(A, 1)), A, b)
+
+        BandedMatrices.positively_banded_matmatmul!(C::StridedMatrix{T}, tA::Char, tB::Char, A::$Typ{T}, B::StridedMatrix{T}) where {T <: BlasFloat} =


this function should be only defined when the types are BLAS compatible

randy3k · 2017-10-31T18:20:15Z

src/generic/linalg.jl

+
+        BandedMatrices.positively_banded_matmatmul!(C::StridedMatrix{T}, tA::Char, tB::Char, A::$Typ{T}, B::StridedMatrix{T}) where {T <: BlasFloat} =
+            BandedMatrices._positively_banded_matmatmul!(C, tA, tB, A, B)
+        BandedMatrices.positively_banded_matmatmul!(C::StridedMatrix{T}, tA::Char, tB::Char, A::StridedMatrix{T}, B::$Typ{T}) where {T <: BlasFloat} =


same here, this function should be only defined when the types are BLAS compatible

randy3k · 2017-10-31T18:21:45Z

src/generic/linalg.jl

+        end
+
+        # use BLAS routine for positively banded BLASBandedMatrix
+        BandedMatrices.positively_banded_matvecmul!(c::StridedVector{T}, tA::Char, A::$Typ{T}, b::StridedVector{T}) where {T <: BandedMatrices.BlasFloat} =


this function should be only defined when the types are BLAS compatible

dlfivefifty · 2017-11-03T15:28:13Z

I think we should turn BLAS compatible into a trait, a la Base.HasLength, and get rid of BLASBandedMatrix (see JuliaLang/julia#24383 (comment) for one reason to get rid of Union types). For example:

struct NotBLAS end
struct IsBLASStridedArray{d} end
const IsBLASStridedMatrix = BLASStridedArray{2}
struct IsBLASBandedMatrix end

blasstructure(::Type{<:AbstractArray}) = NotBLAS()
blasstructure(::Type{SM}) where SM <: StridedMatrix{T} where T <: BlasFloat = IsBLASStridedMatrix()
blasstructure(::Type{BandedMatrix{<:BlasFloat}}) = IsBLASBandedMatrix()
blasstructure(::Type{BandedSubBandedMatrix{<:BlasFloat}}) = IsBLASBandedMatrix()

# if blasstructure returns IsBLASBandedMatrix(), then one is required to override `leadingdimension`
leadingdimension(A::BandedMatrix{<:BlasFloat}) = ...

banded_matmatmul!(C::AbstractMatrix{T}, tA::Char, tB::Char, A::AbstractMatrix{U}, B::AbstractMatrix{V}) where {T, U, V} = banded_matmatmul!(C, tA, tB, A, B, isblasbanded(A), isblasbanded(B))

banded_matmatmul!(C::AbstractMatrix{T}, tA::Char, tB::Char, A::AbstractMatrix{U}, B::AbstractMatrix{V}, blasA, blasB) where {T, U, V} = banded_generic_matmatmul!(C, tA, tB, A, B)

banded_matmatmul!(C::AbstractMatrix{T}, tA::Char, tB::Char, A::AbstractMatrix{U}, B::AbstractMatrix{V}, ::IsBLASBandedMatrix, :IsBLASBandedMatrix) where {T, U, V} = generally_banded_matmatmul!(C, tA, tB, A, B)

This does move away from the design of Base, but I would bet that eventually StridedMatrix in Base becomes a trait.

randy3k · 2017-11-03T15:44:03Z

Cf: JuliaLang/julia#2345 (comment)

dlfivefifty · 2017-11-04T22:10:36Z

I just pushed an update that implements the proposed trait solution. It has simplified the @banded_linalg macros.

If I find the time, I might make a similar change in Base and open a pull request, to at least start a discussion about it. I think if Base's design used traits instead of inheritance to determine BLAS structure, it would greatly simplify the A_*_B! overloads (and remove the need for the @banded_linalg macros completely.

…ces.jl into feat-bandslice

…es that are not <: AbstractBandedMatrix

• SymBandedMatrix{T}(data,...) -> _SymBandedMatrix(data,...) These changes avoid confusion between converting a matrix and specifying the data of a matrix.

