Adjust signatures for bkfact to use Symmetric and Hermitian #22605

andreasnoack · 2017-06-29T01:03:19Z

Factored out of #22601. This is similar to the changes we made to cholfact. The idea is to make Symmetric and Hermitian the default input to the factorization function and that the uplo argument is specified via Symmetric or Hermitian instead of an argument to the factorization function.

ararslan · 2017-06-29T01:21:02Z

base/deprecated.jl

+end
+function bkfact!(A::StridedMatrix, uplo::Symbol, symmetric::Bool = issymmetric(A), rook::Bool = false)
+    depwarn("bkfact with uplo and symmetric arguments deprecated. Please use bkfact!($(symmetric ? "Symmetric(" : "Hermitian(")A, :$uplo))",
+        :bkfact)


This and on the line above should be bkfact! rather than bkfact, right?

tkelman · 2017-06-29T02:32:27Z

base/linalg/bunchkaufman.jl

-    print(io, "successful: $(issuccess(F))")
+function Base.show(io::IO, mime::MIME{Symbol("text/plain")}, B::BunchKaufman)
+    println(io, summary(B))
+    print(io, "\nsuccessful: $(issuccess(B))")


doesn't need to double the newlines?

tkelman · 2017-06-29T02:32:48Z

test/linalg/bunchkaufman.jl

@@ -48,36 +49,40 @@ bimg  = randn(n,2)/2
                εb = eps(abs(float(one(eltyb))))
                ε = max(εa,εb)

-                @testset "(Automatic) Bunch-Kaufman factor of indefinite matrix" begin
-                    bc1 = factorize(asym)
+                # check thactor factorize gives a Bunch-Kaufman


check that ?

tkelman

lgtm if linalg tests pass on CI
(though a news entry might be worth adding now so we don't have to do it later)

Sacha0 · 2017-06-29T20:06:29Z

base/linalg/bunchkaufman.jl

    end
-    BunchKaufman(LD, ipiv, char_uplo(uplo), symmetric, rook, info)
+    BunchKaufman(LD, ipiv, A.uplo, true, rook, info)
 end


If I read the diff correctly (quite possibly not!), the two preceding bkfact! method bodies are identical? If so, perhaps worth DRYing out? (Edit: Could also reduce the five lines associated with the conditional to the single line LD, ipiv, info = rook ? LAPACK.sytrf_rook!(A.uplo, A.data) : LAPACK.sytrf!(A.uplo, A.data)?)

You are right. The HermOrSym one should be removed.

Sacha0 · 2017-06-29T20:13:11Z

base/linalg/bunchkaufman.jl

-            LD, ipiv, info = LAPACK.hetrf!(char_uplo(uplo), A)
-        end
+        LD, ipiv, info = LAPACK.hetrf!(A.uplo, A.data)
+    end


Likewise here, if you feel like condensing the five lines into one :). (LD, ipiv, info = rook ? LAPACK.hetrf_rook!(A.uplo, A.data) : LAPACK.hetrf!(A.uplo, A.data).)

Sacha0 · 2017-06-29T20:17:12Z

base/linalg/bunchkaufman.jl

-        bkfact!(convert(Matrix{promote_type(Float32, typeof(sqrt(one(T))))}, A),
-                uplo, symmetric, rook)
+bkfact(A::AbstractMatrix{T}, rook::Bool=false) where {T} =
+    bkfact!(copy_oftype(A, typeof(sqrt(one(T)))), rook)


Removing the former promotion with Float32 to support e.g. Float16? :)

good eye, missed that change - might want to throw in a test for this bit if there's any behavior that this enables right now

Hm. I had integer and rational input in mind and then sqrt(one(T)) should be sufficient to get a BLAS type but you are right that a matrix with Float16 would stay Float16. Probably not a big problem, though. If you do arithmetic with Float16 you probably would like to avoid that your elements get promoted silently.

so will integer or rational input now behave differently with this?

