2004-05-13 08:41:07 +02:00
|
|
|
`/* Implementation of the MATMUL intrinsic
|
|
|
|
Copyright 2002 Free Software Foundation, Inc.
|
|
|
|
Contributed by Paul Brook <paul@nowt.org>
|
|
|
|
|
2004-08-09 16:34:39 +02:00
|
|
|
This file is part of the GNU Fortran 95 runtime library (libgfortran).
|
2004-05-13 08:41:07 +02:00
|
|
|
|
|
|
|
Libgfortran is free software; you can redistribute it and/or
|
|
|
|
modify it under the terms of the GNU Lesser General Public
|
|
|
|
License as published by the Free Software Foundation; either
|
|
|
|
version 2.1 of the License, or (at your option) any later version.
|
|
|
|
|
|
|
|
Libgfortran is distributed in the hope that it will be useful,
|
|
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
GNU Lesser General Public License for more details.
|
|
|
|
|
|
|
|
You should have received a copy of the GNU Lesser General Public
|
|
|
|
License along with libgfor; see the file COPYING.LIB. If not,
|
|
|
|
write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
|
|
|
|
Boston, MA 02111-1307, USA. */
|
|
|
|
|
|
|
|
#include "config.h"
|
|
|
|
#include <stdlib.h>
|
2004-11-18 09:45:11 +01:00
|
|
|
#include <string.h>
|
2004-05-13 08:41:07 +02:00
|
|
|
#include <assert.h>
|
|
|
|
#include "libgfortran.h"'
|
Makefile.am: Remove references to types.m4.
* Makefile.am: Remove references to types.m4.
* m4/iparm.m4: Merge with types.m4.
* m4/types.m4: Remove.
* m4/cshift1.m4, m4/dotprod.m4, m4/dotprodc.m4, m4/dotprodl.m4,
m4/eoshift1.m4, m4/eoshift3.m4, m4/iforeach.m4, m4/ifunction.m4,
m4/in_pack.m4, m4/in_unpack.m4, m4/iparm.m4, m4/matmul.m4,
m4/matmull.m4, m4/maxloc0.m4, m4/maxloc1.m4, m4/maxval.m4,
m4/minloc0.m4, m4/minloc1.m4, m4/minval.m4, m4/reshape.m4,
m4/shape.m4, m4/specific.m4, m4/specific2.m4, m4/transpose.m4):
Update to use new iparm.m4.
* generated/*.c: Regenerate.
From-SVN: r82003
2004-05-18 21:03:26 +02:00
|
|
|
include(iparm.m4)dnl
|
2004-05-13 08:41:07 +02:00
|
|
|
|
2004-11-18 09:45:11 +01:00
|
|
|
/* This is a C version of the following fortran pseudo-code. The key
|
|
|
|
point is the loop order -- we access all arrays column-first, which
|
|
|
|
improves the performance enough to boost galgel spec score by 50%.
|
|
|
|
|
|
|
|
DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
|
|
|
|
C = 0
|
|
|
|
DO J=1,N
|
|
|
|
DO K=1,COUNT
|
|
|
|
DO I=1,M
|
|
|
|
C(I,J) = C(I,J)+A(I,K)*B(K,J)
|
|
|
|
*/
|
|
|
|
|
2004-05-13 08:41:07 +02:00
|
|
|
void
|
|
|
|
`__matmul_'rtype_code (rtype * retarray, rtype * a, rtype * b)
|
|
|
|
{
|
|
|
|
rtype_name *abase;
|
|
|
|
rtype_name *bbase;
|
|
|
|
rtype_name *dest;
|
2004-11-18 09:45:11 +01:00
|
|
|
|
|
|
|
index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
|
|
|
|
index_type x, y, n, count, xcount, ycount;
|
2004-05-13 08:41:07 +02:00
|
|
|
|
|
|
|
assert (GFC_DESCRIPTOR_RANK (a) == 2
|
|
|
|
|| GFC_DESCRIPTOR_RANK (b) == 2);
|
2004-08-09 16:34:39 +02:00
|
|
|
|
2004-11-18 09:45:11 +01:00
|
|
|
/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
|
|
|
|
|
|
|
|
Either A or B (but not both) can be rank 1:
|
|
|
|
|
|
|
|
o One-dimensional argument A is implicitly treated as a row matrix
|
|
|
|
dimensioned [1,count], so xcount=1.
