]>
Commit | Line | Data |
---|---|---|
1 | /* Implementation of the MATMUL intrinsic | |
2 | Copyright (C) 2002-2017 Free Software Foundation, Inc. | |
3 | Contributed by Paul Brook <paul@nowt.org> | |
4 | ||
5 | This file is part of the GNU Fortran runtime library (libgfortran). | |
6 | ||
7 | Libgfortran is free software; you can redistribute it and/or | |
8 | modify it under the terms of the GNU General Public | |
9 | License as published by the Free Software Foundation; either | |
10 | version 3 of the License, or (at your option) any later version. | |
11 | ||
12 | Libgfortran is distributed in the hope that it will be useful, | |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | GNU General Public License for more details. | |
16 | ||
17 | Under Section 7 of GPL version 3, you are granted additional | |
18 | permissions described in the GCC Runtime Library Exception, version | |
19 | 3.1, as published by the Free Software Foundation. | |
20 | ||
21 | You should have received a copy of the GNU General Public License and | |
22 | a copy of the GCC Runtime Library Exception along with this program; | |
23 | see the files COPYING3 and COPYING.RUNTIME respectively. If not, see | |
24 | <http://www.gnu.org/licenses/>. */ | |
25 | ||
26 | #include "libgfortran.h" | |
27 | #include <string.h> | |
28 | #include <assert.h> | |
29 | ||
30 | ||
31 | #if defined (HAVE_GFC_REAL_16) | |
32 | ||
33 | /* Prototype for the BLAS ?gemm subroutine, a pointer to which can be | |
34 | passed to us by the front-end, in which case we call it for large | |
35 | matrices. */ | |
36 | ||
37 | typedef void (*blas_call)(const char *, const char *, const int *, const int *, | |
38 | const int *, const GFC_REAL_16 *, const GFC_REAL_16 *, | |
39 | const int *, const GFC_REAL_16 *, const int *, | |
40 | const GFC_REAL_16 *, GFC_REAL_16 *, const int *, | |
41 | int, int); | |
42 | ||
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: | |
50 | ||
51 | DIMENSION A(M,COUNT), B(COUNT,N), C(M,N) | |
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 | |
60 | DO I=1,M | |
61 | S = 0 | |
62 | DO K=1,COUNT | |
63 | S = S+A(I,K)*B(K,J) | |
64 | C(I,J) = S | |
65 | ENDIF | |
66 | */ | |
67 | ||
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 | ||
72 | extern void matmul_r16 (gfc_array_r16 * const restrict retarray, | |
73 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
74 | int blas_limit, blas_call gemm); | |
75 | export_proto(matmul_r16); | |
76 | ||
77 | #if defined(HAVE_AVX) && defined(HAVE_AVX2) | |
78 | /* REAL types generate identical code for AVX and AVX2. Only generate | |
79 | an AVX2 function if we are dealing with integer. */ | |
80 | #undef HAVE_AVX2 | |
81 | #endif | |
82 | ||
83 | ||
84 | /* Put exhaustive list of possible architectures here here, ORed together. */ | |
85 | ||
86 | #if defined(HAVE_AVX) || defined(HAVE_AVX2) || defined(HAVE_AVX512F) | |
87 | ||
88 | #ifdef HAVE_AVX | |
89 | static void | |
90 | matmul_r16_avx (gfc_array_r16 * const restrict retarray, | |
91 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
92 | int blas_limit, blas_call gemm) __attribute__((__target__("avx"))); | |
93 | static void | |
94 | matmul_r16_avx (gfc_array_r16 * const restrict retarray, | |
95 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
96 | int blas_limit, blas_call gemm) | |
97 | { | |
98 | const GFC_REAL_16 * restrict abase; | |
99 | const GFC_REAL_16 * restrict bbase; | |
100 | GFC_REAL_16 * restrict dest; | |
101 | ||
102 | index_type rxstride, rystride, axstride, aystride, bxstride, bystride; | |
103 | index_type x, y, n, count, xcount, ycount; | |
104 | ||
105 | assert (GFC_DESCRIPTOR_RANK (a) == 2 | |
106 | || GFC_DESCRIPTOR_RANK (b) == 2); | |
107 | ||
108 | /* C[xcount,ycount] = A[xcount, count] * B[count,ycount] | |
109 | ||
110 | Either A or B (but not both) can be rank 1: | |
111 | ||
112 | o One-dimensional argument A is implicitly treated as a row matrix | |
113 | dimensioned [1,count], so xcount=1. | |
114 | ||
115 | o One-dimensional argument B is implicitly treated as a column matrix | |
116 | dimensioned [count, 1], so ycount=1. | |
117 | */ | |
118 | ||
119 | if (retarray->base_addr == NULL) | |
120 | { | |
121 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
122 | { | |
123 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
124 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1); | |
125 | } | |
126 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
127 | { | |
128 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
129 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
130 | } | |
131 | else | |
132 | { | |
133 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
134 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
135 | ||
136 | GFC_DIMENSION_SET(retarray->dim[1], 0, | |
137 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, | |
138 | GFC_DESCRIPTOR_EXTENT(retarray,0)); | |
139 | } | |
140 | ||
141 | retarray->base_addr | |
142 | = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_REAL_16)); | |
143 | retarray->offset = 0; | |
144 | } | |
145 | else if (unlikely (compile_options.bounds_check)) | |
146 | { | |
147 | index_type ret_extent, arg_extent; | |
148 | ||
149 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
150 | { | |
151 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
152 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
153 | if (arg_extent != ret_extent) | |
154 | runtime_error ("Incorrect extent in return array in" | |
155 | " MATMUL intrinsic: is %ld, should be %ld", | |
156 | (long int) ret_extent, (long int) arg_extent); | |
157 | } | |
158 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
159 | { | |
160 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
161 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
162 | if (arg_extent != ret_extent) | |
163 | runtime_error ("Incorrect extent in return array in" | |
164 | " MATMUL intrinsic: is %ld, should be %ld", | |
165 | (long int) ret_extent, (long int) arg_extent); | |
166 | } | |
167 | else | |
168 | { | |
169 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
170 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
171 | if (arg_extent != ret_extent) | |
172 | runtime_error ("Incorrect extent in return array in" | |
173 | " MATMUL intrinsic for dimension 1:" | |
174 | " is %ld, should be %ld", | |
175 | (long int) ret_extent, (long int) arg_extent); | |
176 | ||
177 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
178 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1); | |
179 | if (arg_extent != ret_extent) | |
180 | runtime_error ("Incorrect extent in return array in" | |
181 | " MATMUL intrinsic for dimension 2:" | |
182 | " is %ld, should be %ld", | |
183 | (long int) ret_extent, (long int) arg_extent); | |
184 | } | |
185 | } | |
186 | ||
187 | ||
188 | if (GFC_DESCRIPTOR_RANK (retarray) == 1) | |
189 | { | |
190 | /* One-dimensional result may be addressed in the code below | |
191 | either as a row or a column matrix. We want both cases to | |
192 | work. */ | |
193 | rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
194 | } | |
195 | else | |
196 | { | |
197 | rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
198 | rystride = GFC_DESCRIPTOR_STRIDE(retarray,1); | |
199 | } | |
200 | ||
201 | ||
202 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
203 | { | |
204 | /* Treat it as a a row matrix A[1,count]. */ | |
205 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
206 | aystride = 1; | |
207 | ||
208 | xcount = 1; | |
209 | count = GFC_DESCRIPTOR_EXTENT(a,0); | |
210 | } | |
211 | else | |
212 | { | |
213 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
214 | aystride = GFC_DESCRIPTOR_STRIDE(a,1); | |
215 | ||
216 | count = GFC_DESCRIPTOR_EXTENT(a,1); | |
217 | xcount = GFC_DESCRIPTOR_EXTENT(a,0); | |
218 | } | |
219 | ||
220 | if (count != GFC_DESCRIPTOR_EXTENT(b,0)) | |
221 | { | |
222 | if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0) | |
223 | runtime_error ("dimension of array B incorrect in MATMUL intrinsic"); | |
224 | } | |
225 | ||
226 | if (GFC_DESCRIPTOR_RANK (b) == 1) | |
227 | { | |
228 | /* Treat it as a column matrix B[count,1] */ | |
229 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
230 | ||
231 | /* bystride should never be used for 1-dimensional b. | |
232 | in case it is we want it to cause a segfault, rather than | |
233 | an incorrect result. */ | |
234 | bystride = 0xDEADBEEF; | |
235 | ycount = 1; | |
236 | } | |
237 | else | |
238 | { | |
239 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
240 | bystride = GFC_DESCRIPTOR_STRIDE(b,1); | |
241 | ycount = GFC_DESCRIPTOR_EXTENT(b,1); | |
242 | } | |
243 | ||
244 | abase = a->base_addr; | |
245 | bbase = b->base_addr; | |
246 | dest = retarray->base_addr; | |
247 | ||
248 | /* Now that everything is set up, we perform the multiplication | |
249 | itself. */ | |
250 | ||
251 | #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x))) | |
252 | #define min(a,b) ((a) <= (b) ? (a) : (b)) | |
253 | #define max(a,b) ((a) >= (b) ? (a) : (b)) | |
254 | ||
255 | if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1) | |
256 | && (bxstride == 1 || bystride == 1) | |
257 | && (((float) xcount) * ((float) ycount) * ((float) count) | |
258 | > POW3(blas_limit))) | |
259 | { | |
260 | const int m = xcount, n = ycount, k = count, ldc = rystride; | |
261 | const GFC_REAL_16 one = 1, zero = 0; | |
262 | const int lda = (axstride == 1) ? aystride : axstride, | |
263 | ldb = (bxstride == 1) ? bystride : bxstride; | |
264 | ||
265 | if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1) | |
266 | { | |
267 | assert (gemm != NULL); | |
268 | gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, | |
269 | &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest, | |
270 | &ldc, 1, 1); | |
271 | return; | |
272 | } | |
273 | } | |
274 | ||
275 | if (rxstride == 1 && axstride == 1 && bxstride == 1) | |
276 | { | |
277 | /* This block of code implements a tuned matmul, derived from | |
278 | Superscalar GEMM-based level 3 BLAS, Beta version 0.1 | |
279 | ||
280 | Bo Kagstrom and Per Ling | |
281 | Department of Computing Science | |
282 | Umea University | |
283 | S-901 87 Umea, Sweden | |
284 | ||
285 | from netlib.org, translated to C, and modified for matmul.m4. */ | |
286 | ||
287 | const GFC_REAL_16 *a, *b; | |
288 | GFC_REAL_16 *c; | |
289 | const index_type m = xcount, n = ycount, k = count; | |
290 | ||
291 | /* System generated locals */ | |
292 | index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset, | |
293 | i1, i2, i3, i4, i5, i6; | |
294 | ||
295 | /* Local variables */ | |
296 | GFC_REAL_16 t1[65536], /* was [256][256] */ | |
297 | f11, f12, f21, f22, f31, f32, f41, f42, | |
298 | f13, f14, f23, f24, f33, f34, f43, f44; | |
299 | index_type i, j, l, ii, jj, ll; | |
300 | index_type isec, jsec, lsec, uisec, ujsec, ulsec; | |
301 | ||
302 | a = abase; | |
303 | b = bbase; | |
304 | c = retarray->base_addr; | |
305 | ||
306 | /* Parameter adjustments */ | |
307 | c_dim1 = rystride; | |
308 | c_offset = 1 + c_dim1; | |
309 | c -= c_offset; | |
310 | a_dim1 = aystride; | |
311 | a_offset = 1 + a_dim1; | |
312 | a -= a_offset; | |
313 | b_dim1 = bystride; | |
314 | b_offset = 1 + b_dim1; | |
315 | b -= b_offset; | |
316 | ||
317 | /* Early exit if possible */ | |
318 | if (m == 0 || n == 0 || k == 0) | |
319 | return; | |
320 | ||
321 | /* Empty c first. */ | |
322 | for (j=1; j<=n; j++) | |
323 | for (i=1; i<=m; i++) | |
324 | c[i + j * c_dim1] = (GFC_REAL_16)0; | |
325 | ||
326 | /* Start turning the crank. */ | |
327 | i1 = n; | |
328 | for (jj = 1; jj <= i1; jj += 512) | |
329 | { | |
330 | /* Computing MIN */ | |
331 | i2 = 512; | |
332 | i3 = n - jj + 1; | |
333 | jsec = min(i2,i3); | |
334 | ujsec = jsec - jsec % 4; | |
335 | i2 = k; | |
336 | for (ll = 1; ll <= i2; ll += 256) | |
337 | { | |
338 | /* Computing MIN */ | |
339 | i3 = 256; | |
340 | i4 = k - ll + 1; | |
341 | lsec = min(i3,i4); | |
342 | ulsec = lsec - lsec % 2; | |
343 | ||
344 | i3 = m; | |
345 | for (ii = 1; ii <= i3; ii += 256) | |
346 | { | |
347 | /* Computing MIN */ | |
348 | i4 = 256; | |
349 | i5 = m - ii + 1; | |
350 | isec = min(i4,i5); | |
351 | uisec = isec - isec % 2; | |
352 | i4 = ll + ulsec - 1; | |
353 | for (l = ll; l <= i4; l += 2) | |
354 | { | |
355 | i5 = ii + uisec - 1; | |
356 | for (i = ii; i <= i5; i += 2) | |
357 | { | |
358 | t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] = | |
