OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: ce3f668c99
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'ce3f668c99'
${ noResults }
OpenBLAS/kernel/mips/dgemm_kernel_8x4_msa.c

1739 lines
49 kB
Raw Blame History 

  1. /*******************************************************************************

  2. Copyright (c) 2016, The OpenBLAS Project

  3. All rights reserved.

  4. Redistribution and use in source and binary forms, with or without

  5. modification, are permitted provided that the following conditions are

  6. met:

  7. 1. Redistributions of source code must retain the above copyright

  8. notice, this list of conditions and the following disclaimer.

  9. 2. Redistributions in binary form must reproduce the above copyright

  10. notice, this list of conditions and the following disclaimer in

  11. the documentation and/or other materials provided with the

  12. distribution.

  13. 3. Neither the name of the OpenBLAS project nor the names of

  14. its contributors may be used to endorse or promote products

  15. derived from this software without specific prior written permission.

  16. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"

  17. AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE

  18. IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE

  19. ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE

  20. LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL

  21. DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR

  22. SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER

  23. CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,

  24. OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE

  25. USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

  26. *******************************************************************************/

  27. 

  28. #include "common.h"

  29. #include "macros_msa.h"

  30. 

  31. static void __attribute__ ((noinline))

  32. dgemmkernel_8x4_core_msa(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha,

  33.                          FLOAT *A, FLOAT *B, FLOAT *C, BLASLONG ldc

  34. #ifdef TRMMKERNEL

  35.           , BLASLONG offset

  36. #endif

  37.                          )

  38. {

  39.     BLASLONG i, j, l, temp;

  40. #if defined(TRMMKERNEL)

  41.     BLASLONG off;

  42. #endif

  43.     FLOAT *pc0, *pc1, *pc2, *pc3, *pa0, *pb0;

  44.     v2f64 v_alpha = {alpha, alpha};

  45.     v2f64 src_a0, src_a1, src_a2, src_a3, src_b, src_b0, src_b1;

  46.     v2f64 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;

  47.     v2f64 res0, res1, res2, res3, res4, res5, res6, res7;

  48.     v2f64 res8, res9, res10, res11, res12, res13, res14, res15;

  49. 

  50. #if defined(TRMMKERNEL) && !defined(LEFT)

  51.     off = -offset;

  52. #endif

  53. 

  54.     for (j = (n >> 2); j--;)

  55.     {

  56.         pc0 = C;

  57.         pc1 = pc0 + ldc;

  58.         pc2 = pc1 + ldc;

  59.         pc3 = pc2 + ldc;

  60. 

  61. #if defined(TRMMKERNEL) && defined(LEFT)

  62.         off = offset;

  63. #endif

  64. 

  65.         pa0 = A;

  66. 

  67.         for (i = (m >> 3); i--;)

  68.         {

  69. #if defined(TRMMKERNEL)

  70. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  71.             pb0 = B;

  72. #else

  73.             pa0 += off * 8;

  74.             pb0 = B + off * 4;

  75. #endif

  76. 

  77. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  78.             temp = k - off;

  79. #elif defined(LEFT)

  80.             temp = off + 8; // number of values in A

  81. #else

  82.             temp = off + 4; // number of values in B

  83. #endif

  84. #else  // #if !defined(TRMMKERNEL)

  85.             pb0 = B;

  86.             temp = k;

  87. #endif

  88. 

  89. #ifdef ENABLE_PREFETCH

  90.             __asm__ __volatile__(

  91.                 "pref   0,   64(%[pa0])   \n\t"

  92.                 "pref   0,   96(%[pa0])   \n\t"

  93.                 "pref   0,  128(%[pa0])   \n\t"

  94.                 "pref   0,  160(%[pa0])   \n\t"

  95.                 "pref   0,   32(%[pb0])   \n\t"

  96.                 "pref   0,   64(%[pb0])   \n\t"

  97.                 "pref   0,   96(%[pb0])   \n\t"

  98. 

  99.                 :

  100.                 : [pa0] "r" (pa0), [pb0] "r" (pb0)

  101.             );

  102. #endif

  103. 

  104.             LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  105.             LD_DP2_INC(pb0, 2, src_b0, src_b1);

  106. 

