]> git.ipfire.org Git - thirdparty/gcc.git/blame - libgfortran/generated/matmul_c16.c
projects / thirdparty / gcc.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
re PR tree-optimization/87059 (internal compiler error: in set_value_range, at tree...
[thirdparty/gcc.git] / libgfortran / generated / matmul_c16.c
CommitLineData
644cb69f 1/* Implementation of the MATMUL intrinsic
85ec4feb 2 Copyright (C) 2002-2018 Free Software Foundation, Inc.
644cb69f
FXC		 3 Contributed by Paul Brook <paul@nowt.org>
4
21d1335b 5This file is part of the GNU Fortran runtime library (libgfortran).
644cb69f
FXC		 6
7Libgfortran is free software; you can redistribute it and/or
8modify it under the terms of the GNU General Public
9License as published by the Free Software Foundation; either
748086b7 10version 3 of the License, or (at your option) any later version.
644cb69f
FXC		 11
12Libgfortran is distributed in the hope that it will be useful,
13but WITHOUT ANY WARRANTY; without even the implied warranty of
14MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15GNU General Public License for more details.
16
748086b7
JJ		 17Under Section 7 of GPL version 3, you are granted additional
18permissions described in the GCC Runtime Library Exception, version
193.1, as published by the Free Software Foundation.
20
21You should have received a copy of the GNU General Public License and
22a copy of the GCC Runtime Library Exception along with this program;
23see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
24<http://www.gnu.org/licenses/>. */
644cb69f 25
36ae8a61 26#include "libgfortran.h"
644cb69f
FXC		 27#include <string.h>
28#include <assert.h>
36ae8a61 29
644cb69f
FXC		 30
31#if defined (HAVE_GFC_COMPLEX_16)
32
5a0aad31 33/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
5d70ab07 34 passed to us by the front-end, in which case we call it for large
5a0aad31
FXC		 35 matrices. */
36
37typedef void (*blas_call)(const char *, const char *, const int *, const int *,
38 const int *, const GFC_COMPLEX_16 *, const GFC_COMPLEX_16 *,
39 const int *, const GFC_COMPLEX_16 *, const int *,
40 const GFC_COMPLEX_16 *, GFC_COMPLEX_16 *, const int *,
41 int, int);
42
1524f80b
RS		 43/* The order of loops is different in the case of plain matrix
44 multiplication C=MATMUL(A,B), and in the frequent special case where
45 the argument A is the temporary result of a TRANSPOSE intrinsic:
46 C=MATMUL(TRANSPOSE(A),B). Transposed temporaries are detected by
47 looking at their strides.
48
49 The equivalent Fortran pseudo-code is:
644cb69f
FXC		 50
51 DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
1524f80b
RS		 52 IF (.NOT.IS_TRANSPOSED(A)) THEN
53 C = 0
54 DO J=1,N
55 DO K=1,COUNT
56 DO I=1,M
57 C(I,J) = C(I,J)+A(I,K)*B(K,J)
58 ELSE
59 DO J=1,N
644cb69f 60 DO I=1,M
1524f80b
RS		 61 S = 0
62 DO K=1,COUNT
5a0aad31 63 S = S+A(I,K)*B(K,J)
1524f80b
RS		 64 C(I,J) = S
65 ENDIF
644cb69f
FXC		 66*/
67
5a0aad31
FXC		 68/* If try_blas is set to a nonzero value, then the matmul function will
69 see if there is a way to perform the matrix multiplication by a call
70 to the BLAS gemm function. */
71
85206901 72extern void matmul_c16 (gfc_array_c16 * const restrict retarray,
5a0aad31
FXC		 73 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
74 int blas_limit, blas_call gemm);
644cb69f
FXC		 75export_proto(matmul_c16);
76
31cfd832
TK		 77/* Put exhaustive list of possible architectures here here, ORed together. */
78
79#if defined(HAVE_AVX) || defined(HAVE_AVX2) || defined(HAVE_AVX512F)
80
81#ifdef HAVE_AVX
82static void
83matmul_c16_avx (gfc_array_c16 * const restrict retarray,
84 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
85 int blas_limit, blas_call gemm) __attribute__((__target__("avx")));
86static void
87matmul_c16_avx (gfc_array_c16 * const restrict retarray,
88 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
89 int blas_limit, blas_call gemm)
90{
91 const GFC_COMPLEX_16 * restrict abase;
92 const GFC_COMPLEX_16 * restrict bbase;
93 GFC_COMPLEX_16 * restrict dest;
94
95 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
96 index_type x, y, n, count, xcount, ycount;
97
98 assert (GFC_DESCRIPTOR_RANK (a) == 2
99 || GFC_DESCRIPTOR_RANK (b) == 2);
100
101/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
102
103 Either A or B (but not both) can be rank 1:
104
105 o One-dimensional argument A is implicitly treated as a row matrix
106 dimensioned [1,count], so xcount=1.
107
108 o One-dimensional argument B is implicitly treated as a column matrix
109 dimensioned [count, 1], so ycount=1.
110*/
111
112 if (retarray->base_addr == NULL)
113 {
114 if (GFC_DESCRIPTOR_RANK (a) == 1)
115 {
116 GFC_DIMENSION_SET(retarray->dim[0], 0,
117 GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
118 }
119 else if (GFC_DESCRIPTOR_RANK (b) == 1)
120 {
121 GFC_DIMENSION_SET(retarray->dim[0], 0,
122 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
123 }
124 else
125 {
126 GFC_DIMENSION_SET(retarray->dim[0], 0,
127 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
128
129 GFC_DIMENSION_SET(retarray->dim[1], 0,
130 GFC_DESCRIPTOR_EXTENT(b,1) - 1,
131 GFC_DESCRIPTOR_EXTENT(retarray,0));
132 }
133
134 retarray->base_addr
135 = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_COMPLEX_16));
136 retarray->offset = 0;
137 }
138 else if (unlikely (compile_options.bounds_check))
139 {
140 index_type ret_extent, arg_extent;
141
142 if (GFC_DESCRIPTOR_RANK (a) == 1)
143 {
144 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
145 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
146 if (arg_extent != ret_extent)
147 runtime_error ("Incorrect extent in return array in"
148 " MATMUL intrinsic: is %ld, should be %ld",
149 (long int) ret_extent, (long int) arg_extent);
150 }
151 else if (GFC_DESCRIPTOR_RANK (b) == 1)
152 {
153 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
154 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
155 if (arg_extent != ret_extent)
156 runtime_error ("Incorrect extent in return array in"
157 " MATMUL intrinsic: is %ld, should be %ld",
158 (long int) ret_extent, (long int) arg_extent);
159 }
160 else
161 {
162 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
163 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
164 if (arg_extent != ret_extent)
165 runtime_error ("Incorrect extent in return array in"
166 " MATMUL intrinsic for dimension 1:"
167 " is %ld, should be %ld",
168 (long int) ret_extent, (long int) arg_extent);
169
170 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
171 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
172 if (arg_extent != ret_extent)
173 runtime_error ("Incorrect extent in return array in"
174 " MATMUL intrinsic for dimension 2:"
175 " is %ld, should be %ld",
176 (long int) ret_extent, (long int) arg_extent);
177 }
178 }
179
180
181 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
182 {
183 /* One-dimensional result may be addressed in the code below
184 either as a row or a column matrix. We want both cases to
185 work. */
186 rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
187 }
188 else
189 {
190 rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
191 rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
192 }
193
194
195 if (GFC_DESCRIPTOR_RANK (a) == 1)
196 {
197 /* Treat it as a a row matrix A[1,count]. */
198 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
199 aystride = 1;
200
201 xcount = 1;
202 count = GFC_DESCRIPTOR_EXTENT(a,0);
203 }
204 else
205 {
206 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
207 aystride = GFC_DESCRIPTOR_STRIDE(a,1);
208
209 count = GFC_DESCRIPTOR_EXTENT(a,1);
210 xcount = GFC_DESCRIPTOR_EXTENT(a,0);
211 }
212
213 if (count != GFC_DESCRIPTOR_EXTENT(b,0))
214 {
215 if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
216 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
217 }
218
219 if (GFC_DESCRIPTOR_RANK (b) == 1)
220 {
221 /* Treat it as a column matrix B[count,1] */
222 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
223
224 /* bystride should never be used for 1-dimensional b.
6ce6a84a
TK		 225 The value is only used for calculation of the
226 memory by the buffer. */
227 bystride = 256;
31cfd832
TK		 228 ycount = 1;
229 }
230 else
231 {
232 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
233 bystride = GFC_DESCRIPTOR_STRIDE(b,1);
234 ycount = GFC_DESCRIPTOR_EXTENT(b,1);
235 }
236
237 abase = a->base_addr;
238 bbase = b->base_addr;
239 dest = retarray->base_addr;
240
241 /* Now that everything is set up, we perform the multiplication
242 itself. */
243
244#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
245#define min(a,b) ((a) <= (b) ? (a) : (b))
246#define max(a,b) ((a) >= (b) ? (a) : (b))
247
248 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
249 && (bxstride == 1 || bystride == 1)
250 && (((float) xcount) * ((float) ycount) * ((float) count)
251 > POW3(blas_limit)))
252 {
253 const int m = xcount, n = ycount, k = count, ldc = rystride;
254 const GFC_COMPLEX_16 one = 1, zero = 0;
255 const int lda = (axstride == 1) ? aystride : axstride,
256 ldb = (bxstride == 1) ? bystride : bxstride;
257
258 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
259 {
260 assert (gemm != NULL);
261 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m,
262 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
263 &ldc, 1, 1);
264 return;
265 }
266 }
267
268 if (rxstride == 1 && axstride == 1 && bxstride == 1)
269 {
270 /* This block of code implements a tuned matmul, derived from
271 Superscalar GEMM-based level 3 BLAS, Beta version 0.1
272
273 Bo Kagstrom and Per Ling
274 Department of Computing Science
275 Umea University
276 S-901 87 Umea, Sweden
277
278 from netlib.org, translated to C, and modified for matmul.m4. */
279
280 const GFC_COMPLEX_16 *a, *b;
281 GFC_COMPLEX_16 *c;
282 const index_type m = xcount, n = ycount, k = count;
283
284 /* System generated locals */
285 index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
286 i1, i2, i3, i4, i5, i6;
287
288 /* Local variables */
fd991039 289 GFC_COMPLEX_16 f11, f12, f21, f22, f31, f32, f41, f42,
31cfd832
TK		 290 f13, f14, f23, f24, f33, f34, f43, f44;
291 index_type i, j, l, ii, jj, ll;
292 index_type isec, jsec, lsec, uisec, ujsec, ulsec;
8e5f30dc 293 GFC_COMPLEX_16 *t1;
31cfd832
TK		 294
295 a = abase;
296 b = bbase;
297 c = retarray->base_addr;
298
299 /* Parameter adjustments */
300 c_dim1 = rystride;
301 c_offset = 1 + c_dim1;
302 c -= c_offset;
303 a_dim1 = aystride;
304 a_offset = 1 + a_dim1;
305 a -= a_offset;
306 b_dim1 = bystride;
307 b_offset = 1 + b_dim1;
308 b -= b_offset;
309
bbf97416
TK		 310 /* Empty c first. */
311 for (j=1; j<=n; j++)
312 for (i=1; i<=m; i++)
313 c[i + j * c_dim1] = (GFC_COMPLEX_16)0;
314
31cfd832
TK		 315 /* Early exit if possible */
316 if (m == 0 || n == 0 || k == 0)
317 return;
318
fd991039
TK		 319 /* Adjust size of t1 to what is needed. */
320 index_type t1_dim;
e889aa0a 321 t1_dim = (a_dim1 - (ycount > 1)) * 256 + b_dim1;
fd991039
TK		 322 if (t1_dim > 65536)
323 t1_dim = 65536;
324
8e5f30dc 325 t1 = malloc (t1_dim * sizeof(GFC_COMPLEX_16));
fd991039 326
31cfd832
TK		 327 /* Start turning the crank. */
328 i1 = n;
329 for (jj = 1; jj <= i1; jj += 512)
330 {
331 /* Computing MIN */
332 i2 = 512;
333 i3 = n - jj + 1;
334 jsec = min(i2,i3);
335 ujsec = jsec - jsec % 4;
336 i2 = k;
337 for (ll = 1; ll <= i2; ll += 256)
338 {
339 /* Computing MIN */
340 i3 = 256;
341 i4 = k - ll + 1;
342 lsec = min(i3,i4);
343 ulsec = lsec - lsec % 2;
344
345 i3 = m;
346 for (ii = 1; ii <= i3; ii += 256)
347 {
348 /* Computing MIN */
349 i4 = 256;
350 i5 = m - ii + 1;
351 isec = min(i4,i5);
352 uisec = isec - isec % 2;
353 i4 = ll + ulsec - 1;
354 for (l = ll; l <= i4; l += 2)
355 {
356 i5 = ii + uisec - 1;
357 for (i = ii; i <= i5; i += 2)
358 {
359 t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
360 a[i + l * a_dim1];
361 t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
362 a[i + (l + 1) * a_dim1];
363 t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
364 a[i + 1 + l * a_dim1];
365 t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
