1 `/* Implementation of the MATMUL intrinsic
2 Copyright 2002, 2005, 2006, 2007 Free Software Foundation, Inc.
3 Contributed by Paul Brook <paul@nowt.org>
5 This file is part of the GNU Fortran 95 runtime library (libgfortran).
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.
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
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.
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. */
31 #include "libgfortran.h"
38 `#if defined (HAVE_'rtype_name`)
40 /* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
41 passed to us by the front-end, in which case we''`ll call it for large
44 typedef void (*blas_call)(const char *, const char *, const int *, const int *,
45 const int *, const 'rtype_name` *, const 'rtype_name` *,
46 const int *, const 'rtype_name` *, const int *,
47 const 'rtype_name` *, 'rtype_name` *, const int *,
50 /* The order of loops is different in the case of plain matrix
51 multiplication C=MATMUL(A,B), and in the frequent special case where
52 the argument A is the temporary result of a TRANSPOSE intrinsic:
53 C=MATMUL(TRANSPOSE(A),B). Transposed temporaries are detected by
54 looking at their strides.
56 The equivalent Fortran pseudo-code is:
58 DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
59 IF (.NOT.IS_TRANSPOSED(A)) THEN
64 C(I,J) = C(I,J)+A(I,K)*B(K,J)
75 /* If try_blas is set to a nonzero value, then the matmul function will
76 see if there is a way to perform the matrix multiplication by a call
77 to the BLAS gemm function. */
79 extern void matmul_'rtype_code` ('rtype` * const restrict retarray,
80 'rtype` * const restrict a, 'rtype` * const restrict b, int try_blas,
81 int blas_limit, blas_call gemm);
82 export_proto(matmul_'rtype_code`);
85 matmul_'rtype_code` ('rtype` * const restrict retarray,
86 'rtype` * const restrict a, 'rtype` * const restrict b, int try_blas,
87 int blas_limit, blas_call gemm)
89 const 'rtype_name` * restrict abase;
90 const 'rtype_name` * restrict bbase;
91 'rtype_name` * restrict dest;
93 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
94 index_type x, y, n, count, xcount, ycount;
96 assert (GFC_DESCRIPTOR_RANK (a) == 2
97 || GFC_DESCRIPTOR_RANK (b) == 2);
99 /* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
101 Either A or B (but not both) can be rank 1:
103 o One-dimensional argument A is implicitly treated as a row matrix
104 dimensioned [1,count], so xcount=1.
106 o One-dimensional argument B is implicitly treated as a column matrix
107 dimensioned [count, 1], so ycount=1.
110 if (retarray->data == NULL)
112 if (GFC_DESCRIPTOR_RANK (a) == 1)
114 retarray->dim[0].lbound = 0;
115 retarray->dim[0].ubound = b->dim[1].ubound - b->dim[1].lbound;
116 retarray->dim[0].stride = 1;
118 else if (GFC_DESCRIPTOR_RANK (b) == 1)
120 retarray->dim[0].lbound = 0;
121 retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
122 retarray->dim[0].stride = 1;
126 retarray->dim[0].lbound = 0;
127 retarray->dim[0].ubound = a->dim[0].ubound - a->dim[0].lbound;
128 retarray->dim[0].stride = 1;
130 retarray->dim[1].lbound = 0;
131 retarray->dim[1].ubound = b->dim[1].ubound - b->dim[1].lbound;
132 retarray->dim[1].stride = retarray->dim[0].ubound+1;
136 = internal_malloc_size (sizeof ('rtype_name`) * size0 ((array_t *) retarray));
137 retarray->offset = 0;
140 sinclude(`matmul_asm_'rtype_code`.m4')dnl
142 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
144 /* One-dimensional result may be addressed in the code below
145 either as a row or a column matrix. We want both cases to
147 rxstride = rystride = retarray->dim[0].stride;
151 rxstride = retarray->dim[0].stride;
152 rystride = retarray->dim[1].stride;
156 if (GFC_DESCRIPTOR_RANK (a) == 1)
158 /* Treat it as a a row matrix A[1,count]. */
159 axstride = a->dim[0].stride;
163 count = a->dim[0].ubound + 1 - a->dim[0].lbound;
167 axstride = a->dim[0].stride;
168 aystride = a->dim[1].stride;
170 count = a->dim[1].ubound + 1 - a->dim[1].lbound;
171 xcount = a->dim[0].ubound + 1 - a->dim[0].lbound;
174 if (count != b->dim[0].ubound + 1 - b->dim[0].lbound)
175 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
177 if (GFC_DESCRIPTOR_RANK (b) == 1)
179 /* Treat it as a column matrix B[count,1] */
180 bxstride = b->dim[0].stride;
182 /* bystride should never be used for 1-dimensional b.
