libgfortran/m4/matmul_internal.m4

   1 `void
   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)
   5 {
   6   const 'rtype_name` * restrict abase;
   7   const 'rtype_name` * restrict bbase;
   8   'rtype_name` * restrict dest;
   9
  10   index_type rxstride, rystride, axstride, aystride, bxstride, bystride;
  11   index_type x, y, n, count, xcount, ycount;
  12
  13   assert (GFC_DESCRIPTOR_RANK (a) == 2
  14           || GFC_DESCRIPTOR_RANK (b) == 2);
  15
  16 /* C[xcount,ycount] = A[xcount, count] * B[count,ycount]
  17
  18    Either A or B (but not both) can be rank 1:
  19
  20    o One-dimensional argument A is implicitly treated as a row matrix
  21      dimensioned [1,count], so xcount=1.
  22
  23    o One-dimensional argument B is implicitly treated as a column matrix
  24      dimensioned [count, 1], so ycount=1.
  25 */
  26
  27   if (retarray->base_addr == NULL)
  28     {
  29       if (GFC_DESCRIPTOR_RANK (a) == 1)
  30         {
  31           GFC_DIMENSION_SET(retarray->dim[0], 0,
  32                             GFC_DESCRIPTOR_EXTENT(b,1) - 1, 1);
  33         }
  34       else if (GFC_DESCRIPTOR_RANK (b) == 1)
  35         {
  36           GFC_DIMENSION_SET(retarray->dim[0], 0,
  37                             GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
  38         }
  39       else
  40         {
  41           GFC_DIMENSION_SET(retarray->dim[0], 0,
  42                             GFC_DESCRIPTOR_EXTENT(a,0) - 1, 1);
  43
  44           GFC_DIMENSION_SET(retarray->dim[1], 0,
  45                             GFC_DESCRIPTOR_EXTENT(b,1) - 1,
  46                             GFC_DESCRIPTOR_EXTENT(retarray,0));
  47         }
  48
  49       retarray->base_addr
  50         = xmallocarray (size0 ((array_t *) retarray), sizeof ('rtype_name`));
  51       retarray->offset = 0;
  52     }
  53   else if (unlikely (compile_options.bounds_check))
  54     {
  55       index_type ret_extent, arg_extent;
  56
  57       if (GFC_DESCRIPTOR_RANK (a) == 1)
  58         {
  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 ("Array bound mismatch for dimension 1 of "
  63                            "array (%ld/%ld) ",
  64                            (long int) ret_extent, (long int) arg_extent);
  65         }
  66       else if (GFC_DESCRIPTOR_RANK (b) == 1)
  67         {
  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 ("Array bound mismatch for dimension 1 of "
  72                            "array (%ld/%ld) ",
  73                            (long int) ret_extent, (long int) arg_extent);
  74         }
  75       else
  76         {
  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 ("Array bound mismatch for dimension 1 of "
  81                            "array (%ld/%ld) ",
  82                            (long int) ret_extent, (long int) arg_extent);
  83
  84           arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
  85           ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
  86           if (arg_extent != ret_extent)
  87             runtime_error ("Array bound mismatch for dimension 2 of "
  88                            "array (%ld/%ld) ",
  89                            (long int) ret_extent, (long int) arg_extent);
  90         }
  91     }
  92 '
  93 sinclude(`matmul_asm_'rtype_code`.m4')dnl
  94 `
  95   if (GFC_DESCRIPTOR_RANK (retarray) == 1)
  96     {
  97       /* One-dimensional result may be addressed in the code below
  98          either as a row or a column matrix. We want both cases to
  99          work. */
 100       rxstride = rystride = GFC_DESCRIPTOR_STRIDE(retarray,0);
 101     }
 102   else
 103     {
 104       rxstride = GFC_DESCRIPTOR_STRIDE(retarray,0);
 105       rystride = GFC_DESCRIPTOR_STRIDE(retarray,1);
 106     }
 107
 108
 109   if (GFC_DESCRIPTOR_RANK (a) == 1)
 110     {
 111       /* Treat it as a a row matrix A[1,count]. */
 112       axstride = GFC_DESCRIPTOR_STRIDE(a,0);
 113       aystride = 1;
 114
 115       xcount = 1;
 116       count = GFC_DESCRIPTOR_EXTENT(a,0);
 117     }
 118   else
 119     {
 120       axstride = GFC_DESCRIPTOR_STRIDE(a,0);
 121       aystride = GFC_DESCRIPTOR_STRIDE(a,1);
 122
 123       count = GFC_DESCRIPTOR_EXTENT(a,1);
 124       xcount = GFC_DESCRIPTOR_EXTENT(a,0);
 125     }
 126
 127   if (count != GFC_DESCRIPTOR_EXTENT(b,0))
 128     {
 129       if (count > 0 || GFC_DESCRIPTOR_EXTENT(b,0) > 0)
 130         runtime_error ("Incorrect extent in argument B in MATMUL intrinsic "
 131                        "in dimension 1: is %ld, should be %ld",
 132                        (long int) GFC_DESCRIPTOR_EXTENT(b,0), (long int) count);
 133     }
 134
 135   if (GFC_DESCRIPTOR_RANK (b) == 1)
 136     {
 137       /* Treat it as a column matrix B[count,1] */
 138       bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
 139
 140       /* bystride should never be used for 1-dimensional b.