366 a[i + 1 + (l + 1) * a_dim1];
367 }
368 if (uisec < isec)
369 {
370 t1[l - ll + 1 + (isec << 8) - 257] =
371 a[ii + isec - 1 + l * a_dim1];
372 t1[l - ll + 2 + (isec << 8) - 257] =
373 a[ii + isec - 1 + (l + 1) * a_dim1];
374 }
375 }
376 if (ulsec < lsec)
377 {
378 i4 = ii + isec - 1;
379 for (i = ii; i<= i4; ++i)
380 {
381 t1[lsec + ((i - ii + 1) << 8) - 257] =
382 a[i + (ll + lsec - 1) * a_dim1];
383 }
384 }
385
386 uisec = isec - isec % 4;
387 i4 = jj + ujsec - 1;
388 for (j = jj; j <= i4; j += 4)
389 {
390 i5 = ii + uisec - 1;
391 for (i = ii; i <= i5; i += 4)
392 {
393 f11 = c[i + j * c_dim1];
394 f21 = c[i + 1 + j * c_dim1];
395 f12 = c[i + (j + 1) * c_dim1];
396 f22 = c[i + 1 + (j + 1) * c_dim1];
397 f13 = c[i + (j + 2) * c_dim1];
398 f23 = c[i + 1 + (j + 2) * c_dim1];
399 f14 = c[i + (j + 3) * c_dim1];
400 f24 = c[i + 1 + (j + 3) * c_dim1];
401 f31 = c[i + 2 + j * c_dim1];
402 f41 = c[i + 3 + j * c_dim1];
403 f32 = c[i + 2 + (j + 1) * c_dim1];
404 f42 = c[i + 3 + (j + 1) * c_dim1];
405 f33 = c[i + 2 + (j + 2) * c_dim1];
406 f43 = c[i + 3 + (j + 2) * c_dim1];
407 f34 = c[i + 2 + (j + 3) * c_dim1];
408 f44 = c[i + 3 + (j + 3) * c_dim1];
409 i6 = ll + lsec - 1;
410 for (l = ll; l <= i6; ++l)
411 {
412 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
413 * b[l + j * b_dim1];
414 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
415 * b[l + j * b_dim1];
416 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
417 * b[l + (j + 1) * b_dim1];
418 f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
419 * b[l + (j + 1) * b_dim1];
420 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
421 * b[l + (j + 2) * b_dim1];
422 f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
423 * b[l + (j + 2) * b_dim1];
424 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
425 * b[l + (j + 3) * b_dim1];
426 f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
427 * b[l + (j + 3) * b_dim1];
428 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
429 * b[l + j * b_dim1];
430 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
431 * b[l + j * b_dim1];
432 f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
433 * b[l + (j + 1) * b_dim1];
434 f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
435 * b[l + (j + 1) * b_dim1];
436 f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
437 * b[l + (j + 2) * b_dim1];
438 f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
439 * b[l + (j + 2) * b_dim1];
440 f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
441 * b[l + (j + 3) * b_dim1];
442 f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
443 * b[l + (j + 3) * b_dim1];
444 }
445 c[i + j * c_dim1] = f11;
446 c[i + 1 + j * c_dim1] = f21;
447 c[i + (j + 1) * c_dim1] = f12;
448 c[i + 1 + (j + 1) * c_dim1] = f22;
449 c[i + (j + 2) * c_dim1] = f13;
450 c[i + 1 + (j + 2) * c_dim1] = f23;
451 c[i + (j + 3) * c_dim1] = f14;
452 c[i + 1 + (j + 3) * c_dim1] = f24;
453 c[i + 2 + j * c_dim1] = f31;
454 c[i + 3 + j * c_dim1] = f41;
455 c[i + 2 + (j + 1) * c_dim1] = f32;
456 c[i + 3 + (j + 1) * c_dim1] = f42;
457 c[i + 2 + (j + 2) * c_dim1] = f33;
458 c[i + 3 + (j + 2) * c_dim1] = f43;
459 c[i + 2 + (j + 3) * c_dim1] = f34;
460 c[i + 3 + (j + 3) * c_dim1] = f44;
461 }
462 if (uisec < isec)
463 {
464 i5 = ii + isec - 1;
465 for (i = ii + uisec; i <= i5; ++i)
466 {
467 f11 = c[i + j * c_dim1];
468 f12 = c[i + (j + 1) * c_dim1];
469 f13 = c[i + (j + 2) * c_dim1];
470 f14 = c[i + (j + 3) * c_dim1];
471 i6 = ll + lsec - 1;
472 for (l = ll; l <= i6; ++l)
473 {
474 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
475 257] * b[l + j * b_dim1];
476 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
477 257] * b[l + (j + 1) * b_dim1];
478 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
479 257] * b[l + (j + 2) * b_dim1];
480 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
481 257] * b[l + (j + 3) * b_dim1];
482 }
483 c[i + j * c_dim1] = f11;
484 c[i + (j + 1) * c_dim1] = f12;
485 c[i + (j + 2) * c_dim1] = f13;
486 c[i + (j + 3) * c_dim1] = f14;
487 }
488 }
489 }
490 if (ujsec < jsec)
491 {
492 i4 = jj + jsec - 1;
493 for (j = jj + ujsec; j <= i4; ++j)
494 {
495 i5 = ii + uisec - 1;
496 for (i = ii; i <= i5; i += 4)
497 {
498 f11 = c[i + j * c_dim1];
499 f21 = c[i + 1 + j * c_dim1];
500 f31 = c[i + 2 + j * c_dim1];
501 f41 = c[i + 3 + j * c_dim1];
502 i6 = ll + lsec - 1;
503 for (l = ll; l <= i6; ++l)
504 {
505 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
506 257] * b[l + j * b_dim1];
507 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
508 257] * b[l + j * b_dim1];
509 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
510 257] * b[l + j * b_dim1];
511 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
512 257] * b[l + j * b_dim1];
513 }
514 c[i + j * c_dim1] = f11;
515 c[i + 1 + j * c_dim1] = f21;
516 c[i + 2 + j * c_dim1] = f31;
517 c[i + 3 + j * c_dim1] = f41;
518 }
519 i5 = ii + isec - 1;
520 for (i = ii + uisec; i <= i5; ++i)
521 {
522 f11 = c[i + j * c_dim1];
523 i6 = ll + lsec - 1;
524 for (l = ll; l <= i6; ++l)
525 {
526 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
527 257] * b[l + j * b_dim1];
528 }
529 c[i + j * c_dim1] = f11;
530 }
531 }
532 }
533 }
534 }
535 }
8e5f30dc 536 free(t1);
31cfd832
TK		 537 return;
538 }
539 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
540 {
541 if (GFC_DESCRIPTOR_RANK (a) != 1)
542 {
543 const GFC_COMPLEX_16 *restrict abase_x;
544 const GFC_COMPLEX_16 *restrict bbase_y;
545 GFC_COMPLEX_16 *restrict dest_y;
546 GFC_COMPLEX_16 s;
547
548 for (y = 0; y < ycount; y++)
549 {
550 bbase_y = &bbase[y*bystride];
551 dest_y = &dest[y*rystride];
552 for (x = 0; x < xcount; x++)
553 {
554 abase_x = &abase[x*axstride];
555 s = (GFC_COMPLEX_16) 0;
556 for (n = 0; n < count; n++)
557 s += abase_x[n] * bbase_y[n];
558 dest_y[x] = s;
559 }
560 }
561 }
562 else
563 {
564 const GFC_COMPLEX_16 *restrict bbase_y;
565 GFC_COMPLEX_16 s;
566
567 for (y = 0; y < ycount; y++)
568 {
569 bbase_y = &bbase[y*bystride];
570 s = (GFC_COMPLEX_16) 0;
571 for (n = 0; n < count; n++)
572 s += abase[n*axstride] * bbase_y[n];
573 dest[y*rystride] = s;
574 }
575 }
576 }
577 else if (axstride < aystride)
578 {
579 for (y = 0; y < ycount; y++)
580 for (x = 0; x < xcount; x++)
581 dest[x*rxstride + y*rystride] = (GFC_COMPLEX_16)0;
582
583 for (y = 0; y < ycount; y++)
584 for (n = 0; n < count; n++)
585 for (x = 0; x < xcount; x++)
586 /* dest[x,y] += a[x,n] * b[n,y] */
587 dest[x*rxstride + y*rystride] +=
588 abase[x*axstride + n*aystride] *
589 bbase[n*bxstride + y*bystride];
590 }
591 else if (GFC_DESCRIPTOR_RANK (a) == 1)
592 {
593 const GFC_COMPLEX_16 *restrict bbase_y;
594 GFC_COMPLEX_16 s;
595
596 for (y = 0; y < ycount; y++)
597 {
598 bbase_y = &bbase[y*bystride];
599 s = (GFC_COMPLEX_16) 0;
600 for (n = 0; n < count; n++)
601 s += abase[n*axstride] * bbase_y[n*bxstride];
602 dest[y*rxstride] = s;
603 }
604 }
605 else
606 {
607 const GFC_COMPLEX_16 *restrict abase_x;
608 const GFC_COMPLEX_16 *restrict bbase_y;
609 GFC_COMPLEX_16 *restrict dest_y;
610 GFC_COMPLEX_16 s;
611
612 for (y = 0; y < ycount; y++)
613 {
614 bbase_y = &bbase[y*bystride];
615 dest_y = &dest[y*rystride];
616 for (x = 0; x < xcount; x++)
617 {
618 abase_x = &abase[x*axstride];
619 s = (GFC_COMPLEX_16) 0;
620 for (n = 0; n < count; n++)
621 s += abase_x[n*aystride] * bbase_y[n*bxstride];
622 dest_y[x*rxstride] = s;
623 }
624 }
625 }
626}
627#undef POW3
628#undef min
629#undef max
630
631#endif /* HAVE_AVX */
632
633#ifdef HAVE_AVX2
634static void
635matmul_c16_avx2 (gfc_array_c16 * const restrict retarray,
636 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
6d03bdcc 637 int blas_limit, blas_call gemm) __attribute__((__target__("avx2,fma")));
31cfd832
TK		 638static void
639matmul_c16_avx2 (gfc_array_c16 * const restrict retarray,
640 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
641 int blas_limit, blas_call gemm)
642{
643 const GFC_COMPLEX_16 * restrict abase;
644 const GFC_COMPLEX_16 * restrict bbase;
645 GFC_COMPLEX_16 * restrict dest;
646
647 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
648 index_type x, y, n, count, xcount, ycount;
649
650 assert (GFC_DESCRIPTOR_RANK (a) == 2
651 || GFC_DESCRIPTOR_RANK (b) == 2);
652
653/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
654
655 Either A or B (but not both) can be rank 1:
656
657 o One-dimensional argument A is implicitly treated as a row matrix
658 dimensioned [1,count], so xcount=1.
659
660 o One-dimensional argument B is implicitly treated as a column matrix
661 dimensioned [count, 1], so ycount=1.
662*/
663
664 if (retarray->base_addr == NULL)
665 {
666 if (GFC_DESCRIPTOR_RANK (a) == 1)
667 {
668 GFC_DIMENSION_SET(retarray->dim[0], 0,
669 GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
670 }
671 else if (GFC_DESCRIPTOR_RANK (b) == 1)
672 {
673 GFC_DIMENSION_SET(retarray->dim[0], 0,
674 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
675 }
676 else
677 {
678 GFC_DIMENSION_SET(retarray->dim[0], 0,
679 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
680
681 GFC_DIMENSION_SET(retarray->dim[1], 0,
682 GFC_DESCRIPTOR_EXTENT(b,1) - 1,
683 GFC_DESCRIPTOR_EXTENT(retarray,0));
684 }
685
686 retarray->base_addr
687 = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_COMPLEX_16));
688 retarray->offset = 0;
689 }
690 else if (unlikely (compile_options.bounds_check))
691 {
692 index_type ret_extent, arg_extent;
693
694 if (GFC_DESCRIPTOR_RANK (a) == 1)
695 {
696 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
697 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
698 if (arg_extent != ret_extent)
699 runtime_error ("Incorrect extent in return array in"
700 " MATMUL intrinsic: is %ld, should be %ld",
701 (long int) ret_extent, (long int) arg_extent);
702 }
703 else if (GFC_DESCRIPTOR_RANK (b) == 1)
704 {
705 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
706 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
707 if (arg_extent != ret_extent)
708 runtime_error ("Incorrect extent in return array in"
709 " MATMUL intrinsic: is %ld, should be %ld",
710 (long int) ret_extent, (long int) arg_extent);
711 }
712 else
713 {
714 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
715 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
716 if (arg_extent != ret_extent)
717 runtime_error ("Incorrect extent in return array in"
718 " MATMUL intrinsic for dimension 1:"
719 " is %ld, should be %ld",
720 (long int) ret_extent, (long int) arg_extent);
721
722 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
723 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
724 if (arg_extent != ret_extent)
725 runtime_error ("Incorrect extent in return array in"
726 " MATMUL intrinsic for dimension 2:"
727 " is %ld, should be %ld",
728 (long int) ret_extent, (long int) arg_extent);
729 }
730 }
731
732
733 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
734 {
735 /* One-dimensional result may be addressed in the code below
736 either as a row or a column matrix. We want both cases to
737 work. */
738 rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
739 }
740 else
741 {
742 rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
743 rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
744 }
745
746
747 if (GFC_DESCRIPTOR_RANK (a) == 1)
748 {
749 /* Treat it as a a row matrix A[1,count]. */
750 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
751 aystride = 1;
752
753 xcount = 1;
754 count = GFC_DESCRIPTOR_EXTENT(a,0);
755 }
756 else
757 {
758 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
759 aystride = GFC_DESCRIPTOR_STRIDE(a,1);
760
761 count = GFC_DESCRIPTOR_EXTENT(a,1);
762 xcount = GFC_DESCRIPTOR_EXTENT(a,0);
763 }
764
765 if (count != GFC_DESCRIPTOR_EXTENT(b,0))
766 {
767 if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
768 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
769 }
770
771 if (GFC_DESCRIPTOR_RANK (b) == 1)
772 {
773 /* Treat it as a column matrix B[count,1] */
774 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
775
776 /* bystride should never be used for 1-dimensional b.
