Add solve based on OpenBLAS #710

zoziha · 2023-05-06T09:16:40Z

Use pkg-config to find openblas and define USE_OPENBLAS;
BUILD_OPENBLAS is off by default because stdlib has few BLAS-based functions yet.
Routines based on BLAS and LAPACK often only support single or double precision floating-point types;
Add BLAS_common.fypp to set the BLAS related common variables;
Add a solve function to stdlib_linalg, and test it;
Discusses and handles the _gesv return value info.
Test it in CI (MINGW✔, Linux-gfortran✔, Linux-ifort, macOS-gfortran);

This PR is an attempt to irrationally generate enthusiasm for BLAS integration into stdlib. Close #709 .

Syntax

x = [[stdlib_linalg(module):solve(interface)]](a,b)

St-Maxwell · 2023-05-07T03:39:21Z

src/stdlib_linalg.fypp

+    #:set CMPLX_KINDS_TYPES_GESV = list(zip(BLAS_CMPLX_KINDS, CMPLX_TYPES, BLAS_CMPLX_GESV))
+    #:set RC_KINDS_TYPES_GESV = REAL_KINDS_TYPES_GESV + CMPLX_KINDS_TYPES_GESV
+    #:for k1, t1, g1 in RC_KINDS_TYPES_GESV
+    module function solve_${t1[0]}$${k1}$(a, b) result(x)


I would prefer a subroutine-style interface, e.g.

subroutine solve(X, A, B)

I agree with the above. It makes more sense to use a subroutine, so that the code stays performant when large matrices are involved.

Also, would it make sense to call the procedure something like linsolve, instead of just solve, anticipating that in the future a nonlinear solver will be added as well?

St-Maxwell · 2023-05-07T03:50:49Z

README.md

@@ -133,6 +133,7 @@ Important options are
  Compiling with maximum rank 15 can be resource intensive and requires at least 16 GB of memory to allow parallel compilation or 4 GB memory for sequential compilation.
 - `-DBUILD_SHARED_LIBS` set to `on` in case you want link your application dynamically against the standard library (default: `off`).
 - `-DBUILD_TESTING` set to `off` in case you want to disable the stdlib tests (default: `on`).
+- `-DBUILD_OPENBLAS` set to `on` in case you want to use routines based on OpenBLAS (default: `off`).


BUILD_WITH_OPENBLAS may be better

jvdp1

Thank you. I think it will be an important step for stdlib. My main concern is about the API. Should it be a simple wrapper, or something more complete (similar to @awvwgk BLAS interfaces).
See also #450 for some discussions.

jvdp1 · 2023-05-07T06:22:46Z

README.md

@@ -133,6 +133,7 @@ Important options are
  Compiling with maximum rank 15 can be resource intensive and requires at least 16 GB of memory to allow parallel compilation or 4 GB memory for sequential compilation.
 - `-DBUILD_SHARED_LIBS` set to `on` in case you want link your application dynamically against the standard library (default: `off`).
 - `-DBUILD_TESTING` set to `off` in case you want to disable the stdlib tests (default: `on`).
+- `-DBUILD_OPENBLAS` set to `on` in case you want to use routines based on OpenBLAS (default: `off`).


fpm should be also able to compile the project.

jvdp1 · 2023-05-07T06:24:01Z

doc/specs/stdlib_linalg.md

+efficient low level implementations of standard linear algebra algorithms.
+
+If possible, highly optimized libraries that take advantage of specialized processor functionality are preferred.
+Examples of such libraries are [OpenBLAS][1], [oneAPI MKL(TM)][2].


Suggested change

Examples of such libraries are [OpenBLAS][1], [oneAPI MKL(TM)][2].

Examples of such libraries are [OpenBLAS][1], [Intel oneAPI MKL(TM)][2].

jvdp1 · 2023-05-07T06:24:34Z

doc/specs/stdlib_linalg.md

@@ -425,3 +436,43 @@ Specifically, upper Hessenberg matrices satisfy `a_ij = 0` when `j < i-1`, and l
 ```fortran
 {!example/linalg/example_is_hessenberg.f90!}
 ```
+
+## `solve` - Solve a-linear-matrix-equation


Suggested change

## `solve` - Solve a-linear-matrix-equation

## `solve` - Solve a linear matrix equation

jvdp1 · 2023-05-07T06:25:44Z

doc/specs/stdlib_linalg.md

+
+Solve a linear matrix equation, or system of linear scalar equations.
+
+Note: `solve` supports only single and double precision floating point types.


This "note" will be implicitely mentioned through the description of the arguments. Therefore, it could be removed

jvdp1 · 2023-05-07T06:27:08Z

doc/specs/stdlib_linalg.md

+
+### Class
+
+Impure function.


For such procedures, I would prefera subroutine over a function, especially for efficienct with large vectors/matrices.

