`/* Implementation of the MATMUL intrinsic
- Copyright 2002 Free Software Foundation, Inc.
+ Copyright 2002, 2005 Free Software Foundation, Inc.
Contributed by Paul Brook <paul@nowt.org>
-This file is part of the GNU Fortran 95 runtime library (libgfor).
+This file is part of the GNU Fortran 95 runtime library (libgfortran).
Libgfortran is free software; you can redistribute it and/or
-modify it under the terms of the GNU Lesser General Public
+modify it under the terms of the GNU General Public
License as published by the Free Software Foundation; either
-version 2.1 of the License, or (at your option) any later version.
+version 2 of the License, or (at your option) any later version.
+
+In addition to the permissions in the GNU General Public License, the
+Free Software Foundation gives you unlimited permission to link the
+compiled version of this file into combinations with other programs,
+and to distribute those combinations without any restriction coming
+from the use of this file. (The General Public License restrictions
+do apply in other respects; for example, they cover modification of
+the file, and distribution when not linked into a combine
+executable.)
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.
+GNU 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. */
+You should have received a copy of the GNU General Public
+License along with libgfortran; see the file COPYING. If not,
+write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+Boston, MA 02110-1301, USA. */
#include "config.h"
#include <stdlib.h>
+#include <string.h>
#include <assert.h>
#include "libgfortran.h"'
-include(types.m4)dnl
-define(rtype_code, regexp(file, `_\([irc][0-9]+\)\.', `\1'))dnl
-define(rtype_letter,substr(rtype_code, 0, 1))dnl
-define(rtype_kind, substr(rtype_code, 1))dnl
-define(rtype,get_arraytype(rtype_letter,rtype_kind))dnl
-define(rtype_name, get_typename(rtype_letter, rtype_kind))dnl
-
-/* Dimensions: retarray(x,y) a(x, count) b(count,y).
- Either a or b can be rank 1. In this case x or y is 1. */
+include(iparm.m4)dnl
+
+`#if defined (HAVE_'rtype_name`)'
+
+/* The order of loops is different in the case of plain matrix
+ multiplication C=MATMUL(A,B), and in the frequent special case where
+ the argument A is the temporary result of a TRANSPOSE intrinsic:
+ C=MATMUL(TRANSPOSE(A),B). Transposed temporaries are detected by
+ looking at their strides.
+
+ The equivalent Fortran pseudo-code is:
+
+ DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
+ IF (.NOT.IS_TRANSPOSED(A)) THEN
+ 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)
+ ELSE
+ DO J=1,N
+ DO I=1,M
+ S = 0
+ DO K=1,COUNT
+ S = S+A(I,K)+B(K,J)
+ C(I,J) = S
+ ENDIF
+*/
+
+extern void matmul_`'rtype_code (rtype * const restrict retarray,
+ rtype * const restrict a, rtype * const restrict b);
+export_proto(matmul_`'rtype_code);
+
void
-`__matmul_'rtype_code (rtype * retarray, rtype * a, rtype * b)
+matmul_`'rtype_code (rtype * const restrict retarray,
+ rtype * const restrict a, rtype * const restrict b)
{
- rtype_name *abase;
- rtype_name *bbase;
- rtype_name *dest;
- rtype_name res;
- index_type rxstride;
- index_type rystride;
- index_type xcount;
- index_type ycount;
- index_type xstride;
- index_type ystride;
- index_type x;
- index_type y;
-
- rtype_name *pa;
- rtype_name *pb;
- index_type astride;
- index_type bstride;
- index_type count;
- index_type n;
+ const rtype_name * restrict abase;
+ const rtype_name * restrict bbase;
+ rtype_name * restrict dest;
+
+ index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
+ index_type x, y, n, count, xcount, ycount;
assert (GFC_DESCRIPTOR_RANK (a) == 2
|| GFC_DESCRIPTOR_RANK (b) == 2);
- abase = a->data;
- bbase = b->data;
- dest = retarray->data;
+
+/* 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.