6ce6a84a
TK		 777 The value is only used for calculation of the
778 memory by the buffer. */
779 bystride = 256;
31cfd832
TK		 780 ycount = 1;
781 }
782 else
783 {
784 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
785 bystride = GFC_DESCRIPTOR_STRIDE(b,1);
786 ycount = GFC_DESCRIPTOR_EXTENT(b,1);
787 }
788
789 abase = a->base_addr;
790 bbase = b->base_addr;
791 dest = retarray->base_addr;
792
793 /* Now that everything is set up, we perform the multiplication
794 itself. */
795
796#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
797#define min(a,b) ((a) <= (b) ? (a) : (b))
798#define max(a,b) ((a) >= (b) ? (a) : (b))
799
800 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
801 && (bxstride == 1 || bystride == 1)
802 && (((float) xcount) * ((float) ycount) * ((float) count)
803 > POW3(blas_limit)))
804 {
805 const int m = xcount, n = ycount, k = count, ldc = rystride;
806 const GFC_COMPLEX_16 one = 1, zero = 0;
807 const int lda = (axstride == 1) ? aystride : axstride,
808 ldb = (bxstride == 1) ? bystride : bxstride;
809
810 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
811 {
812 assert (gemm != NULL);
813 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m,
814 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
815 &ldc, 1, 1);
816 return;
817 }
818 }
819
820 if (rxstride == 1 && axstride == 1 && bxstride == 1)
821 {
822 /* This block of code implements a tuned matmul, derived from
823 Superscalar GEMM-based level 3 BLAS, Beta version 0.1
824
825 Bo Kagstrom and Per Ling
826 Department of Computing Science
827 Umea University
828 S-901 87 Umea, Sweden
829
830 from netlib.org, translated to C, and modified for matmul.m4. */
831
832 const GFC_COMPLEX_16 *a, *b;
833 GFC_COMPLEX_16 *c;
834 const index_type m = xcount, n = ycount, k = count;
835
836 /* System generated locals */
837 index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
838 i1, i2, i3, i4, i5, i6;
839
840 /* Local variables */
fd991039 841 GFC_COMPLEX_16 f11, f12, f21, f22, f31, f32, f41, f42,
31cfd832
TK		 842 f13, f14, f23, f24, f33, f34, f43, f44;
843 index_type i, j, l, ii, jj, ll;
844 index_type isec, jsec, lsec, uisec, ujsec, ulsec;
8e5f30dc 845 GFC_COMPLEX_16 *t1;
31cfd832
TK		 846
847 a = abase;
848 b = bbase;
849 c = retarray->base_addr;
850
851 /* Parameter adjustments */
852 c_dim1 = rystride;
853 c_offset = 1 + c_dim1;
854 c -= c_offset;
855 a_dim1 = aystride;
856 a_offset = 1 + a_dim1;
857 a -= a_offset;
858 b_dim1 = bystride;
859 b_offset = 1 + b_dim1;
860 b -= b_offset;
861
bbf97416
TK		 862 /* Empty c first. */
863 for (j=1; j<=n; j++)
864 for (i=1; i<=m; i++)
865 c[i + j * c_dim1] = (GFC_COMPLEX_16)0;
866
31cfd832
TK		 867 /* Early exit if possible */
868 if (m == 0 || n == 0 || k == 0)
869 return;
870
fd991039
TK		 871 /* Adjust size of t1 to what is needed. */
872 index_type t1_dim;
e889aa0a 873 t1_dim = (a_dim1 - (ycount > 1)) * 256 + b_dim1;
fd991039
TK		 874 if (t1_dim > 65536)
875 t1_dim = 65536;
876
8e5f30dc 877 t1 = malloc (t1_dim * sizeof(GFC_COMPLEX_16));
fd991039 878
31cfd832
TK		 879 /* Start turning the crank. */
880 i1 = n;
881 for (jj = 1; jj <= i1; jj += 512)
882 {
883 /* Computing MIN */
884 i2 = 512;
885 i3 = n - jj + 1;
886 jsec = min(i2,i3);
887 ujsec = jsec - jsec % 4;
888 i2 = k;
889 for (ll = 1; ll <= i2; ll += 256)
890 {
891 /* Computing MIN */
892 i3 = 256;
893 i4 = k - ll + 1;
894 lsec = min(i3,i4);
895 ulsec = lsec - lsec % 2;
896
897 i3 = m;
898 for (ii = 1; ii <= i3; ii += 256)
899 {
900 /* Computing MIN */
901 i4 = 256;
902 i5 = m - ii + 1;
903 isec = min(i4,i5);
904 uisec = isec - isec % 2;
905 i4 = ll + ulsec - 1;
906 for (l = ll; l <= i4; l += 2)
907 {
908 i5 = ii + uisec - 1;
909 for (i = ii; i <= i5; i += 2)
910 {
911 t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
912 a[i + l * a_dim1];
913 t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
914 a[i + (l + 1) * a_dim1];
915 t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
916 a[i + 1 + l * a_dim1];
917 t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
918 a[i + 1 + (l + 1) * a_dim1];
919 }
920 if (uisec < isec)
921 {
922 t1[l - ll + 1 + (isec << 8) - 257] =
923 a[ii + isec - 1 + l * a_dim1];
924 t1[l - ll + 2 + (isec << 8) - 257] =
925 a[ii + isec - 1 + (l + 1) * a_dim1];
926 }
927 }
928 if (ulsec < lsec)
929 {
930 i4 = ii + isec - 1;
931 for (i = ii; i<= i4; ++i)
932 {
933 t1[lsec + ((i - ii + 1) << 8) - 257] =
934 a[i + (ll + lsec - 1) * a_dim1];
935 }
936 }
937
938 uisec = isec - isec % 4;
939 i4 = jj + ujsec - 1;
940 for (j = jj; j <= i4; j += 4)
941 {
942 i5 = ii + uisec - 1;
943 for (i = ii; i <= i5; i += 4)
944 {
945 f11 = c[i + j * c_dim1];
946 f21 = c[i + 1 + j * c_dim1];
947 f12 = c[i + (j + 1) * c_dim1];
948 f22 = c[i + 1 + (j + 1) * c_dim1];
949 f13 = c[i + (j + 2) * c_dim1];
950 f23 = c[i + 1 + (j + 2) * c_dim1];
951 f14 = c[i + (j + 3) * c_dim1];
952 f24 = c[i + 1 + (j + 3) * c_dim1];
953 f31 = c[i + 2 + j * c_dim1];
954 f41 = c[i + 3 + j * c_dim1];
955 f32 = c[i + 2 + (j + 1) * c_dim1];
956 f42 = c[i + 3 + (j + 1) * c_dim1];
957 f33 = c[i + 2 + (j + 2) * c_dim1];
958 f43 = c[i + 3 + (j + 2) * c_dim1];
959 f34 = c[i + 2 + (j + 3) * c_dim1];
960 f44 = c[i + 3 + (j + 3) * c_dim1];
961 i6 = ll + lsec - 1;
962 for (l = ll; l <= i6; ++l)
963 {
964 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
965 * b[l + j * b_dim1];
966 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
967 * b[l + j * b_dim1];
968 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
969 * b[l + (j + 1) * b_dim1];
970 f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
971 * b[l + (j + 1) * b_dim1];
972 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
973 * b[l + (j + 2) * b_dim1];
974 f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
975 * b[l + (j + 2) * b_dim1];
976 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
977 * b[l + (j + 3) * b_dim1];
978 f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
979 * b[l + (j + 3) * b_dim1];
980 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
981 * b[l + j * b_dim1];
982 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
983 * b[l + j * b_dim1];
984 f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
985 * b[l + (j + 1) * b_dim1];
986 f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
987 * b[l + (j + 1) * b_dim1];
988 f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
989 * b[l + (j + 2) * b_dim1];
990 f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
991 * b[l + (j + 2) * b_dim1];
992 f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
993 * b[l + (j + 3) * b_dim1];
994 f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
995 * b[l + (j + 3) * b_dim1];
996 }
997 c[i + j * c_dim1] = f11;
998 c[i + 1 + j * c_dim1] = f21;
999 c[i + (j + 1) * c_dim1] = f12;
1000 c[i + 1 + (j + 1) * c_dim1] = f22;
1001 c[i + (j + 2) * c_dim1] = f13;
1002 c[i + 1 + (j + 2) * c_dim1] = f23;
1003 c[i + (j + 3) * c_dim1] = f14;
1004 c[i + 1 + (j + 3) * c_dim1] = f24;
1005 c[i + 2 + j * c_dim1] = f31;
1006 c[i + 3 + j * c_dim1] = f41;
1007 c[i + 2 + (j + 1) * c_dim1] = f32;
1008 c[i + 3 + (j + 1) * c_dim1] = f42;
1009 c[i + 2 + (j + 2) * c_dim1] = f33;
1010 c[i + 3 + (j + 2) * c_dim1] = f43;
1011 c[i + 2 + (j + 3) * c_dim1] = f34;
1012 c[i + 3 + (j + 3) * c_dim1] = f44;
1013 }
1014 if (uisec < isec)
1015 {
1016 i5 = ii + isec - 1;
1017 for (i = ii + uisec; i <= i5; ++i)
1018 {
1019 f11 = c[i + j * c_dim1];
1020 f12 = c[i + (j + 1) * c_dim1];
1021 f13 = c[i + (j + 2) * c_dim1];
1022 f14 = c[i + (j + 3) * c_dim1];
1023 i6 = ll + lsec - 1;
1024 for (l = ll; l <= i6; ++l)
1025 {
1026 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1027 257] * b[l + j * b_dim1];
1028 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1029 257] * b[l + (j + 1) * b_dim1];
1030 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1031 257] * b[l + (j + 2) * b_dim1];
1032 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1033 257] * b[l + (j + 3) * b_dim1];
1034 }
1035 c[i + j * c_dim1] = f11;
1036 c[i + (j + 1) * c_dim1] = f12;
1037 c[i + (j + 2) * c_dim1] = f13;
1038 c[i + (j + 3) * c_dim1] = f14;
1039 }
1040 }
1041 }
1042 if (ujsec < jsec)
1043 {
1044 i4 = jj + jsec - 1;
1045 for (j = jj + ujsec; j <= i4; ++j)
1046 {
1047 i5 = ii + uisec - 1;
1048 for (i = ii; i <= i5; i += 4)
1049 {
1050 f11 = c[i + j * c_dim1];
1051 f21 = c[i + 1 + j * c_dim1];
1052 f31 = c[i + 2 + j * c_dim1];
1053 f41 = c[i + 3 + j * c_dim1];
1054 i6 = ll + lsec - 1;
1055 for (l = ll; l <= i6; ++l)
1056 {
1057 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1058 257] * b[l + j * b_dim1];
1059 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
1060 257] * b[l + j * b_dim1];
1061 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
1062 257] * b[l + j * b_dim1];
1063 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
1064 257] * b[l + j * b_dim1];
1065 }
1066 c[i + j * c_dim1] = f11;
1067 c[i + 1 + j * c_dim1] = f21;
1068 c[i + 2 + j * c_dim1] = f31;
1069 c[i + 3 + j * c_dim1] = f41;
1070 }
1071 i5 = ii + isec - 1;
1072 for (i = ii + uisec; i <= i5; ++i)
1073 {
1074 f11 = c[i + j * c_dim1];
1075 i6 = ll + lsec - 1;
1076 for (l = ll; l <= i6; ++l)
1077 {
1078 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1079 257] * b[l + j * b_dim1];
1080 }
1081 c[i + j * c_dim1] = f11;
1082 }
1083 }
1084 }
1085 }
1086 }
1087 }
8e5f30dc 1088 free(t1);
31cfd832
TK		 1089 return;
1090 }
1091 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
1092 {
1093 if (GFC_DESCRIPTOR_RANK (a) != 1)
1094 {
1095 const GFC_COMPLEX_16 *restrict abase_x;
1096 const GFC_COMPLEX_16 *restrict bbase_y;
1097 GFC_COMPLEX_16 *restrict dest_y;
1098 GFC_COMPLEX_16 s;
1099
1100 for (y = 0; y < ycount; y++)
1101 {
1102 bbase_y = &bbase[y*bystride];
1103 dest_y = &dest[y*rystride];
1104 for (x = 0; x < xcount; x++)
1105 {
1106 abase_x = &abase[x*axstride];
1107 s = (GFC_COMPLEX_16) 0;
1108 for (n = 0; n < count; n++)
1109 s += abase_x[n] * bbase_y[n];
1110 dest_y[x] = s;
1111 }
1112 }
1113 }
1114 else
1115 {
1116 const GFC_COMPLEX_16 *restrict bbase_y;
1117 GFC_COMPLEX_16 s;
1118
1119 for (y = 0; y < ycount; y++)
1120 {
1121 bbase_y = &bbase[y*bystride];
1122 s = (GFC_COMPLEX_16) 0;
1123 for (n = 0; n < count; n++)
1124 s += abase[n*axstride] * bbase_y[n];
1125 dest[y*rystride] = s;
1126 }
1127 }
1128 }
1129 else if (axstride < aystride)
1130 {
1131 for (y = 0; y < ycount; y++)
1132 for (x = 0; x < xcount; x++)
1133 dest[x*rxstride + y*rystride] = (GFC_COMPLEX_16)0;
1134
1135 for (y = 0; y < ycount; y++)
1136 for (n = 0; n < count; n++)
1137 for (x = 0; x < xcount; x++)
1138 /* dest[x,y] += a[x,n] * b[n,y] */
1139 dest[x*rxstride + y*rystride] +=
1140 abase[x*axstride + n*aystride] *
1141 bbase[n*bxstride + y*bystride];
1142 }
1143 else if (GFC_DESCRIPTOR_RANK (a) == 1)
1144 {
1145 const GFC_COMPLEX_16 *restrict bbase_y;
1146 GFC_COMPLEX_16 s;
1147
1148 for (y = 0; y < ycount; y++)
1149 {
1150 bbase_y = &bbase[y*bystride];
1151 s = (GFC_COMPLEX_16) 0;
1152 for (n = 0; n < count; n++)
1153 s += abase[n*axstride] * bbase_y[n*bxstride];
1154 dest[y*rxstride] = s;
1155 }
1156 }
1157 else
1158 {
1159 const GFC_COMPLEX_16 *restrict abase_x;
1160 const GFC_COMPLEX_16 *restrict bbase_y;
1161 GFC_COMPLEX_16 *restrict dest_y;
1162 GFC_COMPLEX_16 s;
1163
1164 for (y = 0; y < ycount; y++)
1165 {
1166 bbase_y = &bbase[y*bystride];
1167 dest_y = &dest[y*rystride];
1168 for (x = 0; x < xcount; x++)
1169 {
1170 abase_x = &abase[x*axstride];
1171 s = (GFC_COMPLEX_16) 0;
1172 for (n = 0; n < count; n++)
1173 s += abase_x[n*aystride] * bbase_y[n*bxstride];
1174 dest_y[x*rxstride] = s;
1175 }
1176 }
1177 }
1178}
1179#undef POW3
1180#undef min
1181#undef max
1182
1183#endif /* HAVE_AVX2 */
1184
1185#ifdef HAVE_AVX512F
1186static void
1187matmul_c16_avx512f (gfc_array_c16 * const restrict retarray,
1188 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
1189 int blas_limit, blas_call gemm) __attribute__((__target__("avx512f")));
1190static void
1191matmul_c16_avx512f (gfc_array_c16 * const restrict retarray,
1192 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
1193 int blas_limit, blas_call gemm)
1194{
1195 const GFC_COMPLEX_16 * restrict abase;
1196 const GFC_COMPLEX_16 * restrict bbase;
1197 GFC_COMPLEX_16 * restrict dest;
1198
1199 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
1200 index_type x, y, n, count, xcount, ycount;
1201
1202 assert (GFC_DESCRIPTOR_RANK (a) == 2
1203 || GFC_DESCRIPTOR_RANK (b) == 2);
1204
1205/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
1206
1207 Either A or B (but not both) can be rank 1:
1208
1209 o One-dimensional argument A is implicitly treated as a row matrix
1210 dimensioned [1,count], so xcount=1.
