[thirdparty/gcc.git] / libgfortran / generated / matmul_c16.c

/* Implementation of the MATMUL intrinsic
   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 (libgfortran).

Libgfortran is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public
License as published by the Free Software Foundation; either
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 General Public License for more details.

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"

#if defined (HAVE_GFC_COMPLEX_16)

/* 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)
*/

extern void matmul_c16 (gfc_array_c16 * retarray, gfc_array_c16 * a, gfc_array_c16 * b);
export_proto(matmul_c16);

void
matmul_c16 (gfc_array_c16 * retarray, gfc_array_c16 * a, gfc_array_c16 * b)
{
  GFC_COMPLEX_16 *abase;
  GFC_COMPLEX_16 *bbase;
  GFC_COMPLEX_16 *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 (GFC_COMPLEX_16) * 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)
    {
      /* 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;
    }

  abase = a->data;
  bbase = b->data;
  dest = retarray->data;

  if (rxstride == 1 && axstride == 1 && bxstride == 1)
    {
      GFC_COMPLEX_16 *bbase_y;
      GFC_COMPLEX_16 *dest_y;
      GFC_COMPLEX_16 *abase_n;
      GFC_COMPLEX_16 bbase_yn;

      if (rystride == ycount)
	memset (dest, 0, (sizeof (GFC_COMPLEX_16) * size0((array_t *) retarray)));
      else
	{
	  for (y = 0; y < ycount; y++)
	    for (x = 0; x < xcount; x++)
	      dest[x + y*rystride] = (GFC_COMPLEX_16)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
    {
      for (y = 0; y < ycount; y++)
	for (x = 0; x < xcount; x++)
	  dest[x*rxstride + y*rystride] = (GFC_COMPLEX_16)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];
    }
}

