gcc/libgfortran/m4/matmul.m4

208 lines
5.7 KiB
Plaintext
Raw Normal View History

`/* Implementation of the MATMUL intrinsic
Copyright 2002 Free Software Foundation, Inc.
Contributed by Paul Brook <paul@nowt.org>
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
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>
#include <string.h>
#include <assert.h>
#include "libgfortran.h"'
include(iparm.m4)dnl
/* 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)
*/
acinclude.m4 (LIBGFOR_CHECK_ATTRIBUTE_VISIBILITY): New. * acinclude.m4 (LIBGFOR_CHECK_ATTRIBUTE_VISIBILITY): New. (LIBGFOR_CHECK_ATTRIBUTE_DLLEXPORT): New. (LIBGFOR_CHECK_ATTRIBUTE_ALIAS): New. * configure.ac: Use them. * configure, config.h.in, aclocal.m4: Rebuild. * libgfortran.h (prefix): Remove. (PREFIX, IPREFIX): New. (sym_rename, sym_rename1, sym_rename2): New. (internal_proto, export_proto, export_proto_np): New. (iexport_proto, iexport): New. (iexport_data_proto, iexport_data): New. * intrinsics/abort.c, intrinsics/args.c, intrinsics/associated.c, intrinsics/cpu_time.c, intrinsics/cshift0.c, intrinsics/date_and_time.c, intrinsics/env.c, intrinsics/eoshift0.c, intrinsics/eoshift2.c, intrinsics/etime.c, intrinsics/exit.c, intrinsics/flush.c, intrinsics/fnum.c, intrinsics/getXid.c, intrinsics/getcwd.c, intrinsics/ishftc.c, intrinsics/mvbits.c, intrinsics/pack_generic.c, intrinsics/rand.c, intrinsics/random.c, intrinsics/reshape_generic.c, intrinsics/size.c, intrinsics/spread_generic.c, intrinsics/stat.c, intrinsics/string_intrinsics.c, intrinsics/system.c, intrinsics/system_clock.c, intrinsics/transpose_generic.c, intrinsics/umask.c, intrinsics/unlink.c, intrinsics/unpack_generic.c, io/backspace.c, io/close.c, io/endfile.c, io/inquire.c, io/io.h, io/open.c, io/rewind.c, io/transfer.c, libgfortran.h, m4/cshift1.m4, m4/dotprod.m4, m4/dotprodc.m4, m4/dotprodl.m4, m4/eoshift1.m4, m4/eoshift3.m4, m4/exponent.m4, m4/fraction.m4, m4/iforeach.m4, m4/ifunction.m4, m4/matmul.m4, m4/matmull.m4, m4/nearest.m4, m4/pow.m4, m4/reshape.m4, m4/set_exponent.m4, m4/shape.m4, m4/transpose.m4, runtime/environ.c, runtime/error.c, runtime/in_pack_generic.c, runtime/in_unpack_generic.c, runtime/main.c, runtime/memory.c, runtime/pause.c, runtime/select.c, runtime/stop.c: Use them to mark symbols internal or external. * generated/*: Rebuild. From-SVN: r92045
2004-12-12 09:59:05 +01:00
extern void `__matmul_'rtype_code (rtype * retarray, rtype * a, rtype * b);
export_proto_np(`__matmul_'rtype_code);
void
`__matmul_'rtype_code (rtype * retarray, rtype * a, rtype * b)
{
rtype_name *abase;
rtype_name *bbase;
rtype_name *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);
/* 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 (retarray));
retarray->base = 0;
}
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)
{
/* 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
{
rxstride = retarray->dim[0].stride;
rystride = retarray->dim[1].stride;
}
if (GFC_DESCRIPTOR_RANK (a) == 1)
{
/* 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
{
axstride = a->dim[0].stride;
aystride = a->dim[1].stride;
count = a->dim[1].ubound + 1 - a->dim[1].lbound;
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)
{
/* 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
{
bxstride = b->dim[0].stride;
bystride = b->dim[1].stride;
ycount = b->dim[1].ubound + 1 - b->dim[1].lbound;
}
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)
{
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];
}
}