 141          The value is only used for calculation of the
 142          memory by the buffer.  */
 143       bystride = 256;
 144       ycount = 1;
 145     }
 146   else
 147     {
 148       bxstride = GFC_DESCRIPTOR_STRIDE(b,0);
 149       bystride = GFC_DESCRIPTOR_STRIDE(b,1);
 150       ycount = GFC_DESCRIPTOR_EXTENT(b,1);
 151     }
 152
 153   abase = a->base_addr;
 154   bbase = b->base_addr;
 155   dest = retarray->base_addr;
 156
 157   /* Now that everything is set up, we perform the multiplication
 158      itself.  */
 159
 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))
 163
 164   if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
 165       && (bxstride == 1 || bystride == 1)
 166       && (((float) xcount) * ((float) ycount) * ((float) count)
 167           > POW3(blas_limit)))
 168     {
 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;
 173
 174       if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
 175         {
 176           assert (gemm != NULL);
 177           const char *transa, *transb;
 178           if (try_blas & 2)
 179             transa = "C";
 180           else
 181             transa = axstride == 1 ? "N" : "T";
 182
 183           if (try_blas & 4)
 184             transb = "C";
 185           else
 186             transb = bxstride == 1 ? "N" : "T";
 187
 188           gemm (transa, transb , &m,
 189                 &n, &k, &one, abase, &lda, bbase, &ldb, &zero, dest,
 190                 &ldc, 1, 1);
 191           return;
 192         }
 193     }
 194
 195   if (rxstride == 1 && axstride == 1 && bxstride == 1)
 196     {
 197       /* This block of code implements a tuned matmul, derived from
 198          Superscalar GEMM-based level 3 BLAS,  Beta version 0.1
 199
 200                Bo Kagstrom and Per Ling
 201                Department of Computing Science
 202                Umea University
 203                S-901 87 Umea, Sweden
 204
 205          from netlib.org, translated to C, and modified for matmul.m4.  */
 206
 207       const 'rtype_name` *a, *b;
 208       'rtype_name` *c;
 209       const index_type m = xcount, n = ycount, k = count;
 210
 211       /* System generated locals */
 212       index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
 213                  i1, i2, i3, i4, i5, i6;
 214
 215       /* Local variables */
 216       'rtype_name` f11, f12, f21, f22, f31, f32, f41, f42,
 217                  f13, f14, f23, f24, f33, f34, f43, f44;
 218       index_type i, j, l, ii, jj, ll;
 219       index_type isec, jsec, lsec, uisec, ujsec, ulsec;
 220       'rtype_name` *t1;
 221
 222       a = abase;
 223       b = bbase;
 224       c = retarray->base_addr;
 225
 226       /* Parameter adjustments */
 227       c_dim1 = rystride;
 228       c_offset = 1 + c_dim1;
 229       c -= c_offset;
 230       a_dim1 = aystride;
 231       a_offset = 1 + a_dim1;
 232       a -= a_offset;
 233       b_dim1 = bystride;
 234       b_offset = 1 + b_dim1;
 235       b -= b_offset;
 236
 237       /* Empty c first.  */
 238       for (j=1; j<=n; j++)
 239         for (i=1; i<=m; i++)
 240           c[i + j * c_dim1] = ('rtype_name`)0;
 241
 242       /* Early exit if possible */
 243       if (m == 0 || n == 0 || k == 0)
 244         return;
 245
 246       /* Adjust size of t1 to what is needed.  */
 247       index_type t1_dim, a_sz;
 248       if (aystride == 1)
 249         a_sz = rystride;
 250       else
 251         a_sz = a_dim1;
 252
 253       t1_dim = a_sz * 256 + b_dim1;
 254       if (t1_dim > 65536)
 255         t1_dim = 65536;
 256
 257       t1 = malloc (t1_dim * sizeof('rtype_name`));
 258
 259       /* Start turning the crank. */
 260       i1 = n;
 261       for (jj = 1; jj <= i1; jj += 512)
 262         {
 263           /* Computing MIN */
 264           i2 = 512;
 265           i3 = n - jj + 1;
 266           jsec = min(i2,i3);
 267           ujsec = jsec - jsec % 4;
 268           i2 = k;
 269           for (ll = 1; ll <= i2; ll += 256)
 270             {
 271               /* Computing MIN */
 272               i3 = 256;
 273               i4 = k - ll + 1;
 274               lsec = min(i3,i4);
 275               ulsec = lsec - lsec % 2;
 276
 277               i3 = m;
 278               for (ii = 1; ii <= i3; ii += 256)
 279                 {
 280                   /* Computing MIN */
 281                   i4 = 256;
 282                   i5 = m - ii + 1;
 283                   isec = min(i4,i5);
 284                   uisec = isec - isec % 2;
 285                   i4 = ll + ulsec - 1;
 286                   for (l = ll; l <= i4; l += 2)
 287                     {
 288                       i5 = ii + uisec - 1;