359 | a[i + l * a_dim1]; | |
360 | t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] = | |
361 | a[i + (l + 1) * a_dim1]; | |
362 | t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] = | |
363 | a[i + 1 + l * a_dim1]; | |
364 | t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] = | |
365 | a[i + 1 + (l + 1) * a_dim1]; | |
366 | } | |
367 | if (uisec < isec) | |
368 | { | |
369 | t1[l - ll + 1 + (isec << 8) - 257] = | |
370 | a[ii + isec - 1 + l * a_dim1]; | |
371 | t1[l - ll + 2 + (isec << 8) - 257] = | |
372 | a[ii + isec - 1 + (l + 1) * a_dim1]; | |
373 | } | |
374 | } | |
375 | if (ulsec < lsec) | |
376 | { | |
377 | i4 = ii + isec - 1; | |
378 | for (i = ii; i<= i4; ++i) | |
379 | { | |
380 | t1[lsec + ((i - ii + 1) << 8) - 257] = | |
381 | a[i + (ll + lsec - 1) * a_dim1]; | |
382 | } | |
383 | } | |
384 | ||
385 | uisec = isec - isec % 4; | |
386 | i4 = jj + ujsec - 1; | |
387 | for (j = jj; j <= i4; j += 4) | |
388 | { | |
389 | i5 = ii + uisec - 1; | |
390 | for (i = ii; i <= i5; i += 4) | |
391 | { | |
392 | f11 = c[i + j * c_dim1]; | |
393 | f21 = c[i + 1 + j * c_dim1]; | |
394 | f12 = c[i + (j + 1) * c_dim1]; | |
395 | f22 = c[i + 1 + (j + 1) * c_dim1]; | |
396 | f13 = c[i + (j + 2) * c_dim1]; | |
397 | f23 = c[i + 1 + (j + 2) * c_dim1]; | |
398 | f14 = c[i + (j + 3) * c_dim1]; | |
399 | f24 = c[i + 1 + (j + 3) * c_dim1]; | |
400 | f31 = c[i + 2 + j * c_dim1]; | |
401 | f41 = c[i + 3 + j * c_dim1]; | |
402 | f32 = c[i + 2 + (j + 1) * c_dim1]; | |
403 | f42 = c[i + 3 + (j + 1) * c_dim1]; | |
404 | f33 = c[i + 2 + (j + 2) * c_dim1]; | |
405 | f43 = c[i + 3 + (j + 2) * c_dim1]; | |
406 | f34 = c[i + 2 + (j + 3) * c_dim1]; | |
407 | f44 = c[i + 3 + (j + 3) * c_dim1]; | |
408 | i6 = ll + lsec - 1; | |
409 | for (l = ll; l <= i6; ++l) | |
410 | { | |
411 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
412 | * b[l + j * b_dim1]; | |
413 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
414 | * b[l + j * b_dim1]; | |
415 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
416 | * b[l + (j + 1) * b_dim1]; | |
417 | f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
418 | * b[l + (j + 1) * b_dim1]; | |
419 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
420 | * b[l + (j + 2) * b_dim1]; | |
421 | f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
422 | * b[l + (j + 2) * b_dim1]; | |
423 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
424 | * b[l + (j + 3) * b_dim1]; | |
425 | f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
426 | * b[l + (j + 3) * b_dim1]; | |
427 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
428 | * b[l + j * b_dim1]; | |
429 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
430 | * b[l + j * b_dim1]; | |
431 | f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
432 | * b[l + (j + 1) * b_dim1]; | |
433 | f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
434 | * b[l + (j + 1) * b_dim1]; | |
435 | f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
436 | * b[l + (j + 2) * b_dim1]; | |
437 | f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
438 | * b[l + (j + 2) * b_dim1]; | |
439 | f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
440 | * b[l + (j + 3) * b_dim1]; | |
441 | f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
442 | * b[l + (j + 3) * b_dim1]; | |
443 | } | |
444 | c[i + j * c_dim1] = f11; | |
445 | c[i + 1 + j * c_dim1] = f21; | |
446 | c[i + (j + 1) * c_dim1] = f12; | |
447 | c[i + 1 + (j + 1) * c_dim1] = f22; | |
448 | c[i + (j + 2) * c_dim1] = f13; | |
449 | c[i + 1 + (j + 2) * c_dim1] = f23; | |
450 | c[i + (j + 3) * c_dim1] = f14; | |
451 | c[i + 1 + (j + 3) * c_dim1] = f24; | |
452 | c[i + 2 + j * c_dim1] = f31; | |
453 | c[i + 3 + j * c_dim1] = f41; | |
454 | c[i + 2 + (j + 1) * c_dim1] = f32; | |
455 | c[i + 3 + (j + 1) * c_dim1] = f42; | |
456 | c[i + 2 + (j + 2) * c_dim1] = f33; | |
457 | c[i + 3 + (j + 2) * c_dim1] = f43; | |
458 | c[i + 2 + (j + 3) * c_dim1] = f34; | |
459 | c[i + 3 + (j + 3) * c_dim1] = f44; | |
460 | } | |
461 | if (uisec < isec) | |
462 | { | |
463 | i5 = ii + isec - 1; | |
464 | for (i = ii + uisec; i <= i5; ++i) | |
465 | { | |
466 | f11 = c[i + j * c_dim1]; | |
467 | f12 = c[i + (j + 1) * c_dim1]; | |
468 | f13 = c[i + (j + 2) * c_dim1]; | |
469 | f14 = c[i + (j + 3) * c_dim1]; | |
470 | i6 = ll + lsec - 1; | |
471 | for (l = ll; l <= i6; ++l) | |
472 | { | |
473 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
474 | 257] * b[l + j * b_dim1]; | |
475 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
476 | 257] * b[l + (j + 1) * b_dim1]; | |
477 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
478 | 257] * b[l + (j + 2) * b_dim1]; | |
479 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
480 | 257] * b[l + (j + 3) * b_dim1]; | |
481 | } | |
482 | c[i + j * c_dim1] = f11; | |
483 | c[i + (j + 1) * c_dim1] = f12; | |
484 | c[i + (j + 2) * c_dim1] = f13; | |
485 | c[i + (j + 3) * c_dim1] = f14; | |
486 | } | |
487 | } | |
488 | } | |
489 | if (ujsec < jsec) | |
490 | { | |
491 | i4 = jj + jsec - 1; | |
492 | for (j = jj + ujsec; j <= i4; ++j) | |
493 | { | |
494 | i5 = ii + uisec - 1; | |
495 | for (i = ii; i <= i5; i += 4) | |
496 | { | |
497 | f11 = c[i + j * c_dim1]; | |
498 | f21 = c[i + 1 + j * c_dim1]; | |
499 | f31 = c[i + 2 + j * c_dim1]; | |
500 | f41 = c[i + 3 + j * c_dim1]; | |
501 | i6 = ll + lsec - 1; | |
502 | for (l = ll; l <= i6; ++l) | |
503 | { | |
504 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
505 | 257] * b[l + j * b_dim1]; | |
506 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - | |
507 | 257] * b[l + j * b_dim1]; | |
508 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - | |
509 | 257] * b[l + j * b_dim1]; | |
510 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - | |
511 | 257] * b[l + j * b_dim1]; | |
512 | } | |
513 | c[i + j * c_dim1] = f11; | |
514 | c[i + 1 + j * c_dim1] = f21; | |
515 | c[i + 2 + j * c_dim1] = f31; | |
516 | c[i + 3 + j * c_dim1] = f41; | |
517 | } | |
518 | i5 = ii + isec - 1; | |
519 | for (i = ii + uisec; i <= i5; ++i) | |
520 | { | |
521 | f11 = c[i + j * c_dim1]; | |
522 | i6 = ll + lsec - 1; | |
523 | for (l = ll; l <= i6; ++l) | |
524 | { | |
525 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
526 | 257] * b[l + j * b_dim1]; | |
527 | } | |
528 | c[i + j * c_dim1] = f11; | |
529 | } | |
530 | } | |
531 | } | |
532 | } | |
533 | } | |
534 | } | |
535 | return; | |
536 | } | |
537 | else if (rxstride == 1 && aystride == 1 && bxstride == 1) | |
538 | { | |
539 | if (GFC_DESCRIPTOR_RANK (a) != 1) | |
540 | { | |
541 | const GFC_REAL_16 *restrict abase_x; | |
542 | const GFC_REAL_16 *restrict bbase_y; | |
543 | GFC_REAL_16 *restrict dest_y; | |
544 | GFC_REAL_16 s; | |
545 | ||
546 | for (y = 0; y < ycount; y++) | |
547 | { | |
548 | bbase_y = &bbase[y*bystride]; | |
549 | dest_y = &dest[y*rystride]; | |
550 | for (x = 0; x < xcount; x++) | |
551 | { | |
552 | abase_x = &abase[x*axstride]; | |
553 | s = (GFC_REAL_16) 0; | |
554 | for (n = 0; n < count; n++) | |
555 | s += abase_x[n] * bbase_y[n]; | |
556 | dest_y[x] = s; | |
557 | } | |
558 | } | |
559 | } | |
560 | else | |
561 | { | |
562 | const GFC_REAL_16 *restrict bbase_y; | |
563 | GFC_REAL_16 s; | |
564 | ||
565 | for (y = 0; y < ycount; y++) | |
566 | { | |
567 | bbase_y = &bbase[y*bystride]; | |
568 | s = (GFC_REAL_16) 0; | |
569 | for (n = 0; n < count; n++) | |
570 | s += abase[n*axstride] * bbase_y[n]; | |
571 | dest[y*rystride] = s; | |
572 | } | |
573 | } | |
574 | } | |
575 | else if (axstride < aystride) | |
576 | { | |
577 | for (y = 0; y < ycount; y++) | |
578 | for (x = 0; x < xcount; x++) | |
579 | dest[x*rxstride + y*rystride] = (GFC_REAL_16)0; | |
580 | ||
581 | for (y = 0; y < ycount; y++) | |
582 | for (n = 0; n < count; n++) | |
583 | for (x = 0; x < xcount; x++) | |
584 | /* dest[x,y] += a[x,n] * b[n,y] */ | |
585 | dest[x*rxstride + y*rystride] += | |
586 | abase[x*axstride + n*aystride] * | |
587 | bbase[n*bxstride + y*bystride]; | |
588 | } | |
589 | else if (GFC_DESCRIPTOR_RANK (a) == 1) | |
590 | { | |
591 | const GFC_REAL_16 *restrict bbase_y; | |
592 | GFC_REAL_16 s; | |
593 | ||
594 | for (y = 0; y < ycount; y++) | |
595 | { | |
596 | bbase_y = &bbase[y*bystride]; | |
597 | s = (GFC_REAL_16) 0; | |
598 | for (n = 0; n < count; n++) | |
599 | s += abase[n*axstride] * bbase_y[n*bxstride]; | |
600 | dest[y*rxstride] = s; | |
601 | } | |
602 | } | |
603 | else | |
604 | { | |
605 | const GFC_REAL_16 *restrict abase_x; | |
606 | const GFC_REAL_16 *restrict bbase_y; | |
607 | GFC_REAL_16 *restrict dest_y; | |
608 | GFC_REAL_16 s; | |
609 | ||
610 | for (y = 0; y < ycount; y++) | |
611 | { | |
612 | bbase_y = &bbase[y*bystride]; | |
613 | dest_y = &dest[y*rystride]; | |
614 | for (x = 0; x < xcount; x++) | |
615 | { | |
616 | abase_x = &abase[x*axstride]; | |
617 | s = (GFC_REAL_16) 0; | |
618 | for (n = 0; n < count; n++) | |
619 | s += abase_x[n*aystride] * bbase_y[n*bxstride]; | |
620 | dest_y[x*rxstride] = s; | |
621 | } | |
622 | } | |
623 | } | |
624 | } | |
625 | #undef POW3 | |
626 | #undef min | |
627 | #undef max | |
628 | ||
629 | #endif /* HAVE_AVX */ | |
630 | ||
631 | #ifdef HAVE_AVX2 | |
632 | static void | |
633 | matmul_r16_avx2 (gfc_array_r16 * const restrict retarray, | |
634 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
635 | int blas_limit, blas_call gemm) __attribute__((__target__("avx2"))); | |
636 | static void | |
637 | matmul_r16_avx2 (gfc_array_r16 * const restrict retarray, | |
638 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
639 | int blas_limit, blas_call gemm) | |
640 | { | |
641 | const GFC_REAL_16 * restrict abase; | |
642 | const GFC_REAL_16 * restrict bbase; | |
643 | GFC_REAL_16 * restrict dest; | |
644 | ||
645 | index_type rxstride, rystride, axstride, aystride, bxstride, bystride; | |
646 | index_type x, y, n, count, xcount, ycount; | |
647 | ||
648 | assert (GFC_DESCRIPTOR_RANK (a) == 2 | |
649 | || GFC_DESCRIPTOR_RANK (b) == 2); | |
650 | ||
651 | /* C[xcount,ycount] = A[xcount, count] * B[count,ycount] | |
652 | ||
653 | Either A or B (but not both) can be rank 1: | |
654 | ||
655 | o One-dimensional argument A is implicitly treated as a row matrix | |
656 | dimensioned [1,count], so xcount=1. | |
657 | ||
658 | o One-dimensional argument B is implicitly treated as a column matrix | |
659 | dimensioned [count, 1], so ycount=1. | |
660 | */ | |
661 | ||
662 | if (retarray->base_addr == NULL) | |
663 | { | |
664 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
665 | { | |
666 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
667 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1); | |
668 | } | |
669 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
670 | { | |
671 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
672 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
673 | } | |
674 | else | |
675 | { | |
676 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
677 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
678 | ||
679 | GFC_DIMENSION_SET(retarray->dim[1], 0, | |
680 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, | |
681 | GFC_DESCRIPTOR_EXTENT(retarray,0)); | |
682 | } | |
683 | ||
684 | retarray->base_addr | |
685 | = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_REAL_16)); | |
686 | retarray->offset = 0; | |
687 | } | |
688 | else if (unlikely (compile_options.bounds_check)) | |
689 | { | |
690 | index_type ret_extent, arg_extent; | |
691 | ||
692 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
693 | { | |
694 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
695 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
696 | if (arg_extent != ret_extent) | |
697 | runtime_error ("Incorrect extent in return array in" | |
698 | " MATMUL intrinsic: is %ld, should be %ld", | |
699 | (long int) ret_extent, (long int) arg_extent); | |
700 | } | |
701 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
702 | { | |
703 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
704 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
705 | if (arg_extent != ret_extent) | |
706 | runtime_error ("Incorrect extent in return array in" | |
707 | " MATMUL intrinsic: is %ld, should be %ld", | |
708 | (long int) ret_extent, (long int) arg_extent); | |
709 | } | |
710 | else | |
711 | { | |
712 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
713 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