… banded matrix interface for Base types

SymBandedMatrix{T}(n,m,a,b) -> SymBandedMatrix{T}(uninitialized,n,m,a,b)

• fix BlasStrided -> BandedLayout

Sheehan Olver and others added 4 commits October 24, 2017 12:17

add special time for views of Band

f5cc302

export isbanded

ff27d5d

reorganize files, macro for lin alg overrides

24b63cf

move macro to generic

98e12ad

randy3k reviewed Oct 31, 2017

View reviewed changes

add BlasBanded, BlasSymBanded and BlasStrided traits to dispatch linalg

32417c0

dlfivefifty changed the title ~~add special type for views of Band~~ add special type for views of Band, macro to implement banded linear algebra Nov 4, 2017

dlfivefifty and others added 20 commits November 12, 2017 16:59

Merge branch 'master' of https://github.com/JuliaMatrices/BandedMatri…

3d258a0

…ces.jl into feat-bandslice

using Test in 0.7

f80767f

Add @banded and @banded_interface macros for adding new banded matric…

96b95ac

…es that are not <: AbstractBandedMatrix

start using Zeros and Eye

acb5b84

Deprecate BandedMatrix(data,...) for BandedMatrix{T}(data,...).

35ca220

Fix deprecation warnings

a8d3456

• BandedMatrix{T}(data,...) -> _BandedMatrix(data,...)

532ea57

• SymBandedMatrix{T}(data,...) -> _SymBandedMatrix(data,...) These changes avoid confusion between converting a matrix and specifying the data of a matrix.

Straggler changes to _BandedMatrix

c20884e

REQUIRE FillArrays v0.0.1

94735b4

At BandedMatrix(A::AbstractMatrix, bnds) that projects, and implement…

1301823

… banded matrix interface for Base types

tests pass again

e191ddb

Merge fillarrays

7bf0fed

REQUIRE Compat 0.38

93cdcac

Add constructor BandedMatrix{T}((n,m), (l,u))

a6391cb

BandedMatrix{T}(n,m,a,b) -> BandedMatrix{T}(uninitialized,n,m,a,b)

7bbeb50

SymBandedMatrix{T}(n,m,a,b) -> SymBandedMatrix{T}(uninitialized,n,m,a,b)

use uninitialized

e6ecc97

test BBandedMatrix{T}(uninitialized, (n,m), (a,b)) constructor

fab2e25

blasstructure -> memorylayout

8685b81

• αA_mul_B_plus_βC! -> scalemul!

fcc4ccf

• fix BlasStrided -> BandedLayout

Add fill! and shorthand BandError

60ced52

dlfivefifty and others added 4 commits January 8, 2018 12:58

add banded_copy!

422c71d

scalemul! -> mul!

ca26051

reorder mul! arguments

0529245

Update MemoryLayout

c316132

dlfivefifty merged commit cc4e07c into master Feb 23, 2018

dlfivefifty deleted the feat-bandslice branch February 23, 2018 10:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add special type for views of Band, macro to implement banded linear algebra #37

add special type for views of Band, macro to implement banded linear algebra #37

dlfivefifty commented Oct 24, 2017

randy3k left a comment •

edited

Loading

randy3k Oct 31, 2017

randy3k Oct 31, 2017 •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading

dlfivefifty commented Nov 3, 2017

randy3k commented Nov 3, 2017

dlfivefifty commented Nov 4, 2017

add special type for views of Band, macro to implement banded linear algebra #37

add special type for views of Band, macro to implement banded linear algebra #37

Conversation

dlfivefifty commented Oct 24, 2017

randy3k left a comment • edited Loading

Choose a reason for hiding this comment

randy3k Oct 31, 2017

Choose a reason for hiding this comment

randy3k Oct 31, 2017 • edited Loading

Choose a reason for hiding this comment

randy3k Oct 31, 2017 • edited Loading

Choose a reason for hiding this comment

randy3k Oct 31, 2017 • edited Loading

Choose a reason for hiding this comment

randy3k Oct 31, 2017 • edited Loading

Choose a reason for hiding this comment

dlfivefifty commented Nov 3, 2017

randy3k commented Nov 3, 2017

dlfivefifty commented Nov 4, 2017

randy3k left a comment •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading

randy3k Oct 31, 2017 •

edited

Loading