 289                       for (i = ii; i <= i5; i += 2)
 290                         {
 291                           t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
 292                                         a[i + l * a_dim1];
 293                           t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
 294                                         a[i + (l + 1) * a_dim1];
 295                           t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
 296                                         a[i + 1 + l * a_dim1];
 297                           t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
 298                                         a[i + 1 + (l + 1) * a_dim1];
 299                         }
 300                       if (uisec < isec)
 301                         {
 302                           t1[l - ll + 1 + (isec << 8) - 257] =
 303                                     a[ii + isec - 1 + l * a_dim1];
 304                           t1[l - ll + 2 + (isec << 8) - 257] =
 305                                     a[ii + isec - 1 + (l + 1) * a_dim1];
 306                         }
 307                     }
 308                   if (ulsec < lsec)
 309                     {
 310                       i4 = ii + isec - 1;
 311                       for (i = ii; i<= i4; ++i)
 312                         {
 313                           t1[lsec + ((i - ii + 1) << 8) - 257] =
 314                                     a[i + (ll + lsec - 1) * a_dim1];
 315                         }
 316                     }
 317
 318                   uisec = isec - isec % 4;
 319                   i4 = jj + ujsec - 1;
 320                   for (j = jj; j <= i4; j += 4)
 321                     {
 322                       i5 = ii + uisec - 1;
 323                       for (i = ii; i <= i5; i += 4)
 324                         {
 325                           f11 = c[i + j * c_dim1];
 326                           f21 = c[i + 1 + j * c_dim1];
 327                           f12 = c[i + (j + 1) * c_dim1];
 328                           f22 = c[i + 1 + (j + 1) * c_dim1];
 329                           f13 = c[i + (j + 2) * c_dim1];
 330                           f23 = c[i + 1 + (j + 2) * c_dim1];
 331                           f14 = c[i + (j + 3) * c_dim1];
 332                           f24 = c[i + 1 + (j + 3) * c_dim1];
 333                           f31 = c[i + 2 + j * c_dim1];
 334                           f41 = c[i + 3 + j * c_dim1];
 335                           f32 = c[i + 2 + (j + 1) * c_dim1];
 336                           f42 = c[i + 3 + (j + 1) * c_dim1];
 337                           f33 = c[i + 2 + (j + 2) * c_dim1];
 338                           f43 = c[i + 3 + (j + 2) * c_dim1];
 339                           f34 = c[i + 2 + (j + 3) * c_dim1];
 340                           f44 = c[i + 3 + (j + 3) * c_dim1];
 341                           i6 = ll + lsec - 1;
 342                           for (l = ll; l <= i6; ++l)
 343                             {
 344                               f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
 345                                       * b[l + j * b_dim1];
 346                               f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
 347                                       * b[l + j * b_dim1];
 348                               f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
 349                                       * b[l + (j + 1) * b_dim1];
 350                               f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
 351                                       * b[l + (j + 1) * b_dim1];
 352                               f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
 353                                       * b[l + (j + 2) * b_dim1];
 354                               f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
 355                                       * b[l + (j + 2) * b_dim1];
 356                               f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
 357                                       * b[l + (j + 3) * b_dim1];
 358                               f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
 359                                       * b[l + (j + 3) * b_dim1];
 360                               f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
 361                                       * b[l + j * b_dim1];
 362                               f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
 363                                       * b[l + j * b_dim1];
 364                               f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