714 | if (arg_extent != ret_extent) | |
715 | runtime_error ("Incorrect extent in return array in" | |
716 | " MATMUL intrinsic for dimension 1:" | |
717 | " is %ld, should be %ld", | |
718 | (long int) ret_extent, (long int) arg_extent); | |
719 | ||
720 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
721 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1); | |
722 | if (arg_extent != ret_extent) | |
723 | runtime_error ("Incorrect extent in return array in" | |
724 | " MATMUL intrinsic for dimension 2:" | |
725 | " is %ld, should be %ld", | |
726 | (long int) ret_extent, (long int) arg_extent); | |
727 | } | |
728 | } | |
729 | ||
730 | ||
731 | if (GFC_DESCRIPTOR_RANK (retarray) == 1) | |
732 | { | |
733 | /* One-dimensional result may be addressed in the code below | |
734 | either as a row or a column matrix. We want both cases to | |
735 | work. */ | |
736 | rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
737 | } | |
738 | else | |
739 | { | |
740 | rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
741 | rystride = GFC_DESCRIPTOR_STRIDE(retarray,1); | |
742 | } | |
743 | ||
744 | ||
745 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
746 | { | |
747 | /* Treat it as a a row matrix A[1,count]. */ | |
748 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
749 | aystride = 1; | |
750 | ||
751 | xcount = 1; | |
752 | count = GFC_DESCRIPTOR_EXTENT(a,0); | |
753 | } | |
754 | else | |
755 | { | |
756 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
757 | aystride = GFC_DESCRIPTOR_STRIDE(a,1); | |
758 | ||
759 | count = GFC_DESCRIPTOR_EXTENT(a,1); | |
760 | xcount = GFC_DESCRIPTOR_EXTENT(a,0); | |
761 | } | |
762 | ||
763 | if (count != GFC_DESCRIPTOR_EXTENT(b,0)) | |
764 | { | |
765 | if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0) | |
766 | runtime_error ("dimension of array B incorrect in MATMUL intrinsic"); | |
767 | } | |
768 | ||
769 | if (GFC_DESCRIPTOR_RANK (b) == 1) | |
770 | { | |
771 | /* Treat it as a column matrix B[count,1] */ | |
772 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
773 | ||
774 | /* bystride should never be used for 1-dimensional b. | |
775 | in case it is we want it to cause a segfault, rather than | |
776 | an incorrect result. */ | |
777 | bystride = 0xDEADBEEF; | |
778 | ycount = 1; | |
779 | } | |
780 | else | |
781 | { | |
782 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
783 | bystride = GFC_DESCRIPTOR_STRIDE(b,1); | |
784 | ycount = GFC_DESCRIPTOR_EXTENT(b,1); | |
785 | } | |
786 | ||
787 | abase = a->base_addr; | |
788 | bbase = b->base_addr; | |
789 | dest = retarray->base_addr; | |
790 | ||
791 | /* Now that everything is set up, we perform the multiplication | |
792 | itself. */ | |
793 | ||
794 | #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x))) | |
795 | #define min(a,b) ((a) <= (b) ? (a) : (b)) | |
796 | #define max(a,b) ((a) >= (b) ? (a) : (b)) | |
797 | ||
798 | if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1) | |
799 | && (bxstride == 1 || bystride == 1) | |
800 | && (((float) xcount) * ((float) ycount) * ((float) count) | |
801 | > POW3(blas_limit))) | |
802 | { | |
803 | const int m = xcount, n = ycount, k = count, ldc = rystride; | |
804 | const GFC_REAL_16 one = 1, zero = 0; | |
805 | const int lda = (axstride == 1) ? aystride : axstride, | |
806 | ldb = (bxstride == 1) ? bystride : bxstride; | |
807 | ||
808 | if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1) | |
809 | { | |
810 | assert (gemm != NULL); | |
811 | gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, | |
812 | &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest, | |
813 | &ldc, 1, 1); | |
814 | return; | |
815 | } | |
816 | } | |
817 | ||
818 | if (rxstride == 1 && axstride == 1 && bxstride == 1) | |
819 | { | |
820 | /* This block of code implements a tuned matmul, derived from | |
821 | Superscalar GEMM-based level 3 BLAS, Beta version 0.1 | |
822 | ||
823 | Bo Kagstrom and Per Ling | |
824 | Department of Computing Science | |
825 | Umea University | |
826 | S-901 87 Umea, Sweden | |
827 | ||
828 | from netlib.org, translated to C, and modified for matmul.m4. */ | |
829 | ||
830 | const GFC_REAL_16 *a, *b; | |
831 | GFC_REAL_16 *c; | |
832 | const index_type m = xcount, n = ycount, k = count; | |
833 | ||
834 | /* System generated locals */ | |
835 | index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset, | |
836 | i1, i2, i3, i4, i5, i6; | |
837 | ||
838 | /* Local variables */ | |
839 | GFC_REAL_16 t1[65536], /* was [256][256] */ | |
840 | f11, f12, f21, f22, f31, f32, f41, f42, | |
841 | f13, f14, f23, f24, f33, f34, f43, f44; | |
842 | index_type i, j, l, ii, jj, ll; | |
843 | index_type isec, jsec, lsec, uisec, ujsec, ulsec; | |
844 | ||
845 | a = abase; | |
846 | b = bbase; | |
847 | c = retarray->base_addr; | |
848 | ||
849 | /* Parameter adjustments */ | |
850 | c_dim1 = rystride; | |
851 | c_offset = 1 + c_dim1; | |
852 | c -= c_offset; | |
853 | a_dim1 = aystride; | |
854 | a_offset = 1 + a_dim1; | |
855 | a -= a_offset; | |
856 | b_dim1 = bystride; | |
857 | b_offset = 1 + b_dim1; | |
858 | b -= b_offset; | |
859 | ||
860 | /* Early exit if possible */ | |
861 | if (m == 0 || n == 0 || k == 0) | |
862 | return; | |
863 | ||
864 | /* Empty c first. */ | |
865 | for (j=1; j<=n; j++) | |
866 | for (i=1; i<=m; i++) | |
867 | c[i + j * c_dim1] = (GFC_REAL_16)0; | |
868 | ||
869 | /* Start turning the crank. */ | |
870 | i1 = n; | |
871 | for (jj = 1; jj <= i1; jj += 512) | |
872 | { | |
873 | /* Computing MIN */ | |
874 | i2 = 512; | |
875 | i3 = n - jj + 1; | |
876 | jsec = min(i2,i3); | |
877 | ujsec = jsec - jsec % 4; | |
878 | i2 = k; | |
879 | for (ll = 1; ll <= i2; ll += 256) | |
880 | { | |
881 | /* Computing MIN */ | |
882 | i3 = 256; | |
883 | i4 = k - ll + 1; | |
884 | lsec = min(i3,i4); | |
885 | ulsec = lsec - lsec % 2; | |
886 | ||
887 | i3 = m; | |
888 | for (ii = 1; ii <= i3; ii += 256) | |
889 | { | |
890 | /* Computing MIN */ | |
891 | i4 = 256; | |
892 | i5 = m - ii + 1; | |
893 | isec = min(i4,i5); | |
894 | uisec = isec - isec % 2; | |
895 | i4 = ll + ulsec - 1; | |
896 | for (l = ll; l <= i4; l += 2) | |
897 | { | |
898 | i5 = ii + uisec - 1; | |
899 | for (i = ii; i <= i5; i += 2) | |
900 | { | |
901 | t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] = | |
902 | a[i + l * a_dim1]; | |
903 | t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] = | |
904 | a[i + (l + 1) * a_dim1]; | |
905 | t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] = | |
906 | a[i + 1 + l * a_dim1]; | |
907 | t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] = | |
908 | a[i + 1 + (l + 1) * a_dim1]; | |
909 | } | |
910 | if (uisec < isec) | |
911 | { | |
912 | t1[l - ll + 1 + (isec << 8) - 257] = | |
913 | a[ii + isec - 1 + l * a_dim1]; | |
914 | t1[l - ll + 2 + (isec << 8) - 257] = | |
915 | a[ii + isec - 1 + (l + 1) * a_dim1]; | |
916 | } | |
917 | } | |
918 | if (ulsec < lsec) | |
919 | { | |
920 | i4 = ii + isec - 1; | |
921 | for (i = ii; i<= i4; ++i) | |
922 | { | |
923 | t1[lsec + ((i - ii + 1) << 8) - 257] = | |
924 | a[i + (ll + lsec - 1) * a_dim1]; | |
925 | } | |
926 | } | |
927 | ||
928 | uisec = isec - isec % 4; | |
929 | i4 = jj + ujsec - 1; | |
930 | for (j = jj; j <= i4; j += 4) | |
931 | { | |
932 | i5 = ii + uisec - 1; | |
933 | for (i = ii; i <= i5; i += 4) | |
934 | { | |
935 | f11 = c[i + j * c_dim1]; | |
936 | f21 = c[i + 1 + j * c_dim1]; | |
937 | f12 = c[i + (j + 1) * c_dim1]; | |
938 | f22 = c[i + 1 + (j + 1) * c_dim1]; | |
939 | f13 = c[i + (j + 2) * c_dim1]; | |
940 | f23 = c[i + 1 + (j + 2) * c_dim1]; | |
941 | f14 = c[i + (j + 3) * c_dim1]; | |
942 | f24 = c[i + 1 + (j + 3) * c_dim1]; | |
943 | f31 = c[i + 2 + j * c_dim1]; | |
944 | f41 = c[i + 3 + j * c_dim1]; | |
945 | f32 = c[i + 2 + (j + 1) * c_dim1]; | |
946 | f42 = c[i + 3 + (j + 1) * c_dim1]; | |
947 | f33 = c[i + 2 + (j + 2) * c_dim1]; | |
948 | f43 = c[i + 3 + (j + 2) * c_dim1]; | |
949 | f34 = c[i + 2 + (j + 3) * c_dim1]; | |
950 | f44 = c[i + 3 + (j + 3) * c_dim1]; | |
951 | i6 = ll + lsec - 1; | |
952 | for (l = ll; l <= i6; ++l) | |
953 | { | |
954 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
955 | * b[l + j * b_dim1]; | |
956 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
957 | * b[l + j * b_dim1]; | |
958 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
959 | * b[l + (j + 1) * b_dim1]; | |
960 | f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
961 | * b[l + (j + 1) * b_dim1]; | |
962 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
963 | * b[l + (j + 2) * b_dim1]; | |
964 | f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
965 | * b[l + (j + 2) * b_dim1]; | |
966 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
967 | * b[l + (j + 3) * b_dim1]; | |
968 | f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
969 | * b[l + (j + 3) * b_dim1]; | |
970 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
971 | * b[l + j * b_dim1]; | |
972 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
973 | * b[l + j * b_dim1]; | |
974 | f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
975 | * b[l + (j + 1) * b_dim1]; | |
976 | f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
977 | * b[l + (j + 1) * b_dim1]; | |
978 | f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
979 | * b[l + (j + 2) * b_dim1]; | |
980 | f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
981 | * b[l + (j + 2) * b_dim1]; | |
982 | f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
983 | * b[l + (j + 3) * b_dim1]; | |
984 | f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
985 | * b[l + (j + 3) * b_dim1]; | |
986 | } | |
987 | c[i + j * c_dim1] = f11; | |
988 | c[i + 1 + j * c_dim1] = f21; | |
989 | c[i + (j + 1) * c_dim1] = f12; | |
990 | c[i + 1 + (j + 1) * c_dim1] = f22; | |
991 | c[i + (j + 2) * c_dim1] = f13; | |
992 | c[i + 1 + (j + 2) * c_dim1] = f23; | |
993 | c[i + (j + 3) * c_dim1] = f14; | |
994 | c[i + 1 + (j + 3) * c_dim1] = f24; | |
995 | c[i + 2 + j * c_dim1] = f31; | |
996 | c[i + 3 + j * c_dim1] = f41; | |
997 | c[i + 2 + (j + 1) * c_dim1] = f32; | |
998 | c[i + 3 + (j + 1) * c_dim1] = f42; | |
999 | c[i + 2 + (j + 2) * c_dim1] = f33; | |
1000 | c[i + 3 + (j + 2) * c_dim1] = f43; | |
1001 | c[i + 2 + (j + 3) * c_dim1] = f34; | |
1002 | c[i + 3 + (j + 3) * c_dim1] = f44; | |
1003 | } | |
1004 | if (uisec < isec) | |
1005 | { | |
1006 | i5 = ii + isec - 1; | |
1007 | for (i = ii + uisec; i <= i5; ++i) | |
1008 | { | |
1009 | f11 = c[i + j * c_dim1]; | |
1010 | f12 = c[i + (j + 1) * c_dim1]; | |
1011 | f13 = c[i + (j + 2) * c_dim1]; | |
1012 | f14 = c[i + (j + 3) * c_dim1]; | |
1013 | i6 = ll + lsec - 1; | |
1014 | for (l = ll; l <= i6; ++l) | |
1015 | { | |
1016 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1017 | 257] * b[l + j * b_dim1]; | |
1018 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1019 | 257] * b[l + (j + 1) * b_dim1]; | |
1020 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1021 | 257] * b[l + (j + 2) * b_dim1]; | |
1022 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1023 | 257] * b[l + (j + 3) * b_dim1]; | |
1024 | } | |
1025 | c[i + j * c_dim1] = f11; | |
1026 | c[i + (j + 1) * c_dim1] = f12; | |
1027 | c[i + (j + 2) * c_dim1] = f13; | |
1028 | c[i + (j + 3) * c_dim1] = f14; | |
1029 | } | |
1030 | } | |
1031 | } | |
1032 | if (ujsec < jsec) | |
1033 | { | |
1034 | i4 = jj + jsec - 1; | |
1035 | for (j = jj + ujsec; j <= i4; ++j) | |
1036 | { | |
1037 | i5 = ii + uisec - 1; | |
1038 | for (i = ii; i <= i5; i += 4) | |
1039 | { | |
1040 | f11 = c[i + j * c_dim1]; | |
1041 | f21 = c[i + 1 + j * c_dim1]; | |
1042 | f31 = c[i + 2 + j * c_dim1]; | |
1043 | f41 = c[i + 3 + j * c_dim1]; | |
1044 | i6 = ll + lsec - 1; | |
1045 | for (l = ll; l <= i6; ++l) | |
1046 | { | |
1047 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1048 | 257] * b[l + j * b_dim1]; | |
1049 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - | |
1050 | 257] * b[l + j * b_dim1]; | |
1051 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - | |
1052 | 257] * b[l + j * b_dim1]; | |
1053 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - | |
1054 | 257] * b[l + j * b_dim1]; | |
1055 | } | |
1056 | c[i + j * c_dim1] = f11; | |
1057 | c[i + 1 + j * c_dim1] = f21; | |
1058 | c[i + 2 + j * c_dim1] = f31; | |
1059 | c[i + 3 + j * c_dim1] = f41; | |
1060 | } | |
1061 | i5 = ii + isec - 1; | |
1062 | for (i = ii + uisec; i <= i5; ++i) | |
1063 | { | |
1064 | f11 = c[i + j * c_dim1]; | |