Subroutines are usually faster than functions, functions are usually easier to use but may cause performance degradation. If you need high performance, you should use subroutines. If you need simpler code, you should use a function.

Intuitively, solving a system of AX = B linear equations, the normal form X = solve(A, B) will naturally be understood by the average user and will be surprised by call solve(A, B). My personal recommendation is to think of solve as a function dedicated to solving systems of linear equations, rather than for performance, which can be done with _gesv routines.

(Not recommended) Perhaps we could provide a solve_ subroutine that implements the in-place version of solve, like this:

subroutine solve_(a, b) real, intent(inout) :: a(:, :), b(:, :) integer :: ipiv(size(a, 1)), info call sgesv(size(a, 1), size(b, 2), a, size(a, 1), ipiv, b, size(b, 1), info) end subroutine solve_

(Recommended) Or for performance-sensitive users, it is recommended to just use the low-level interface _gesv for programming by borrowing the implementation of solve?

jvdp1 · 2023-05-07T06:29:10Z

doc/specs/stdlib_linalg.md

+
+### Synatx
+
+`x = [[stdlib_linalg(module):solve(interface)]](a,b)`


Could b be a vector too?

jvdp1 · 2023-05-07T06:30:18Z

doc/specs/stdlib_linalg.md

+
+Experimental
+
+### Description


It might be good to add somewhere to which BLAS procedure it is linked with

zoziha · 2023-05-07T07:43:22Z

Should it be a simple wrapper, or something more complete (similar to @awvwgk BLAS interfaces).

@jvdp1 I'm not worried about this, and it shouldn't be too complicated to switch if the full BLAS interfaces library is available later.

jvdp1 · 2023-12-25T22:19:00Z

Related to #749 (ping @perazz )

jvdp1 · 2024-04-02T19:31:51Z

This PR should be reviewed considering latest developments of @perazz on BLAS/LAPACK

jvdp1 · 2024-05-11T08:54:57Z

Closed in favour of #806 .
Thank you to all contributors to this PR.

Add solve based on OpenBLAS

25b7fd9

zoziha marked this pull request as draft May 6, 2023 09:16

zoziha added 2 commits May 6, 2023 18:09

Fix fpm-CI

a508dc5

Use OpenBLAS in Linux/MINGW CI

2aedbcd

St-Maxwell reviewed May 7, 2023

View reviewed changes

jvdp1 reviewed May 7, 2023

View reviewed changes

jvdp1 requested review from certik, arjenmarkus and awvwgk May 7, 2023 06:42

perazz mentioned this pull request Mar 11, 2024

Provide BLAS, LAPACK backends and interfaces #772

Merged

3 tasks

perazz mentioned this pull request Apr 25, 2024

linalg: solve #806

Merged

7 tasks

jvdp1 closed this May 11, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add solve based on OpenBLAS #710

Add solve based on OpenBLAS #710

zoziha commented May 6, 2023 •

edited by jvdp1

Loading

St-Maxwell May 7, 2023

cval26 Jun 30, 2023

St-Maxwell May 7, 2023

jvdp1 left a comment •

edited

Loading

jvdp1 May 7, 2023

jvdp1 May 7, 2023

jvdp1 May 7, 2023

jvdp1 May 7, 2023

jvdp1 May 7, 2023

zoziha May 7, 2023 •

edited

Loading

jvdp1 May 7, 2023

jvdp1 May 7, 2023

zoziha commented May 7, 2023 •

edited

Loading

jvdp1 commented Dec 25, 2023

jvdp1 commented Apr 2, 2024

jvdp1 commented May 11, 2024

	Examples of such libraries are [OpenBLAS][1], [oneAPI MKL(TM)][2].
	Examples of such libraries are [OpenBLAS][1], [Intel oneAPI MKL(TM)][2].

	## `solve` - Solve a-linear-matrix-equation
	## `solve` - Solve a linear matrix equation


		Solve a linear matrix equation, or system of linear scalar equations.

		Note: `solve` supports only single and double precision floating point types.


		### Synatx

		`x = [[stdlib_linalg(module):solve(interface)]](a,b)`

Add solve based on OpenBLAS #710

Add solve based on OpenBLAS #710

Conversation

zoziha commented May 6, 2023 • edited by jvdp1 Loading

Syntax

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jvdp1 left a comment • 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

zoziha May 7, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zoziha commented May 7, 2023 • edited Loading

jvdp1 commented Dec 25, 2023

jvdp1 commented Apr 2, 2024

jvdp1 commented May 11, 2024

zoziha commented May 6, 2023 •

edited by jvdp1

Loading

jvdp1 left a comment •

edited

Loading

zoziha May 7, 2023 •

edited

Loading

zoziha commented May 7, 2023 •

edited

Loading