1211
1212 o One-dimensional argument B is implicitly treated as a column matrix
1213 dimensioned [count, 1], so ycount=1.
1214*/
1215
1216 if (retarray->base_addr == NULL)
1217 {
1218 if (GFC_DESCRIPTOR_RANK (a) == 1)
1219 {
1220 GFC_DIMENSION_SET(retarray->dim[0], 0,
1221 GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
1222 }
1223 else if (GFC_DESCRIPTOR_RANK (b) == 1)
1224 {
1225 GFC_DIMENSION_SET(retarray->dim[0], 0,
1226 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
1227 }
1228 else
1229 {
1230 GFC_DIMENSION_SET(retarray->dim[0], 0,
1231 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
1232
1233 GFC_DIMENSION_SET(retarray->dim[1], 0,
1234 GFC_DESCRIPTOR_EXTENT(b,1) - 1,
1235 GFC_DESCRIPTOR_EXTENT(retarray,0));
1236 }
1237
1238 retarray->base_addr
1239 = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_COMPLEX_16));
1240 retarray->offset = 0;
1241 }
1242 else if (unlikely (compile_options.bounds_check))
1243 {
1244 index_type ret_extent, arg_extent;
1245
1246 if (GFC_DESCRIPTOR_RANK (a) == 1)
1247 {
1248 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
1249 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
1250 if (arg_extent != ret_extent)
1251 runtime_error ("Incorrect extent in return array in"
1252 " MATMUL intrinsic: is %ld, should be %ld",
1253 (long int) ret_extent, (long int) arg_extent);
1254 }
1255 else if (GFC_DESCRIPTOR_RANK (b) == 1)
1256 {
1257 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
1258 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
1259 if (arg_extent != ret_extent)
1260 runtime_error ("Incorrect extent in return array in"
1261 " MATMUL intrinsic: is %ld, should be %ld",
1262 (long int) ret_extent, (long int) arg_extent);
1263 }
1264 else
1265 {
1266 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
1267 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
1268 if (arg_extent != ret_extent)
1269 runtime_error ("Incorrect extent in return array in"
1270 " MATMUL intrinsic for dimension 1:"
1271 " is %ld, should be %ld",
1272 (long int) ret_extent, (long int) arg_extent);
1273
1274 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
1275 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
1276 if (arg_extent != ret_extent)
1277 runtime_error ("Incorrect extent in return array in"
1278 " MATMUL intrinsic for dimension 2:"
1279 " is %ld, should be %ld",
1280 (long int) ret_extent, (long int) arg_extent);
1281 }
1282 }
1283
1284
1285 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
1286 {
1287 /* One-dimensional result may be addressed in the code below
1288 either as a row or a column matrix. We want both cases to
1289 work. */
1290 rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
1291 }
1292 else
1293 {
1294 rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
1295 rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
1296 }
1297
1298
1299 if (GFC_DESCRIPTOR_RANK (a) == 1)
1300 {
1301 /* Treat it as a a row matrix A[1,count]. */
1302 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
1303 aystride = 1;
1304
1305 xcount = 1;
1306 count = GFC_DESCRIPTOR_EXTENT(a,0);
1307 }
1308 else
1309 {
1310 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
1311 aystride = GFC_DESCRIPTOR_STRIDE(a,1);
1312
1313 count = GFC_DESCRIPTOR_EXTENT(a,1);
1314 xcount = GFC_DESCRIPTOR_EXTENT(a,0);
1315 }
1316
1317 if (count != GFC_DESCRIPTOR_EXTENT(b,0))
1318 {
1319 if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
1320 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
1321 }
1322
1323 if (GFC_DESCRIPTOR_RANK (b) == 1)
1324 {
1325 /* Treat it as a column matrix B[count,1] */
1326 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
1327
1328 /* bystride should never be used for 1-dimensional b.
6ce6a84a
TK		 1329 The value is only used for calculation of the
1330 memory by the buffer. */
1331 bystride = 256;
31cfd832
TK		 1332 ycount = 1;
1333 }
1334 else
1335 {
1336 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
1337 bystride = GFC_DESCRIPTOR_STRIDE(b,1);
1338 ycount = GFC_DESCRIPTOR_EXTENT(b,1);
1339 }
1340
1341 abase = a->base_addr;
1342 bbase = b->base_addr;
1343 dest = retarray->base_addr;
1344
1345 /* Now that everything is set up, we perform the multiplication
1346 itself. */
1347
1348#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
1349#define min(a,b) ((a) <= (b) ? (a) : (b))
1350#define max(a,b) ((a) >= (b) ? (a) : (b))
1351
1352 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
1353 && (bxstride == 1 || bystride == 1)
1354 && (((float) xcount) * ((float) ycount) * ((float) count)
1355 > POW3(blas_limit)))
1356 {
1357 const int m = xcount, n = ycount, k = count, ldc = rystride;
1358 const GFC_COMPLEX_16 one = 1, zero = 0;
1359 const int lda = (axstride == 1) ? aystride : axstride,
1360 ldb = (bxstride == 1) ? bystride : bxstride;
1361
1362 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
1363 {
1364 assert (gemm != NULL);
1365 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m,
1366 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
1367 &ldc, 1, 1);
1368 return;
1369 }
1370 }
1371
1372 if (rxstride == 1 && axstride == 1 && bxstride == 1)
1373 {
1374 /* This block of code implements a tuned matmul, derived from
1375 Superscalar GEMM-based level 3 BLAS, Beta version 0.1
1376
1377 Bo Kagstrom and Per Ling
1378 Department of Computing Science
1379 Umea University
1380 S-901 87 Umea, Sweden
1381
1382 from netlib.org, translated to C, and modified for matmul.m4. */
1383
1384 const GFC_COMPLEX_16 *a, *b;
1385 GFC_COMPLEX_16 *c;
1386 const index_type m = xcount, n = ycount, k = count;
1387
1388 /* System generated locals */
1389 index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
1390 i1, i2, i3, i4, i5, i6;
1391
1392 /* Local variables */
fd991039 1393 GFC_COMPLEX_16 f11, f12, f21, f22, f31, f32, f41, f42,
31cfd832
TK		 1394 f13, f14, f23, f24, f33, f34, f43, f44;
1395 index_type i, j, l, ii, jj, ll;
1396 index_type isec, jsec, lsec, uisec, ujsec, ulsec;
8e5f30dc 1397 GFC_COMPLEX_16 *t1;
31cfd832
TK		 1398
1399 a = abase;
1400 b = bbase;
1401 c = retarray->base_addr;
1402
1403 /* Parameter adjustments */
1404 c_dim1 = rystride;
1405 c_offset = 1 + c_dim1;
1406 c -= c_offset;
1407 a_dim1 = aystride;
1408 a_offset = 1 + a_dim1;
1409 a -= a_offset;
1410 b_dim1 = bystride;
1411 b_offset = 1 + b_dim1;
1412 b -= b_offset;
1413
bbf97416
TK		 1414 /* Empty c first. */
1415 for (j=1; j<=n; j++)
1416 for (i=1; i<=m; i++)
1417 c[i + j * c_dim1] = (GFC_COMPLEX_16)0;
1418
31cfd832
TK		 1419 /* Early exit if possible */
1420 if (m == 0 || n == 0 || k == 0)
1421 return;
1422
fd991039
TK		 1423 /* Adjust size of t1 to what is needed. */
1424 index_type t1_dim;
e889aa0a 1425 t1_dim = (a_dim1 - (ycount > 1)) * 256 + b_dim1;
fd991039
TK		 1426 if (t1_dim > 65536)
1427 t1_dim = 65536;
1428
8e5f30dc 1429 t1 = malloc (t1_dim * sizeof(GFC_COMPLEX_16));
fd991039 1430
31cfd832
TK		 1431 /* Start turning the crank. */
1432 i1 = n;
1433 for (jj = 1; jj <= i1; jj += 512)
1434 {
1435 /* Computing MIN */
1436 i2 = 512;
1437 i3 = n - jj + 1;
1438 jsec = min(i2,i3);
1439 ujsec = jsec - jsec % 4;
1440 i2 = k;
1441 for (ll = 1; ll <= i2; ll += 256)
1442 {
1443 /* Computing MIN */
1444 i3 = 256;
1445 i4 = k - ll + 1;
1446 lsec = min(i3,i4);
1447 ulsec = lsec - lsec % 2;
1448
1449 i3 = m;
1450 for (ii = 1; ii <= i3; ii += 256)
1451 {
1452 /* Computing MIN */
1453 i4 = 256;
1454 i5 = m - ii + 1;
1455 isec = min(i4,i5);
1456 uisec = isec - isec % 2;
1457 i4 = ll + ulsec - 1;
1458 for (l = ll; l <= i4; l += 2)
1459 {
1460 i5 = ii + uisec - 1;
1461 for (i = ii; i <= i5; i += 2)
1462 {
1463 t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
1464 a[i + l * a_dim1];
1465 t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
1466 a[i + (l + 1) * a_dim1];
1467 t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
1468 a[i + 1 + l * a_dim1];
1469 t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
1470 a[i + 1 + (l + 1) * a_dim1];
1471 }
1472 if (uisec < isec)
1473 {
1474 t1[l - ll + 1 + (isec << 8) - 257] =
1475 a[ii + isec - 1 + l * a_dim1];
1476 t1[l - ll + 2 + (isec << 8) - 257] =
1477 a[ii + isec - 1 + (l + 1) * a_dim1];
1478 }
1479 }
1480 if (ulsec < lsec)
1481 {
1482 i4 = ii + isec - 1;
1483 for (i = ii; i<= i4; ++i)
1484 {
1485 t1[lsec + ((i - ii + 1) << 8) - 257] =
1486 a[i + (ll + lsec - 1) * a_dim1];
1487 }
1488 }
1489
1490 uisec = isec - isec % 4;
1491 i4 = jj + ujsec - 1;
1492 for (j = jj; j <= i4; j += 4)
1493 {
1494 i5 = ii + uisec - 1;
1495 for (i = ii; i <= i5; i += 4)
1496 {
1497 f11 = c[i + j * c_dim1];
1498 f21 = c[i + 1 + j * c_dim1];
1499 f12 = c[i + (j + 1) * c_dim1];
1500 f22 = c[i + 1 + (j + 1) * c_dim1];
1501 f13 = c[i + (j + 2) * c_dim1];
1502 f23 = c[i + 1 + (j + 2) * c_dim1];
1503 f14 = c[i + (j + 3) * c_dim1];
1504 f24 = c[i + 1 + (j + 3) * c_dim1];
1505 f31 = c[i + 2 + j * c_dim1];
1506 f41 = c[i + 3 + j * c_dim1];
1507 f32 = c[i + 2 + (j + 1) * c_dim1];
1508 f42 = c[i + 3 + (j + 1) * c_dim1];
1509 f33 = c[i + 2 + (j + 2) * c_dim1];
1510 f43 = c[i + 3 + (j + 2) * c_dim1];
1511 f34 = c[i + 2 + (j + 3) * c_dim1];
1512 f44 = c[i + 3 + (j + 3) * c_dim1];
1513 i6 = ll + lsec - 1;
1514 for (l = ll; l <= i6; ++l)
1515 {
1516 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
1517 * b[l + j * b_dim1];
1518 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
1519 * b[l + j * b_dim1];
1520 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
1521 * b[l + (j + 1) * b_dim1];
1522 f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
1523 * b[l + (j + 1) * b_dim1];
1524 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
1525 * b[l + (j + 2) * b_dim1];
1526 f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
1527 * b[l + (j + 2) * b_dim1];
1528 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
1529 * b[l + (j + 3) * b_dim1];
1530 f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
1531 * b[l + (j + 3) * b_dim1];
1532 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
1533 * b[l + j * b_dim1];
1534 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
1535 * b[l + j * b_dim1];
1536 f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
1537 * b[l + (j + 1) * b_dim1];
1538 f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
1539 * b[l + (j + 1) * b_dim1];
1540 f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
1541 * b[l + (j + 2) * b_dim1];
1542 f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
1543 * b[l + (j + 2) * b_dim1];
1544 f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