 365                                       * b[l + (j + 1) * b_dim1];
 366                               f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
 367                                       * b[l + (j + 1) * b_dim1];
 368                               f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
 369                                       * b[l + (j + 2) * b_dim1];
 370                               f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
 371                                       * b[l + (j + 2) * b_dim1];
 372                               f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
 373                                       * b[l + (j + 3) * b_dim1];
 374                               f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
 375                                       * b[l + (j + 3) * b_dim1];
 376                             }
 377                           c[i + j * c_dim1] = f11;
 378                           c[i + 1 + j * c_dim1] = f21;
 379                           c[i + (j + 1) * c_dim1] = f12;
 380                           c[i + 1 + (j + 1) * c_dim1] = f22;
 381                           c[i + (j + 2) * c_dim1] = f13;
 382                           c[i + 1 + (j + 2) * c_dim1] = f23;
 383                           c[i + (j + 3) * c_dim1] = f14;
 384                           c[i + 1 + (j + 3) * c_dim1] = f24;
 385                           c[i + 2 + j * c_dim1] = f31;
 386                           c[i + 3 + j * c_dim1] = f41;
 387                           c[i + 2 + (j + 1) * c_dim1] = f32;
 388                           c[i + 3 + (j + 1) * c_dim1] = f42;
 389                           c[i + 2 + (j + 2) * c_dim1] = f33;
 390                           c[i + 3 + (j + 2) * c_dim1] = f43;
 391                           c[i + 2 + (j + 3) * c_dim1] = f34;
 392                           c[i + 3 + (j + 3) * c_dim1] = f44;
 393                         }
 394                       if (uisec < isec)
 395                         {
 396                           i5 = ii + isec - 1;
 397                           for (i = ii + uisec; i <= i5; ++i)
 398                             {
 399                               f11 = c[i + j * c_dim1];
 400                               f12 = c[i + (j + 1) * c_dim1];
 401                               f13 = c[i + (j + 2) * c_dim1];
 402                               f14 = c[i + (j + 3) * c_dim1];
 403                               i6 = ll + lsec - 1;
 404                               for (l = ll; l <= i6; ++l)
 405                                 {
 406                                   f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
 407                                           257] * b[l + j * b_dim1];
 408                                   f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
 409                                           257] * b[l + (j + 1) * b_dim1];
 410                                   f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
 411                                           257] * b[l + (j + 2) * b_dim1];
 412                                   f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
 413                                           257] * b[l + (j + 3) * b_dim1];
 414                                 }
 415                               c[i + j * c_dim1] = f11;
 416                               c[i + (j + 1) * c_dim1] = f12;
 417                               c[i + (j + 2) * c_dim1] = f13;
 418                               c[i + (j + 3) * c_dim1] = f14;
 419                             }
 420                         }
 421                     }
 422                   if (ujsec < jsec)
 423                     {
 424                       i4 = jj + jsec - 1;
 425                       for (j = jj + ujsec; j <= i4; ++j)
 426                         {
 427                           i5 = ii + uisec - 1;
 428                           for (i = ii; i <= i5; i += 4)
 429                             {
 430                               f11 = c[i + j * c_dim1];
 431                               f21 = c[i + 1 + j * c_dim1];
 432                               f31 = c[i + 2 + j * c_dim1];
 433                               f41 = c[i + 3 + j * c_dim1];
 434                               i6 = ll + lsec - 1;
 435                               for (l = ll; l <= i6; ++l)
 436                                 {
 437                                   f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
 438                                           257] * b[l + j * b_dim1];
 439                                   f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
 440                                           257] * b[l + j * b_dim1];
 441                                   f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
 442                                           257] * b[l + j * b_dim1];
 443                                   f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
 444                                           257] * b[l + j * b_dim1];
 445                                 }
 446                               c[i + j * c_dim1] = f11;
 447                               c[i + 1 + j * c_dim1] = f21;
 448                               c[i + 2 + j * c_dim1] = f31;
 449                               c[i + 3 + j * c_dim1] = f41;
 450                             }
 451                           i5 = ii + isec - 1;
 452                           for (i = ii + uisec; i <= i5; ++i)
 453                             {
 454                               f11 = c[i + j * c_dim1];
 455                               i6 = ll + lsec - 1;
 456                               for (l = ll; l <= i6; ++l)
 457                                 {
 458                                   f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
 459                                           257] * b[l + j * b_dim1];
 460                                 }
 461                               c[i + j * c_dim1] = f11;
 462                             }
 463                         }
 464                     }
 465                 }
 466             }
 467         }
 468       free(t1);
 469       return;
 470     }
 471   else if (rxstride == 1 && aystride == 1 && bxstride == 1)
 472     {
 473       if (GFC_DESCRIPTOR_RANK (a) != 1)
 474         {
 475           const 'rtype_name` *restrict abase_x;
 476           const 'rtype_name` *restrict bbase_y;
 477           'rtype_name` *restrict dest_y;
 478           'rtype_name` s;
 479
 480           for (y = 0; y < ycount; y++)
 481             {
 482               bbase_y = &bbase[y*bystride];
 483               dest_y = &dest[y*rystride];
 484               for (x = 0; x < xcount; x++)
 485                 {
 486                   abase_x = &abase[x*axstride];
 487                   s = ('rtype_name`) 0;
 488                   for (n = 0; n < count; n++)
 489                     s += abase_x[n] * bbase_y[n];
 490                   dest_y[x] = s;
 491                 }
 492             }
 493         }
 494       else
 495         {
 496           const 'rtype_name` *restrict bbase_y;
 497           'rtype_name` s;
 498
 499           for (y = 0; y < ycount; y++)
 500             {
 501               bbase_y = &bbase[y*bystride];
 502               s = ('rtype_name`) 0;
 503               for (n = 0; n < count; n++)
 504                 s += abase[n*axstride] * bbase_y[n];
 505               dest[y*rystride] = s;
 506             }
 507         }
 508     }
 509   else if (axstride < aystride)
 510     {
 511       for (y = 0; y < ycount; y++)
 512         for (x = 0; x < xcount; x++)
 513           dest[x*rxstride + y*rystride] = ('rtype_name`)0;
 514
 515       for (y = 0; y < ycount; y++)
 516         for (n = 0; n < count; n++)
 517           for (x = 0; x < xcount; x++)
 518             /* dest[x,y] += a[x,n] * b[n,y] */
 519             dest[x*rxstride + y*rystride] +=
 520                                         abase[x*axstride + n*aystride] *
 521                                         bbase[n*bxstride + y*bystride];
 522     }
 523   else if (GFC_DESCRIPTOR_RANK (a) == 1)
 524     {
 525       const 'rtype_name` *restrict bbase_y;
 526       'rtype_name` s;
 527
 528       for (y = 0; y < ycount; y++)
 529         {
 530           bbase_y = &bbase[y*bystride];
 531           s = ('rtype_name`) 0;
 532           for (n = 0; n < count; n++)
 533             s += abase[n*axstride] * bbase_y[n*bxstride];
 534           dest[y*rxstride] = s;
 535         }
 536     }
 537   else
 538     {
 539       const 'rtype_name` *restrict abase_x;
 540       const 'rtype_name` *restrict bbase_y;
 541       'rtype_name` *restrict dest_y;
 542       'rtype_name` s;
 543
 544       for (y = 0; y < ycount; y++)
 545         {
 546           bbase_y = &bbase[y*bystride];
 547           dest_y = &dest[y*rystride];
 548           for (x = 0; x < xcount; x++)
 549             {
 550               abase_x = &abase[x*axstride];
 551               s = ('rtype_name`) 0;
 552               for (n = 0; n < count; n++)
 553                 s += abase_x[n*aystride] * bbase_y[n*bxstride];
 554               dest_y[x*rxstride] = s;
 555             }
 556         }
 557     }
 558 }
 559 #undef POW3
 560 #undef min
 561 #undef max
 562 '