New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[LAPACK] Interface lacpy! and add a Julia version copytrito! #51909

Merged

ViralBShah merged 10 commits into JuliaLang:master from amontoison:lapack_lacpy

Dec 4, 2023

Contributor

amontoison commented Oct 28, 2023

The PR adds a function lacpy! to easily copy a triangular part of a square or rectangular dense matrix.

jishnub added the linear algebra label

Contributor

jishnub commented Oct 28, 2023 •

edited

Loading

It'll probably be better to specialise copyto! for triangular matrices instead of adding another function. The new method may call LAPACK.lacpy! internally. I wonder how the LAPACK call performs compared to the Julia implementation?

Contributor Author

amontoison commented Oct 28, 2023 •

edited

Loading

Triangular wrappers only work with square matrices. It's a major issue when we want to copy the L / R factor of an LQ / QR decomposition.
But you are right, we can still specialize copyto! for triangular square matrices.

For the performance of the LAPACK version, I don't think that we have a difference for the backend OpenBLAS.
We could have a gain if we use the Intel MKL backend.

Member

ViralBShah commented Oct 28, 2023

Slightly Off-topic here, but we need to have some sort of a global preference for picking a non-default BLAS.

dkarrasch reviewed

View reviewed changes

Member

dkarrasch left a comment

I agree with @jishnub on both points: (a) do we need the LAPACK version at all, or would it suffice to have the generic one, and (b) can't we have a more Julian name for the generic function (that may or may not call LAPACK for strided matrices of BlasFloats elements? Perhaps we could target at something closer to LinearAlgebra.copytri! (it copies a triangle into the opposite triangle of the same matrix), perhaps copytotri!?

stdlib/LinearAlgebra/src/generic.jl Outdated

Comment on lines 1904 to 1908

+Copies all or part of a matrix A to another matrix B.
+uplo specifies the part of the matrix A to be copied to B.
+Set `uplo = 'L'` for the lower triangular part, `uplo = 'U'`
+for the upper triangular part, any other character for all
+the matrix A.

Member

dkarrasch Oct 29, 2023

I'm not 100% sure, but I think the arguments/variables should be wrapped in ``, like A and uplo.

stdlib/LinearAlgebra/src/generic.jl Outdated

+4
+```
+"""
+function lacpy!(uplo::AbstractChar, A::AbstractMatrix, B::AbstractMatrix)

Member

dkarrasch Oct 29, 2023

This cries for require_one_based_indexing, or checks on axes. The following code may then be annotated with @inbounds.

Contributor Author

amontoison commented Oct 30, 2023

I like the idea copytotri!.
copytotri! is a generic Julia version that doesn't call lacpy!.
I propose to keep lacpy! because it's already interfaced.

amontoison force-pushed the lapack_lacpy branch from 7c4be2e to bd64930 Compare

October 30, 2023 20:37

amontoison changed the title ~~[LAPACK] Interface lacpy! and add a generic version~~ [LAPACK] Interface lacpy! and add a Julia version copytotri!

amontoison requested a review from dkarrasch

October 30, 2023 20:38

amontoison force-pushed the lapack_lacpy branch from bd64930 to 99ae394 Compare

October 30, 2023 20:39

amontoison mentioned this pull request

Move Block-GMRES in Krylov.jl JuliaSmoothOptimizers/Krylov.jl#831

Closed

dkarrasch reviewed

View reviewed changes

stdlib/LinearAlgebra/src/generic.jl Outdated

+function copytotri!(uplo::AbstractChar, A::AbstractMatrix, B::AbstractMatrix)
+ require_one_based_indexing(A, B)
+ LinearAlgebra.BLAS.chkuplo(uplo)
+ m, n = size(A)

Member

dkarrasch Oct 31, 2023

You definitely need to check the size of B as well, otherwise (if B is smaller than A) some indices might be out-of-bounds. Judging by the LAPACK version below, you do want to allow that B is larger than A, right?