1065 | i6 = ll + lsec - 1; | |
1066 | for (l = ll; l <= i6; ++l) | |
1067 | { | |
1068 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1069 | 257] * b[l + j * b_dim1]; | |
1070 | } | |
1071 | c[i + j * c_dim1] = f11; | |
1072 | } | |
1073 | } | |
1074 | } | |
1075 | } | |
1076 | } | |
1077 | } | |
1078 | return; | |
1079 | } | |
1080 | else if (rxstride == 1 && aystride == 1 && bxstride == 1) | |
1081 | { | |
1082 | if (GFC_DESCRIPTOR_RANK (a) != 1) | |
1083 | { | |
1084 | const GFC_REAL_16 *restrict abase_x; | |
1085 | const GFC_REAL_16 *restrict bbase_y; | |
1086 | GFC_REAL_16 *restrict dest_y; | |
1087 | GFC_REAL_16 s; | |
1088 | ||
1089 | for (y = 0; y < ycount; y++) | |
1090 | { | |
1091 | bbase_y = &bbase[y*bystride]; | |
1092 | dest_y = &dest[y*rystride]; | |
1093 | for (x = 0; x < xcount; x++) | |
1094 | { | |
1095 | abase_x = &abase[x*axstride]; | |
1096 | s = (GFC_REAL_16) 0; | |
1097 | for (n = 0; n < count; n++) | |
1098 | s += abase_x[n] * bbase_y[n]; | |
1099 | dest_y[x] = s; | |
1100 | } | |
1101 | } | |
1102 | } | |
1103 | else | |
1104 | { | |
1105 | const GFC_REAL_16 *restrict bbase_y; | |
1106 | GFC_REAL_16 s; | |
1107 | ||
1108 | for (y = 0; y < ycount; y++) | |
1109 | { | |
1110 | bbase_y = &bbase[y*bystride]; | |
1111 | s = (GFC_REAL_16) 0; | |
1112 | for (n = 0; n < count; n++) | |
1113 | s += abase[n*axstride] * bbase_y[n]; | |
1114 | dest[y*rystride] = s; | |
1115 | } | |
1116 | } | |
1117 | } | |
1118 | else if (axstride < aystride) | |
1119 | { | |
1120 | for (y = 0; y < ycount; y++) | |
1121 | for (x = 0; x < xcount; x++) | |
1122 | dest[x*rxstride + y*rystride] = (GFC_REAL_16)0; | |
1123 | ||
1124 | for (y = 0; y < ycount; y++) | |
1125 | for (n = 0; n < count; n++) | |
1126 | for (x = 0; x < xcount; x++) | |
1127 | /* dest[x,y] += a[x,n] * b[n,y] */ | |
1128 | dest[x*rxstride + y*rystride] += | |
1129 | abase[x*axstride + n*aystride] * | |
1130 | bbase[n*bxstride + y*bystride]; | |
1131 | } | |
1132 | else if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1133 | { | |
1134 | const GFC_REAL_16 *restrict bbase_y; | |
1135 | GFC_REAL_16 s; | |
1136 | ||
1137 | for (y = 0; y < ycount; y++) | |
1138 | { | |
1139 | bbase_y = &bbase[y*bystride]; | |
1140 | s = (GFC_REAL_16) 0; | |
1141 | for (n = 0; n < count; n++) | |
1142 | s += abase[n*axstride] * bbase_y[n*bxstride]; | |
1143 | dest[y*rxstride] = s; | |
1144 | } | |
1145 | } | |
1146 | else | |
1147 | { | |
1148 | const GFC_REAL_16 *restrict abase_x; | |
1149 | const GFC_REAL_16 *restrict bbase_y; | |
1150 | GFC_REAL_16 *restrict dest_y; | |
1151 | GFC_REAL_16 s; | |
1152 | ||
1153 | for (y = 0; y < ycount; y++) | |
1154 | { | |
1155 | bbase_y = &bbase[y*bystride]; | |
1156 | dest_y = &dest[y*rystride]; | |
1157 | for (x = 0; x < xcount; x++) | |
1158 | { | |
1159 | abase_x = &abase[x*axstride]; | |
1160 | s = (GFC_REAL_16) 0; | |
1161 | for (n = 0; n < count; n++) | |
1162 | s += abase_x[n*aystride] * bbase_y[n*bxstride]; | |
1163 | dest_y[x*rxstride] = s; | |
1164 | } | |
1165 | } | |
1166 | } | |
1167 | } | |
1168 | #undef POW3 | |
1169 | #undef min | |
1170 | #undef max | |
1171 | ||
1172 | #endif /* HAVE_AVX2 */ | |
1173 | ||
1174 | #ifdef HAVE_AVX512F | |
1175 | static void | |
1176 | matmul_r16_avx512f (gfc_array_r16 * const restrict retarray, | |
1177 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
1178 | int blas_limit, blas_call gemm) __attribute__((__target__("avx512f"))); | |
1179 | static void | |
1180 | matmul_r16_avx512f (gfc_array_r16 * const restrict retarray, | |
1181 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
1182 | int blas_limit, blas_call gemm) | |
1183 | { | |
1184 | const GFC_REAL_16 * restrict abase; | |
1185 | const GFC_REAL_16 * restrict bbase; | |
1186 | GFC_REAL_16 * restrict dest; | |
1187 | ||
1188 | index_type rxstride, rystride, axstride, aystride, bxstride, bystride; | |
1189 | index_type x, y, n, count, xcount, ycount; | |
1190 | ||
1191 | assert (GFC_DESCRIPTOR_RANK (a) == 2 | |
1192 | || GFC_DESCRIPTOR_RANK (b) == 2); | |
1193 | ||
1194 | /* C[xcount,ycount] = A[xcount, count] * B[count,ycount] | |
1195 | ||
1196 | Either A or B (but not both) can be rank 1: | |
1197 | ||
1198 | o One-dimensional argument A is implicitly treated as a row matrix | |
1199 | dimensioned [1,count], so xcount=1. | |
1200 | ||
1201 | o One-dimensional argument B is implicitly treated as a column matrix | |
1202 | dimensioned [count, 1], so ycount=1. | |
1203 | */ | |
1204 | ||
1205 | if (retarray->base_addr == NULL) | |
1206 | { | |
1207 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1208 | { | |
1209 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
1210 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1); | |
1211 | } | |
1212 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
1213 | { | |
1214 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
1215 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
1216 | } | |
1217 | else | |
1218 | { | |
1219 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
1220 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
1221 | ||
1222 | GFC_DIMENSION_SET(retarray->dim[1], 0, | |
1223 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, | |
1224 | GFC_DESCRIPTOR_EXTENT(retarray,0)); | |
1225 | } | |
1226 | ||
1227 | retarray->base_addr | |
1228 | = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_REAL_16)); | |
1229 | retarray->offset = 0; | |
1230 | } | |
1231 | else if (unlikely (compile_options.bounds_check)) | |
1232 | { | |
1233 | index_type ret_extent, arg_extent; | |
1234 | ||
1235 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1236 | { | |
1237 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
1238 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
1239 | if (arg_extent != ret_extent) | |
1240 | runtime_error ("Incorrect extent in return array in" | |
1241 | " MATMUL intrinsic: is %ld, should be %ld", | |
1242 | (long int) ret_extent, (long int) arg_extent); | |
1243 | } | |
1244 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
1245 | { | |
1246 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
1247 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
1248 | if (arg_extent != ret_extent) | |
1249 | runtime_error ("Incorrect extent in return array in" | |
1250 | " MATMUL intrinsic: is %ld, should be %ld", | |
1251 | (long int) ret_extent, (long int) arg_extent); | |
1252 | } | |
1253 | else | |
1254 | { | |
1255 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
1256 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
1257 | if (arg_extent != ret_extent) | |
1258 | runtime_error ("Incorrect extent in return array in" | |
1259 | " MATMUL intrinsic for dimension 1:" | |
1260 | " is %ld, should be %ld", | |
1261 | (long int) ret_extent, (long int) arg_extent); | |
1262 | ||
1263 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
1264 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1); | |
1265 | if (arg_extent != ret_extent) | |
1266 | runtime_error ("Incorrect extent in return array in" | |
1267 | " MATMUL intrinsic for dimension 2:" | |
1268 | " is %ld, should be %ld", | |
1269 | (long int) ret_extent, (long int) arg_extent); | |
1270 | } | |
1271 | } | |
1272 | ||
1273 | ||
1274 | if (GFC_DESCRIPTOR_RANK (retarray) == 1) | |
1275 | { | |
1276 | /* One-dimensional result may be addressed in the code below | |
1277 | either as a row or a column matrix. We want both cases to | |
1278 | work. */ | |
1279 | rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
1280 | } | |
1281 | else | |
1282 | { | |
1283 | rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
1284 | rystride = GFC_DESCRIPTOR_STRIDE(retarray,1); | |
1285 | } | |
1286 | ||
1287 | ||
1288 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1289 | { | |
1290 | /* Treat it as a a row matrix A[1,count]. */ | |
1291 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
1292 | aystride = 1; | |
1293 | ||
1294 | xcount = 1; | |
1295 | count = GFC_DESCRIPTOR_EXTENT(a,0); | |
1296 | } | |
1297 | else | |
1298 | { | |
1299 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
1300 | aystride = GFC_DESCRIPTOR_STRIDE(a,1); | |
1301 | ||
1302 | count = GFC_DESCRIPTOR_EXTENT(a,1); | |
1303 | xcount = GFC_DESCRIPTOR_EXTENT(a,0); | |
1304 | } | |
1305 | ||
1306 | if (count != GFC_DESCRIPTOR_EXTENT(b,0)) | |
1307 | { | |
1308 | if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0) | |
1309 | runtime_error ("dimension of array B incorrect in MATMUL intrinsic"); | |
1310 | } | |
1311 | ||
1312 | if (GFC_DESCRIPTOR_RANK (b) == 1) | |
1313 | { | |
1314 | /* Treat it as a column matrix B[count,1] */ | |
1315 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
1316 | ||
1317 | /* bystride should never be used for 1-dimensional b. | |
1318 | in case it is we want it to cause a segfault, rather than | |
1319 | an incorrect result. */ | |
1320 | bystride = 0xDEADBEEF; | |
1321 | ycount = 1; | |
1322 | } | |
1323 | else | |
1324 | { | |
1325 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
1326 | bystride = GFC_DESCRIPTOR_STRIDE(b,1); | |
1327 | ycount = GFC_DESCRIPTOR_EXTENT(b,1); | |
1328 | } | |
1329 | ||
1330 | abase = a->base_addr; | |
1331 | bbase = b->base_addr; | |
1332 | dest = retarray->base_addr; | |
1333 | ||
1334 | /* Now that everything is set up, we perform the multiplication | |
1335 | itself. */ | |
1336 | ||
1337 | #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x))) | |
1338 | #define min(a,b) ((a) <= (b) ? (a) : (b)) | |
1339 | #define max(a,b) ((a) >= (b) ? (a) : (b)) | |
1340 | ||
1341 | if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1) | |
1342 | && (bxstride == 1 || bystride == 1) | |
1343 | && (((float) xcount) * ((float) ycount) * ((float) count) | |
1344 | > POW3(blas_limit))) | |
1345 | { | |
1346 | const int m = xcount, n = ycount, k = count, ldc = rystride; | |
1347 | const GFC_REAL_16 one = 1, zero = 0; | |
1348 | const int lda = (axstride == 1) ? aystride : axstride, | |
1349 | ldb = (bxstride == 1) ? bystride : bxstride; | |
1350 | ||
1351 | if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1) | |
1352 | { | |
1353 | assert (gemm != NULL); | |
1354 | gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, | |
1355 | &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest, | |
1356 | &ldc, 1, 1); | |
1357 | return; | |
1358 | } | |
1359 | } | |
1360 | ||
1361 | if (rxstride == 1 && axstride == 1 && bxstride == 1) | |
1362 | { | |
1363 | /* This block of code implements a tuned matmul, derived from | |
1364 | Superscalar GEMM-based level 3 BLAS, Beta version 0.1 | |
1365 | ||
1366 | Bo Kagstrom and Per Ling | |
1367 | Department of Computing Science | |
1368 | Umea University | |
1369 | S-901 87 Umea, Sweden | |
1370 | ||
1371 | from netlib.org, translated to C, and modified for matmul.m4. */ | |
1372 | ||
1373 | const GFC_REAL_16 *a, *b; | |
1374 | GFC_REAL_16 *c; | |
1375 | const index_type m = xcount, n = ycount, k = count; | |
1376 | ||
1377 | /* System generated locals */ | |
1378 | index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset, | |
1379 | i1, i2, i3, i4, i5, i6; | |
1380 | ||
1381 | /* Local variables */ | |
1382 | GFC_REAL_16 t1[65536], /* was [256][256] */ | |
1383 | f11, f12, f21, f22, f31, f32, f41, f42, | |
1384 | f13, f14, f23, f24, f33, f34, f43, f44; | |
1385 | index_type i, j, l, ii, jj, ll; | |
1386 | index_type isec, jsec, lsec, uisec, ujsec, ulsec; | |
1387 | ||
1388 | a = abase; | |
1389 | b = bbase; | |
1390 | c = retarray->base_addr; | |
1391 | ||
1392 | /* Parameter adjustments */ | |
1393 | c_dim1 = rystride; | |
1394 | c_offset = 1 + c_dim1; | |
1395 | c -= c_offset; | |
1396 | a_dim1 = aystride; | |
1397 | a_offset = 1 + a_dim1; | |
1398 | a -= a_offset; | |
1399 | b_dim1 = bystride; | |
1400 | b_offset = 1 + b_dim1; | |
1401 | b -= b_offset; | |
1402 | ||
1403 | /* Early exit if possible */ | |
1404 | if (m == 0 || n == 0 || k == 0) | |
1405 | return; | |
1406 | ||
1407 | /* Empty c first. */ | |
1408 | for (j=1; j<=n; j++) | |
1409 | for (i=1; i<=m; i++) | |
1410 | c[i + j * c_dim1] = (GFC_REAL_16)0; | |
1411 | ||
1412 | /* Start turning the crank. */ | |
1413 | i1 = n; | |
1414 | for (jj = 1; jj <= i1; jj += 512) | |
1415 | { | |
1416 | /* Computing MIN */ | |
1417 | i2 = 512; | |
1418 | i3 = n - jj + 1; | |
1419 | jsec = min(i2,i3); | |
1420 | ujsec = jsec - jsec % 4; | |
1421 | i2 = k; | |
1422 | for (ll = 1; ll <= i2; ll += 256) | |
1423 | { | |
1424 | /* Computing MIN */ | |