No. They are unaffected since sqrt(one(T)) isa Float64 for T in (Rational{Int},Int). A Matrix{Float16} will now fail instead of being promoted. I think that is reasonable because people using Float16 most likely prefer an error over silent promotion and I really don't think anybody was ever called bkfact with a Matrix{Float16} in serious code.

Float16 is used somewhat frequently in the gpu community. What will the error be in that case? We silently promote for intermediate calculations pretty often and convert back to the destination type when an implementation isn't available in a specific element type.

We don't do this for factorizations and there hasn't been expressed any demand for this so I think we good for now. We can revisit if Float16 users start asking for this behavior.

can you answer the question at least?

don't all the other factorizations have generic implementations so they wouldn't error?

You'll get a MethodError in bkfact!. Not all factorizations have generic fallbacks, e.g. eig and svd.

please rebase out the commit from a different pr

ararslan added the domain:linear algebra Linear algebra label Jun 29, 2017

ararslan reviewed Jun 29, 2017

View reviewed changes

tkelman added kind:deprecation This change introduces or involves a deprecation needs news A NEWS entry is required for this change labels Jun 29, 2017

tkelman reviewed Jun 29, 2017

View reviewed changes

andreasnoack force-pushed the anj/bkfactsig branch from 7f9e516 to 3563f5e Compare June 29, 2017 18:50

tkelman previously approved these changes Jun 29, 2017

View reviewed changes

tkelman removed the needs news A NEWS entry is required for this change label Jun 29, 2017

Sacha0 reviewed Jun 29, 2017

View reviewed changes

andreasnoack force-pushed the anj/bkfactsig branch from 5addd1d to 2a89f9a Compare June 29, 2017 21:08

andreasnoack force-pushed the anj/bkfactsig branch from 2a89f9a to b668aa9 Compare June 29, 2017 23:45

andreasnoack added 2 commits June 29, 2017 19:47

Adjust signatures for bkfact to use Symmetric and Hermitian

31cdb97

Add NEWS.md entry

d605eea

andreasnoack force-pushed the anj/bkfactsig branch from b668aa9 to d605eea Compare June 29, 2017 23:48

andreasnoack merged commit 88e7fbc into master Jun 30, 2017

andreasnoack deleted the anj/bkfactsig branch June 30, 2017 18:59

cormullion mentioned this pull request Mar 18, 2018

NEWS.md is getting a bit untidy #26508

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adjust signatures for bkfact to use Symmetric and Hermitian #22605

Adjust signatures for bkfact to use Symmetric and Hermitian #22605

andreasnoack commented Jun 29, 2017

ararslan Jun 29, 2017

tkelman Jun 29, 2017

tkelman Jun 29, 2017

tkelman left a comment •

edited

Loading

Sacha0 Jun 29, 2017 •

edited

Loading

andreasnoack Jun 29, 2017

Sacha0 Jun 29, 2017

Sacha0 Jun 29, 2017

tkelman Jun 29, 2017 •

edited

Loading

andreasnoack Jun 29, 2017 •

edited

Loading

tkelman Jun 30, 2017

andreasnoack Jun 30, 2017

tkelman Jun 30, 2017

andreasnoack Jun 30, 2017

tkelman Jun 30, 2017

tkelman Jun 30, 2017

andreasnoack Jun 30, 2017

Adjust signatures for bkfact to use Symmetric and Hermitian #22605

Adjust signatures for bkfact to use Symmetric and Hermitian #22605

Conversation

andreasnoack commented Jun 29, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tkelman left a comment • edited Loading

Choose a reason for hiding this comment

Sacha0 Jun 29, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tkelman Jun 29, 2017 • edited Loading

Choose a reason for hiding this comment

andreasnoack Jun 29, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tkelman left a comment •

edited

Loading

Sacha0 Jun 29, 2017 •

edited

Loading

tkelman Jun 29, 2017 •

edited

Loading

andreasnoack Jun 29, 2017 •

edited

Loading