  107.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  108.             res0 = src_a0 * src_b;

  109.             res1 = src_a1 * src_b;

  110.             res2 = src_a2 * src_b;

  111.             res3 = src_a3 * src_b;

  112. 

  113.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  114.             res4 = src_a0 * src_b;

  115.             res5 = src_a1 * src_b;

  116.             res6 = src_a2 * src_b;

  117.             res7 = src_a3 * src_b;

  118. 

  119.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  120.             res8  = src_a0 * src_b;

  121.             res9  = src_a1 * src_b;

  122.             res10 = src_a2 * src_b;

  123.             res11 = src_a3 * src_b;

  124. 

  125.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  126.             res12 = src_a0 * src_b;

  127.             res13 = src_a1 * src_b;

  128.             res14 = src_a2 * src_b;

  129.             res15 = src_a3 * src_b;

  130. 

  131.             for (l = ((temp - 1) >> 1); l--;)

  132.             {

  133. #ifdef ENABLE_PREFETCH

  134.                 __asm__ __volatile__(

  135.                     "pref   0,  128(%[pa0])   \n\t"

  136.                     "pref   0,  160(%[pa0])   \n\t"

  137.                     "pref   0,  192(%[pa0])   \n\t"

  138.                     "pref   0,  224(%[pa0])   \n\t"

  139. 

  140.                     :

  141.                     : [pa0] "r" (pa0)

  142.                 );

  143. #endif

  144.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  145.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  146. 

  147.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  148.                 res0 += src_a0 * src_b;

  149.                 res1 += src_a1 * src_b;

  150.                 res2 += src_a2 * src_b;

  151.                 res3 += src_a3 * src_b;

  152. 

  153.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  154.                 res4 += src_a0 * src_b;

  155.                 res5 += src_a1 * src_b;

  156.                 res6 += src_a2 * src_b;

  157.                 res7 += src_a3 * src_b;

  158. 

  159.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  160.                 res8  += src_a0 * src_b;

  161.                 res9  += src_a1 * src_b;

  162.                 res10 += src_a2 * src_b;

  163.                 res11 += src_a3 * src_b;

  164. 

  165.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  166.                 res12 += src_a0 * src_b;

  167.                 res13 += src_a1 * src_b;

  168.                 res14 += src_a2 * src_b;

  169.                 res15 += src_a3 * src_b;

  170. 

  171.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  172. #ifdef ENABLE_PREFETCH

  173.                 __asm__ __volatile__(

  174.                     "pref   0,   64(%[pb0])   \n\t"

  175.                     "pref   0,   96(%[pb0])   \n\t"

  176. 

  177.                     :

  178.                     : [pb0] "r" (pb0)

  179.                 );

  180. #endif

  181.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  182. 

  183.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  184.                 res0 += src_a0 * src_b;

  185.                 res1 += src_a1 * src_b;

  186.                 res2 += src_a2 * src_b;

  187.                 res3 += src_a3 * src_b;

  188. 

  189.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  190.                 res4 += src_a0 * src_b;

  191.                 res5 += src_a1 * src_b;

  192.                 res6 += src_a2 * src_b;

  193.                 res7 += src_a3 * src_b;

  194. 

  195.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  196.                 res8  += src_a0 * src_b;

  197.                 res9  += src_a1 * src_b;

  198.                 res10 += src_a2 * src_b;

  199.                 res11 += src_a3 * src_b;

  200. 

  201.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  202.                 res12 += src_a0 * src_b;

  203.                 res13 += src_a1 * src_b;

  204.                 res14 += src_a2 * src_b;

  205.                 res15 += src_a3 * src_b;

  206.             }

  207. 

  208.             if ((temp - 1) & 1)

  209.             {

  210.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  211.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  212. 

  213.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  214.                 res0 += src_a0 * src_b;

  215.                 res1 += src_a1 * src_b;

  216.                 res2 += src_a2 * src_b;

  217.                 res3 += src_a3 * src_b;

  218. 

  219.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  220.                 res4 += src_a0 * src_b;

  221.                 res5 += src_a1 * src_b;

  222.                 res6 += src_a2 * src_b;

  223.                 res7 += src_a3 * src_b;

  224. 

  225.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  226.                 res8  += src_a0 * src_b;

  227.                 res9  += src_a1 * src_b;

  228.                 res10 += src_a2 * src_b;

  229.                 res11 += src_a3 * src_b;

  230. 

