mirror/gcc
2
0
Fork 0
mirror of git://gcc.gnu.org/git/gcc.git synced 2025-04-23 01:40:32 +08:00
Code Issues Packages Projects Releases Wiki Activity

re PR libfortran/51119 (MATMUL slow for large matrices)

Browse Source 
2017-02-26  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR fortran/51119
	* options.c (gfc_post_options): Set default limit for matmul
	inlining to 30.
	* invoke.texi: Document change.

2017-02-26  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR fortran/51119
	* gfortran.dg/inline_matmul_1.f90: Scan optimized dump instead
	of original.
	* gfortran.dg/inline_matmul_11.f90: Likewise.
	* gfortran.dg/inline_matmul_9.f90: Likewise.
	* gfortran.dg/matmul_13.f90: New test.
	* gfortran.dg/matmul_14.f90: New test.

From-SVN: r245745
This commit is contained in:
Thomas Koenig 2017-02-26 13:22:43 +00:00
parent 462a7b5685
commit c7b608a9f7
9 changed files with 58 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
gcc/fortran/ChangeLog
View File

@ -1,3 +1,10 @@
2017-02-26 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/51119
* options.c (gfc_post_options): Set default limit for matmul
inlining to 30.
* invoke.texi: Document change.
2017-02-25 Dominique d'Humieres <dominiq@lps.ens.fr>
PR fortran/79601

5
gcc/fortran/invoke.texi
View File

@ -1629,9 +1629,8 @@ for matrices with size up to @var{n}. If the matrices involved are not
square, the size comparison is performed using the geometric mean of
the dimensions of the argument and result matrices.
The default value for @var{n} is the value specified for
@code{-fblas-matmul-limit} if this option is specified, or unlimitited
otherwise.
The default value for @var{n} is 30. The @code{-fblas-matmul-limit}
can be used to change this value.
@item -frecursive
@opindex @code{frecursive}

12
gcc/fortran/options.c
View File

@ -388,10 +388,16 @@ gfc_post_options (const char **pfilename)
if (!flag_automatic)
flag_max_stack_var_size = 0;
/* If we call BLAS directly, only inline up to the BLAS limit. */
/* If the user did not specify an inline matmul limit, inline up to the BLAS
limit or up to 30 if no external BLAS is specified. */
if (flag_external_blas && flag_inline_matmul_limit < 0)
flag_inline_matmul_limit = flag_blas_matmul_limit;
if (flag_inline_matmul_limit < 0)
{
if (flag_external_blas)
flag_inline_matmul_limit = flag_blas_matmul_limit;
else
flag_inline_matmul_limit = 30;
}
/* Optimization implies front end optimization, unless the user
specified it directly. */

10
gcc/testsuite/ChangeLog
View File

@ -1,3 +1,13 @@
2017-02-26 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/51119
* gfortran.dg/inline_matmul_1.f90: Scan optimized dump instead
of original.
* gfortran.dg/inline_matmul_11.f90: Likewise.
* gfortran.dg/inline_matmul_9.f90: Likewise.
* gfortran.dg/matmul_13.f90: New test.
* gfortran.dg/matmul_14.f90: New test.
2017-02-25 Jakub Jelinek <jakub@redhat.com>
PR middle-end/79396

4
gcc/testsuite/gfortran.dg/inline_matmul_1.f90
View File

@ -1,5 +1,5 @@
! { dg-do run }
! { dg-options "-ffrontend-optimize -fdump-tree-original -Wrealloc-lhs" }
! { dg-options "-ffrontend-optimize -fdump-tree-optimized -Wrealloc-lhs" }
! PR 37131 - check basic functionality of inlined matmul, making
! sure that the library is not called, with and without reallocation.
@ -149,4 +149,4 @@ program main
end program main
! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "original" } }
! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "optimized" } }

4
gcc/testsuite/gfortran.dg/inline_matmul_11.f90
View File

@ -1,5 +1,5 @@
! { dg-do run }
! { dg-additional-options "-ffrontend-optimize -fdump-tree-original" }
! { dg-additional-options "-ffrontend-optimize -fdump-tree-optimized" }
! PR fortran/66176 - inline conjg for matml.
program main
complex, dimension(3,2) :: a
@ -29,4 +29,4 @@ program main
c = matmul(conjg(a), b)
if (any(conjg(c) /= res2)) call abort
end program main
! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "original" } }
! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "optimized" } }

4
gcc/testsuite/gfortran.dg/inline_matmul_9.f90
View File

@ -1,5 +1,5 @@
! { dg-do run }
! { dg-options "-ffrontend-optimize -fdump-tree-original" }
! { dg-options "-ffrontend-optimize -fdump-tree-optimized" }
! PR 66041 - this used to ICE with an incomplete fix for the PR.
program main
implicit none
@ -21,4 +21,4 @@ program main
if (any (c2-reshape([248., -749.],shape(c2)) /= 0.)) call abort
end program main
! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "original" } }
! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "optimized" } }

12
gcc/testsuite/gfortran.dg/matmul_13.f90 Normal file
View File

@ -0,0 +1,12 @@
! { dg-do compile }
! { dg-options "-O3 -fdump-tree-optimized" }
! Check that the default limit of 30 for inlining matmul applies.
program main
integer, parameter :: n = 31
real, dimension(n,n) :: a, b, c
call random_number(a)
call random_number(b)
c = matmul(a,b)
print *,sum(c)
end program main
! { dg-final { scan-tree-dump-times "_gfortran_matmul_r4" 1 "optimized" } }

12
gcc/testsuite/gfortran.dg/matmul_14.f90 Normal file
View File

@ -0,0 +1,12 @@
! { dg-do compile }
! { dg-options "-O3 -fdump-tree-optimized" }
! Check that the default limit of 30 for inlining matmul applies.
program main
integer, parameter :: n = 30
real, dimension(n,n) :: a, b, c
call random_number(a)
call random_number(b)
c = matmul(a,b)
print *,sum(c)
end program main
! { dg-final { scan-tree-dump-times "_gfortran_matmul_r4" 0 "optimized" } }