1545 * b[l + (j + 3) * b_dim1];
1546 f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
1547 * b[l + (j + 3) * b_dim1];
1548 }
1549 c[i + j * c_dim1] = f11;
1550 c[i + 1 + j * c_dim1] = f21;
1551 c[i + (j + 1) * c_dim1] = f12;
1552 c[i + 1 + (j + 1) * c_dim1] = f22;
1553 c[i + (j + 2) * c_dim1] = f13;
1554 c[i + 1 + (j + 2) * c_dim1] = f23;
1555 c[i + (j + 3) * c_dim1] = f14;
1556 c[i + 1 + (j + 3) * c_dim1] = f24;
1557 c[i + 2 + j * c_dim1] = f31;
1558 c[i + 3 + j * c_dim1] = f41;
1559 c[i + 2 + (j + 1) * c_dim1] = f32;
1560 c[i + 3 + (j + 1) * c_dim1] = f42;
1561 c[i + 2 + (j + 2) * c_dim1] = f33;
1562 c[i + 3 + (j + 2) * c_dim1] = f43;
1563 c[i + 2 + (j + 3) * c_dim1] = f34;
1564 c[i + 3 + (j + 3) * c_dim1] = f44;
1565 }
1566 if (uisec < isec)
1567 {
1568 i5 = ii + isec - 1;
1569 for (i = ii + uisec; i <= i5; ++i)
1570 {
1571 f11 = c[i + j * c_dim1];
1572 f12 = c[i + (j + 1) * c_dim1];
1573 f13 = c[i + (j + 2) * c_dim1];
1574 f14 = c[i + (j + 3) * c_dim1];
1575 i6 = ll + lsec - 1;
1576 for (l = ll; l <= i6; ++l)
1577 {
1578 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1579 257] * b[l + j * b_dim1];
1580 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1581 257] * b[l + (j + 1) * b_dim1];
1582 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1583 257] * b[l + (j + 2) * b_dim1];
1584 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1585 257] * b[l + (j + 3) * b_dim1];
1586 }
1587 c[i + j * c_dim1] = f11;
1588 c[i + (j + 1) * c_dim1] = f12;
1589 c[i + (j + 2) * c_dim1] = f13;
1590 c[i + (j + 3) * c_dim1] = f14;
1591 }
1592 }
1593 }
1594 if (ujsec < jsec)
1595 {
1596 i4 = jj + jsec - 1;
1597 for (j = jj + ujsec; j <= i4; ++j)
1598 {
1599 i5 = ii + uisec - 1;
1600 for (i = ii; i <= i5; i += 4)
1601 {
1602 f11 = c[i + j * c_dim1];
1603 f21 = c[i + 1 + j * c_dim1];
1604 f31 = c[i + 2 + j * c_dim1];
1605 f41 = c[i + 3 + j * c_dim1];
1606 i6 = ll + lsec - 1;
1607 for (l = ll; l <= i6; ++l)
1608 {
1609 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1610 257] * b[l + j * b_dim1];
1611 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
1612 257] * b[l + j * b_dim1];
1613 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
1614 257] * b[l + j * b_dim1];
1615 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
1616 257] * b[l + j * b_dim1];
1617 }
1618 c[i + j * c_dim1] = f11;
1619 c[i + 1 + j * c_dim1] = f21;
1620 c[i + 2 + j * c_dim1] = f31;
1621 c[i + 3 + j * c_dim1] = f41;
1622 }
1623 i5 = ii + isec - 1;
1624 for (i = ii + uisec; i <= i5; ++i)
1625 {
1626 f11 = c[i + j * c_dim1];
1627 i6 = ll + lsec - 1;
1628 for (l = ll; l <= i6; ++l)
1629 {
1630 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
1631 257] * b[l + j * b_dim1];
1632 }
1633 c[i + j * c_dim1] = f11;
1634 }
1635 }
1636 }
1637 }
1638 }
1639 }
8e5f30dc 1640 free(t1);
31cfd832
TK		 1641 return;
1642 }
1643 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
1644 {
1645 if (GFC_DESCRIPTOR_RANK (a) != 1)
1646 {
1647 const GFC_COMPLEX_16 *restrict abase_x;
1648 const GFC_COMPLEX_16 *restrict bbase_y;
1649 GFC_COMPLEX_16 *restrict dest_y;
1650 GFC_COMPLEX_16 s;
1651
1652 for (y = 0; y < ycount; y++)
1653 {
1654 bbase_y = &bbase[y*bystride];
1655 dest_y = &dest[y*rystride];
1656 for (x = 0; x < xcount; x++)
1657 {
1658 abase_x = &abase[x*axstride];
1659 s = (GFC_COMPLEX_16) 0;
1660 for (n = 0; n < count; n++)
1661 s += abase_x[n] * bbase_y[n];
1662 dest_y[x] = s;
1663 }
1664 }
1665 }
1666 else
1667 {
1668 const GFC_COMPLEX_16 *restrict bbase_y;
1669 GFC_COMPLEX_16 s;
1670
1671 for (y = 0; y < ycount; y++)
1672 {
1673 bbase_y = &bbase[y*bystride];
1674 s = (GFC_COMPLEX_16) 0;
1675 for (n = 0; n < count; n++)
1676 s += abase[n*axstride] * bbase_y[n];
1677 dest[y*rystride] = s;
1678 }
1679 }
1680 }
1681 else if (axstride < aystride)
1682 {
1683 for (y = 0; y < ycount; y++)
1684 for (x = 0; x < xcount; x++)
1685 dest[x*rxstride + y*rystride] = (GFC_COMPLEX_16)0;
1686
1687 for (y = 0; y < ycount; y++)
1688 for (n = 0; n < count; n++)
1689 for (x = 0; x < xcount; x++)
1690 /* dest[x,y] += a[x,n] * b[n,y] */
1691 dest[x*rxstride + y*rystride] +=
1692 abase[x*axstride + n*aystride] *
1693 bbase[n*bxstride + y*bystride];
1694 }
1695 else if (GFC_DESCRIPTOR_RANK (a) == 1)
1696 {
1697 const GFC_COMPLEX_16 *restrict bbase_y;
1698 GFC_COMPLEX_16 s;
1699
1700 for (y = 0; y < ycount; y++)
1701 {
1702 bbase_y = &bbase[y*bystride];
1703 s = (GFC_COMPLEX_16) 0;
1704 for (n = 0; n < count; n++)
1705 s += abase[n*axstride] * bbase_y[n*bxstride];
1706 dest[y*rxstride] = s;
1707 }
1708 }
1709 else
1710 {
1711 const GFC_COMPLEX_16 *restrict abase_x;
1712 const GFC_COMPLEX_16 *restrict bbase_y;
1713 GFC_COMPLEX_16 *restrict dest_y;
1714 GFC_COMPLEX_16 s;
1715
1716 for (y = 0; y < ycount; y++)
1717 {
1718 bbase_y = &bbase[y*bystride];
1719 dest_y = &dest[y*rystride];
1720 for (x = 0; x < xcount; x++)
1721 {
1722 abase_x = &abase[x*axstride];
1723 s = (GFC_COMPLEX_16) 0;
1724 for (n = 0; n < count; n++)
1725 s += abase_x[n*aystride] * bbase_y[n*bxstride];
1726 dest_y[x*rxstride] = s;
1727 }
1728 }
1729 }
1730}
1731#undef POW3
1732#undef min
1733#undef max
1734
1735#endif /* HAVE_AVX512F */
1736
1d5cf7fc
TK		 1737/* AMD-specifix funtions with AVX128 and FMA3/FMA4. */
1738
1739#if defined(HAVE_AVX) && defined(HAVE_FMA3) && defined(HAVE_AVX128)
1740void
1741matmul_c16_avx128_fma3 (gfc_array_c16 * const restrict retarray,
1742 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
1743 int blas_limit, blas_call gemm) __attribute__((__target__("avx,fma")));
1744internal_proto(matmul_c16_avx128_fma3);
1745#endif
1746
1747#if defined(HAVE_AVX) && defined(HAVE_FMA4) && defined(HAVE_AVX128)
1748void
1749matmul_c16_avx128_fma4 (gfc_array_c16 * const restrict retarray,
1750 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
1751 int blas_limit, blas_call gemm) __attribute__((__target__("avx,fma4")));
1752internal_proto(matmul_c16_avx128_fma4);
1753#endif
1754
31cfd832
TK		 1755/* Function to fall back to if there is no special processor-specific version. */
1756static void
1757matmul_c16_vanilla (gfc_array_c16 * const restrict retarray,
1758 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
1759 int blas_limit, blas_call gemm)
1760{
1761 const GFC_COMPLEX_16 * restrict abase;
1762 const GFC_COMPLEX_16 * restrict bbase;
1763 GFC_COMPLEX_16 * restrict dest;
1764
1765 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
1766 index_type x, y, n, count, xcount, ycount;
1767
1768 assert (GFC_DESCRIPTOR_RANK (a) == 2
1769 || GFC_DESCRIPTOR_RANK (b) == 2);
1770
1771/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
1772
1773 Either A or B (but not both) can be rank 1:
1774
1775 o One-dimensional argument A is implicitly treated as a row matrix
1776 dimensioned [1,count], so xcount=1.
1777
1778 o One-dimensional argument B is implicitly treated as a column matrix
1779 dimensioned [count, 1], so ycount=1.
1780*/
1781
1782 if (retarray->base_addr == NULL)
1783 {
1784 if (GFC_DESCRIPTOR_RANK (a) == 1)
1785 {
1786 GFC_DIMENSION_SET(retarray->dim[0], 0,
1787 GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
1788 }
1789 else if (GFC_DESCRIPTOR_RANK (b) == 1)
1790 {
1791 GFC_DIMENSION_SET(retarray->dim[0], 0,
1792 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
1793 }
1794 else
1795 {
1796 GFC_DIMENSION_SET(retarray->dim[0], 0,
1797 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
1798
1799 GFC_DIMENSION_SET(retarray->dim[1], 0,
1800 GFC_DESCRIPTOR_EXTENT(b,1) - 1,
1801 GFC_DESCRIPTOR_EXTENT(retarray,0));
1802 }
1803
1804 retarray->base_addr
1805 = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_COMPLEX_16));
1806 retarray->offset = 0;
1807 }
1808 else if (unlikely (compile_options.bounds_check))
1809 {
1810 index_type ret_extent, arg_extent;
1811
1812 if (GFC_DESCRIPTOR_RANK (a) == 1)
1813 {
1814 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
1815 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
1816 if (arg_extent != ret_extent)
1817 runtime_error ("Incorrect extent in return array in"
1818 " MATMUL intrinsic: is %ld, should be %ld",
1819 (long int) ret_extent, (long int) arg_extent);
1820 }
1821 else if (GFC_DESCRIPTOR_RANK (b) == 1)
1822 {
1823 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
1824 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
1825 if (arg_extent != ret_extent)
1826 runtime_error ("Incorrect extent in return array in"
1827 " MATMUL intrinsic: is %ld, should be %ld",
1828 (long int) ret_extent, (long int) arg_extent);
1829 }
1830 else
1831 {
1832 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
1833 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
1834 if (arg_extent != ret_extent)
1835 runtime_error ("Incorrect extent in return array in"
1836 " MATMUL intrinsic for dimension 1:"
1837 " is %ld, should be %ld",
1838 (long int) ret_extent, (long int) arg_extent);
1839
1840 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
1841 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
1842 if (arg_extent != ret_extent)
1843 runtime_error ("Incorrect extent in return array in"
1844 " MATMUL intrinsic for dimension 2:"
1845 " is %ld, should be %ld",
1846 (long int) ret_extent, (long int) arg_extent);
1847 }
1848 }
1849
1850
1851 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
1852 {
1853 /* One-dimensional result may be addressed in the code below
1854 either as a row or a column matrix. We want both cases to
1855 work. */
1856 rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
1857 }
1858 else
1859 {
1860 rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
1861 rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
1862 }
1863
1864
1865 if (GFC_DESCRIPTOR_RANK (a) == 1)
1866 {
1867 /* Treat it as a a row matrix A[1,count]. */
1868 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
1869 aystride = 1;
1870
1871 xcount = 1;
1872 count = GFC_DESCRIPTOR_EXTENT(a,0);
1873 }
1874 else
1875 {
1876 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
1877 aystride = GFC_DESCRIPTOR_STRIDE(a,1);
1878
1879 count = GFC_DESCRIPTOR_EXTENT(a,1);
1880 xcount = GFC_DESCRIPTOR_EXTENT(a,0);
1881 }
1882
1883 if (count != GFC_DESCRIPTOR_EXTENT(b,0))
1884 {
1885 if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
1886 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
1887 }
1888
1889 if (GFC_DESCRIPTOR_RANK (b) == 1)
1890 {
1891 /* Treat it as a column matrix B[count,1] */
1892 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
1893
1894 /* bystride should never be used for 1-dimensional b.