  231.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  232.                 res12 += src_a0 * src_b;

  233.                 res13 += src_a1 * src_b;

  234.                 res14 += src_a2 * src_b;

  235.                 res15 += src_a3 * src_b;

  236.             }

  237. 

  238. #ifdef ENABLE_PREFETCH

  239.             __asm__ __volatile__(

  240.                 "pref   0,  64(%[pc0])   \n\t"

  241.                 "pref   0,  64(%[pc1])   \n\t"

  242.                 "pref   0,  64(%[pc2])   \n\t"

  243.                 "pref   0,  64(%[pc3])   \n\t"

  244. 

  245.                 :

  246.                 : [pc0] "r" (pc0), [pc1] "r" (pc1),

  247.                   [pc2] "r" (pc2), [pc3] "r" (pc3)

  248.             );

  249. #endif

  250. 

  251. #if defined(TRMMKERNEL)

  252.             dst0 = res0 * v_alpha;

  253.             dst1 = res1 * v_alpha;

  254.             dst2 = res2 * v_alpha;

  255.             dst3 = res3 * v_alpha;

  256.             dst4 = res4 * v_alpha;

  257.             dst5 = res5 * v_alpha;

  258.             dst6 = res6 * v_alpha;

  259.             dst7 = res7 * v_alpha;

  260. #else

  261.             LD_DP4(pc0, 2, dst0, dst1, dst2, dst3);

  262.             LD_DP4(pc1, 2, dst4, dst5, dst6, dst7);

  263. 

  264.             dst0 += res0 * v_alpha;

  265.             dst1 += res1 * v_alpha;

  266.             dst2 += res2 * v_alpha;

  267.             dst3 += res3 * v_alpha;

  268.             dst4 += res4 * v_alpha;

  269.             dst5 += res5 * v_alpha;

  270.             dst6 += res6 * v_alpha;

  271.             dst7 += res7 * v_alpha;

  272. #endif

  273.             ST_DP4_INC(dst0, dst1, dst2, dst3, pc0, 2);

  274.             ST_DP4_INC(dst4, dst5, dst6, dst7, pc1, 2);

  275. 

  276. #if defined(TRMMKERNEL)

  277.             dst0 = res8 * v_alpha;

  278.             dst1 = res9 * v_alpha;

  279.             dst2 = res10 * v_alpha;

  280.             dst3 = res11 * v_alpha;

  281.             dst4 = res12 * v_alpha;

  282.             dst5 = res13 * v_alpha;

  283.             dst6 = res14 * v_alpha;

  284.             dst7 = res15 * v_alpha;

  285. #else

  286.             LD_DP4(pc2, 2, dst0, dst1, dst2, dst3);

  287.             LD_DP4(pc3, 2, dst4, dst5, dst6, dst7);

  288. 

  289.             dst0 += res8 * v_alpha;

  290.             dst1 += res9 * v_alpha;

  291.             dst2 += res10 * v_alpha;

  292.             dst3 += res11 * v_alpha;

  293.             dst4 += res12 * v_alpha;

  294.             dst5 += res13 * v_alpha;

  295.             dst6 += res14 * v_alpha;

  296.             dst7 += res15 * v_alpha;

  297. #endif

  298.             ST_DP4_INC(dst0, dst1, dst2, dst3, pc2, 2);

  299.             ST_DP4_INC(dst4, dst5, dst6, dst7, pc3, 2);

  300. 

  301. #if defined(TRMMKERNEL)

  302. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  303.             temp = k - off;

  304. #ifdef LEFT

  305.             temp -= 8; // number of values in A

  306. #else

  307.             temp -= 4; // number of values in B

  308. #endif

  309.             pa0 += temp * 8;

  310.             pb0 += temp * 4;

  311. #endif

  312. 

  313. #ifdef LEFT

  314.             off += 8; // number of values in A

  315. #endif

  316. #endif  // #if defined(TRMMKERNEL)

  317.         }

  318. 

  319. #if defined(TRMMKERNEL) && !defined(LEFT)

  320.         off += 4; // number of values in A

  321. #endif

  322. 

  323.         B += (k << 2);

  324.         C += (ldc << 2);

  325.     }

  326. }

  327. 