183 in case it is we want it to cause a segfault, rather than
184 an incorrect result. */
185 bystride = 0xDEADBEEF;
190 bxstride = b->dim[0].stride;
191 bystride = b->dim[1].stride;
192 ycount = b->dim[1].ubound + 1 - b->dim[1].lbound;
197 dest = retarray->data;
200 /* Now that everything is set up, we''`re performing the multiplication
203 #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
205 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
206 && (bxstride == 1 || bystride == 1)
207 && (((float) xcount) * ((float) ycount) * ((float) count)
210 const int m = xcount, n = ycount, k = count, ldc = rystride;
211 const 'rtype_name` one = 1, zero = 0;
212 const int lda = (axstride == 1) ? aystride : axstride,
213 ldb = (bxstride == 1) ? bystride : bxstride;
215 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
217 assert (gemm != NULL);
218 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
219 &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
224 if (rxstride == 1 && axstride == 1 && bxstride == 1)
226 const 'rtype_name` * restrict bbase_y;
227 'rtype_name` * restrict dest_y;
228 const 'rtype_name` * restrict abase_n;
229 'rtype_name` bbase_yn;
231 if (rystride == xcount)
232 memset (dest, 0, (sizeof ('rtype_name`) * xcount * ycount));
235 for (y = 0; y < ycount; y++)
236 for (x = 0; x < xcount; x++)
237 dest[x + y*rystride] = ('rtype_name`)0;
240 for (y = 0; y < ycount; y++)
242 bbase_y = bbase + y*bystride;
243 dest_y = dest + y*rystride;
244 for (n = 0; n < count; n++)
246 abase_n = abase + n*aystride;
247 bbase_yn = bbase_y[n];
248 for (x = 0; x < xcount; x++)
250 dest_y[x] += abase_n[x] * bbase_yn;
255 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
257 if (GFC_DESCRIPTOR_RANK (a) != 1)
259 const 'rtype_name` *restrict abase_x;
260 const 'rtype_name` *restrict bbase_y;
261 'rtype_name` *restrict dest_y;
264 for (y = 0; y < ycount; y++)
266 bbase_y = &bbase[y*bystride];
267 dest_y = &dest[y*rystride];
268 for (x = 0; x < xcount; x++)
270 abase_x = &abase[x*axstride];
271 s = ('rtype_name`) 0;
272 for (n = 0; n < count; n++)
273 s += abase_x[n] * bbase_y[n];
280 const 'rtype_name` *restrict bbase_y;
283 for (y = 0; y < ycount; y++)
285 bbase_y = &bbase[y*bystride];
286 s = ('rtype_name`) 0;
287 for (n = 0; n < count; n++)
288 s += abase[n*axstride] * bbase_y[n];
289 dest[y*rystride] = s;
293 else if (axstride < aystride)
295 for (y = 0; y < ycount; y++)
296 for (x = 0; x < xcount; x++)
297 dest[x*rxstride + y*rystride] = ('rtype_name`)0;
299 for (y = 0; y < ycount; y++)
300 for (n = 0; n < count; n++)
301 for (x = 0; x < xcount; x++)
302 /* dest[x,y] += a[x,n] * b[n,y] */
303 dest[x*rxstride + y*rystride] += abase[x*axstride + n*aystride] * bbase[n*bxstride + y*bystride];
305 else if (GFC_DESCRIPTOR_RANK (a) == 1)
307 const 'rtype_name` *restrict bbase_y;
310 for (y = 0; y < ycount; y++)
312 bbase_y = &bbase[y*bystride];
313 s = ('rtype_name`) 0;
314 for (n = 0; n < count; n++)
315 s += abase[n*axstride] * bbase_y[n*bxstride];
316 dest[y*rxstride] = s;
321 const 'rtype_name` *restrict abase_x;
322 const 'rtype_name` *restrict bbase_y;
323 'rtype_name` *restrict dest_y;
326 for (y = 0; y < ycount; y++)
328 bbase_y = &bbase[y*bystride];
329 dest_y = &dest[y*rystride];
330 for (x = 0; x < xcount; x++)
332 abase_x = &abase[x*axstride];
333 s = ('rtype_name`) 0;
334 for (n = 0; n < count; n++)
335 s += abase_x[n*aystride] * bbase_y[n*bxstride];
336 dest_y[x*rxstride] = s;