1425 | i3 = 256; | |
1426 | i4 = k - ll + 1; | |
1427 | lsec = min(i3,i4); | |
1428 | ulsec = lsec - lsec % 2; | |
1429 | ||
1430 | i3 = m; | |
1431 | for (ii = 1; ii <= i3; ii += 256) | |
1432 | { | |
1433 | /* Computing MIN */ | |
1434 | i4 = 256; | |
1435 | i5 = m - ii + 1; | |
1436 | isec = min(i4,i5); | |
1437 | uisec = isec - isec % 2; | |
1438 | i4 = ll + ulsec - 1; | |
1439 | for (l = ll; l <= i4; l += 2) | |
1440 | { | |
1441 | i5 = ii + uisec - 1; | |
1442 | for (i = ii; i <= i5; i += 2) | |
1443 | { | |
1444 | t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] = | |
1445 | a[i + l * a_dim1]; | |
1446 | t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] = | |
1447 | a[i + (l + 1) * a_dim1]; | |
1448 | t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] = | |
1449 | a[i + 1 + l * a_dim1]; | |
1450 | t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] = | |
1451 | a[i + 1 + (l + 1) * a_dim1]; | |
1452 | } | |
1453 | if (uisec < isec) | |
1454 | { | |
1455 | t1[l - ll + 1 + (isec << 8) - 257] = | |
1456 | a[ii + isec - 1 + l * a_dim1]; | |
1457 | t1[l - ll + 2 + (isec << 8) - 257] = | |
1458 | a[ii + isec - 1 + (l + 1) * a_dim1]; | |
1459 | } | |
1460 | } | |
1461 | if (ulsec < lsec) | |
1462 | { | |
1463 | i4 = ii + isec - 1; | |
1464 | for (i = ii; i<= i4; ++i) | |
1465 | { | |
1466 | t1[lsec + ((i - ii + 1) << 8) - 257] = | |
1467 | a[i + (ll + lsec - 1) * a_dim1]; | |
1468 | } | |
1469 | } | |
1470 | ||
1471 | uisec = isec - isec % 4; | |
1472 | i4 = jj + ujsec - 1; | |
1473 | for (j = jj; j <= i4; j += 4) | |
1474 | { | |
1475 | i5 = ii + uisec - 1; | |
1476 | for (i = ii; i <= i5; i += 4) | |
1477 | { | |
1478 | f11 = c[i + j * c_dim1]; | |
1479 | f21 = c[i + 1 + j * c_dim1]; | |
1480 | f12 = c[i + (j + 1) * c_dim1]; | |
1481 | f22 = c[i + 1 + (j + 1) * c_dim1]; | |
1482 | f13 = c[i + (j + 2) * c_dim1]; | |
1483 | f23 = c[i + 1 + (j + 2) * c_dim1]; | |
1484 | f14 = c[i + (j + 3) * c_dim1]; | |
1485 | f24 = c[i + 1 + (j + 3) * c_dim1]; | |
1486 | f31 = c[i + 2 + j * c_dim1]; | |
1487 | f41 = c[i + 3 + j * c_dim1]; | |
1488 | f32 = c[i + 2 + (j + 1) * c_dim1]; | |
1489 | f42 = c[i + 3 + (j + 1) * c_dim1]; | |
1490 | f33 = c[i + 2 + (j + 2) * c_dim1]; | |
1491 | f43 = c[i + 3 + (j + 2) * c_dim1]; | |
1492 | f34 = c[i + 2 + (j + 3) * c_dim1]; | |
1493 | f44 = c[i + 3 + (j + 3) * c_dim1]; | |
1494 | i6 = ll + lsec - 1; | |
1495 | for (l = ll; l <= i6; ++l) | |
1496 | { | |
1497 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
1498 | * b[l + j * b_dim1]; | |
1499 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
1500 | * b[l + j * b_dim1]; | |
1501 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
1502 | * b[l + (j + 1) * b_dim1]; | |
1503 | f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
1504 | * b[l + (j + 1) * b_dim1]; | |
1505 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
1506 | * b[l + (j + 2) * b_dim1]; | |
1507 | f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
1508 | * b[l + (j + 2) * b_dim1]; | |
1509 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
1510 | * b[l + (j + 3) * b_dim1]; | |
1511 | f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
1512 | * b[l + (j + 3) * b_dim1]; | |
1513 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
1514 | * b[l + j * b_dim1]; | |
1515 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
1516 | * b[l + j * b_dim1]; | |
1517 | f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
1518 | * b[l + (j + 1) * b_dim1]; | |
1519 | f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
1520 | * b[l + (j + 1) * b_dim1]; | |
1521 | f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
1522 | * b[l + (j + 2) * b_dim1]; | |
1523 | f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
1524 | * b[l + (j + 2) * b_dim1]; | |
1525 | f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
1526 | * b[l + (j + 3) * b_dim1]; | |
1527 | f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
1528 | * b[l + (j + 3) * b_dim1]; | |
1529 | } | |
1530 | c[i + j * c_dim1] = f11; | |
1531 | c[i + 1 + j * c_dim1] = f21; | |
1532 | c[i + (j + 1) * c_dim1] = f12; | |
1533 | c[i + 1 + (j + 1) * c_dim1] = f22; | |
1534 | c[i + (j + 2) * c_dim1] = f13; | |
1535 | c[i + 1 + (j + 2) * c_dim1] = f23; | |
1536 | c[i + (j + 3) * c_dim1] = f14; | |
1537 | c[i + 1 + (j + 3) * c_dim1] = f24; | |
1538 | c[i + 2 + j * c_dim1] = f31; | |
1539 | c[i + 3 + j * c_dim1] = f41; | |
1540 | c[i + 2 + (j + 1) * c_dim1] = f32; | |
1541 | c[i + 3 + (j + 1) * c_dim1] = f42; | |
1542 | c[i + 2 + (j + 2) * c_dim1] = f33; | |
1543 | c[i + 3 + (j + 2) * c_dim1] = f43; | |
1544 | c[i + 2 + (j + 3) * c_dim1] = f34; | |
1545 | c[i + 3 + (j + 3) * c_dim1] = f44; | |
1546 | } | |
1547 | if (uisec < isec) | |
1548 | { | |
1549 | i5 = ii + isec - 1; | |
1550 | for (i = ii + uisec; i <= i5; ++i) | |
1551 | { | |
1552 | f11 = c[i + j * c_dim1]; | |
1553 | f12 = c[i + (j + 1) * c_dim1]; | |
1554 | f13 = c[i + (j + 2) * c_dim1]; | |
1555 | f14 = c[i + (j + 3) * c_dim1]; | |
1556 | i6 = ll + lsec - 1; | |
1557 | for (l = ll; l <= i6; ++l) | |
1558 | { | |
1559 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1560 | 257] * b[l + j * b_dim1]; | |
1561 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1562 | 257] * b[l + (j + 1) * b_dim1]; | |
1563 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1564 | 257] * b[l + (j + 2) * b_dim1]; | |
1565 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1566 | 257] * b[l + (j + 3) * b_dim1]; | |
1567 | } | |
1568 | c[i + j * c_dim1] = f11; | |
1569 | c[i + (j + 1) * c_dim1] = f12; | |
1570 | c[i + (j + 2) * c_dim1] = f13; | |
1571 | c[i + (j + 3) * c_dim1] = f14; | |
1572 | } | |
1573 | } | |
1574 | } | |
1575 | if (ujsec < jsec) | |
1576 | { | |
1577 | i4 = jj + jsec - 1; | |
1578 | for (j = jj + ujsec; j <= i4; ++j) | |
1579 | { | |
1580 | i5 = ii + uisec - 1; | |
1581 | for (i = ii; i <= i5; i += 4) | |
1582 | { | |
1583 | f11 = c[i + j * c_dim1]; | |
1584 | f21 = c[i + 1 + j * c_dim1]; | |
1585 | f31 = c[i + 2 + j * c_dim1]; | |
1586 | f41 = c[i + 3 + j * c_dim1]; | |
1587 | i6 = ll + lsec - 1; | |
1588 | for (l = ll; l <= i6; ++l) | |
1589 | { | |
1590 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1591 | 257] * b[l + j * b_dim1]; | |
1592 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - | |
1593 | 257] * b[l + j * b_dim1]; | |
1594 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - | |
1595 | 257] * b[l + j * b_dim1]; | |
1596 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - | |
1597 | 257] * b[l + j * b_dim1]; | |
1598 | } | |
1599 | c[i + j * c_dim1] = f11; | |
1600 | c[i + 1 + j * c_dim1] = f21; | |
1601 | c[i + 2 + j * c_dim1] = f31; | |
1602 | c[i + 3 + j * c_dim1] = f41; | |
1603 | } | |
1604 | i5 = ii + isec - 1; | |
1605 | for (i = ii + uisec; i <= i5; ++i) | |
1606 | { | |
1607 | f11 = c[i + j * c_dim1]; | |
1608 | i6 = ll + lsec - 1; | |
1609 | for (l = ll; l <= i6; ++l) | |
1610 | { | |
1611 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
1612 | 257] * b[l + j * b_dim1]; | |
1613 | } | |
1614 | c[i + j * c_dim1] = f11; | |
1615 | } | |
1616 | } | |
1617 | } | |
1618 | } | |
1619 | } | |
1620 | } | |
1621 | return; | |
1622 | } | |
1623 | else if (rxstride == 1 && aystride == 1 && bxstride == 1) | |
1624 | { | |
1625 | if (GFC_DESCRIPTOR_RANK (a) != 1) | |
1626 | { | |
1627 | const GFC_REAL_16 *restrict abase_x; | |
1628 | const GFC_REAL_16 *restrict bbase_y; | |
1629 | GFC_REAL_16 *restrict dest_y; | |
1630 | GFC_REAL_16 s; | |
1631 | ||
1632 | for (y = 0; y < ycount; y++) | |
1633 | { | |
1634 | bbase_y = &bbase[y*bystride]; | |
1635 | dest_y = &dest[y*rystride]; | |
1636 | for (x = 0; x < xcount; x++) | |
1637 | { | |
1638 | abase_x = &abase[x*axstride]; | |
1639 | s = (GFC_REAL_16) 0; | |
1640 | for (n = 0; n < count; n++) | |
1641 | s += abase_x[n] * bbase_y[n]; | |
1642 | dest_y[x] = s; | |
1643 | } | |
1644 | } | |
1645 | } | |
1646 | else | |
1647 | { | |
1648 | const GFC_REAL_16 *restrict bbase_y; | |
1649 | GFC_REAL_16 s; | |
1650 | ||
1651 | for (y = 0; y < ycount; y++) | |
1652 | { | |
1653 | bbase_y = &bbase[y*bystride]; | |
1654 | s = (GFC_REAL_16) 0; | |
1655 | for (n = 0; n < count; n++) | |
1656 | s += abase[n*axstride] * bbase_y[n]; | |
1657 | dest[y*rystride] = s; | |
1658 | } | |
1659 | } | |
1660 | } | |
1661 | else if (axstride < aystride) | |
1662 | { | |
1663 | for (y = 0; y < ycount; y++) | |
1664 | for (x = 0; x < xcount; x++) | |
1665 | dest[x*rxstride + y*rystride] = (GFC_REAL_16)0; | |
1666 | ||
1667 | for (y = 0; y < ycount; y++) | |
1668 | for (n = 0; n < count; n++) | |
1669 | for (x = 0; x < xcount; x++) | |
1670 | /* dest[x,y] += a[x,n] * b[n,y] */ | |
1671 | dest[x*rxstride + y*rystride] += | |
1672 | abase[x*axstride + n*aystride] * | |
1673 | bbase[n*bxstride + y*bystride]; | |
1674 | } | |
1675 | else if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1676 | { | |
1677 | const GFC_REAL_16 *restrict bbase_y; | |
1678 | GFC_REAL_16 s; | |
1679 | ||
1680 | for (y = 0; y < ycount; y++) | |
1681 | { | |
1682 | bbase_y = &bbase[y*bystride]; | |
1683 | s = (GFC_REAL_16) 0; | |
1684 | for (n = 0; n < count; n++) | |
1685 | s += abase[n*axstride] * bbase_y[n*bxstride]; | |
1686 | dest[y*rxstride] = s; | |
1687 | } | |
1688 | } | |
1689 | else | |
1690 | { | |
1691 | const GFC_REAL_16 *restrict abase_x; | |
1692 | const GFC_REAL_16 *restrict bbase_y; | |
1693 | GFC_REAL_16 *restrict dest_y; | |
1694 | GFC_REAL_16 s; | |
1695 | ||
1696 | for (y = 0; y < ycount; y++) | |
1697 | { | |
1698 | bbase_y = &bbase[y*bystride]; | |
1699 | dest_y = &dest[y*rystride]; | |
1700 | for (x = 0; x < xcount; x++) | |
1701 | { | |
1702 | abase_x = &abase[x*axstride]; | |
1703 | s = (GFC_REAL_16) 0; | |
1704 | for (n = 0; n < count; n++) | |
1705 | s += abase_x[n*aystride] * bbase_y[n*bxstride]; | |
1706 | dest_y[x*rxstride] = s; | |
1707 | } | |
1708 | } | |
1709 | } | |
1710 | } | |
1711 | #undef POW3 | |
1712 | #undef min | |
1713 | #undef max | |
1714 | ||
1715 | #endif /* HAVE_AVX512F */ | |
1716 | ||
1717 | /* Function to fall back to if there is no special processor-specific version. */ | |
1718 | static void | |
1719 | matmul_r16_vanilla (gfc_array_r16 * const restrict retarray, | |
1720 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
1721 | int blas_limit, blas_call gemm) | |
1722 | { | |
1723 | const GFC_REAL_16 * restrict abase; | |
1724 | const GFC_REAL_16 * restrict bbase; | |
1725 | GFC_REAL_16 * restrict dest; | |
1726 | ||
1727 | index_type rxstride, rystride, axstride, aystride, bxstride, bystride; | |
1728 | index_type x, y, n, count, xcount, ycount; | |
1729 | ||
1730 | assert (GFC_DESCRIPTOR_RANK (a) == 2 | |
1731 | || GFC_DESCRIPTOR_RANK (b) == 2); | |
1732 | ||
1733 | /* C[xcount,ycount] = A[xcount, count] * B[count,ycount] | |
1734 | ||
1735 | Either A or B (but not both) can be rank 1: | |
1736 | ||
1737 | o One-dimensional argument A is implicitly treated as a row matrix | |
1738 | dimensioned [1,count], so xcount=1. | |
1739 | ||
1740 | o One-dimensional argument B is implicitly treated as a column matrix | |
1741 | dimensioned [count, 1], so ycount=1. | |
1742 | */ | |
1743 | ||
1744 | if (retarray->base_addr == NULL) | |
1745 | { | |
1746 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1747 | { | |
1748 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
1749 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1); | |
1750 | } | |
1751 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
1752 | { | |
1753 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
1754 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
1755 | } | |
1756 | else | |
1757 | { | |
1758 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
1759 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
1760 | ||
1761 | GFC_DIMENSION_SET(retarray->dim[1], 0, | |
1762 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, | |
1763 | GFC_DESCRIPTOR_EXTENT(retarray,0)); | |
1764 | } | |
1765 | ||
1766 | retarray->base_addr | |
1767 | = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_REAL_16)); | |
1768 | retarray->offset = 0; | |
1769 | } | |
1770 | else if (unlikely (compile_options.bounds_check)) | |
1771 | { | |
1772 | index_type ret_extent, arg_extent; | |
1773 | ||
1774 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1775 | { | |
1776 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
1777 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
1778 | if (arg_extent != ret_extent) | |
