[patch, fortran] Inline BLAS calls to support conjg(transpose(a))

[Thomas Koenig]

Hello world,

this patch generates direct calls to *GEMM when -fexternal-blas is
specified. This allows to handle arguments to conjugate and transposed
elements, which is quite a common use case.

While looking at the code, I found that the inline limit checks were not
correctly handled for cases except for A2B2. This is also fixed.

In order to check all cases at runtime, I simply copied the reference
BLAS routines to the test cases, why they are *.f instead of *.f90.

Regarding the bounds checking: I added three new test cases, but
as for checking everything, that would be a bit too much. The code
is clear enough that I think the other cases should be OK.

OK for trunk?

Regards

Thomas

2018-09-18 Thomas Koenig <***@gcc.gnu.org>

PR fortran/29550
* gfortran.h (gfc_expr): Add external_blas flag.
* frontend-passes.c (matrix_case): Add case A2TB2T.
(optimize_namespace): Handle flag_external_blas by
calling call_external_blas.
(get_array_inq_function): Add argument okind. If
it is nonzero, use it as the kind of argument
to be used.
(inline_limit_check): Remove m_case argument, add
limit argument instead. Remove assert about m_case.
Set the limit for inlining from the limit argument.
(matmul_lhs_realloc): Handle case A2TB2T.
(inline_matmul_assign): Handle inline limit for other cases with
two rank-two matrices. Remove no-op calls to inline_limit_check.
(call_external_blas): New function.
* trans-intrinsic.c (gfc_conv_intrinsic_funcall): Do not add
argument to external BLAS if external_blas is already set.

2018-09-18 Thomas Koenig <***@gcc.gnu.org>

PR fortran/29550
* gfortran.dg/inline_matmul_13.f90: Adjust count for
_gfortran_matmul.
* gfortran.dg/inline_matmul_16.f90: Likewise.
* gfortran.dg/promotion_2.f90: Add -fblas-matmul-limit=1. Scan
for dgemm instead of dgemm_. Add call to random_number to make
standard conforming.
* gfortran.dg/matmul_blas_1.f90: New test.
* gfortran.dg/matmul_bounds_14.f: New test.
* gfortran.dg/matmul_bounds_15.f: New test.
* gfortran.dg/matmul_bounds_16.f: New test.

[Paul Richard Thomas]

Hi Thomas,

This fine, except for one niggle. Rather than having the blas source
in each testcase, why don't you put all the functions in a blas file
and use the dg-additional-sources mechanism?

Cheers

Paul

Post by Thomas Koenig
Hello world,
this patch generates direct calls to *GEMM when -fexternal-blas is
specified. This allows to handle arguments to conjugate and transposed
elements, which is quite a common use case.
While looking at the code, I found that the inline limit checks were not
correctly handled for cases except for A2B2. This is also fixed.
In order to check all cases at runtime, I simply copied the reference
BLAS routines to the test cases, why they are *.f instead of *.f90.
Regarding the bounds checking: I added three new test cases, but
as for checking everything, that would be a bit too much. The code
is clear enough that I think the other cases should be OK.
OK for trunk?
Regards
Thomas
PR fortran/29550
* gfortran.h (gfc_expr): Add external_blas flag.
* frontend-passes.c (matrix_case): Add case A2TB2T.
(optimize_namespace): Handle flag_external_blas by
calling call_external_blas.
(get_array_inq_function): Add argument okind. If
it is nonzero, use it as the kind of argument
to be used.
(inline_limit_check): Remove m_case argument, add
limit argument instead. Remove assert about m_case.
Set the limit for inlining from the limit argument.
(matmul_lhs_realloc): Handle case A2TB2T.
(inline_matmul_assign): Handle inline limit for other cases with
two rank-two matrices. Remove no-op calls to inline_limit_check.
(call_external_blas): New function.
* trans-intrinsic.c (gfc_conv_intrinsic_funcall): Do not add
argument to external BLAS if external_blas is already set.
PR fortran/29550
* gfortran.dg/inline_matmul_13.f90: Adjust count for
_gfortran_matmul.
* gfortran.dg/inline_matmul_16.f90: Likewise.
* gfortran.dg/promotion_2.f90: Add -fblas-matmul-limit=1. Scan
for dgemm instead of dgemm_. Add call to random_number to make
standard conforming.
* gfortran.dg/matmul_blas_1.f90: New test.
* gfortran.dg/matmul_bounds_14.f: New test.
* gfortran.dg/matmul_bounds_15.f: New test.
* gfortran.dg/matmul_bounds_16.f: New test.

[Thomas Koenig]

Hi Paul,

Post by Paul Richard Thomas
This fine, except for one niggle. Rather than having the blas source
in each testcase, why don't you put all the functions in a blas file
and use the dg-additional-sources mechanism?

Good idea, I have added this. Committed as r264411.

Regards

Thomas

[Dominique d'Humières]

Post by Thomas Koenig
Committed as r264411.

Nope!-(

Dominique

[Thomas Koenig]

Hi Dominique,

Post by Dominique d'Humières

Post by Thomas Koenig
Committed as r264411.

Nope!-(

Well, I made r264412.

Admittedly, the ChangeLogs only would have had a limited effect :-)

Thanks for the heads-up!

Thomas

[Bernhard Reutner-Fischer]

On Tue, 18 Sep 2018 at 22:19, Thomas Koenig <***@netcologne.de> wrote:

s/mamtul/matmul/

thanks,