]> git.ipfire.org Git - thirdparty/gcc.git/blob - libgfortran/m4/matmul.m4
re PR libfortran/19308 (I/O library should support more real and integer kinds)
[thirdparty/gcc.git] / libgfortran / m4 / matmul.m4
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 include(iparm.m4)dnl
37
38 `#if defined (HAVE_'rtype_name`)'
39
40 /* This is a C version of the following fortran pseudo-code. The key
41 point is the loop order -- we access all arrays column-first, which
42 improves the performance enough to boost galgel spec score by 50%.
43
44 DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
45 C = 0
46 DO J=1,N
47 DO K=1,COUNT
48 DO I=1,M
49 C(I,J) = C(I,J)+A(I,K)*B(K,J)
50 */
51
52 extern void matmul_`'rtype_code (rtype * retarray, rtype * a, rtype * b);
53 export_proto(matmul_`'rtype_code);
54
55 void
56 matmul_`'rtype_code (rtype * retarray, rtype * a, rtype * b)
57 {
58 rtype_name *abase;
59 rtype_name *bbase;
60 rtype_name *dest;
61
62 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
63 index_type x, y, n, count, xcount, ycount;
64
65 assert (GFC_DESCRIPTOR_RANK (a) == 2
66 || GFC_DESCRIPTOR_RANK (b) == 2);
67
68 /* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
69
70 Either A or B (but not both) can be rank 1:
71
72 o One-dimensional argument A is implicitly treated as a row matrix
73 dimensioned [1,count], so xcount=1.
74
75 o One-dimensional argument B is implicitly treated as a column matrix
76 dimensioned [count, 1], so ycount=1.
77 */
78
79 if (retarray->data == NULL)
80 {
81 if (GFC_DESCRIPTOR_RANK (a) == 1)
82 {
83 retarray->dim[0].lbound = 0;
84 retarray->dim[0].ubound = b->dim[1].ubound - b->dim[1].lbound;
85 retarray->dim[0].stride = 1;
86 }
87 else if (GFC_DESCRIPTOR_RANK (b) == 1)
88 {
89 retarray->dim[0].lbound = 0;
90 retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
91 retarray->dim[0].stride = 1;
92 }
93 else
94 {
95 retarray->dim[0].lbound = 0;
96 retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
97 retarray->dim[0].stride = 1;
98
99 retarray->dim[1].lbound = 0;
100 retarray->dim[1].ubound = b->dim[1].ubound - b->dim[1].lbound;
101 retarray->dim[1].stride = retarray->dim[0].ubound+1;
102 }
103
104 retarray->data
105 = internal_malloc_size (sizeof (rtype_name) * size0 ((array_t *) retarray));
106 retarray->offset = 0;
107 }
108
109 abase = a->data;
110 bbase = b->data;
111 dest = retarray->data;
112
113 if (retarray->dim[0].stride == 0)
114 retarray->dim[0].stride = 1;
115 if (a->dim[0].stride == 0)
116 a->dim[0].stride = 1;
117 if (b->dim[0].stride == 0)
118 b->dim[0].stride = 1;
119
120 sinclude(`matmul_asm_'rtype_code`.m4')dnl
121
122 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
123 {
124 /* One-dimensional result may be addressed in the code below
125 either as a row or a column matrix. We want both cases to
126 work. */
127 rxstride = rystride = retarray->dim[0].stride;
128 }
129 else
130 {
131 rxstride = retarray->dim[0].stride;
132 rystride = retarray->dim[1].stride;
133 }
134
135
136 if (GFC_DESCRIPTOR_RANK (a) == 1)
137 {
138 /* Treat it as a a row matrix A[1,count]. */
139 axstride = a->dim[0].stride;
140 aystride = 1;
141
142 xcount = 1;
143 count = a->dim[0].ubound + 1 - a->dim[0].lbound;
144 }
145 else
146 {
147 axstride = a->dim[0].stride;
148 aystride = a->dim[1].stride;
149
150 count = a->dim[1].ubound + 1 - a->dim[1].lbound;
151 xcount = a->dim[0].ubound + 1 - a->dim[0].lbound;
152 }
153
154 assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
155
156 if (GFC_DESCRIPTOR_RANK (b) == 1)
157 {
158 /* Treat it as a column matrix B[count,1] */
159 bxstride = b->dim[0].stride;
160
161 /* bystride should never be used for 1-dimensional b.
162 in case it is we want it to cause a segfault, rather than
163 an incorrect result. */
164 bystride = 0xDEADBEEF;
165 ycount = 1;
166 }
167 else
168 {
169 bxstride = b->dim[0].stride;
170 bystride = b->dim[1].stride;
171 ycount = b->dim[1].ubound + 1 - b->dim[1].lbound;
172 }
173
174 abase = a->data;
175 bbase = b->data;
176 dest = retarray->data;
177
178 if (rxstride == 1 && axstride == 1 && bxstride == 1)
179 {
180 rtype_name *bbase_y;
181 rtype_name *dest_y;
182 rtype_name *abase_n;
183 rtype_name bbase_yn;
184
185 if (rystride == ycount)
186 memset (dest, 0, (sizeof (rtype_name) * size0((array_t *) retarray)));
187 else
188 {
189 for (y = 0; y < ycount; y++)
190 for (x = 0; x < xcount; x++)
191 dest[x + y*rystride] = (rtype_name)0;
192 }
193
194 for (y = 0; y < ycount; y++)
195 {
196 bbase_y = bbase + y*bystride;
197 dest_y = dest + y*rystride;
198 for (n = 0; n < count; n++)
199 {
200 abase_n = abase + n*aystride;
201 bbase_yn = bbase_y[n];
202 for (x = 0; x < xcount; x++)
203 {
204 dest_y[x] += abase_n[x] * bbase_yn;
205 }
206 }
207 }
208 }
209 else
210 {
211 for (y = 0; y < ycount; y++)
212 for (x = 0; x < xcount; x++)
213 dest[x*rxstride + y*rystride] = (rtype_name)0;
214
215 for (y = 0; y < ycount; y++)
216 for (n = 0; n < count; n++)
217 for (x = 0; x < xcount; x++)
218 /* dest[x,y] += a[x,n] * b[n,y] */
219 dest[x*rxstride + y*rystride] += abase[x*axstride + n*aystride] * bbase[n*bxstride + y*bystride];
220 }
221 }
222
223 #endif