#endif
Commit	Line	Data
644cb69f FXC	1	/* Implementation of the MATMUL intrinsic
	2	Copyright 2002, 2005 Free Software Foundation, Inc.
	3	Contributed by Paul Brook <paul@nowt.org>
	4
	5	This file is part of the GNU Fortran 95 runtime library (libgfortran).
	6
	7	Libgfortran is free software; you can redistribute it and/or
	8	modify it under the terms of the GNU General Public
	9	License as published by the Free Software Foundation; either
	10	version 2 of the License, or (at your option) any later version.
	11
	12	In addition to the permissions in the GNU General Public License, the
	13	Free Software Foundation gives you unlimited permission to link the
	14	compiled version of this file into combinations with other programs,
	15	and to distribute those combinations without any restriction coming
	16	from the use of this file. (The General Public License restrictions
	17	do apply in other respects; for example, they cover modification of
	18	the file, and distribution when not linked into a combine
	19	executable.)
	20
	21	Libgfortran is distributed in the hope that it will be useful,
	22	but WITHOUT ANY WARRANTY; without even the implied warranty of
	23	MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	24	GNU General Public License for more details.
	25
	26	You should have received a copy of the GNU General Public
	27	License along with libgfortran; see the file COPYING. If not,
	28	write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
	29	Boston, MA 02110-1301, USA. */
	30
	31	#include "config.h"
	32	#include <stdlib.h>
	33	#include <string.h>
	34	#include <assert.h>
	35	#include "libgfortran.h"
	36
	37	#if defined (HAVE_GFC_COMPLEX_16)
	38
	39	/* This is a C version of the following fortran pseudo-code. The key
	40	point is the loop order -- we access all arrays column-first, which
	41	improves the performance enough to boost galgel spec score by 50%.
	42
	43	DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
	44	C = 0
	45	DO J=1,N
	46	DO K=1,COUNT
	47	DO I=1,M
	48	C(I,J) = C(I,J)+A(I,K)*B(K,J)
	49	*/
	50
	51	extern void matmul_c16 (gfc_array_c16 * retarray, gfc_array_c16 * a, gfc_array_c16 * b);
	52	export_proto(matmul_c16);
	53
	54	void
	55	matmul_c16 (gfc_array_c16 * retarray, gfc_array_c16 * a, gfc_array_c16 * b)
	56	{
	57	GFC_COMPLEX_16 *abase;
	58	GFC_COMPLEX_16 *bbase;
	59	GFC_COMPLEX_16 *dest;
	60
	61	index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
	62	index_type x, y, n, count, xcount, ycount;
	63
	64	assert (GFC_DESCRIPTOR_RANK (a) == 2
65	\|\| GFC_DESCRIPTOR_RANK (b) == 2);
66
67	/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
68
69	Either A or B (but not both) can be rank 1:
70
71	o One-dimensional argument A is implicitly treated as a row matrix
72	dimensioned [1,count], so xcount=1.
73
74	o One-dimensional argument B is implicitly treated as a column matrix
75	dimensioned [count, 1], so ycount=1.
76	*/
77
78	if (retarray->data == NULL)
79	{
80	if (GFC_DESCRIPTOR_RANK (a) == 1)
81	{
82	retarray->dim[0].lbound = 0;
83	retarray->dim[0].ubound = b->dim[1].ubound - b->dim[1].lbound;
84	retarray->dim[0].stride = 1;
85	}
86	else if (GFC_DESCRIPTOR_RANK (b) == 1)
87	{
88	retarray->dim[0].lbound = 0;
89	retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
90	retarray->dim[0].stride = 1;
91	}
92	else
93	{
94	retarray->dim[0].lbound = 0;
95	retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
96	retarray->dim[0].stride = 1;
97
98	retarray->dim[1].lbound = 0;
99	retarray->dim[1].ubound = b->dim[1].ubound - b->dim[1].lbound;
100	retarray->dim[1].stride = retarray->dim[0].ubound+1;
101	}
102
103	retarray->data
104	= internal_malloc_size (sizeof (GFC_COMPLEX_16) * size0 ((array_t *) retarray));
105	retarray->offset = 0;
106	}
107
108	abase = a->data;
109	bbase = b->data;
110	dest = retarray->data;
111
112	if (retarray->dim[0].stride == 0)
113	retarray->dim[0].stride = 1;
114	if (a->dim[0].stride == 0)
115	a->dim[0].stride = 1;
116	if (b->dim[0].stride == 0)
117	b->dim[0].stride = 1;
118
119
120	if (GFC_DESCRIPTOR_RANK (retarray) == 1)
121	{
122	/* One-dimensional result may be addressed in the code below
123	either as a row or a column matrix. We want both cases to
124	work. */
125	rxstride = rystride = retarray->dim[0].stride;
126	}
127	else
128	{
129	rxstride = retarray->dim[0].stride;
130	rystride = retarray->dim[1].stride;
131	}
132
133
134	if (GFC_DESCRIPTOR_RANK (a) == 1)
135	{
136	/* Treat it as a a row matrix A[1,count]. */
137	axstride = a->dim[0].stride;
138	aystride = 1;
139
140	xcount = 1;
141	count = a->dim[0].ubound + 1 - a->dim[0].lbound;
142	}
143	else
144	{
145	axstride = a->dim[0].stride;
146	aystride = a->dim[1].stride;
147
148	count = a->dim[1].ubound + 1 - a->dim[1].lbound;
149	xcount = a->dim[0].ubound + 1 - a->dim[0].lbound;
150	}
151
152	assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
153
154	if (GFC_DESCRIPTOR_RANK (b) == 1)
155	{
156	/* Treat it as a column matrix B[count,1] */
157	bxstride = b->dim[0].stride;
158
159	/* bystride should never be used for 1-dimensional b.
160	in case it is we want it to cause a segfault, rather than
161	an incorrect result. */
162	bystride = 0xDEADBEEF;
163	ycount = 1;
164	}
165	else
166	{
167	bxstride = b->dim[0].stride;
168	bystride = b->dim[1].stride;
169	ycount = b->dim[1].ubound + 1 - b->dim[1].lbound;
170	}
171
172	abase = a->data;
173	bbase = b->data;
174	dest = retarray->data;
175
176	if (rxstride == 1 && axstride == 1 && bxstride == 1)
177	{
178	GFC_COMPLEX_16 *bbase_y;
179	GFC_COMPLEX_16 *dest_y;
180	GFC_COMPLEX_16 *abase_n;
181	GFC_COMPLEX_16 bbase_yn;
182
183	if (rystride == ycount)
184	memset (dest, 0, (sizeof (GFC_COMPLEX_16) * size0((array_t *) retarray)));
185	else
186	{
187	for (y = 0; y < ycount; y++)
188	for (x = 0; x < xcount; x++)
189	dest[x + y*rystride] = (GFC_COMPLEX_16)0;
190	}
191
192	for (y = 0; y < ycount; y++)
193	{
194	bbase_y = bbase + y*bystride;
195	dest_y = dest + y*rystride;
196	for (n = 0; n < count; n++)
197	{
198	abase_n = abase + n*aystride;
199	bbase_yn = bbase_y[n];
200	for (x = 0; x < xcount; x++)
201	{
202	dest_y[x] += abase_n[x] * bbase_yn;
203	}
204	}
205	}
206	}
207	else
208	{
209	for (y = 0; y < ycount; y++)
210	for (x = 0; x < xcount; x++)
211	dest[xrxstride + yrystride] = (GFC_COMPLEX_16)0;
212
213	for (y = 0; y < ycount; y++)
214	for (n = 0; n < count; n++)
215	for (x = 0; x < xcount; x++)
216	/* dest[x,y] += a[x,n] * b[n,y] */
217	dest[xrxstride + yrystride] += abase[xaxstride + naystride] * bbase[nbxstride + ybystride];
218	}
219	}
220
221	#endif