+ */
+
+ 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_size (sizeof (rtype_name) * size0 ((array_t *) retarray));
+ retarray->offset = 0;
+ }
if (retarray->dim[0].stride == 0)
retarray->dim[0].stride = 1;
+
+ /* This prevents constifying the input arguments. */
if (a->dim[0].stride == 0)
a->dim[0].stride = 1;
if (b->dim[0].stride == 0)
if (GFC_DESCRIPTOR_RANK (retarray) == 1)
{
- rxstride = retarray->dim[0].stride;
- rystride = rxstride;
+ /* 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;
}
else
{
rystride = retarray->dim[1].stride;
}
- /* If we have rank 1 parameters, zero the absent stride, and set the size to
- one. */
+
if (GFC_DESCRIPTOR_RANK (a) == 1)
{
- astride = a->dim[0].stride;
- count = a->dim[0].ubound + 1 - a->dim[0].lbound;
- xstride = 0;
- rxstride = 0;
+ /* Treat it as a a row matrix A[1,count]. */
+ axstride = a->dim[0].stride;
+ aystride = 1;
+
xcount = 1;
+ count = a->dim[0].ubound + 1 - a->dim[0].lbound;
}
else
{
- astride = a->dim[1].stride;
+ axstride = a->dim[0].stride;
+ aystride = a->dim[1].stride;
+
count = a->dim[1].ubound + 1 - a->dim[1].lbound;
- xstride = a->dim[0].stride;
xcount = a->dim[0].ubound + 1 - a->dim[0].lbound;
}
+
+ assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
+
if (GFC_DESCRIPTOR_RANK (b) == 1)
{
- bstride = b->dim[0].stride;
- assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
- ystride = 0;
- rystride = 0;
+ /* 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;
ycount = 1;
}
else
{
- bstride = b->dim[0].stride;
- assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
- ystride = b->dim[1].stride;
+ bxstride = b->dim[0].stride;
+ bystride = b->dim[1].stride;
ycount = b->dim[1].ubound + 1 - b->dim[1].lbound;
}
- for (y = 0; y < ycount; y++)
+ abase = a->data;
+ bbase = b->data;
+ dest = retarray->data;
+
+ if (rxstride == 1 && axstride == 1 && bxstride == 1)
{
- for (x = 0; x < xcount; x++)
- {
- /* Do the summation for this element. For real and integer types
- this is the same as DOT_PRODUCT. For complex types we use do
- a*b, not conjg(a)*b. */
- pa = abase;
- pb = bbase;
- res = 0;
-
- for (n = 0; n < count; n++)
- {
- res += *pa * *pb;
- pa += astride;
- pb += bstride;
- }
-
- *dest = res;
-
- dest += rxstride;
- abase += xstride;
- }
- abase -= xstride * xcount;
- bbase += ystride;
- dest += rystride - (rxstride * xcount);
+ const rtype_name * restrict bbase_y;
+ rtype_name * restrict dest_y;
+ const rtype_name * restrict abase_n;
+ rtype_name bbase_yn;
+
+ if (rystride == xcount)
+ memset (dest, 0, (sizeof (rtype_name) * xcount * ycount));
+ else
+ {
+ for (y = 0; y < ycount; y++)
+ for (x = 0; x < xcount; x++)
+ dest[x + y*rystride] = (rtype_name)0;
+ }
+
+ 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 if (rxstride == 1 && aystride == 1 && bxstride == 1)
+ {
+ const rtype_name *restrict abase_x;
+ const rtype_name *restrict bbase_y;
+ rtype_name *restrict dest_y;
+ rtype_name s;
+
+ for (y = 0; y < ycount; y++)
+ {
+ bbase_y = &bbase[y*bystride];
+ dest_y = &dest[y*rystride];
+ for (x = 0; x < xcount; x++)
+ {
+ abase_x = &abase[x*axstride];
+ s = (rtype_name) 0;
+ for (n = 0; n < count; n++)
+ s += abase_x[n] * bbase_y[n];
+ dest_y[x] = s;
+ }
+ }
+ }
+ else if (axstride < aystride)
+ {
+ 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];
+ }
+ else
+ {
+ const rtype_name *restrict abase_x;
+ const rtype_name *restrict bbase_y;
+ rtype_name *restrict dest_y;
+ rtype_name s;
+
+ for (y = 0; y < ycount; y++)
+ {
+ bbase_y = &bbase[y*bystride];
+ dest_y = &dest[y*rystride];
+ for (x = 0; x < xcount; x++)
+ {
+ abase_x = &abase[x*axstride];
+ s = (rtype_name) 0;
+ for (n = 0; n < count; n++)
+ s += abase_x[n*aystride] * bbase_y[n*bxstride];
+ dest_y[x*rxstride] = s;
+ }
+ }
}
}
+#endif