1779 | runtime_error ("Incorrect extent in return array in" | |
1780 | " MATMUL intrinsic: is %ld, should be %ld", | |
1781 | (long int) ret_extent, (long int) arg_extent); | |
1782 | } | |
1783 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
1784 | { | |
1785 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
1786 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
1787 | if (arg_extent != ret_extent) | |
1788 | runtime_error ("Incorrect extent in return array in" | |
1789 | " MATMUL intrinsic: is %ld, should be %ld", | |
1790 | (long int) ret_extent, (long int) arg_extent); | |
1791 | } | |
1792 | else | |
1793 | { | |
1794 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
1795 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
1796 | if (arg_extent != ret_extent) | |
1797 | runtime_error ("Incorrect extent in return array in" | |
1798 | " MATMUL intrinsic for dimension 1:" | |
1799 | " is %ld, should be %ld", | |
1800 | (long int) ret_extent, (long int) arg_extent); | |
1801 | ||
1802 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
1803 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1); | |
1804 | if (arg_extent != ret_extent) | |
1805 | runtime_error ("Incorrect extent in return array in" | |
1806 | " MATMUL intrinsic for dimension 2:" | |
1807 | " is %ld, should be %ld", | |
1808 | (long int) ret_extent, (long int) arg_extent); | |
1809 | } | |
1810 | } | |
1811 | ||
1812 | ||
1813 | if (GFC_DESCRIPTOR_RANK (retarray) == 1) | |
1814 | { | |
1815 | /* One-dimensional result may be addressed in the code below | |
1816 | either as a row or a column matrix. We want both cases to | |
1817 | work. */ | |
1818 | rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
1819 | } | |
1820 | else | |
1821 | { | |
1822 | rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
1823 | rystride = GFC_DESCRIPTOR_STRIDE(retarray,1); | |
1824 | } | |
1825 | ||
1826 | ||
1827 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
1828 | { | |
1829 | /* Treat it as a a row matrix A[1,count]. */ | |
1830 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
1831 | aystride = 1; | |
1832 | ||
1833 | xcount = 1; | |
1834 | count = GFC_DESCRIPTOR_EXTENT(a,0); | |
1835 | } | |
1836 | else | |
1837 | { | |
1838 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
1839 | aystride = GFC_DESCRIPTOR_STRIDE(a,1); | |
1840 | ||
1841 | count = GFC_DESCRIPTOR_EXTENT(a,1); | |
1842 | xcount = GFC_DESCRIPTOR_EXTENT(a,0); | |
1843 | } | |
1844 | ||
1845 | if (count != GFC_DESCRIPTOR_EXTENT(b,0)) | |
1846 | { | |
1847 | if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0) | |
1848 | runtime_error ("dimension of array B incorrect in MATMUL intrinsic"); | |
1849 | } | |
1850 | ||
1851 | if (GFC_DESCRIPTOR_RANK (b) == 1) | |
1852 | { | |
1853 | /* Treat it as a column matrix B[count,1] */ | |
1854 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
1855 | ||
1856 | /* bystride should never be used for 1-dimensional b. | |
1857 | in case it is we want it to cause a segfault, rather than | |
1858 | an incorrect result. */ | |
1859 | bystride = 0xDEADBEEF; | |
1860 | ycount = 1; | |
1861 | } | |
1862 | else | |
1863 | { | |
1864 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
1865 | bystride = GFC_DESCRIPTOR_STRIDE(b,1); | |
1866 | ycount = GFC_DESCRIPTOR_EXTENT(b,1); | |
1867 | } | |
1868 | ||
1869 | abase = a->base_addr; | |
1870 | bbase = b->base_addr; | |
1871 | dest = retarray->base_addr; | |
1872 | ||
1873 | /* Now that everything is set up, we perform the multiplication | |
1874 | itself. */ | |
1875 | ||
1876 | #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x))) | |
1877 | #define min(a,b) ((a) <= (b) ? (a) : (b)) | |
1878 | #define max(a,b) ((a) >= (b) ? (a) : (b)) | |
1879 | ||
1880 | if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1) | |
1881 | && (bxstride == 1 || bystride == 1) | |
1882 | && (((float) xcount) * ((float) ycount) * ((float) count) | |
1883 | > POW3(blas_limit))) | |
1884 | { | |
1885 | const int m = xcount, n = ycount, k = count, ldc = rystride; | |
1886 | const GFC_REAL_16 one = 1, zero = 0; | |
1887 | const int lda = (axstride == 1) ? aystride : axstride, | |
1888 | ldb = (bxstride == 1) ? bystride : bxstride; | |
1889 | ||
1890 | if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1) | |
1891 | { | |
1892 | assert (gemm != NULL); | |
1893 | gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, | |
1894 | &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest, | |
1895 | &ldc, 1, 1); | |
1896 | return; | |
1897 | } | |
1898 | } | |
1899 | ||
1900 | if (rxstride == 1 && axstride == 1 && bxstride == 1) | |
1901 | { | |
1902 | /* This block of code implements a tuned matmul, derived from | |
1903 | Superscalar GEMM-based level 3 BLAS, Beta version 0.1 | |
1904 | ||
1905 | Bo Kagstrom and Per Ling | |
1906 | Department of Computing Science | |
1907 | Umea University | |
1908 | S-901 87 Umea, Sweden | |
1909 | ||
1910 | from netlib.org, translated to C, and modified for matmul.m4. */ | |
1911 | ||
1912 | const GFC_REAL_16 *a, *b; | |
1913 | GFC_REAL_16 *c; | |
1914 | const index_type m = xcount, n = ycount, k = count; | |
1915 | ||
1916 | /* System generated locals */ | |
1917 | index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset, | |
1918 | i1, i2, i3, i4, i5, i6; | |
1919 | ||
1920 | /* Local variables */ | |
1921 | GFC_REAL_16 t1[65536], /* was [256][256] */ | |
1922 | f11, f12, f21, f22, f31, f32, f41, f42, | |
1923 | f13, f14, f23, f24, f33, f34, f43, f44; | |
1924 | index_type i, j, l, ii, jj, ll; | |
1925 | index_type isec, jsec, lsec, uisec, ujsec, ulsec; | |
1926 | ||
1927 | a = abase; | |
1928 | b = bbase; | |
1929 | c = retarray->base_addr; | |
1930 | ||
1931 | /* Parameter adjustments */ | |
1932 | c_dim1 = rystride; | |
1933 | c_offset = 1 + c_dim1; | |
1934 | c -= c_offset; | |
1935 | a_dim1 = aystride; | |
1936 | a_offset = 1 + a_dim1; | |
1937 | a -= a_offset; | |
1938 | b_dim1 = bystride; | |
1939 | b_offset = 1 + b_dim1; | |
1940 | b -= b_offset; | |
1941 | ||
1942 | /* Early exit if possible */ | |
1943 | if (m == 0 || n == 0 || k == 0) | |
1944 | return; | |
1945 | ||
1946 | /* Empty c first. */ | |
1947 | for (j=1; j<=n; j++) | |
1948 | for (i=1; i<=m; i++) | |
1949 | c[i + j * c_dim1] = (GFC_REAL_16)0; | |
1950 | ||
1951 | /* Start turning the crank. */ | |
1952 | i1 = n; | |
1953 | for (jj = 1; jj <= i1; jj += 512) | |
1954 | { | |
1955 | /* Computing MIN */ | |
1956 | i2 = 512; | |
1957 | i3 = n - jj + 1; | |
1958 | jsec = min(i2,i3); | |
1959 | ujsec = jsec - jsec % 4; | |
1960 | i2 = k; | |
1961 | for (ll = 1; ll <= i2; ll += 256) | |
1962 | { | |
1963 | /* Computing MIN */ | |
1964 | i3 = 256; | |
1965 | i4 = k - ll + 1; | |
1966 | lsec = min(i3,i4); | |
1967 | ulsec = lsec - lsec % 2; | |
1968 | ||
1969 | i3 = m; | |
1970 | for (ii = 1; ii <= i3; ii += 256) | |
1971 | { | |
1972 | /* Computing MIN */ | |
1973 | i4 = 256; | |
1974 | i5 = m - ii + 1; | |
1975 | isec = min(i4,i5); | |
1976 | uisec = isec - isec % 2; | |
1977 | i4 = ll + ulsec - 1; | |
1978 | for (l = ll; l <= i4; l += 2) | |
1979 | { | |
1980 | i5 = ii + uisec - 1; | |
1981 | for (i = ii; i <= i5; i += 2) | |
1982 | { | |
1983 | t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] = | |
1984 | a[i + l * a_dim1]; | |
1985 | t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] = | |
1986 | a[i + (l + 1) * a_dim1]; | |
1987 | t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] = | |
1988 | a[i + 1 + l * a_dim1]; | |
1989 | t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] = | |
1990 | a[i + 1 + (l + 1) * a_dim1]; | |
1991 | } | |
1992 | if (uisec < isec) | |
1993 | { | |
1994 | t1[l - ll + 1 + (isec << 8) - 257] = | |
1995 | a[ii + isec - 1 + l * a_dim1]; | |
1996 | t1[l - ll + 2 + (isec << 8) - 257] = | |
1997 | a[ii + isec - 1 + (l + 1) * a_dim1]; | |
1998 | } | |
1999 | } | |
2000 | if (ulsec < lsec) | |
2001 | { | |
2002 | i4 = ii + isec - 1; | |
2003 | for (i = ii; i<= i4; ++i) | |
2004 | { | |
2005 | t1[lsec + ((i - ii + 1) << 8) - 257] = | |
2006 | a[i + (ll + lsec - 1) * a_dim1]; | |
2007 | } | |
2008 | } | |
2009 | ||
2010 | uisec = isec - isec % 4; | |
2011 | i4 = jj + ujsec - 1; | |
2012 | for (j = jj; j <= i4; j += 4) | |
2013 | { | |
2014 | i5 = ii + uisec - 1; | |
2015 | for (i = ii; i <= i5; i += 4) | |
2016 | { | |
2017 | f11 = c[i + j * c_dim1]; | |
2018 | f21 = c[i + 1 + j * c_dim1]; | |
2019 | f12 = c[i + (j + 1) * c_dim1]; | |
2020 | f22 = c[i + 1 + (j + 1) * c_dim1]; | |
2021 | f13 = c[i + (j + 2) * c_dim1]; | |
2022 | f23 = c[i + 1 + (j + 2) * c_dim1]; | |
2023 | f14 = c[i + (j + 3) * c_dim1]; | |
2024 | f24 = c[i + 1 + (j + 3) * c_dim1]; | |
2025 | f31 = c[i + 2 + j * c_dim1]; | |
2026 | f41 = c[i + 3 + j * c_dim1]; | |
2027 | f32 = c[i + 2 + (j + 1) * c_dim1]; | |
2028 | f42 = c[i + 3 + (j + 1) * c_dim1]; | |
2029 | f33 = c[i + 2 + (j + 2) * c_dim1]; | |
2030 | f43 = c[i + 3 + (j + 2) * c_dim1]; | |
2031 | f34 = c[i + 2 + (j + 3) * c_dim1]; | |
2032 | f44 = c[i + 3 + (j + 3) * c_dim1]; | |
2033 | i6 = ll + lsec - 1; | |
2034 | for (l = ll; l <= i6; ++l) | |
2035 | { | |
2036 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2037 | * b[l + j * b_dim1]; | |
2038 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2039 | * b[l + j * b_dim1]; | |
2040 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2041 | * b[l + (j + 1) * b_dim1]; | |
2042 | f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2043 | * b[l + (j + 1) * b_dim1]; | |
2044 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2045 | * b[l + (j + 2) * b_dim1]; | |
2046 | f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2047 | * b[l + (j + 2) * b_dim1]; | |
2048 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2049 | * b[l + (j + 3) * b_dim1]; | |
2050 | f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2051 | * b[l + (j + 3) * b_dim1]; | |
2052 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2053 | * b[l + j * b_dim1]; | |
2054 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2055 | * b[l + j * b_dim1]; | |
2056 | f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2057 | * b[l + (j + 1) * b_dim1]; | |
2058 | f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2059 | * b[l + (j + 1) * b_dim1]; | |
2060 | f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2061 | * b[l + (j + 2) * b_dim1]; | |
2062 | f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2063 | * b[l + (j + 2) * b_dim1]; | |
2064 | f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2065 | * b[l + (j + 3) * b_dim1]; | |
2066 | f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2067 | * b[l + (j + 3) * b_dim1]; | |
2068 | } | |
2069 | c[i + j * c_dim1] = f11; | |
2070 | c[i + 1 + j * c_dim1] = f21; | |
2071 | c[i + (j + 1) * c_dim1] = f12; | |
2072 | c[i + 1 + (j + 1) * c_dim1] = f22; | |
2073 | c[i + (j + 2) * c_dim1] = f13; | |
2074 | c[i + 1 + (j + 2) * c_dim1] = f23; | |
2075 | c[i + (j + 3) * c_dim1] = f14; | |
2076 | c[i + 1 + (j + 3) * c_dim1] = f24; | |
2077 | c[i + 2 + j * c_dim1] = f31; | |
2078 | c[i + 3 + j * c_dim1] = f41; | |
2079 | c[i + 2 + (j + 1) * c_dim1] = f32; | |
2080 | c[i + 3 + (j + 1) * c_dim1] = f42; | |
2081 | c[i + 2 + (j + 2) * c_dim1] = f33; | |
2082 | c[i + 3 + (j + 2) * c_dim1] = f43; | |
2083 | c[i + 2 + (j + 3) * c_dim1] = f34; | |
2084 | c[i + 3 + (j + 3) * c_dim1] = f44; | |
2085 | } | |
2086 | if (uisec < isec) | |
2087 | { | |
2088 | i5 = ii + isec - 1; | |
2089 | for (i = ii + uisec; i <= i5; ++i) | |
2090 | { | |
2091 | f11 = c[i + j * c_dim1]; | |
2092 | f12 = c[i + (j + 1) * c_dim1]; | |
2093 | f13 = c[i + (j + 2) * c_dim1]; | |
2094 | f14 = c[i + (j + 3) * c_dim1]; | |
2095 | i6 = ll + lsec - 1; | |
2096 | for (l = ll; l <= i6; ++l) | |
2097 | { | |
2098 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2099 | 257] * b[l + j * b_dim1]; | |
2100 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2101 | 257] * b[l + (j + 1) * b_dim1]; | |
2102 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2103 | 257] * b[l + (j + 2) * b_dim1]; | |
2104 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2105 | 257] * b[l + (j + 3) * b_dim1]; | |
2106 | } | |
2107 | c[i + j * c_dim1] = f11; | |
2108 | c[i + (j + 1) * c_dim1] = f12; | |
2109 | c[i + (j + 2) * c_dim1] = f13; | |
2110 | c[i + (j + 3) * c_dim1] = f14; | |
2111 | } | |
2112 | } | |
2113 | } | |
2114 | if (ujsec < jsec) | |
2115 | { | |
2116 | i4 = jj + jsec - 1; | |
2117 | for (j = jj + ujsec; j <= i4; ++j) | |
