]>
Commit | Line | Data |
---|---|---|
6de9cd9a DN |
1 | /* Implementation of the MATMUL intrinsic |
2 | Copyright 2002 Free Software Foundation, Inc. | |
3 | Contributed by Paul Brook <paul@nowt.org> | |
4 | ||
883c9d4d | 5 | This file is part of the GNU Fortran 95 runtime library (libgfortran). |
6de9cd9a DN |
6 | |
7 | Libgfortran is free software; you can redistribute it and/or | |
57dea9f6 | 8 | modify it under the terms of the GNU General Public |
6de9cd9a | 9 | License as published by the Free Software Foundation; either |
57dea9f6 TM |
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.) | |
6de9cd9a DN |
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 | |
57dea9f6 | 24 | GNU General Public License for more details. |
6de9cd9a | 25 | |
57dea9f6 TM |
26 | You should have received a copy of the GNU General Public |
27 | License along with libgfortran; see the file COPYING. If not, | |
6de9cd9a DN |
28 | write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330, |
29 | Boston, MA 02111-1307, USA. */ | |
30 | ||
31 | #include "config.h" | |
32 | #include <stdlib.h> | |
410d3bba | 33 | #include <string.h> |
6de9cd9a DN |
34 | #include <assert.h> |
35 | #include "libgfortran.h" | |
36 | ||
410d3bba VL |
37 | /* This is a C version of the following fortran pseudo-code. The key |
38 | point is the loop order -- we access all arrays column-first, which | |
39 | improves the performance enough to boost galgel spec score by 50%. | |
40 | ||
41 | DIMENSION A(M,COUNT), B(COUNT,N), C(M,N) | |
42 | C = 0 | |
43 | DO J=1,N | |
44 | DO K=1,COUNT | |
45 | DO I=1,M | |
46 | C(I,J) = C(I,J)+A(I,K)*B(K,J) | |
47 | */ | |
48 | ||
7f68c75f RH |
49 | extern void matmul_i4 (gfc_array_i4 * retarray, gfc_array_i4 * a, gfc_array_i4 * b); |
50 | export_proto(matmul_i4); | |
7d7b8bfe | 51 | |
6de9cd9a | 52 | void |
7f68c75f | 53 | matmul_i4 (gfc_array_i4 * retarray, gfc_array_i4 * a, gfc_array_i4 * b) |
6de9cd9a DN |
54 | { |
55 | GFC_INTEGER_4 *abase; | |
56 | GFC_INTEGER_4 *bbase; | |
57 | GFC_INTEGER_4 *dest; | |
410d3bba VL |
58 | |
59 | index_type rxstride, rystride, axstride, aystride, bxstride, bystride; | |
60 | index_type x, y, n, count, xcount, ycount; | |
6de9cd9a DN |
61 | |
62 | assert (GFC_DESCRIPTOR_RANK (a) == 2 | |
63 | || GFC_DESCRIPTOR_RANK (b) == 2); | |
883c9d4d | 64 | |
410d3bba VL |
65 | /* C[xcount,ycount] = A[xcount, count] * B[count,ycount] |
66 | ||
67 | Either A or B (but not both) can be rank 1: | |
68 | ||
69 | o One-dimensional argument A is implicitly treated as a row matrix | |
70 | dimensioned [1,count], so xcount=1. | |
71 | ||
72 | o One-dimensional argument B is implicitly treated as a column matrix | |
73 | dimensioned [count, 1], so ycount=1. | |
74 | */ | |
75 | ||
883c9d4d VL |
76 | if (retarray->data == NULL) |
77 | { | |
78 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
79 | { | |
80 | retarray->dim[0].lbound = 0; | |
81 | retarray->dim[0].ubound = b->dim[1].ubound - b->dim[1].lbound; | |
82 | retarray->dim[0].stride = 1; | |
83 | } | |
84 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
85 | { | |
86 | retarray->dim[0].lbound = 0; | |
87 | retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound; | |
88 | retarray->dim[0].stride = 1; | |
89 | } | |
90 | else | |
91 | { | |
92 | retarray->dim[0].lbound = 0; | |
93 | retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound; | |
94 | retarray->dim[0].stride = 1; | |
95 | ||
96 | retarray->dim[1].lbound = 0; | |
97 | retarray->dim[1].ubound = b->dim[1].ubound - b->dim[1].lbound; | |
98 | retarray->dim[1].stride = retarray->dim[0].ubound+1; | |
99 | } | |
100 | ||
07d3cebe RH |