6ce6a84a
TK		 1895 The value is only used for calculation of the
1896 memory by the buffer. */
1897 bystride = 256;
31cfd832
TK		 1898 ycount = 1;
1899 }
1900 else
1901 {
1902 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
1903 bystride = GFC_DESCRIPTOR_STRIDE(b,1);
1904 ycount = GFC_DESCRIPTOR_EXTENT(b,1);
1905 }
1906
1907 abase = a->base_addr;
1908 bbase = b->base_addr;
1909 dest = retarray->base_addr;
1910
1911 /* Now that everything is set up, we perform the multiplication
1912 itself. */
1913
1914#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
1915#define min(a,b) ((a) <= (b) ? (a) : (b))
1916#define max(a,b) ((a) >= (b) ? (a) : (b))
1917
1918 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
1919 && (bxstride == 1 || bystride == 1)
1920 && (((float) xcount) * ((float) ycount) * ((float) count)
1921 > POW3(blas_limit)))
1922 {
1923 const int m = xcount, n = ycount, k = count, ldc = rystride;
1924 const GFC_COMPLEX_16 one = 1, zero = 0;
1925 const int lda = (axstride == 1) ? aystride : axstride,
1926 ldb = (bxstride == 1) ? bystride : bxstride;
1927
1928 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
1929 {
1930 assert (gemm != NULL);
1931 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m,
1932 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
1933 &ldc, 1, 1);
1934 return;
1935 }
1936 }
1937
1938 if (rxstride == 1 && axstride == 1 && bxstride == 1)
1939 {
1940 /* This block of code implements a tuned matmul, derived from
1941 Superscalar GEMM-based level 3 BLAS, Beta version 0.1
1942
1943 Bo Kagstrom and Per Ling
1944 Department of Computing Science
1945 Umea University
1946 S-901 87 Umea, Sweden
1947
1948 from netlib.org, translated to C, and modified for matmul.m4. */
1949
1950 const GFC_COMPLEX_16 *a, *b;
1951 GFC_COMPLEX_16 *c;
1952 const index_type m = xcount, n = ycount, k = count;
1953
1954 /* System generated locals */
1955 index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
1956 i1, i2, i3, i4, i5, i6;
1957
1958 /* Local variables */
fd991039 1959 GFC_COMPLEX_16 f11, f12, f21, f22, f31, f32, f41, f42,
31cfd832
TK		 1960 f13, f14, f23, f24, f33, f34, f43, f44;
1961 index_type i, j, l, ii, jj, ll;
1962 index_type isec, jsec, lsec, uisec, ujsec, ulsec;
8e5f30dc 1963 GFC_COMPLEX_16 *t1;
31cfd832
TK		 1964
1965 a = abase;
1966 b = bbase;
1967 c = retarray->base_addr;
1968
1969 /* Parameter adjustments */
1970 c_dim1 = rystride;
1971 c_offset = 1 + c_dim1;
1972 c -= c_offset;
1973 a_dim1 = aystride;
1974 a_offset = 1 + a_dim1;
1975 a -= a_offset;
1976 b_dim1 = bystride;
1977 b_offset = 1 + b_dim1;
1978 b -= b_offset;
1979
bbf97416
TK		 1980 /* Empty c first. */
1981 for (j=1; j<=n; j++)
1982 for (i=1; i<=m; i++)
1983 c[i + j * c_dim1] = (GFC_COMPLEX_16)0;
1984
31cfd832
TK		 1985 /* Early exit if possible */
1986 if (m == 0 || n == 0 || k == 0)
1987 return;
1988
fd991039
TK		 1989 /* Adjust size of t1 to what is needed. */
1990 index_type t1_dim;
e889aa0a 1991 t1_dim = (a_dim1 - (ycount > 1)) * 256 + b_dim1;
fd991039
TK		 1992 if (t1_dim > 65536)
1993 t1_dim = 65536;
1994
8e5f30dc 1995 t1 = malloc (t1_dim * sizeof(GFC_COMPLEX_16));
fd991039 1996
31cfd832
TK		 1997 /* Start turning the crank. */
1998 i1 = n;
1999 for (jj = 1; jj <= i1; jj += 512)
2000 {
2001 /* Computing MIN */
2002 i2 = 512;
2003 i3 = n - jj + 1;
2004 jsec = min(i2,i3);
2005 ujsec = jsec - jsec % 4;
2006 i2 = k;
2007 for (ll = 1; ll <= i2; ll += 256)
2008 {
2009 /* Computing MIN */
2010 i3 = 256;
2011 i4 = k - ll + 1;
2012 lsec = min(i3,i4);
2013 ulsec = lsec - lsec % 2;
2014
2015 i3 = m;
2016 for (ii = 1; ii <= i3; ii += 256)
2017 {
2018 /* Computing MIN */
2019 i4 = 256;
2020 i5 = m - ii + 1;
2021 isec = min(i4,i5);
2022 uisec = isec - isec % 2;
2023 i4 = ll + ulsec - 1;
2024 for (l = ll; l <= i4; l += 2)
2025 {
2026 i5 = ii + uisec - 1;
2027 for (i = ii; i <= i5; i += 2)
2028 {
2029 t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
2030 a[i + l * a_dim1];
2031 t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
2032 a[i + (l + 1) * a_dim1];
2033 t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
2034 a[i + 1 + l * a_dim1];
2035 t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
2036 a[i + 1 + (l + 1) * a_dim1];
2037 }
2038 if (uisec < isec)
2039 {
2040 t1[l - ll + 1 + (isec << 8) - 257] =
2041 a[ii + isec - 1 + l * a_dim1];
2042 t1[l - ll + 2 + (isec << 8) - 257] =
2043 a[ii + isec - 1 + (l + 1) * a_dim1];
2044 }
2045 }
2046 if (ulsec < lsec)
2047 {
2048 i4 = ii + isec - 1;
2049 for (i = ii; i<= i4; ++i)
2050 {
2051 t1[lsec + ((i - ii + 1) << 8) - 257] =
2052 a[i + (ll + lsec - 1) * a_dim1];
2053 }
2054 }
2055
2056 uisec = isec - isec % 4;
2057 i4 = jj + ujsec - 1;
2058 for (j = jj; j <= i4; j += 4)
2059 {
2060 i5 = ii + uisec - 1;
2061 for (i = ii; i <= i5; i += 4)
2062 {
2063 f11 = c[i + j * c_dim1];
2064 f21 = c[i + 1 + j * c_dim1];
2065 f12 = c[i + (j + 1) * c_dim1];
2066 f22 = c[i + 1 + (j + 1) * c_dim1];
2067 f13 = c[i + (j + 2) * c_dim1];
2068 f23 = c[i + 1 + (j + 2) * c_dim1];
2069 f14 = c[i + (j + 3) * c_dim1];
2070 f24 = c[i + 1 + (j + 3) * c_dim1];
2071 f31 = c[i + 2 + j * c_dim1];
2072 f41 = c[i + 3 + j * c_dim1];
2073 f32 = c[i + 2 + (j + 1) * c_dim1];
2074 f42 = c[i + 3 + (j + 1) * c_dim1];
2075 f33 = c[i + 2 + (j + 2) * c_dim1];
2076 f43 = c[i + 3 + (j + 2) * c_dim1];
2077 f34 = c[i + 2 + (j + 3) * c_dim1];
2078 f44 = c[i + 3 + (j + 3) * c_dim1];
2079 i6 = ll + lsec - 1;
2080 for (l = ll; l <= i6; ++l)
2081 {
2082 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2083 * b[l + j * b_dim1];
2084 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2085 * b[l + j * b_dim1];
2086 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2087 * b[l + (j + 1) * b_dim1];
2088 f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2089 * b[l + (j + 1) * b_dim1];
2090 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2091 * b[l + (j + 2) * b_dim1];
2092 f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2093 * b[l + (j + 2) * b_dim1];
2094 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2095 * b[l + (j + 3) * b_dim1];
2096 f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2097 * b[l + (j + 3) * b_dim1];
2098 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2099 * b[l + j * b_dim1];
2100 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2101 * b[l + j * b_dim1];
2102 f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2103 * b[l + (j + 1) * b_dim1];
2104 f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2105 * b[l + (j + 1) * b_dim1];
2106 f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2107 * b[l + (j + 2) * b_dim1];
2108 f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2109 * b[l + (j + 2) * b_dim1];
2110 f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2111 * b[l + (j + 3) * b_dim1];
2112 f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2113 * b[l + (j + 3) * b_dim1];
2114 }
2115 c[i + j * c_dim1] = f11;
2116 c[i + 1 + j * c_dim1] = f21;
2117 c[i + (j + 1) * c_dim1] = f12;
2118 c[i + 1 + (j + 1) * c_dim1] = f22;
2119 c[i + (j + 2) * c_dim1] = f13;
2120 c[i + 1 + (j + 2) * c_dim1] = f23;
2121 c[i + (j + 3) * c_dim1] = f14;
2122 c[i + 1 + (j + 3) * c_dim1] = f24;
2123 c[i + 2 + j * c_dim1] = f31;
2124 c[i + 3 + j * c_dim1] = f41;
2125 c[i + 2 + (j + 1) * c_dim1] = f32;
2126 c[i + 3 + (j + 1) * c_dim1] = f42;
2127 c[i + 2 + (j + 2) * c_dim1] = f33;
2128 c[i + 3 + (j + 2) * c_dim1] = f43;
2129 c[i + 2 + (j + 3) * c_dim1] = f34;
2130 c[i + 3 + (j + 3) * c_dim1] = f44;
2131 }
2132 if (uisec < isec)
2133 {
2134 i5 = ii + isec - 1;
2135 for (i = ii + uisec; i <= i5; ++i)
2136 {
2137 f11 = c[i + j * c_dim1];
2138 f12 = c[i + (j + 1) * c_dim1];
2139 f13 = c[i + (j + 2) * c_dim1];
2140 f14 = c[i + (j + 3) * c_dim1];
2141 i6 = ll + lsec - 1;
2142 for (l = ll; l <= i6; ++l)
2143 {
2144 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2145 257] * b[l + j * b_dim1];
2146 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2147 257] * b[l + (j + 1) * b_dim1];
2148 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2149 257] * b[l + (j + 2) * b_dim1];
2150 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2151 257] * b[l + (j + 3) * b_dim1];
2152 }
2153 c[i + j * c_dim1] = f11;
2154 c[i + (j + 1) * c_dim1] = f12;
2155 c[i + (j + 2) * c_dim1] = f13;
2156 c[i + (j + 3) * c_dim1] = f14;
2157 }
2158 }
2159 }
2160 if (ujsec < jsec)
2161 {
2162 i4 = jj + jsec - 1;
2163 for (j = jj + ujsec; j <= i4; ++j)
2164 {
2165 i5 = ii + uisec - 1;
2166 for (i = ii; i <= i5; i += 4)
2167 {
2168 f11 = c[i + j * c_dim1];
2169 f21 = c[i + 1 + j * c_dim1];
2170 f31 = c[i + 2 + j * c_dim1];
2171 f41 = c[i + 3 + j * c_dim1];
2172 i6 = ll + lsec - 1;
2173 for (l = ll; l <= i6; ++l)
2174 {
2175 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2176 257] * b[l + j * b_dim1];
2177 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
2178 257] * b[l + j * b_dim1];
2179 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
2180 257] * b[l + j * b_dim1];
2181 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
2182 257] * b[l + j * b_dim1];
2183 }
2184 c[i + j * c_dim1] = f11;
2185 c[i + 1 + j * c_dim1] = f21;
2186 c[i + 2 + j * c_dim1] = f31;
2187 c[i + 3 + j * c_dim1] = f41;
2188 }
2189 i5 = ii + isec - 1;
2190 for (i = ii + uisec; i <= i5; ++i)
2191 {
2192 f11 = c[i + j * c_dim1];
2193 i6 = ll + lsec - 1;
2194 for (l = ll; l <= i6; ++l)
2195 {
2196 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2197 257] * b[l + j * b_dim1];
2198 }
2199 c[i + j * c_dim1] = f11;
2200 }
2201 }
2202 }
2203 }
2204 }
2205 }
8e5f30dc 2206 free(t1);
31cfd832
TK		 2207 return;
2208 }
2209 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
2210 {
2211 if (GFC_DESCRIPTOR_RANK (a) != 1)
2212 {
2213 const GFC_COMPLEX_16 *restrict abase_x;
2214 const GFC_COMPLEX_16 *restrict bbase_y;
