2 'matmul_name` ('rtype` * const restrict retarray,
3 'rtype` * const restrict a, 'rtype` * const restrict b, int try_blas,
4 int blas_limit, blas_call gemm)
6 const 'rtype_name` * restrict abase;
7 const 'rtype_name` * restrict bbase;
8 'rtype_name` * restrict dest;
10 index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
11 index_type x, y, n, count, xcount, ycount;
13 assert (GFC_DESCRIPTOR_RANK (a) == 2
14 || GFC_DESCRIPTOR_RANK (b) == 2);
16 /* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
18 Either A or B (but not both) can be rank 1:
20 o One-dimensional argument A is implicitly treated as a row matrix
21 dimensioned [1,count], so xcount=1.
23 o One-dimensional argument B is implicitly treated as a column matrix
24 dimensioned [count, 1], so ycount=1.
27 if (retarray->base_addr == NULL)
29 if (GFC_DESCRIPTOR_RANK (a) == 1)
31 GFC_DIMENSION_SET(retarray->dim[0], 0,
32 GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
34 else if (GFC_DESCRIPTOR_RANK (b) == 1)
36 GFC_DIMENSION_SET(retarray->dim[0], 0,
37 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
41 GFC_DIMENSION_SET(retarray->dim[0], 0,
42 GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
44 GFC_DIMENSION_SET(retarray->dim[1], 0,
45 GFC_DESCRIPTOR_EXTENT(b,1) - 1,
46 GFC_DESCRIPTOR_EXTENT(retarray,0));
50 = xmallocarray (size0 ((array_t *) retarray), sizeof ('rtype_name`));
53 else if (unlikely (compile_options.bounds_check))
55 index_type ret_extent, arg_extent;
57 if (GFC_DESCRIPTOR_RANK (a) == 1)
59 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
60 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
61 if (arg_extent != ret_extent)
62 runtime_error ("Incorrect extent in return array in"
63 " MATMUL intrinsic: is %ld, should be %ld",
64 (long int) ret_extent, (long int) arg_extent);
66 else if (GFC_DESCRIPTOR_RANK (b) == 1)
68 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
69 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
70 if (arg_extent != ret_extent)
71 runtime_error ("Incorrect extent in return array in"
72 " MATMUL intrinsic: is %ld, should be %ld",
73 (long int) ret_extent, (long int) arg_extent);
77 arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
78 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
79 if (arg_extent != ret_extent)
80 runtime_error ("Incorrect extent in return array in"
81 " MATMUL intrinsic for dimension 1:"
82 " is %ld, should be %ld",
83 (long int) ret_extent, (long int) arg_extent);
85 arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
86 ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
87 if (arg_extent != ret_extent)
88 runtime_error ("Incorrect extent in return array in"
89 " MATMUL intrinsic for dimension 2:"
90 " is %ld, should be %ld",
91 (long int) ret_extent, (long int) arg_extent);
95 sinclude(`matmul_asm_'rtype_code`.m4')dnl
97 if (GFC_DESCRIPTOR_RANK (retarray) == 1)
99 /* One-dimensional result may be addressed in the code below
100 either as a row or a column matrix. We want both cases to
102 rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
106 rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
107 rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
111 if (GFC_DESCRIPTOR_RANK (a) == 1)
113 /* Treat it as a a row matrix A[1,count]. */
114 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
118 count = GFC_DESCRIPTOR_EXTENT(a,0);
122 axstride = GFC_DESCRIPTOR_STRIDE(a,0);
123 aystride = GFC_DESCRIPTOR_STRIDE(a,1);
125 count = GFC_DESCRIPTOR_EXTENT(a,1);
126 xcount = GFC_DESCRIPTOR_EXTENT(a,0);
129 if (count != GFC_DESCRIPTOR_EXTENT(b,0))
131 if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
132 runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
135 if (GFC_DESCRIPTOR_RANK (b) == 1)
137 /* Treat it as a column matrix B[count,1] */
138 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
140 /* bystride should never be used for 1-dimensional b.
141 in case it is we want it to cause a segfault, rather than
142 an incorrect result. */
143 bystride = 0xDEADBEEF;
148 bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
149 bystride = GFC_DESCRIPTOR_STRIDE(b,1);
150 ycount = GFC_DESCRIPTOR_EXTENT(b,1);
153 abase = a->base_addr;
154 bbase = b->base_addr;
155 dest = retarray->base_addr;
157 /* Now that everything is set up, we perform the multiplication
160 #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
161 #define min(a,b) ((a) <= (b) ? (a) : (b))
162 #define max(a,b) ((a) >= (b) ? (a) : (b))
164 if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
165 && (bxstride == 1 || bystride == 1)
166 && (((float) xcount) * ((float) ycount) * ((float) count)
169 const int m = xcount, n = ycount, k = count, ldc = rystride;
170 const 'rtype_name` one = 1, zero = 0;
171 const int lda = (axstride == 1) ? aystride : axstride,
172 ldb = (bxstride == 1) ? bystride : bxstride;
174 if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
176 assert (gemm != NULL);
177 gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m,
178 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
184 if (rxstride == 1 && axstride == 1 && bxstride == 1)
186 /* This block of code implements a tuned matmul, derived from
187 Superscalar GEMM-based level 3 BLAS, Beta version 0.1
189 Bo Kagstrom and Per Ling
190 Department of Computing Science
192 S-901 87 Umea, Sweden
194 from netlib.org, translated to C, and modified for matmul.m4. */
196 const 'rtype_name` *a, *b;
198 const index_type m = xcount, n = ycount, k = count;
200 /* System generated locals */
201 index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
202 i1, i2, i3, i4, i5, i6;
204 /* Local variables */
205 'rtype_name` t1[65536], /* was [256][256] */
206 f11, f12, f21, f22, f31, f32, f41, f42,
207 f13, f14, f23, f24, f33, f34, f43, f44;
208 index_type i, j, l, ii, jj, ll;
209 index_type isec, jsec, lsec, uisec, ujsec, ulsec;
213 c = retarray->base_addr;
215 /* Parameter adjustments */
217 c_offset = 1 + c_dim1;
220 a_offset = 1 + a_dim1;
223 b_offset = 1 + b_dim1;
226 /* Early exit if possible */
227 if (m == 0 || n == 0 || k == 0)
233 c[i + j * c_dim1] = ('rtype_name`)0;
235 /* Start turning the crank. */
237 for (jj = 1; jj <= i1; jj += 512)
243 ujsec = jsec - jsec % 4;
245 for (ll = 1; ll <= i2; ll += 256)
251 ulsec = lsec - lsec % 2;
254 for (ii = 1; ii <= i3; ii += 256)
260 uisec = isec - isec % 2;
262 for (l = ll; l <= i4; l += 2)
265 for (i = ii; i <= i5; i += 2)
267 t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
269 t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
270 a[i + (l + 1) * a_dim1];
271 t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
272 a[i + 1 + l * a_dim1];
273 t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
274 a[i + 1 + (l + 1) * a_dim1];
278 t1[l - ll + 1 + (isec << 8) - 257] =
279 a[ii + isec - 1 + l * a_dim1];
280 t1[l - ll + 2 + (isec << 8) - 257] =
281 a[ii + isec - 1 + (l + 1) * a_dim1];
287 for (i = ii; i<= i4; ++i)
289 t1[lsec + ((i - ii + 1) << 8) - 257] =
290 a[i + (ll + lsec - 1) * a_dim1];
294 uisec = isec - isec % 4;
296 for (j = jj; j <= i4; j += 4)
299 for (i = ii; i <= i5; i += 4)
301 f11 = c[i + j * c_dim1];
302 f21 = c[i + 1 + j * c_dim1];
303 f12 = c[i + (j + 1) * c_dim1];
304 f22 = c[i + 1 + (j + 1) * c_dim1];
305 f13 = c[i + (j + 2) * c_dim1];
306 f23 = c[i + 1 + (j + 2) * c_dim1];
307 f14 = c[i + (j + 3) * c_dim1];
308 f24 = c[i + 1 + (j + 3) * c_dim1];
309 f31 = c[i + 2 + j * c_dim1];
310 f41 = c[i + 3 + j * c_dim1];
311 f32 = c[i + 2 + (j + 1) * c_dim1];
312 f42 = c[i + 3 + (j + 1) * c_dim1];
313 f33 = c[i + 2 + (j + 2) * c_dim1];
314 f43 = c[i + 3 + (j + 2) * c_dim1];
315 f34 = c[i + 2 + (j + 3) * c_dim1];
316 f44 = c[i + 3 + (j + 3) * c_dim1];
318 for (l = ll; l <= i6; ++l)
320 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
322 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
324 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
325 * b[l + (j + 1) * b_dim1];
326 f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
327 * b[l + (j + 1) * b_dim1];
328 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
329 * b[l + (j + 2) * b_dim1];
330 f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
331 * b[l + (j + 2) * b_dim1];
332 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
333 * b[l + (j + 3) * b_dim1];
334 f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
335 * b[l + (j + 3) * b_dim1];
336 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
338 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
340 f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
341 * b[l + (j + 1) * b_dim1];
342 f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
343 * b[l + (j + 1) * b_dim1];
344 f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
345 * b[l + (j + 2) * b_dim1];
346 f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
347 * b[l + (j + 2) * b_dim1];
348 f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
349 * b[l + (j + 3) * b_dim1];
350 f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
351 * b[l + (j + 3) * b_dim1];
353 c[i + j * c_dim1] = f11;
354 c[i + 1 + j * c_dim1] = f21;
355 c[i + (j + 1) * c_dim1] = f12;
356 c[i + 1 + (j + 1) * c_dim1] = f22;
357 c[i + (j + 2) * c_dim1] = f13;
358 c[i + 1 + (j + 2) * c_dim1] = f23;
359 c[i + (j + 3) * c_dim1] = f14;
360 c[i + 1 + (j + 3) * c_dim1] = f24;
361 c[i + 2 + j * c_dim1] = f31;
362 c[i + 3 + j * c_dim1] = f41;
363 c[i + 2 + (j + 1) * c_dim1] = f32;
364 c[i + 3 + (j + 1) * c_dim1] = f42;
365 c[i + 2 + (j + 2) * c_dim1] = f33;
366 c[i + 3 + (j + 2) * c_dim1] = f43;
367 c[i + 2 + (j + 3) * c_dim1] = f34;
368 c[i + 3 + (j + 3) * c_dim1] = f44;
373 for (i = ii + uisec; i <= i5; ++i)
375 f11 = c[i + j * c_dim1];
376 f12 = c[i + (j + 1) * c_dim1];
377 f13 = c[i + (j + 2) * c_dim1];
378 f14 = c[i + (j + 3) * c_dim1];
380 for (l = ll; l <= i6; ++l)
382 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
383 257] * b[l + j * b_dim1];
384 f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
385 257] * b[l + (j + 1) * b_dim1];
386 f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
387 257] * b[l + (j + 2) * b_dim1];
388 f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
389 257] * b[l + (j + 3) * b_dim1];
391 c[i + j * c_dim1] = f11;
392 c[i + (j + 1) * c_dim1] = f12;
393 c[i + (j + 2) * c_dim1] = f13;
394 c[i + (j + 3) * c_dim1] = f14;
401 for (j = jj + ujsec; j <= i4; ++j)
404 for (i = ii; i <= i5; i += 4)
406 f11 = c[i + j * c_dim1];
407 f21 = c[i + 1 + j * c_dim1];
408 f31 = c[i + 2 + j * c_dim1];
409 f41 = c[i + 3 + j * c_dim1];
411 for (l = ll; l <= i6; ++l)
413 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
414 257] * b[l + j * b_dim1];
415 f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
416 257] * b[l + j * b_dim1];
417 f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
418 257] * b[l + j * b_dim1];
419 f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
420 257] * b[l + j * b_dim1];
422 c[i + j * c_dim1] = f11;
423 c[i + 1 + j * c_dim1] = f21;
424 c[i + 2 + j * c_dim1] = f31;
425 c[i + 3 + j * c_dim1] = f41;
428 for (i = ii + uisec; i <= i5; ++i)
430 f11 = c[i + j * c_dim1];
432 for (l = ll; l <= i6; ++l)
434 f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
435 257] * b[l + j * b_dim1];
437 c[i + j * c_dim1] = f11;
446 else if (rxstride == 1 && aystride == 1 && bxstride == 1)
448 if (GFC_DESCRIPTOR_RANK (a) != 1)
450 const 'rtype_name` *restrict abase_x;
451 const 'rtype_name` *restrict bbase_y;
452 'rtype_name` *restrict dest_y;
455 for (y = 0; y < ycount; y++)
457 bbase_y = &bbase[y*bystride];
458 dest_y = &dest[y*rystride];
459 for (x = 0; x < xcount; x++)
461 abase_x = &abase[x*axstride];
462 s = ('rtype_name`) 0;
463 for (n = 0; n < count; n++)
464 s += abase_x[n] * bbase_y[n];
471 const 'rtype_name` *restrict bbase_y;
474 for (y = 0; y < ycount; y++)
476 bbase_y = &bbase[y*bystride];
477 s = ('rtype_name`) 0;
478 for (n = 0; n < count; n++)
479 s += abase[n*axstride] * bbase_y[n];
480 dest[y*rystride] = s;
484 else if (axstride < aystride)
486 for (y = 0; y < ycount; y++)
487 for (x = 0; x < xcount; x++)
488 dest[x*rxstride + y*rystride] = ('rtype_name`)0;
490 for (y = 0; y < ycount; y++)
491 for (n = 0; n < count; n++)
492 for (x = 0; x < xcount; x++)
493 /* dest[x,y] += a[x,n] * b[n,y] */
494 dest[x*rxstride + y*rystride] +=
495 abase[x*axstride + n*aystride] *
496 bbase[n*bxstride + y*bystride];
498 else if (GFC_DESCRIPTOR_RANK (a) == 1)
500 const 'rtype_name` *restrict bbase_y;
503 for (y = 0; y < ycount; y++)
505 bbase_y = &bbase[y*bystride];
506 s = ('rtype_name`) 0;
507 for (n = 0; n < count; n++)
508 s += abase[n*axstride] * bbase_y[n*bxstride];
509 dest[y*rxstride] = s;
514 const 'rtype_name` *restrict abase_x;
515 const 'rtype_name` *restrict bbase_y;
516 'rtype_name` *restrict dest_y;
519 for (y = 0; y < ycount; y++)
521 bbase_y = &bbase[y*bystride];
522 dest_y = &dest[y*rystride];
523 for (x = 0; x < xcount; x++)
525 abase_x = &abase[x*axstride];
526 s = ('rtype_name`) 0;
527 for (n = 0; n < count; n++)
528 s += abase_x[n*aystride] * bbase_y[n*bxstride];
529 dest_y[x*rxstride] = s;