2118 | { | |
2119 | i5 = ii + uisec - 1; | |
2120 | for (i = ii; i <= i5; i += 4) | |
2121 | { | |
2122 | f11 = c[i + j * c_dim1]; | |
2123 | f21 = c[i + 1 + j * c_dim1]; | |
2124 | f31 = c[i + 2 + j * c_dim1]; | |
2125 | f41 = c[i + 3 + j * c_dim1]; | |
2126 | i6 = ll + lsec - 1; | |
2127 | for (l = ll; l <= i6; ++l) | |
2128 | { | |
2129 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2130 | 257] * b[l + j * b_dim1]; | |
2131 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - | |
2132 | 257] * b[l + j * b_dim1]; | |
2133 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - | |
2134 | 257] * b[l + j * b_dim1]; | |
2135 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - | |
2136 | 257] * b[l + j * b_dim1]; | |
2137 | } | |
2138 | c[i + j * c_dim1] = f11; | |
2139 | c[i + 1 + j * c_dim1] = f21; | |
2140 | c[i + 2 + j * c_dim1] = f31; | |
2141 | c[i + 3 + j * c_dim1] = f41; | |
2142 | } | |
2143 | i5 = ii + isec - 1; | |
2144 | for (i = ii + uisec; i <= i5; ++i) | |
2145 | { | |
2146 | f11 = c[i + j * c_dim1]; | |
2147 | i6 = ll + lsec - 1; | |
2148 | for (l = ll; l <= i6; ++l) | |
2149 | { | |
2150 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2151 | 257] * b[l + j * b_dim1]; | |
2152 | } | |
2153 | c[i + j * c_dim1] = f11; | |
2154 | } | |
2155 | } | |
2156 | } | |
2157 | } | |
2158 | } | |
2159 | } | |
2160 | return; | |
2161 | } | |
2162 | else if (rxstride == 1 && aystride == 1 && bxstride == 1) | |
2163 | { | |
2164 | if (GFC_DESCRIPTOR_RANK (a) != 1) | |
2165 | { | |
2166 | const GFC_REAL_16 *restrict abase_x; | |
2167 | const GFC_REAL_16 *restrict bbase_y; | |
2168 | GFC_REAL_16 *restrict dest_y; | |
2169 | GFC_REAL_16 s; | |
2170 | ||
2171 | for (y = 0; y < ycount; y++) | |
2172 | { | |
2173 | bbase_y = &bbase[y*bystride]; | |
2174 | dest_y = &dest[y*rystride]; | |
2175 | for (x = 0; x < xcount; x++) | |
2176 | { | |
2177 | abase_x = &abase[x*axstride]; | |
2178 | s = (GFC_REAL_16) 0; | |
2179 | for (n = 0; n < count; n++) | |
2180 | s += abase_x[n] * bbase_y[n]; | |
2181 | dest_y[x] = s; | |
2182 | } | |
2183 | } | |
2184 | } | |
2185 | else | |
2186 | { | |
2187 | const GFC_REAL_16 *restrict bbase_y; | |
2188 | GFC_REAL_16 s; | |
2189 | ||
2190 | for (y = 0; y < ycount; y++) | |
2191 | { | |
2192 | bbase_y = &bbase[y*bystride]; | |
2193 | s = (GFC_REAL_16) 0; | |
2194 | for (n = 0; n < count; n++) | |
2195 | s += abase[n*axstride] * bbase_y[n]; | |
2196 | dest[y*rystride] = s; | |
2197 | } | |
2198 | } | |
2199 | } | |
2200 | else if (axstride < aystride) | |
2201 | { | |
2202 | for (y = 0; y < ycount; y++) | |
2203 | for (x = 0; x < xcount; x++) | |
2204 | dest[x*rxstride + y*rystride] = (GFC_REAL_16)0; | |
2205 | ||
2206 | for (y = 0; y < ycount; y++) | |
2207 | for (n = 0; n < count; n++) | |
2208 | for (x = 0; x < xcount; x++) | |
2209 | /* dest[x,y] += a[x,n] * b[n,y] */ | |
2210 | dest[x*rxstride + y*rystride] += | |
2211 | abase[x*axstride + n*aystride] * | |
2212 | bbase[n*bxstride + y*bystride]; | |
2213 | } | |
2214 | else if (GFC_DESCRIPTOR_RANK (a) == 1) | |
2215 | { | |
2216 | const GFC_REAL_16 *restrict bbase_y; | |
2217 | GFC_REAL_16 s; | |
2218 | ||
2219 | for (y = 0; y < ycount; y++) | |
2220 | { | |
2221 | bbase_y = &bbase[y*bystride]; | |
2222 | s = (GFC_REAL_16) 0; | |
2223 | for (n = 0; n < count; n++) | |
2224 | s += abase[n*axstride] * bbase_y[n*bxstride]; | |
2225 | dest[y*rxstride] = s; | |
2226 | } | |
2227 | } | |
2228 | else | |
2229 | { | |
2230 | const GFC_REAL_16 *restrict abase_x; | |
2231 | const GFC_REAL_16 *restrict bbase_y; | |
2232 | GFC_REAL_16 *restrict dest_y; | |
2233 | GFC_REAL_16 s; | |
2234 | ||
2235 | for (y = 0; y < ycount; y++) | |
2236 | { | |
2237 | bbase_y = &bbase[y*bystride]; | |
2238 | dest_y = &dest[y*rystride]; | |
2239 | for (x = 0; x < xcount; x++) | |
2240 | { | |
2241 | abase_x = &abase[x*axstride]; | |
2242 | s = (GFC_REAL_16) 0; | |
2243 | for (n = 0; n < count; n++) | |
2244 | s += abase_x[n*aystride] * bbase_y[n*bxstride]; | |
2245 | dest_y[x*rxstride] = s; | |
2246 | } | |
2247 | } | |
2248 | } | |
2249 | } | |
2250 | #undef POW3 | |
2251 | #undef min | |
2252 | #undef max | |
2253 | ||
2254 | ||
2255 | /* Compiling main function, with selection code for the processor. */ | |
2256 | ||
2257 | /* Currently, this is i386 only. Adjust for other architectures. */ | |
2258 | ||
2259 | #include <config/i386/cpuinfo.h> | |
2260 | void matmul_r16 (gfc_array_r16 * const restrict retarray, | |
2261 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
2262 | int blas_limit, blas_call gemm) | |
2263 | { | |
2264 | static void (*matmul_p) (gfc_array_r16 * const restrict retarray, | |
2265 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
2266 | int blas_limit, blas_call gemm) = NULL; | |
2267 | ||
2268 | if (matmul_p == NULL) | |
2269 | { | |
2270 | matmul_p = matmul_r16_vanilla; | |
2271 | if (__cpu_model.__cpu_vendor == VENDOR_INTEL) | |
2272 | { | |
2273 | /* Run down the available processors in order of preference. */ | |
2274 | #ifdef HAVE_AVX512F | |
2275 | if (__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX512F)) | |
2276 | { | |
2277 | matmul_p = matmul_r16_avx512f; | |
2278 | goto tailcall; | |
2279 | } | |
2280 | ||
2281 | #endif /* HAVE_AVX512F */ | |
2282 | ||
2283 | #ifdef HAVE_AVX2 | |
2284 | if (__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX2)) | |
2285 | { | |
2286 | matmul_p = matmul_r16_avx2; | |
2287 | goto tailcall; | |
2288 | } | |
2289 | ||
2290 | #endif | |
2291 | ||
2292 | #ifdef HAVE_AVX | |
2293 | if (__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX)) | |
2294 | { | |
2295 | matmul_p = matmul_r16_avx; | |
2296 | goto tailcall; | |
2297 | } | |
2298 | #endif /* HAVE_AVX */ | |
2299 | } | |
2300 | } | |
2301 | ||
2302 | tailcall: | |
2303 | (*matmul_p) (retarray, a, b, try_blas, blas_limit, gemm); | |
2304 | } | |
2305 | ||
2306 | #else /* Just the vanilla function. */ | |
2307 | ||
2308 | void | |
2309 | matmul_r16 (gfc_array_r16 * const restrict retarray, | |
2310 | gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas, | |
2311 | int blas_limit, blas_call gemm) | |
2312 | { | |
2313 | const GFC_REAL_16 * restrict abase; | |
2314 | const GFC_REAL_16 * restrict bbase; | |
2315 | GFC_REAL_16 * restrict dest; | |
2316 | ||
2317 | index_type rxstride, rystride, axstride, aystride, bxstride, bystride; | |
2318 | index_type x, y, n, count, xcount, ycount; | |
2319 | ||
2320 | assert (GFC_DESCRIPTOR_RANK (a) == 2 | |
2321 | || GFC_DESCRIPTOR_RANK (b) == 2); | |
2322 | ||
2323 | /* C[xcount,ycount] = A[xcount, count] * B[count,ycount] | |
2324 | ||
2325 | Either A or B (but not both) can be rank 1: | |
2326 | ||
2327 | o One-dimensional argument A is implicitly treated as a row matrix | |
2328 | dimensioned [1,count], so xcount=1. | |
2329 | ||
2330 | o One-dimensional argument B is implicitly treated as a column matrix | |
2331 | dimensioned [count, 1], so ycount=1. | |
2332 | */ | |
2333 | ||
2334 | if (retarray->base_addr == NULL) | |
2335 | { | |
2336 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
2337 | { | |
2338 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
2339 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1); | |
2340 | } | |
2341 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
2342 | { | |
2343 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
2344 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
2345 | } | |
2346 | else | |
2347 | { | |
2348 | GFC_DIMENSION_SET(retarray->dim[0], 0, | |
2349 | GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1); | |
2350 | ||
2351 | GFC_DIMENSION_SET(retarray->dim[1], 0, | |
2352 | GFC_DESCRIPTOR_EXTENT(b,1) - 1, | |
2353 | GFC_DESCRIPTOR_EXTENT(retarray,0)); | |
2354 | } | |
2355 | ||
2356 | retarray->base_addr | |
2357 | = xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_REAL_16)); | |
2358 | retarray->offset = 0; | |
2359 | } | |
2360 | else if (unlikely (compile_options.bounds_check)) | |
2361 | { | |
2362 | index_type ret_extent, arg_extent; | |
2363 | ||
2364 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
2365 | { | |
2366 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
2367 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
2368 | if (arg_extent != ret_extent) | |
2369 | runtime_error ("Incorrect extent in return array in" | |
2370 | " MATMUL intrinsic: is %ld, should be %ld", | |
2371 | (long int) ret_extent, (long int) arg_extent); | |
2372 | } | |
2373 | else if (GFC_DESCRIPTOR_RANK (b) == 1) | |
2374 | { | |
2375 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
2376 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
2377 | if (arg_extent != ret_extent) | |
2378 | runtime_error ("Incorrect extent in return array in" | |
2379 | " MATMUL intrinsic: is %ld, should be %ld", | |
2380 | (long int) ret_extent, (long int) arg_extent); | |
2381 | } | |
2382 | else | |
2383 | { | |
2384 | arg_extent = GFC_DESCRIPTOR_EXTENT(a,0); | |
2385 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0); | |
2386 | if (arg_extent != ret_extent) | |
2387 | runtime_error ("Incorrect extent in return array in" | |
2388 | " MATMUL intrinsic for dimension 1:" | |
2389 | " is %ld, should be %ld", | |
2390 | (long int) ret_extent, (long int) arg_extent); | |
2391 | ||
2392 | arg_extent = GFC_DESCRIPTOR_EXTENT(b,1); | |
2393 | ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1); | |
2394 | if (arg_extent != ret_extent) | |
2395 | runtime_error ("Incorrect extent in return array in" | |
2396 | " MATMUL intrinsic for dimension 2:" | |
2397 | " is %ld, should be %ld", | |
2398 | (long int) ret_extent, (long int) arg_extent); | |
2399 | } | |
2400 | } | |
2401 | ||
2402 | ||
2403 | if (GFC_DESCRIPTOR_RANK (retarray) == 1) | |
2404 | { | |
2405 | /* One-dimensional result may be addressed in the code below | |
2406 | either as a row or a column matrix. We want both cases to | |
2407 | work. */ | |
2408 | rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
2409 | } | |
2410 | else | |
2411 | { | |
2412 | rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0); | |
2413 | rystride = GFC_DESCRIPTOR_STRIDE(retarray,1); | |
2414 | } | |
2415 | ||
2416 | ||
2417 | if (GFC_DESCRIPTOR_RANK (a) == 1) | |
2418 | { | |
2419 | /* Treat it as a a row matrix A[1,count]. */ | |
2420 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
2421 | aystride = 1; | |
2422 | ||
2423 | xcount = 1; | |
2424 | count = GFC_DESCRIPTOR_EXTENT(a,0); | |
2425 | } | |
2426 | else | |
2427 | { | |
2428 | axstride = GFC_DESCRIPTOR_STRIDE(a,0); | |
2429 | aystride = GFC_DESCRIPTOR_STRIDE(a,1); | |
2430 | ||
2431 | count = GFC_DESCRIPTOR_EXTENT(a,1); | |
2432 | xcount = GFC_DESCRIPTOR_EXTENT(a,0); | |
2433 | } | |
2434 | ||
2435 | if (count != GFC_DESCRIPTOR_EXTENT(b,0)) | |
2436 | { | |
2437 | if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0) | |
2438 | runtime_error ("dimension of array B incorrect in MATMUL intrinsic"); | |
2439 | } | |
2440 | ||
2441 | if (GFC_DESCRIPTOR_RANK (b) == 1) | |
2442 | { | |
2443 | /* Treat it as a column matrix B[count,1] */ | |
2444 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
2445 | ||
2446 | /* bystride should never be used for 1-dimensional b. | |
2447 | in case it is we want it to cause a segfault, rather than | |
2448 | an incorrect result. */ | |
2449 | bystride = 0xDEADBEEF; | |
2450 | ycount = 1; | |
2451 | } | |
2452 | else | |
2453 | { | |
2454 | bxstride = GFC_DESCRIPTOR_STRIDE(b,0); | |
2455 | bystride = GFC_DESCRIPTOR_STRIDE(b,1); | |
2456 | ycount = GFC_DESCRIPTOR_EXTENT(b,1); | |
2457 | } | |
2458 | ||
2459 | abase = a->base_addr; | |
2460 | bbase = b->base_addr; | |
2461 | dest = retarray->base_addr; | |
2462 | ||
2463 | /* Now that everything is set up, we perform the multiplication | |
2464 | itself. */ | |
2465 | ||
2466 | #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x))) | |
2467 | #define min(a,b) ((a) <= (b) ? (a) : (b)) | |
2468 | #define max(a,b) ((a) >= (b) ? (a) : (b)) | |
2469 | ||
2470 | if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1) | |
2471 | && (bxstride == 1 || bystride == 1) | |
2472 | && (((float) xcount) * ((float) ycount) * ((float) count) | |
2473 | > POW3(blas_limit))) | |
2474 | { | |
2475 | const int m = xcount, n = ycount, k = count, ldc = rystride; | |
2476 | const GFC_REAL_16 one = 1, zero = 0; | |
2477 | const int lda = (axstride == 1) ? aystride : axstride, | |
2478 | ldb = (bxstride == 1) ? bystride : bxstride; | |
2479 | ||