Contributor Author

amontoison Nov 1, 2023

Yes, we should allow B to be larger than A.

stdlib/LinearAlgebra/src/generic.jl Outdated Show resolved Hide resolved

stdlib/LinearAlgebra/src/generic.jl Outdated Show resolved Hide resolved

amontoison force-pushed the lapack_lacpy branch from 1614c4a to 2db81ef Compare

November 1, 2023 03:05

amontoison commented

View reviewed changes

stdlib/LinearAlgebra/src/lapack.jl Outdated Show resolved Hide resolved

amontoison force-pushed the lapack_lacpy branch from 574c85d to 43ba8cb Compare

November 2, 2023 23:28

amontoison requested a review from jishnub

November 4, 2023 04:42

amontoison force-pushed the lapack_lacpy branch 2 times, most recently from e5c00a0 to 58b513d Compare

November 11, 2023 00:31

Contributor Author

amontoison commented Nov 11, 2023

@dkarrasch @jishnub The errors on CI are unrelated to this PR.
Do you have additional comments?

Member

ViralBShah commented Nov 19, 2023

I'm rerunning the CI. If no other concerns, we should merge.

jishnub reviewed

View reviewed changes

stdlib/LinearAlgebra/src/generic.jl Outdated Show resolved Hide resolved

jishnub reviewed

View reviewed changes

stdlib/LinearAlgebra/src/generic.jl Outdated Show resolved Hide resolved

amontoison commented

View reviewed changes

stdlib/LinearAlgebra/src/generic.jl Outdated Show resolved Hide resolved

amontoison force-pushed the lapack_lacpy branch 2 times, most recently from 4e1b45f to 1d0c2fc Compare

November 20, 2023 22:34

Member

dkarrasch commented Nov 27, 2023 •

edited

Loading

Well, naming is hard. I had no better idea, I just thought that picking the BLAS name for a generic Julia function is not quite right, but other than that I have no strong opinions. Would copytrito! perhaps be better, like "copy a given tri angle to something else"?

Contributor

jishnub commented Nov 27, 2023 •

edited

Loading

This seems better, as copytri! copies the triangular part onto itself, whereas copytrito! copies it to another destination. It also reflects the fact that the structured part of the source is being copied, and not that the source is being copied to a triangular destination.

amontoison added 3 commits

November 27, 2023 13:42


[LAPACK] Interface lacpy! and add a generic version

067af51


Add a comma

ca95d59


Fix the docstring of lacpy!

0709c7b

amontoison and others added 6 commits

November 27, 2023 13:42


Add the function copytotri!

1b032be


Use LinearAlgebra.BLAS.chkuplo

73d1fa0


Update lacpy! and copytotri!

d445e01


Update stdlib/LinearAlgebra/src/lapack.jl

6fb898a


Fix the tests

51076d4


Rename copytotri into copytrito

b53e54b

Contributor Author

amontoison commented Nov 27, 2023

I renamed copytotri! into copytrito! 👍

amontoison force-pushed the lapack_lacpy branch from 1d0c2fc to b53e54b Compare

November 27, 2023 20:21

Contributor Author

amontoison commented Dec 4, 2023

@ViralBShah @dkarrasch @jishnub
Do you have any additional comments? If not, I suggest merging the pull request.


Update stdlib/LinearAlgebra/src/generic.jl

dbb1377

dkarrasch approved these changes

View reviewed changes

amontoison changed the title ~~[LAPACK] Interface lacpy! and add a Julia version copytotri!~~ [LAPACK] Interface lacpy! and add a Julia version copytrito!

ViralBShah merged commit e280387 into JuliaLang:master

6 of 8 checks passed

amontoison deleted the lapack_lacpy branch

December 4, 2023 13:21

amontoison mentioned this pull request

Add a GPU version of copytrito! JuliaGPU/GPUArrays.jl#504

Merged

amontoison mentioned this pull request

Dimension check too restrictive in copytrito! and lacpy! #54546

Closed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment