/* Implementation of the MATMUL intrinsic
- Copyright 2002 Free Software Foundation, Inc.
+ Copyright 2002, 2005, 2006 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
+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 <assert.h>
#include "libgfortran.h"
-/* 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%.
+#if defined (HAVE_GFC_COMPLEX_4)
+
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+ passed to us by the front-end, in which case we'll call it for large
+ matrices. */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+ const int *, const GFC_COMPLEX_4 *, const GFC_COMPLEX_4 *,
+ const int *, const GFC_COMPLEX_4 *, const int *,
+ const GFC_COMPLEX_4 *, GFC_COMPLEX_4 *, const int *,
+ int, int);
+
+/* 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)
- C = 0
- DO J=1,N
- DO K=1,COUNT
+ 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
- C(I,J) = C(I,J)+A(I,K)*B(K,J)
+ S = 0
+ DO K=1,COUNT
+ S = S+A(I,K)*B(K,J)
+ C(I,J) = S
+ ENDIF
*/
-extern void matmul_c4 (gfc_array_c4 * retarray, gfc_array_c4 * a, gfc_array_c4 * b);
+/* If try_blas is set to a nonzero value, then the matmul function will
+ see if there is a way to perform the matrix multiplication by a call
+ to the BLAS gemm function. */
+
+extern void matmul_c4 (gfc_array_c4 * const restrict retarray,
+ gfc_array_c4 * const restrict a, gfc_array_c4 * const restrict b, int try_blas,
+ int blas_limit, blas_call gemm);
export_proto(matmul_c4);
void
-matmul_c4 (gfc_array_c4 * retarray, gfc_array_c4 * a, gfc_array_c4 * b)
+matmul_c4 (gfc_array_c4 * const restrict retarray,
+ gfc_array_c4 * const restrict a, gfc_array_c4 * const restrict b, int try_blas,
+ int blas_limit, blas_call gemm)
{
- GFC_COMPLEX_4 *abase;
- GFC_COMPLEX_4 *bbase;
- GFC_COMPLEX_4 *dest;
+ const GFC_COMPLEX_4 * restrict abase;
+ const GFC_COMPLEX_4 * restrict bbase;
+ GFC_COMPLEX_4 * restrict dest;
index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
index_type x, y, n, count, xcount, ycount;
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 (GFC_COMPLEX_4) * size0 (retarray));
- retarray->base = 0;
+ = internal_malloc_size (sizeof (GFC_COMPLEX_4) * size0 ((array_t *) retarray));
+ retarray->offset = 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;
-
if (GFC_DESCRIPTOR_RANK (retarray) == 1)
{
/* 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;
+ bystride = 0xDEADBEEF;
ycount = 1;
}
else
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;
+
+ /* Now that everything is set up, we're performing the multiplication
+ itself. */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+ if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+ && (bxstride == 1 || bystride == 1)
+ && (((float) xcount) * ((float) ycount) * ((float) count)
+ > POW3(blas_limit)))
+ {
+ const int m = xcount, n = ycount, k = count, ldc = rystride;
+ const GFC_COMPLEX_4 one = 1, zero = 0;
+ const int lda = (axstride == 1) ? aystride : axstride,
+ ldb = (bxstride == 1) ? bystride : bxstride;
+
+ if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+ {
+ assert (gemm != NULL);
+ gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+ &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+ return;
+ }
+ }
+
if (rxstride == 1 && axstride == 1 && bxstride == 1)
{
- GFC_COMPLEX_4 *bbase_y;
- GFC_COMPLEX_4 *dest_y;
- GFC_COMPLEX_4 *abase_n;
+ const GFC_COMPLEX_4 * restrict bbase_y;
+ GFC_COMPLEX_4 * restrict dest_y;
+ const GFC_COMPLEX_4 * restrict abase_n;
GFC_COMPLEX_4 bbase_yn;
- memset (dest, 0, (sizeof (GFC_COMPLEX_4) * size0(retarray)));
+ if (rystride == xcount)
+ memset (dest, 0, (sizeof (GFC_COMPLEX_4) * xcount * ycount));
+ else
+ {
+ for (y = 0; y < ycount; y++)
+ for (x = 0; x < xcount; x++)
+ dest[x + y*rystride] = (GFC_COMPLEX_4)0;
+ }
for (y = 0; y < ycount; y++)
{
}
}
}
- else
+ else if (rxstride == 1 && aystride == 1 && bxstride == 1)
+ {
+ if (GFC_DESCRIPTOR_RANK (a) != 1)
+ {
+ const GFC_COMPLEX_4 *restrict abase_x;
+ const GFC_COMPLEX_4 *restrict bbase_y;
+ GFC_COMPLEX_4 *restrict dest_y;
+ GFC_COMPLEX_4 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 = (GFC_COMPLEX_4) 0;
+ for (n = 0; n < count; n++)
+ s += abase_x[n] * bbase_y[n];
+ dest_y[x] = s;
+ }
+ }
+ }
+ else
+ {
+ const GFC_COMPLEX_4 *restrict bbase_y;
+ GFC_COMPLEX_4 s;
+
+ for (y = 0; y < ycount; y++)
+ {
+ bbase_y = &bbase[y*bystride];
+ s = (GFC_COMPLEX_4) 0;
+ for (n = 0; n < count; n++)
+ s += abase[n*axstride] * bbase_y[n];
+ dest[y*rystride] = s;
+ }
+ }
+ }
+ else if (axstride < aystride)
{
for (y = 0; y < ycount; y++)
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 if (GFC_DESCRIPTOR_RANK (a) == 1)
+ {
+ const GFC_COMPLEX_4 *restrict bbase_y;
+ GFC_COMPLEX_4 s;
+
+ for (y = 0; y < ycount; y++)
+ {
+ bbase_y = &bbase[y*bystride];
+ s = (GFC_COMPLEX_4) 0;
+ for (n = 0; n < count; n++)
+ s += abase[n*axstride] * bbase_y[n*bxstride];
+ dest[y*rxstride] = s;
+ }
+ }
+ else
+ {
+ const GFC_COMPLEX_4 *restrict abase_x;
+ const GFC_COMPLEX_4 *restrict bbase_y;
+ GFC_COMPLEX_4 *restrict dest_y;
+ GFC_COMPLEX_4 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 = (GFC_COMPLEX_4) 0;
+ for (n = 0; n < count; n++)
+ s += abase_x[n*aystride] * bbase_y[n*bxstride];
+ dest_y[x*rxstride] = s;
+ }
+ }
+ }
}
+
+#endif