  328. static void __attribute__ ((noinline))

  329. dgemmkernel_7x4_core_msa(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha,

  330.                          FLOAT *A, FLOAT *B, FLOAT *C, BLASLONG ldc

  331. #ifdef TRMMKERNEL

  332.           , BLASLONG offset

  333. #endif

  334.                          )

  335. {

  336.     BLASLONG j, l, temp;

  337. #if defined(TRMMKERNEL)

  338.     BLASLONG off;

  339. #endif

  340.     FLOAT *pc0, *pc1, *pc2, *pc3, *pa0, *pb0;

  341.     FLOAT tmp0, tmp1, tmp2, tmp3;

  342.     FLOAT a0, b0, b1, b2, b3;

  343.     v2f64 v_alpha = {alpha, alpha};

  344.     v2f64 src_a0, src_a1, src_b, src_b0, src_b1;

  345.     v2f64 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;

  346.     v2f64 res0, res1, res2, res3, res4, res5, res6, res7;

  347. 

  348. #if defined(TRMMKERNEL) && !defined(LEFT)

  349.     off = -offset;

  350. #endif

  351. 

  352.     for (j = (n >> 2); j--;)

  353.     {

  354. #if defined(TRMMKERNEL)

  355.         pc0 = C;

  356.         pc1 = pc0 + ldc;

  357.         pc2 = pc1 + ldc;

  358.         pc3 = pc2 + ldc;

  359. 

  360.         pa0 = A;

  361. 

  362. #if defined(LEFT)

  363.         off = offset;

  364. #endif

  365. 

  366.         for (l = (m >> 3); l--;)

  367.         {

  368. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  369.             pb0 = B;

  370. #else

  371.             pa0 += off * 8;

  372.             pb0 = B + off * 4;

  373. #endif

  374. 

  375. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  376.             temp = k - off;

  377. #elif defined(LEFT)

  378.             temp = off + 8; // number of values in A

  379. #else

  380.             temp = off + 4; // number of values in B

  381. #endif

  382. 

  383.             pc0 += 8;

  384.             pc1 += 8;

  385.             pc2 += 8;

  386.             pc3 += 8;

  387.             pa0 += 8 * temp;

  388.             pb0 += 4 * temp;

  389. 

  390. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  391.             temp = k - off;

  392. #ifdef LEFT

  393.             temp -= 8; // number of values in A

  394. #else

  395.             temp -= 4; // number of values in B

  396. #endif

  397.             pa0 += temp * 8;

  398.             pb0 += temp * 4;

  399. #endif

  400. 

  401. #ifdef LEFT

  402.             off += 8; // number of values in A

  403. #endif

  404.         }

  405. #else  // #if !defined(TRMMKERNEL)

  406.         pc0 = C + 8 * (m >> 3);

  407.         pc1 = pc0 + ldc;

  408.         pc2 = pc1 + ldc;

  409.         pc3 = pc2 + ldc;

  410. 

  411.         pa0 = A + k * 8 * (m >> 3);

  412. #endif

  413. 

  414.         if (m & 4)

  415.         {

  416. #if defined(TRMMKERNEL)

  417. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  418.             pb0 = B;

  419. #else

  420.             pa0 += off * 4;

  421.             pb0 = B + off * 4;

  422. #endif

  423. 

  424. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  425.             temp = k - off;

  426. #elif defined(LEFT)

  427.             temp = off + 4; // number of values in A

  428. #else

  429.             temp = off + 4; // number of values in B

  430. #endif

  431. #else  // #if !defined(TRMMKERNEL)

  432.             pb0 = B;

  433.             temp = k;

  434. #endif

  435. 

  436.             LD_DP2_INC(pa0, 2, src_a0, src_a1);

  437.             LD_DP2_INC(pb0, 2, src_b0, src_b1);

  438. 

  439.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  440.             res0 = src_a0 * src_b;

  441.             res1 = src_a1 * src_b;

  442. 

  443.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  444.             res2 = src_a0 * src_b;

  445.             res3 = src_a1 * src_b;

  446. 

  447.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  448.             res4 = src_a0 * src_b;

  449.             res5 = src_a1 * src_b;

  450. 

  451.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  452.             res6 = src_a0 * src_b;

  453.             res7 = src_a1 * src_b;

  454. 

  455.             for (l = ((temp - 1) >> 1); l--;)

  456.             {

  457.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  458.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  459. 