2215 GFC_COMPLEX_16 *restrict dest_y;
2216 GFC_COMPLEX_16 s;
2217
2218 for (y = 0; y < ycount; y++)
2219 {
2220 bbase_y = &bbase[y*bystride];
2221 dest_y = &dest[y*rystride];
2222 for (x = 0; x < xcount; x++)
2223 {
2224 abase_x = &abase[x*axstride];
2225 s = (GFC_COMPLEX_16) 0;
2226 for (n = 0; n < count; n++)
2227 s += abase_x[n] * bbase_y[n];
2228 dest_y[x] = s;
2229 }
2230 }
2231 }
2232 else
2233 {
2234 const GFC_COMPLEX_16 *restrict bbase_y;
2235 GFC_COMPLEX_16 s;
2236
2237 for (y = 0; y < ycount; y++)
2238 {
2239 bbase_y = &bbase[y*bystride];
2240 s = (GFC_COMPLEX_16) 0;
2241 for (n = 0; n < count; n++)
2242 s += abase[n*axstride] * bbase_y[n];
2243 dest[y*rystride] = s;
2244 }
2245 }
2246 }
2247 else if (axstride < aystride)
2248 {
2249 for (y = 0; y < ycount; y++)
2250 for (x = 0; x < xcount; x++)
2251 dest[x*rxstride + y*rystride] = (GFC_COMPLEX_16)0;
2252
2253 for (y = 0; y < ycount; y++)
2254 for (n = 0; n < count; n++)
2255 for (x = 0; x < xcount; x++)
2256 /* dest[x,y] += a[x,n] * b[n,y] */
2257 dest[x*rxstride + y*rystride] +=
2258 abase[x*axstride + n*aystride] *
2259 bbase[n*bxstride + y*bystride];
2260 }
2261 else if (GFC_DESCRIPTOR_RANK (a) == 1)
2262 {
2263 const GFC_COMPLEX_16 *restrict bbase_y;
2264 GFC_COMPLEX_16 s;
2265
2266 for (y = 0; y < ycount; y++)
2267 {
2268 bbase_y = &bbase[y*bystride];
2269 s = (GFC_COMPLEX_16) 0;
2270 for (n = 0; n < count; n++)
2271 s += abase[n*axstride] * bbase_y[n*bxstride];
2272 dest[y*rxstride] = s;
2273 }
2274 }
2275 else
2276 {
2277 const GFC_COMPLEX_16 *restrict abase_x;
2278 const GFC_COMPLEX_16 *restrict bbase_y;
2279 GFC_COMPLEX_16 *restrict dest_y;
2280 GFC_COMPLEX_16 s;
2281
2282 for (y = 0; y < ycount; y++)
2283 {
2284 bbase_y = &bbase[y*bystride];
2285 dest_y = &dest[y*rystride];
2286 for (x = 0; x < xcount; x++)
2287 {
2288 abase_x = &abase[x*axstride];
2289 s = (GFC_COMPLEX_16) 0;
2290 for (n = 0; n < count; n++)
2291 s += abase_x[n*aystride] * bbase_y[n*bxstride];
2292 dest_y[x*rxstride] = s;
2293 }
2294 }
2295 }
2296}
2297#undef POW3
2298#undef min
2299#undef max
2300
2301
2302/* Compiling main function, with selection code for the processor. */
2303
2304/* Currently, this is i386 only. Adjust for other architectures. */
2305
2306#include <config/i386/cpuinfo.h>
2307void matmul_c16 (gfc_array_c16 * const restrict retarray,
2308 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
2309 int blas_limit, blas_call gemm)
2310{
2311 static void (*matmul_p) (gfc_array_c16 * const restrict retarray,
2312 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
f03e9217
TK		 2313 int blas_limit, blas_call gemm);
2314
2315 void (*matmul_fn) (gfc_array_c16 * const restrict retarray,
2316 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
2317 int blas_limit, blas_call gemm);
31cfd832 2318
f03e9217
TK		 2319 matmul_fn = __atomic_load_n (&matmul_p, __ATOMIC_RELAXED);
2320 if (matmul_fn == NULL)
31cfd832 2321 {
f03e9217 2322 matmul_fn = matmul_c16_vanilla;
31cfd832
TK		 2323 if (__cpu_model.__cpu_vendor == VENDOR_INTEL)
2324 {
2325 /* Run down the available processors in order of preference. */
2326#ifdef HAVE_AVX512F
2327 if (__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX512F))
2328 {
f03e9217
TK		 2329 matmul_fn = matmul_c16_avx512f;
2330 goto store;
31cfd832
TK		 2331 }
2332
2333#endif /* HAVE_AVX512F */
2334
2335#ifdef HAVE_AVX2
6d03bdcc
TK		 2336 if ((__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX2))
2337 && (__cpu_model.__cpu_features[0] & (1 << FEATURE_FMA)))
31cfd832 2338 {
f03e9217
TK		 2339 matmul_fn = matmul_c16_avx2;
2340 goto store;
31cfd832
TK		 2341 }
2342
2343#endif
2344
2345#ifdef HAVE_AVX
2346 if (__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX))
2347 {
f03e9217
TK		 2348 matmul_fn = matmul_c16_avx;
2349 goto store;
31cfd832
TK		 2350 }
2351#endif /* HAVE_AVX */
2352 }
1d5cf7fc
TK		 2353 else if (__cpu_model.__cpu_vendor == VENDOR_AMD)
2354 {
2355#if defined(HAVE_AVX) && defined(HAVE_FMA3) && defined(HAVE_AVX128)
2356 if ((__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX))
2357 && (__cpu_model.__cpu_features[0] & (1 << FEATURE_FMA)))
2358 {
2359 matmul_fn = matmul_c16_avx128_fma3;
2360 goto store;
2361 }
2362#endif
2363#if defined(HAVE_AVX) && defined(HAVE_FMA4) && defined(HAVE_AVX128)
2364 if ((__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX))
2365 && (__cpu_model.__cpu_features[0] & (1 << FEATURE_FMA4)))
2366 {
2367 matmul_fn = matmul_c16_avx128_fma4;
2368 goto store;
2369 }
2370#endif
2371
2372 }
f03e9217
TK		 2373 store:
2374 __atomic_store_n (&matmul_p, matmul_fn, __ATOMIC_RELAXED);
31cfd832
TK		 2375 }
2376
f03e9217 2377 (*matmul_fn) (retarray, a, b, try_blas, blas_limit, gemm);
31cfd832
TK		 2378}
2379
2380#else /* Just the vanilla function. */
2381
644cb69f 2382void
85206901 2383matmul_c16 (gfc_array_c16 * const restrict retarray,
5a0aad31
FXC		 2384 gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
2385 int blas_limit, blas_call gemm)
644cb69f 2386{
85206901
JB		 2387 const GFC_COMPLEX_16 * restrict abase;
2388 const GFC_COMPLEX_16 * restrict bbase;
2389 GFC_COMPLEX_16 * restrict dest;
644cb69f
FXC		 2390
2391 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
2392 index_type x, y, n, count, xcount, ycount;
2393
2394 assert (GFC_DESCRIPTOR_RANK (a) == 2
2395 || GFC_DESCRIPTOR_RANK (b) == 2);
2396
2397/* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
2398
2399 Either A or B (but not both) can be rank 1:
2400
2401 o One-dimensional argument A is implicitly treated as a row matrix
2402 dimensioned [1,count], so xcount=1.
2403
2404 o One-dimensional argument B is implicitly treated as a column matrix
2405 dimensioned [count, 1], so ycount=1.
5d70ab07 2406*/
644cb69f 2407
21d1335b 2408 if (retarray->base_addr == NULL)
644cb69f
FXC		 2409 {
2410 if (GFC_DESCRIPTOR_RANK (a) == 1)
2411 {
dfb55fdc
TK		 2412 GFC_DIMENSION_SET(retarray->dim[0], 0,
2413 GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
644cb69f
FXC		 2414 }
2415 else if (GFC_DESCRIPTOR_RANK (b) == 1)
2416 {
dfb55fdc
TK		 2417 GFC_DIMENSION_SET(retarray->dim[0], 0,
2418 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
644cb69f
FXC		 2419 }
2420 else
2421 {
dfb55fdc
TK		 2422 GFC_DIMENSION_SET(retarray->dim[0], 0,
2423 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
644cb69f 2424
dfb55fdc
TK		 2425 GFC_DIMENSION_SET(retarray->dim[1], 0,
2426 GFC_DESCRIPTOR_EXTENT(b,1) - 1,
2427 GFC_DESCRIPTOR_EXTENT(retarray,0));
644cb69f
FXC		 2428 }
2429
21d1335b 2430 retarray->base_addr
92e6f3a4 2431 = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_COMPLEX_16));
644cb69f
FXC		 2432 retarray->offset = 0;
2433 }
5d70ab07
JD		 2434 else if (unlikely (compile_options.bounds_check))
2435 {
2436 index_type ret_extent, arg_extent;
2437
2438 if (GFC_DESCRIPTOR_RANK (a) == 1)
2439 {
2440 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
2441 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
2442 if (arg_extent != ret_extent)
2443 runtime_error ("Incorrect extent in return array in"
2444 " MATMUL intrinsic: is %ld, should be %ld",
2445 (long int) ret_extent, (long int) arg_extent);
2446 }
2447 else if (GFC_DESCRIPTOR_RANK (b) == 1)
2448 {
2449 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
2450 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
2451 if (arg_extent != ret_extent)
2452 runtime_error ("Incorrect extent in return array in"
2453 " MATMUL intrinsic: is %ld, should be %ld",
2454 (long int) ret_extent, (long int) arg_extent);
2455 }
2456 else
2457 {
2458 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
2459 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
2460 if (arg_extent != ret_extent)
2461 runtime_error ("Incorrect extent in return array in"
2462 " MATMUL intrinsic for dimension 1:"
2463 " is %ld, should be %ld",
2464 (long int) ret_extent, (long int) arg_extent);
2465
2466 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
2467 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
2468 if (arg_extent != ret_extent)
2469 runtime_error ("Incorrect extent in return array in"
2470 " MATMUL intrinsic for dimension 2:"
2471 " is %ld, should be %ld",
2472 (long int) ret_extent, (long int) arg_extent);
2473 }
2474 }
644cb69f 2475
644cb69f
FXC		 2476
2477 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
2478 {
2479 /* One-dimensional result may be addressed in the code below
2480 either as a row or a column matrix. We want both cases to
2481 work. */
dfb55fdc 2482 rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
644cb69f
FXC		 2483 }
2484 else
2485 {
dfb55fdc
TK		 2486 rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
2487 rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
644cb69f
FXC		 2488 }
2489
2490
2491 if (GFC_DESCRIPTOR_RANK (a) == 1)
2492 {
2493 /* Treat it as a a row matrix A[1,count]. */
dfb55fdc 2494 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
644cb69f
FXC		 2495 aystride = 1;
2496
2497 xcount = 1;
dfb55fdc 2498 count = GFC_DESCRIPTOR_EXTENT(a,0);
644cb69f
FXC		 2499 }
2500 else
2501 {
dfb55fdc
TK		 2502 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
2503 aystride = GFC_DESCRIPTOR_STRIDE(a,1);
644cb69f 2504
dfb55fdc
TK		 2505 count = GFC_DESCRIPTOR_EXTENT(a,1);
2506 xcount = GFC_DESCRIPTOR_EXTENT(a,0);
644cb69f
FXC		 2507 }
2508
dfb55fdc 2509 if (count != GFC_DESCRIPTOR_EXTENT(b,0))
7edc89d4 2510 {
dfb55fdc 2511 if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
7edc89d4
TK		 2512 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
2513 }
644cb69f
FXC		 2514
2515 if (GFC_DESCRIPTOR_RANK (b) == 1)
2516 {
2517 /* Treat it as a column matrix B[count,1] */
dfb55fdc 2518 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
644cb69f
FXC		 2519
2520 /* bystride should never be used for 1-dimensional b.