|
|
|
|
|
|
|
|
o One-dimensional argument B is implicitly treated as a column matrix
|
|
|
|
dimensioned [count, 1], so ycount=1.
|
|
|
|
*/
|
|
|
|
|
2004-08-09 16:34:39 +02:00
|
|
|
if (retarray->data == NULL)
|
|
|
|
{
|
|
|
|
if (GFC_DESCRIPTOR_RANK (a) == 1)
|
|
|
|
{
|
|
|
|
retarray->dim[0].lbound = 0;
|
|
|
|
retarray->dim[0].ubound = b->dim[1].ubound - b->dim[1].lbound;
|
|
|
|
retarray->dim[0].stride = 1;
|
|
|
|
}
|
|
|
|
else if (GFC_DESCRIPTOR_RANK (b) == 1)
|
|
|
|
{
|
|
|
|
retarray->dim[0].lbound = 0;
|
|
|
|
retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
|
|
|
|
retarray->dim[0].stride = 1;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
retarray->dim[0].lbound = 0;
|
|
|
|
retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
|
|
|
|
retarray->dim[0].stride = 1;
|
|
|
|
|
|
|
|
retarray->dim[1].lbound = 0;
|
|
|
|
retarray->dim[1].ubound = b->dim[1].ubound - b->dim[1].lbound;
|
|
|
|
retarray->dim[1].stride = retarray->dim[0].ubound+1;
|
|
|
|
}
|
|
|
|
|
|
|
|
retarray->data = internal_malloc (sizeof (rtype_name) * size0 (retarray));
|
|
|
|
retarray->base = 0;
|
|
|
|
}
|
|
|
|
|
2004-05-13 08:41:07 +02:00
|
|
|
abase = a->data;
|
|
|
|
bbase = b->data;
|
|
|
|
dest = retarray->data;
|
|
|
|
|
|
|
|
if (retarray->dim[0].stride == 0)
|
|
|
|
retarray->dim[0].stride = 1;
|
|
|
|
if (a->dim[0].stride == 0)
|
|
|
|
a->dim[0].stride = 1;
|
|
|
|
if (b->dim[0].stride == 0)
|
|
|
|
b->dim[0].stride = 1;
|
|
|
|
|
|
|
|
sinclude(`matmul_asm_'rtype_code`.m4')dnl
|
|
|
|
|
|
|
|
if (GFC_DESCRIPTOR_RANK (retarray) == 1)
|
|
|
|
{
|
2004-11-18 09:45:11 +01:00
|
|
|
/* One-dimensional result may be addressed in the code below
|
|
|
|
either as a row or a column matrix. We want both cases to
|
|
|
|
work. */
|
|
|
|
rxstride = rystride = retarray->dim[0].stride;
|
2004-05-13 08:41:07 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
rxstride = retarray->dim[0].stride;
|
|
|
|
rystride = retarray->dim[1].stride;
|
|
|
|
}
|
|
|
|
|
2004-11-18 09:45:11 +01:00
|
|
|
|
2004-05-13 08:41:07 +02:00
|
|
|
if (GFC_DESCRIPTOR_RANK (a) == 1)
|
|
|
|
{
|
2004-11-18 09:45:11 +01:00
|
|
|
/* Treat it as a a row matrix A[1,count]. */
|
|
|
|
axstride = a->dim[0].stride;
|
|
|
|
aystride = 1;
|
|
|
|
|
2004-05-13 08:41:07 +02:00
|
|
|
xcount = 1;
|
2004-11-18 09:45:11 +01:00
|
|
|
count = a->dim[0].ubound + 1 - a->dim[0].lbound;
|
2004-05-13 08:41:07 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2004-11-18 09:45:11 +01:00
|
|
|
axstride = a->dim[0].stride;
|
|
|
|
aystride = a->dim[1].stride;
|
|
|
|
|
2004-05-13 08:41:07 +02:00
|
|
|
count = a->dim[1].ubound + 1 - a->dim[1].lbound;
|
|
|
|
xcount = a->dim[0].ubound + 1 - a->dim[0].lbound;
|
|
|
|
}
|
2004-11-18 09:45:11 +01:00
|
|
|
|
|
|
|
assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
|
|
|
|
|
2004-05-13 08:41:07 +02:00
|
|
|
if (GFC_DESCRIPTOR_RANK (b) == 1)
|
|
|
|
{
|
2004-11-18 09:45:11 +01:00
|
|
|
/* Treat it as a column matrix B[count,1] */
|
|
|
|
bxstride = b->dim[0].stride;
|
|
|
|
|
|
|
|
/* bystride should never be used for 1-dimensional b.