101 | retarray->data |
102 | = internal_malloc_size (sizeof (GFC_INTEGER_4) * size0 (retarray)); | |
883c9d4d VL |
103 | retarray->base = 0; |
104 | } | |
105 | ||
6de9cd9a DN |
106 | abase = a->data; |
107 | bbase = b->data; | |
108 | dest = retarray->data; | |
109 | ||
110 | if (retarray->dim[0].stride == 0) | |
111 | retarray->dim[0].stride = 1; | |
112 | if (a->dim[0].stride == 0) | |
113 | a->dim[0].stride = 1; | |
114 | if (b->dim[0].stride == 0) | |
115 | b->dim[0].stride = 1; | |
116 | ||
117 | ||
118 | if (GFC_DESCRIPTOR_RANK (retarray) == 1) | |
119 | { | |
410d3bba VL |
120 | /* One-dimensional result may be addressed in the code below |
121 | either as a row or a column matrix. We want both cases to | |
122 | work. */ | |
123 | rxstride = rystride = retarray->dim[0].stride; | |
6de9cd9a DN |
124 | } |
125 | else | |
126 | { | |
127 | rxstride = retarray->dim[0].stride; | |
128 | rystride = retarray->dim[1].stride; | |
129 | } | |
130 | ||
410d3bba | 131 | |
6de9cd9a DN |
132 | if (GFC_DESCRIPTOR_RANK (a) == 1) |
133 | { | |
410d3bba VL |
134 | /* Treat it as a a row matrix A[1,count]. */ |
135 | axstride = a->dim[0].stride; | |
136 | aystride = 1; | |
137 | ||
6de9cd9a | 138 | xcount = 1; |
410d3bba | 139 | count = a->dim[0].ubound + 1 - a->dim[0].lbound; |
6de9cd9a DN |
140 | } |
141 | else | |
142 | { | |
410d3bba VL |
143 | axstride = a->dim[0].stride; |
144 | aystride = a->dim[1].stride; | |
145 | ||
6de9cd9a | 146 | count = a->dim[1].ubound + 1 - a->dim[1].lbound; |
6de9cd9a DN |
147 | xcount = a->dim[0].ubound + 1 - a->dim[0].lbound; |
148 | } | |
410d3bba VL |
149 | |
150 | assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound); | |
151 | ||
6de9cd9a DN |
152 | if (GFC_DESCRIPTOR_RANK (b) == 1) |
153 | { | |
410d3bba VL |
154 | /* Treat it as a column matrix B[count,1] */ |
155 | bxstride = b->dim[0].stride; | |
156 | ||
157 | /* bystride should never be used for 1-dimensional b. | |
158 | in case it is we want it to cause a segfault, rather than | |
159 | an incorrect result. */ | |
160 | bystride = 0xDEADBEEF; | |
6de9cd9a DN |
161 | ycount = 1; |
162 | } | |
163 | else | |
164 | { | |
410d3bba VL |
165 | bxstride = b->dim[0].stride; |
166 | bystride = b->dim[1].stride; | |
6de9cd9a DN |
167 | ycount = b->dim[1].ubound + 1 - b->dim[1].lbound; |
168 | } | |
169 | ||
410d3bba VL |
170 | abase = a->data; |
171 | bbase = b->data; | |
172 | dest = retarray->data; | |
173 | ||
174 | if (rxstride == 1 && axstride == 1 && bxstride == 1) | |
6de9cd9a | 175 | { |
410d3bba VL |
176 | GFC_INTEGER_4 *bbase_y; |
177 | GFC_INTEGER_4 *dest_y; | |
178 | GFC_INTEGER_4 *abase_n; | |
179 | GFC_INTEGER_4 bbase_yn; | |
180 | ||
181 | memset (dest, 0, (sizeof (GFC_INTEGER_4) * size0(retarray))); | |
182 | ||
183 | for (y = 0; y < ycount; y++) | |
184 | { | |
185 | bbase_y = bbase + y*bystride; | |
186 | dest_y = dest + y*rystride; | |
187 | for (n = 0; n < count; n++) | |
188 | { | |
189 | abase_n = abase + n*aystride; | |
190 | bbase_yn = bbase_y[n]; | |
191 | for (x = 0; x < xcount; x++) | |
192 | { | |
193 | dest_y[x] += abase_n[x] * bbase_yn; | |
194 | } | |
195 | } | |
196 | } | |
197 | } | |
198 | else | |
199 | { | |
200 | for (y = 0; y < ycount; y++) | |
201 | for (x = 0; x < xcount; x++) | |
202 | dest[x*rxstride + y*rystride] = (GFC_INTEGER_4)0; | |
203 | ||
204 | for (y = 0; y < ycount; y++) | |
205 | for (n = 0; n < count; n++) | |
206 | for (x = 0; x < xcount; x++) | |
207 | /* dest[x,y] += a[x,n] * b[n,y] */ | |
208 | dest[x*rxstride + y*rystride] += abase[x*axstride + n*aystride] * bbase[n*bxstride + y*bystride]; | |
6de9cd9a DN |
209 | } |
210 | } |