6ce6a84a
TK		 2521 The value is only used for calculation of the
2522 memory by the buffer. */
2523 bystride = 256;
644cb69f
FXC		 2524 ycount = 1;
2525 }
2526 else
2527 {
dfb55fdc
TK		 2528 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
2529 bystride = GFC_DESCRIPTOR_STRIDE(b,1);
2530 ycount = GFC_DESCRIPTOR_EXTENT(b,1);
644cb69f
FXC		 2531 }
2532
21d1335b
TB		 2533 abase = a->base_addr;
2534 bbase = b->base_addr;
2535 dest = retarray->base_addr;
644cb69f 2536
5d70ab07 2537 /* Now that everything is set up, we perform the multiplication
5a0aad31
FXC		 2538 itself. */
2539
2540#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
5d70ab07
JD		 2541#define min(a,b) ((a) <= (b) ? (a) : (b))
2542#define max(a,b) ((a) >= (b) ? (a) : (b))
5a0aad31
FXC		 2543
2544 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
2545 && (bxstride == 1 || bystride == 1)
2546 && (((float) xcount) * ((float) ycount) * ((float) count)
2547 > POW3(blas_limit)))
644cb69f 2548 {
5d70ab07
JD		 2549 const int m = xcount, n = ycount, k = count, ldc = rystride;
2550 const GFC_COMPLEX_16 one = 1, zero = 0;
2551 const int lda = (axstride == 1) ? aystride : axstride,
2552 ldb = (bxstride == 1) ? bystride : bxstride;
644cb69f 2553
5d70ab07 2554 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
644cb69f 2555 {
5d70ab07
JD		 2556 assert (gemm != NULL);
2557 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m,
2558 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
2559 &ldc, 1, 1);
2560 return;
644cb69f 2561 }
5d70ab07 2562 }
644cb69f 2563
5d70ab07
JD		 2564 if (rxstride == 1 && axstride == 1 && bxstride == 1)
2565 {
2566 /* This block of code implements a tuned matmul, derived from
2567 Superscalar GEMM-based level 3 BLAS, Beta version 0.1
2568
2569 Bo Kagstrom and Per Ling
2570 Department of Computing Science
2571 Umea University
2572 S-901 87 Umea, Sweden
2573
2574 from netlib.org, translated to C, and modified for matmul.m4. */
2575
2576 const GFC_COMPLEX_16 *a, *b;
2577 GFC_COMPLEX_16 *c;
2578 const index_type m = xcount, n = ycount, k = count;
2579
2580 /* System generated locals */
2581 index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
2582 i1, i2, i3, i4, i5, i6;
2583
2584 /* Local variables */
fd991039 2585 GFC_COMPLEX_16 f11, f12, f21, f22, f31, f32, f41, f42,
5d70ab07
JD		 2586 f13, f14, f23, f24, f33, f34, f43, f44;
2587 index_type i, j, l, ii, jj, ll;
2588 index_type isec, jsec, lsec, uisec, ujsec, ulsec;
8e5f30dc 2589 GFC_COMPLEX_16 *t1;
5d70ab07
JD		 2590
2591 a = abase;
2592 b = bbase;
2593 c = retarray->base_addr;
2594
2595 /* Parameter adjustments */
2596 c_dim1 = rystride;
2597 c_offset = 1 + c_dim1;
2598 c -= c_offset;
2599 a_dim1 = aystride;
2600 a_offset = 1 + a_dim1;
2601 a -= a_offset;
2602 b_dim1 = bystride;
2603 b_offset = 1 + b_dim1;
2604 b -= b_offset;
2605
bbf97416
TK		 2606 /* Empty c first. */
2607 for (j=1; j<=n; j++)
2608 for (i=1; i<=m; i++)
2609 c[i + j * c_dim1] = (GFC_COMPLEX_16)0;
2610
5d70ab07
JD		 2611 /* Early exit if possible */
2612 if (m == 0 || n == 0 || k == 0)
2613 return;
2614
fd991039
TK		 2615 /* Adjust size of t1 to what is needed. */
2616 index_type t1_dim;
e889aa0a 2617 t1_dim = (a_dim1 - (ycount > 1)) * 256 + b_dim1;
fd991039
TK		 2618 if (t1_dim > 65536)
2619 t1_dim = 65536;
2620
8e5f30dc 2621 t1 = malloc (t1_dim * sizeof(GFC_COMPLEX_16));
fd991039 2622
5d70ab07
JD		 2623 /* Start turning the crank. */
2624 i1 = n;
2625 for (jj = 1; jj <= i1; jj += 512)
644cb69f 2626 {
5d70ab07
JD		 2627 /* Computing MIN */
2628 i2 = 512;
2629 i3 = n - jj + 1;
2630 jsec = min(i2,i3);
2631 ujsec = jsec - jsec % 4;
2632 i2 = k;
2633 for (ll = 1; ll <= i2; ll += 256)
644cb69f 2634 {
5d70ab07
JD		 2635 /* Computing MIN */
2636 i3 = 256;
2637 i4 = k - ll + 1;
2638 lsec = min(i3,i4);
2639 ulsec = lsec - lsec % 2;
2640
2641 i3 = m;
2642 for (ii = 1; ii <= i3; ii += 256)
644cb69f 2643 {
5d70ab07
JD		 2644 /* Computing MIN */
2645 i4 = 256;
2646 i5 = m - ii + 1;
2647 isec = min(i4,i5);
2648 uisec = isec - isec % 2;
2649 i4 = ll + ulsec - 1;
2650 for (l = ll; l <= i4; l += 2)
2651 {
2652 i5 = ii + uisec - 1;
2653 for (i = ii; i <= i5; i += 2)
2654 {
2655 t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
2656 a[i + l * a_dim1];
2657 t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
2658 a[i + (l + 1) * a_dim1];
2659 t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
2660 a[i + 1 + l * a_dim1];
2661 t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
2662 a[i + 1 + (l + 1) * a_dim1];
2663 }
2664 if (uisec < isec)
2665 {
2666 t1[l - ll + 1 + (isec << 8) - 257] =
2667 a[ii + isec - 1 + l * a_dim1];
2668 t1[l - ll + 2 + (isec << 8) - 257] =
2669 a[ii + isec - 1 + (l + 1) * a_dim1];
2670 }
2671 }
2672 if (ulsec < lsec)
2673 {
2674 i4 = ii + isec - 1;
2675 for (i = ii; i<= i4; ++i)
2676 {
2677 t1[lsec + ((i - ii + 1) << 8) - 257] =
2678 a[i + (ll + lsec - 1) * a_dim1];
2679 }
2680 }
2681
2682 uisec = isec - isec % 4;
2683 i4 = jj + ujsec - 1;
2684 for (j = jj; j <= i4; j += 4)
2685 {
2686 i5 = ii + uisec - 1;
2687 for (i = ii; i <= i5; i += 4)
2688 {
2689 f11 = c[i + j * c_dim1];
2690 f21 = c[i + 1 + j * c_dim1];
2691 f12 = c[i + (j + 1) * c_dim1];
2692 f22 = c[i + 1 + (j + 1) * c_dim1];
2693 f13 = c[i + (j + 2) * c_dim1];
2694 f23 = c[i + 1 + (j + 2) * c_dim1];
2695 f14 = c[i + (j + 3) * c_dim1];
2696 f24 = c[i + 1 + (j + 3) * c_dim1];
2697 f31 = c[i + 2 + j * c_dim1];
2698 f41 = c[i + 3 + j * c_dim1];
2699 f32 = c[i + 2 + (j + 1) * c_dim1];
2700 f42 = c[i + 3 + (j + 1) * c_dim1];
2701 f33 = c[i + 2 + (j + 2) * c_dim1];
2702 f43 = c[i + 3 + (j + 2) * c_dim1];
2703 f34 = c[i + 2 + (j + 3) * c_dim1];
2704 f44 = c[i + 3 + (j + 3) * c_dim1];
2705 i6 = ll + lsec - 1;
2706 for (l = ll; l <= i6; ++l)
2707 {
2708 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2709 * b[l + j * b_dim1];
2710 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2711 * b[l + j * b_dim1];
2712 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2713 * b[l + (j + 1) * b_dim1];
2714 f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2715 * b[l + (j + 1) * b_dim1];
2716 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2717 * b[l + (j + 2) * b_dim1];
2718 f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2719 * b[l + (j + 2) * b_dim1];
2720 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
2721 * b[l + (j + 3) * b_dim1];
2722 f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
2723 * b[l + (j + 3) * b_dim1];
2724 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2725 * b[l + j * b_dim1];
2726 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2727 * b[l + j * b_dim1];
2728 f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2729 * b[l + (j + 1) * b_dim1];
2730 f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2731 * b[l + (j + 1) * b_dim1];
2732 f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2733 * b[l + (j + 2) * b_dim1];
2734 f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2735 * b[l + (j + 2) * b_dim1];
2736 f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
2737 * b[l + (j + 3) * b_dim1];
2738 f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
2739 * b[l + (j + 3) * b_dim1];
2740 }
2741 c[i + j * c_dim1] = f11;
2742 c[i + 1 + j * c_dim1] = f21;
2743 c[i + (j + 1) * c_dim1] = f12;
2744 c[i + 1 + (j + 1) * c_dim1] = f22;
2745 c[i + (j + 2) * c_dim1] = f13;
2746 c[i + 1 + (j + 2) * c_dim1] = f23;
2747 c[i + (j + 3) * c_dim1] = f14;
2748 c[i + 1 + (j + 3) * c_dim1] = f24;
2749 c[i + 2 + j * c_dim1] = f31;
2750 c[i + 3 + j * c_dim1] = f41;
2751 c[i + 2 + (j + 1) * c_dim1] = f32;
2752 c[i + 3 + (j + 1) * c_dim1] = f42;
2753 c[i + 2 + (j + 2) * c_dim1] = f33;
2754 c[i + 3 + (j + 2) * c_dim1] = f43;
2755 c[i + 2 + (j + 3) * c_dim1] = f34;
2756 c[i + 3 + (j + 3) * c_dim1] = f44;
2757 }
2758 if (uisec < isec)
2759 {
2760 i5 = ii + isec - 1;
2761 for (i = ii + uisec; i <= i5; ++i)
2762 {
2763 f11 = c[i + j * c_dim1];
2764 f12 = c[i + (j + 1) * c_dim1];
2765 f13 = c[i + (j + 2) * c_dim1];
2766 f14 = c[i + (j + 3) * c_dim1];
2767 i6 = ll + lsec - 1;
2768 for (l = ll; l <= i6; ++l)
2769 {
2770 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2771 257] * b[l + j * b_dim1];
2772 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2773 257] * b[l + (j + 1) * b_dim1];
2774 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2775 257] * b[l + (j + 2) * b_dim1];
2776 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2777 257] * b[l + (j + 3) * b_dim1];
2778 }
2779 c[i + j * c_dim1] = f11;
2780 c[i + (j + 1) * c_dim1] = f12;
2781 c[i + (j + 2) * c_dim1] = f13;
2782 c[i + (j + 3) * c_dim1] = f14;
2783 }
2784 }
2785 }
2786 if (ujsec < jsec)
2787 {
2788 i4 = jj + jsec - 1;
2789 for (j = jj + ujsec; j <= i4; ++j)
2790 {
2791 i5 = ii + uisec - 1;
2792 for (i = ii; i <= i5; i += 4)
2793 {
2794 f11 = c[i + j * c_dim1];
2795 f21 = c[i + 1 + j * c_dim1];
2796 f31 = c[i + 2 + j * c_dim1];
2797 f41 = c[i + 3 + j * c_dim1];
2798 i6 = ll + lsec - 1;
2799 for (l = ll; l <= i6; ++l)
2800 {
2801 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2802 257] * b[l + j * b_dim1];
2803 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
2804 257] * b[l + j * b_dim1];
2805 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
2806 257] * b[l + j * b_dim1];
2807 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
2808 257] * b[l + j * b_dim1];
2809 }
2810 c[i + j * c_dim1] = f11;
2811 c[i + 1 + j * c_dim1] = f21;
2812 c[i + 2 + j * c_dim1] = f31;
2813 c[i + 3 + j * c_dim1] = f41;
2814 }
2815 i5 = ii + isec - 1;
2816 for (i = ii + uisec; i <= i5; ++i)
2817 {
2818 f11 = c[i + j * c_dim1];
2819 i6 = ll + lsec - 1;
2820 for (l = ll; l <= i6; ++l)
2821 {
2822 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
2823 257] * b[l + j * b_dim1];
2824 }
2825 c[i + j * c_dim1] = f11;
2826 }
2827 }
2828 }
644cb69f
FXC		 2829 }
2830 }
2831 }
8e5f30dc 2832 free(t1);
5d70ab07 2833 return;
644cb69f 2834 }
1524f80b
RS		 2835 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
2836 {
a4a11197
PT		 2837 if (GFC_DESCRIPTOR_RANK (a) != 1)
2838 {
2839 const GFC_COMPLEX_16 *restrict abase_x;
2840 const GFC_COMPLEX_16 *restrict bbase_y;
2841 GFC_COMPLEX_16 *restrict dest_y;
2842 GFC_COMPLEX_16 s;
1524f80b 2843
a4a11197
PT		 2844 for (y = 0; y < ycount; y++)
2845 {
2846 bbase_y = &bbase[y*bystride];
2847 dest_y = &dest[y*rystride];
2848 for (x = 0; x < xcount; x++)
2849 {
2850 abase_x = &abase[x*axstride];
2851 s = (GFC_COMPLEX_16) 0;
2852 for (n = 0; n < count; n++)
2853 s += abase_x[n] * bbase_y[n];
2854 dest_y[x] = s;
2855 }
2856 }
2857 }
2858 else
1524f80b 2859 {
a4a11197
PT		 2860 const GFC_COMPLEX_16 *restrict bbase_y;
2861 GFC_COMPLEX_16 s;
2862
2863 for (y = 0; y < ycount; y++)
1524f80b 2864 {
a4a11197 2865 bbase_y = &bbase[y*bystride];
1524f80b
RS		 2866 s = (GFC_COMPLEX_16) 0;
2867 for (n = 0; n < count; n++)
a4a11197
PT		 2868 s += abase[n*axstride] * bbase_y[n];
2869 dest[y*rystride] = s;
1524f80b
RS		 2870 }
2871 }
2872 }
2873 else if (axstride < aystride)
644cb69f
FXC		 2874 {
2875 for (y = 0; y < ycount; y++)
2876 for (x = 0; x < xcount; x++)
2877 dest[x*rxstride + y*rystride] = (GFC_COMPLEX_16)0;
2878
2879 for (y = 0; y < ycount; y++)
2880 for (n = 0; n < count; n++)
2881 for (x = 0; x < xcount; x++)
2882 /* dest[x,y] += a[x,n] * b[n,y] */
5d70ab07
JD		 2883 dest[x*rxstride + y*rystride] +=
2884 abase[x*axstride + n*aystride] *
2885 bbase[n*bxstride + y*bystride];
644cb69f 2886 }
f0e871d6
PT		 2887 else if (GFC_DESCRIPTOR_RANK (a) == 1)
2888 {
2889 const GFC_COMPLEX_16 *restrict bbase_y;
2890 GFC_COMPLEX_16 s;
2891
2892 for (y = 0; y < ycount; y++)
2893 {
2894 bbase_y = &bbase[y*bystride];
2895 s = (GFC_COMPLEX_16) 0;
2896 for (n = 0; n < count; n++)
2897 s += abase[n*axstride] * bbase_y[n*bxstride];
2898 dest[y*rxstride] = s;
2899 }
2900 }
1524f80b
RS		 2901 else
2902 {
2903 const GFC_COMPLEX_16 *restrict abase_x;
2904 const GFC_COMPLEX_16 *restrict bbase_y;
2905 GFC_COMPLEX_16 *restrict dest_y;
2906 GFC_COMPLEX_16 s;
2907
2908 for (y = 0; y < ycount; y++)
2909 {
2910 bbase_y = &bbase[y*bystride];
2911 dest_y = &dest[y*rystride];
2912 for (x = 0; x < xcount; x++)
2913 {
2914 abase_x = &abase[x*axstride];
2915 s = (GFC_COMPLEX_16) 0;
2916 for (n = 0; n < count; n++)
2917 s += abase_x[n*aystride] * bbase_y[n*bxstride];
2918 dest_y[x*rxstride] = s;
2919 }
2920 }
2921 }
644cb69f 2922}
31cfd832
TK		 2923#undef POW3
2924#undef min
2925#undef max
2926
644cb69f 2927#endif
31cfd832
TK		 2928#endif
2929
Mirror of https://gcc.gnu.org/git/gcc.git