|
|
|
|
in case it is we want it to cause a segfault, rather than
|
|
|
|
an incorrect result. */
|
|
|
|
bystride = 0xDEADBEEF;
|
2004-05-13 08:41:07 +02:00
|
|
|
ycount = 1;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2004-11-18 09:45:11 +01:00
|
|
|
bxstride = b->dim[0].stride;
|
|
|
|
bystride = b->dim[1].stride;
|
2004-05-13 08:41:07 +02:00
|
|
|
ycount = b->dim[1].ubound + 1 - b->dim[1].lbound;
|
|
|
|
}
|
|
|
|
|
2004-11-18 09:45:11 +01:00
|
|
|
assert (a->base == 0);
|
|
|
|
assert (b->base == 0);
|
|
|
|
assert (retarray->base == 0);
|
|
|
|
|
|
|
|
abase = a->data;
|
|
|
|
bbase = b->data;
|
|
|
|
dest = retarray->data;
|
|
|
|
|
|
|
|
if (rxstride == 1 && axstride == 1 && bxstride == 1)
|
2004-05-13 08:41:07 +02:00
|
|
|
{
|
2004-11-18 09:45:11 +01:00
|
|
|
rtype_name *bbase_y;
|
|
|
|
rtype_name *dest_y;
|
|
|
|
rtype_name *abase_n;
|
|
|
|
rtype_name bbase_yn;
|
|
|
|
|
|
|
|
memset (dest, 0, (sizeof (rtype_name) * size0(retarray)));
|
|
|
|
|
|
|
|
for (y = 0; y < ycount; y++)
|
|
|
|
{
|
|
|
|
bbase_y = bbase + y*bystride;
|
|
|
|
dest_y = dest + y*rystride;
|
|
|
|
for (n = 0; n < count; n++)
|
|
|
|
{
|
|
|
|
abase_n = abase + n*aystride;
|
|
|
|
bbase_yn = bbase_y[n];
|
|
|
|
for (x = 0; x < xcount; x++)
|
|
|
|
{
|
|
|
|
dest_y[x] += abase_n[x] * bbase_yn;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
for (y = 0; y < ycount; y++)
|
|
|
|
for (x = 0; x < xcount; x++)
|
|
|
|
dest[x*rxstride + y*rystride] = (rtype_name)0;
|
|
|
|
|
|
|
|
for (y = 0; y < ycount; y++)
|
|
|
|
for (n = 0; n < count; n++)
|
|
|
|
for (x = 0; x < xcount; x++)
|
|
|
|
/* dest[x,y] += a[x,n] * b[n,y] */
|
|
|
|
dest[x*rxstride + y*rystride] += abase[x*axstride + n*aystride] * bbase[n*bxstride + y*bystride];
|
2004-05-13 08:41:07 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|