2480 | if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1) | |
2481 | { | |
2482 | assert (gemm != NULL); | |
2483 | gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, | |
2484 | &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest, | |
2485 | &ldc, 1, 1); | |
2486 | return; | |
2487 | } | |
2488 | } | |
2489 | ||
2490 | if (rxstride == 1 && axstride == 1 && bxstride == 1) | |
2491 | { | |
2492 | /* This block of code implements a tuned matmul, derived from | |
2493 | Superscalar GEMM-based level 3 BLAS, Beta version 0.1 | |
2494 | ||
2495 | Bo Kagstrom and Per Ling | |
2496 | Department of Computing Science | |
2497 | Umea University | |
2498 | S-901 87 Umea, Sweden | |
2499 | ||
2500 | from netlib.org, translated to C, and modified for matmul.m4. */ | |
2501 | ||
2502 | const GFC_REAL_16 *a, *b; | |
2503 | GFC_REAL_16 *c; | |
2504 | const index_type m = xcount, n = ycount, k = count; | |
2505 | ||
2506 | /* System generated locals */ | |
2507 | index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset, | |
2508 | i1, i2, i3, i4, i5, i6; | |
2509 | ||
2510 | /* Local variables */ | |
2511 | GFC_REAL_16 t1[65536], /* was [256][256] */ | |
2512 | f11, f12, f21, f22, f31, f32, f41, f42, | |
2513 | f13, f14, f23, f24, f33, f34, f43, f44; | |
2514 | index_type i, j, l, ii, jj, ll; | |
2515 | index_type isec, jsec, lsec, uisec, ujsec, ulsec; | |
2516 | ||
2517 | a = abase; | |
2518 | b = bbase; | |
2519 | c = retarray->base_addr; | |
2520 | ||
2521 | /* Parameter adjustments */ | |
2522 | c_dim1 = rystride; | |
2523 | c_offset = 1 + c_dim1; | |
2524 | c -= c_offset; | |
2525 | a_dim1 = aystride; | |
2526 | a_offset = 1 + a_dim1; | |
2527 | a -= a_offset; | |
2528 | b_dim1 = bystride; | |
2529 | b_offset = 1 + b_dim1; | |
2530 | b -= b_offset; | |
2531 | ||
2532 | /* Early exit if possible */ | |
2533 | if (m == 0 || n == 0 || k == 0) | |
2534 | return; | |
2535 | ||
2536 | /* Empty c first. */ | |
2537 | for (j=1; j<=n; j++) | |
2538 | for (i=1; i<=m; i++) | |
2539 | c[i + j * c_dim1] = (GFC_REAL_16)0; | |
2540 | ||
2541 | /* Start turning the crank. */ | |
2542 | i1 = n; | |
2543 | for (jj = 1; jj <= i1; jj += 512) | |
2544 | { | |
2545 | /* Computing MIN */ | |
2546 | i2 = 512; | |
2547 | i3 = n - jj + 1; | |
2548 | jsec = min(i2,i3); | |
2549 | ujsec = jsec - jsec % 4; | |
2550 | i2 = k; | |
2551 | for (ll = 1; ll <= i2; ll += 256) | |
2552 | { | |
2553 | /* Computing MIN */ | |
2554 | i3 = 256; | |
2555 | i4 = k - ll + 1; | |
2556 | lsec = min(i3,i4); | |
2557 | ulsec = lsec - lsec % 2; | |
2558 | ||
2559 | i3 = m; | |
2560 | for (ii = 1; ii <= i3; ii += 256) | |
2561 | { | |
2562 | /* Computing MIN */ | |
2563 | i4 = 256; | |
2564 | i5 = m - ii + 1; | |
2565 | isec = min(i4,i5); | |
2566 | uisec = isec - isec % 2; | |
2567 | i4 = ll + ulsec - 1; | |
2568 | for (l = ll; l <= i4; l += 2) | |
2569 | { | |
2570 | i5 = ii + uisec - 1; | |
2571 | for (i = ii; i <= i5; i += 2) | |
2572 | { | |
2573 | t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] = | |
2574 | a[i + l * a_dim1]; | |
2575 | t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] = | |
2576 | a[i + (l + 1) * a_dim1]; | |
2577 | t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] = | |
2578 | a[i + 1 + l * a_dim1]; | |
2579 | t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] = | |
2580 | a[i + 1 + (l + 1) * a_dim1]; | |
2581 | } | |
2582 | if (uisec < isec) | |
2583 | { | |
2584 | t1[l - ll + 1 + (isec << 8) - 257] = | |
2585 | a[ii + isec - 1 + l * a_dim1]; | |
2586 | t1[l - ll + 2 + (isec << 8) - 257] = | |
2587 | a[ii + isec - 1 + (l + 1) * a_dim1]; | |
2588 | } | |
2589 | } | |
2590 | if (ulsec < lsec) | |
2591 | { | |
2592 | i4 = ii + isec - 1; | |
2593 | for (i = ii; i<= i4; ++i) | |
2594 | { | |
2595 | t1[lsec + ((i - ii + 1) << 8) - 257] = | |
2596 | a[i + (ll + lsec - 1) * a_dim1]; | |
2597 | } | |
2598 | } | |
2599 | ||
2600 | uisec = isec - isec % 4; | |
2601 | i4 = jj + ujsec - 1; | |
2602 | for (j = jj; j <= i4; j += 4) | |
2603 | { | |
2604 | i5 = ii + uisec - 1; | |
2605 | for (i = ii; i <= i5; i += 4) | |
2606 | { | |
2607 | f11 = c[i + j * c_dim1]; | |
2608 | f21 = c[i + 1 + j * c_dim1]; | |
2609 | f12 = c[i + (j + 1) * c_dim1]; | |
2610 | f22 = c[i + 1 + (j + 1) * c_dim1]; | |
2611 | f13 = c[i + (j + 2) * c_dim1]; | |
2612 | f23 = c[i + 1 + (j + 2) * c_dim1]; | |
2613 | f14 = c[i + (j + 3) * c_dim1]; | |
2614 | f24 = c[i + 1 + (j + 3) * c_dim1]; | |
2615 | f31 = c[i + 2 + j * c_dim1]; | |
2616 | f41 = c[i + 3 + j * c_dim1]; | |
2617 | f32 = c[i + 2 + (j + 1) * c_dim1]; | |
2618 | f42 = c[i + 3 + (j + 1) * c_dim1]; | |
2619 | f33 = c[i + 2 + (j + 2) * c_dim1]; | |
2620 | f43 = c[i + 3 + (j + 2) * c_dim1]; | |
2621 | f34 = c[i + 2 + (j + 3) * c_dim1]; | |
2622 | f44 = c[i + 3 + (j + 3) * c_dim1]; | |
2623 | i6 = ll + lsec - 1; | |
2624 | for (l = ll; l <= i6; ++l) | |
2625 | { | |
2626 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2627 | * b[l + j * b_dim1]; | |
2628 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2629 | * b[l + j * b_dim1]; | |
2630 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2631 | * b[l + (j + 1) * b_dim1]; | |
2632 | f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2633 | * b[l + (j + 1) * b_dim1]; | |
2634 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2635 | * b[l + (j + 2) * b_dim1]; | |
2636 | f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2637 | * b[l + (j + 2) * b_dim1]; | |
2638 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] | |
2639 | * b[l + (j + 3) * b_dim1]; | |
2640 | f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] | |
2641 | * b[l + (j + 3) * b_dim1]; | |
2642 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2643 | * b[l + j * b_dim1]; | |
2644 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2645 | * b[l + j * b_dim1]; | |
2646 | f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2647 | * b[l + (j + 1) * b_dim1]; | |
2648 | f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2649 | * b[l + (j + 1) * b_dim1]; | |
2650 | f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2651 | * b[l + (j + 2) * b_dim1]; | |
2652 | f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2653 | * b[l + (j + 2) * b_dim1]; | |
2654 | f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257] | |
2655 | * b[l + (j + 3) * b_dim1]; | |
2656 | f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257] | |
2657 | * b[l + (j + 3) * b_dim1]; | |
2658 | } | |
2659 | c[i + j * c_dim1] = f11; | |
2660 | c[i + 1 + j * c_dim1] = f21; | |
2661 | c[i + (j + 1) * c_dim1] = f12; | |
2662 | c[i + 1 + (j + 1) * c_dim1] = f22; | |
2663 | c[i + (j + 2) * c_dim1] = f13; | |
2664 | c[i + 1 + (j + 2) * c_dim1] = f23; | |
2665 | c[i + (j + 3) * c_dim1] = f14; | |
2666 | c[i + 1 + (j + 3) * c_dim1] = f24; | |
2667 | c[i + 2 + j * c_dim1] = f31; | |
2668 | c[i + 3 + j * c_dim1] = f41; | |
2669 | c[i + 2 + (j + 1) * c_dim1] = f32; | |
2670 | c[i + 3 + (j + 1) * c_dim1] = f42; | |
2671 | c[i + 2 + (j + 2) * c_dim1] = f33; | |
2672 | c[i + 3 + (j + 2) * c_dim1] = f43; | |
2673 | c[i + 2 + (j + 3) * c_dim1] = f34; | |
2674 | c[i + 3 + (j + 3) * c_dim1] = f44; | |
2675 | } | |
2676 | if (uisec < isec) | |
2677 | { | |
2678 | i5 = ii + isec - 1; | |
2679 | for (i = ii + uisec; i <= i5; ++i) | |
2680 | { | |
2681 | f11 = c[i + j * c_dim1]; | |
2682 | f12 = c[i + (j + 1) * c_dim1]; | |
2683 | f13 = c[i + (j + 2) * c_dim1]; | |
2684 | f14 = c[i + (j + 3) * c_dim1]; | |
2685 | i6 = ll + lsec - 1; | |
2686 | for (l = ll; l <= i6; ++l) | |
2687 | { | |
2688 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2689 | 257] * b[l + j * b_dim1]; | |
2690 | f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2691 | 257] * b[l + (j + 1) * b_dim1]; | |
2692 | f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2693 | 257] * b[l + (j + 2) * b_dim1]; | |
2694 | f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2695 | 257] * b[l + (j + 3) * b_dim1]; | |
2696 | } | |
2697 | c[i + j * c_dim1] = f11; | |
2698 | c[i + (j + 1) * c_dim1] = f12; | |
2699 | c[i + (j + 2) * c_dim1] = f13; | |
2700 | c[i + (j + 3) * c_dim1] = f14; | |
2701 | } | |
2702 | } | |
2703 | } | |
2704 | if (ujsec < jsec) | |
2705 | { | |
2706 | i4 = jj + jsec - 1; | |
2707 | for (j = jj + ujsec; j <= i4; ++j) | |
2708 | { | |
2709 | i5 = ii + uisec - 1; | |
2710 | for (i = ii; i <= i5; i += 4) | |
2711 | { | |
2712 | f11 = c[i + j * c_dim1]; | |
2713 | f21 = c[i + 1 + j * c_dim1]; | |
2714 | f31 = c[i + 2 + j * c_dim1]; | |
2715 | f41 = c[i + 3 + j * c_dim1]; | |
2716 | i6 = ll + lsec - 1; | |
2717 | for (l = ll; l <= i6; ++l) | |
2718 | { | |
2719 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2720 | 257] * b[l + j * b_dim1]; | |
2721 | f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - | |
2722 | 257] * b[l + j * b_dim1]; | |
2723 | f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - | |
2724 | 257] * b[l + j * b_dim1]; | |
2725 | f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - | |
2726 | 257] * b[l + j * b_dim1]; | |
2727 | } | |
2728 | c[i + j * c_dim1] = f11; | |
2729 | c[i + 1 + j * c_dim1] = f21; | |
2730 | c[i + 2 + j * c_dim1] = f31; | |
2731 | c[i + 3 + j * c_dim1] = f41; | |
2732 | } | |
2733 | i5 = ii + isec - 1; | |
2734 | for (i = ii + uisec; i <= i5; ++i) | |
2735 | { | |
2736 | f11 = c[i + j * c_dim1]; | |
2737 | i6 = ll + lsec - 1; | |
2738 | for (l = ll; l <= i6; ++l) | |
2739 | { | |
2740 | f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - | |
2741 | 257] * b[l + j * b_dim1]; | |
2742 | } | |
2743 | c[i + j * c_dim1] = f11; | |
2744 | } | |
2745 | } | |
2746 | } | |
2747 | } | |
2748 | } | |
2749 | } | |
2750 | return; | |
2751 | } | |
2752 | else if (rxstride == 1 && aystride == 1 && bxstride == 1) | |
2753 | { | |
2754 | if (GFC_DESCRIPTOR_RANK (a) != 1) | |
2755 | { | |
2756 | const GFC_REAL_16 *restrict abase_x; | |
2757 | const GFC_REAL_16 *restrict bbase_y; | |
2758 | GFC_REAL_16 *restrict dest_y; | |
2759 | GFC_REAL_16 s; | |
2760 | ||
2761 | for (y = 0; y < ycount; y++) | |
2762 | { | |
2763 | bbase_y = &bbase[y*bystride]; | |
2764 | dest_y = &dest[y*rystride]; | |
2765 | for (x = 0; x < xcount; x++) | |
2766 | { | |
2767 | abase_x = &abase[x*axstride]; | |
2768 | s = (GFC_REAL_16) 0; | |
2769 | for (n = 0; n < count; n++) | |
2770 | s += abase_x[n] * bbase_y[n]; | |
2771 | dest_y[x] = s; | |
2772 | } | |
2773 | } | |
2774 | } | |
2775 | else | |
2776 | { | |
2777 | const GFC_REAL_16 *restrict bbase_y; | |
2778 | GFC_REAL_16 s; | |
2779 | ||
2780 | for (y = 0; y < ycount; y++) | |
2781 | { | |
2782 | bbase_y = &bbase[y*bystride]; | |
2783 | s = (GFC_REAL_16) 0; | |
2784 | for (n = 0; n < count; n++) | |
2785 | s += abase[n*axstride] * bbase_y[n]; | |
2786 | dest[y*rystride] = s; | |
2787 | } | |
2788 | } | |
2789 | } | |
2790 | else if (axstride < aystride) | |
2791 | { | |
2792 | for (y = 0; y < ycount; y++) | |
2793 | for (x = 0; x < xcount; x++) | |
2794 | dest[x*rxstride + y*rystride] = (GFC_REAL_16)0; | |
2795 | ||
2796 | for (y = 0; y < ycount; y++) | |
2797 | for (n = 0; n < count; n++) | |
2798 | for (x = 0; x < xcount; x++) | |
2799 | /* dest[x,y] += a[x,n] * b[n,y] */ | |
2800 | dest[x*rxstride + y*rystride] += | |
2801 | abase[x*axstride + n*aystride] * | |
2802 | bbase[n*bxstride + y*bystride]; | |
2803 | } | |
2804 | else if (GFC_DESCRIPTOR_RANK (a) == 1) | |
2805 | { | |
2806 | const GFC_REAL_16 *restrict bbase_y; | |
2807 | GFC_REAL_16 s; | |
2808 | ||
2809 | for (y = 0; y < ycount; y++) | |
2810 | { | |
2811 | bbase_y = &bbase[y*bystride]; | |
2812 | s = (GFC_REAL_16) 0; | |
2813 | for (n = 0; n < count; n++) | |
2814 | s += abase[n*axstride] * bbase_y[n*bxstride]; | |
2815 | dest[y*rxstride] = s; | |
2816 | } | |
2817 | } | |
2818 | else | |
2819 | { | |
2820 | const GFC_REAL_16 *restrict abase_x; | |
2821 | const GFC_REAL_16 *restrict bbase_y; | |
2822 | GFC_REAL_16 *restrict dest_y; | |
2823 | GFC_REAL_16 s; | |
2824 | ||
2825 | for (y = 0; y < ycount; y++) | |
2826 | { | |
2827 | bbase_y = &bbase[y*bystride]; | |
2828 | dest_y = &dest[y*rystride]; | |
2829 | for (x = 0; x < xcount; x++) | |
2830 | { | |
2831 | abase_x = &abase[x*axstride]; | |
2832 | s = (GFC_REAL_16) 0; | |
2833 | for (n = 0; n < count; n++) | |
2834 | s += abase_x[n*aystride] * bbase_y[n*bxstride]; | |
2835 | dest_y[x*rxstride] = s; | |
2836 | } | |
2837 | } | |
2838 | } | |
2839 | } | |
2840 | #undef POW3 | |
2841 | #undef min | |
2842 | #undef max | |
2843 | ||
2844 | #endif | |
2845 | #endif | |
2846 |