  460.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  461.                 res0 += src_a0 * src_b;

  462.                 res1 += src_a1 * src_b;

  463. 

  464.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  465.                 res2 += src_a0 * src_b;

  466.                 res3 += src_a1 * src_b;

  467. 

  468.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  469.                 res4 += src_a0 * src_b;

  470.                 res5 += src_a1 * src_b;

  471. 

  472.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  473.                 res6 += src_a0 * src_b;

  474.                 res7 += src_a1 * src_b;

  475. 

  476.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  477.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  478. 

  479.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  480.                 res0 += src_a0 * src_b;

  481.                 res1 += src_a1 * src_b;

  482. 

  483.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  484.                 res2 += src_a0 * src_b;

  485.                 res3 += src_a1 * src_b;

  486. 

  487.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  488.                 res4 += src_a0 * src_b;

  489.                 res5 += src_a1 * src_b;

  490. 

  491.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  492.                 res6 += src_a0 * src_b;

  493.                 res7 += src_a1 * src_b;

  494.             }

  495. 

  496.             if ((temp - 1) & 1)

  497.             {

  498.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  499.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  500. 

  501.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  502.                 res0 += src_a0 * src_b;

  503.                 res1 += src_a1 * src_b;

  504. 

  505.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  506.                 res2 += src_a0 * src_b;

  507.                 res3 += src_a1 * src_b;

  508. 

  509.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  510.                 res4 += src_a0 * src_b;

  511.                 res5 += src_a1 * src_b;

  512. 

  513.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  514.                 res6 += src_a0 * src_b;

  515.                 res7 += src_a1 * src_b;

  516.             }

  517. 

  518. #if defined(TRMMKERNEL)

  519.             dst0 = res0 * v_alpha;

  520.             dst1 = res1 * v_alpha;

  521.             dst2 = res2 * v_alpha;

  522.             dst3 = res3 * v_alpha;

  523.             dst4 = res4 * v_alpha;

  524.             dst5 = res5 * v_alpha;

  525.             dst6 = res6 * v_alpha;

  526.             dst7 = res7 * v_alpha;

  527. #else

  528.             LD_DP2(pc0, 2, dst0, dst1);

  529.             LD_DP2(pc1, 2, dst2, dst3);

  530.             LD_DP2(pc2, 2, dst4, dst5);

  531.             LD_DP2(pc3, 2, dst6, dst7);

  532. 

  533.             dst0 += res0 * v_alpha;

  534.             dst1 += res1 * v_alpha;

  535.             dst2 += res2 * v_alpha;

  536.             dst3 += res3 * v_alpha;

  537.             dst4 += res4 * v_alpha;

  538.             dst5 += res5 * v_alpha;

  539.             dst6 += res6 * v_alpha;

  540.             dst7 += res7 * v_alpha;

  541. #endif

  542.             ST_DP2_INC(dst0, dst1, pc0, 2);

  543.             ST_DP2_INC(dst2, dst3, pc1, 2);

  544.             ST_DP2_INC(dst4, dst5, pc2, 2);

  545.             ST_DP2_INC(dst6, dst7, pc3, 2);

  546. 

  547. #if defined(TRMMKERNEL)

  548. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  549.             temp = k - off;

  550. #ifdef LEFT

  551.             temp -= 4; // number of values in A

  552. #else

  553.             temp -= 4; // number of values in B

  554. #endif

  555.             pa0 += temp * 4;

  556.             pb0 += temp * 4;

  557. #endif

  558. 

  559. #ifdef LEFT

  560.             off += 4; // number of values in A

  561. #endif

  562. #endif  // #if defined(TRMMKERNEL)

  563.         }

  564. 

  565.         if (m & 2)

  566.         {

  567. #if defined(TRMMKERNEL)

  568. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  569.             pb0 = B;

  570. #else

  571.             pa0 += off * 2;

  572.             pb0 = B + off * 4;

  573. #endif

  574. 

  575. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  576.             temp = k - off;

  577. #elif defined(LEFT)

  578.             temp = off + 2; // number of values in A

  579. #else

  580.             temp = off + 4; // number of values in B

  581. #endif

  582. #else  // #if !defined(TRMMKERNEL)

  583.             pb0 = B;

  584.             temp = k;

  585. #endif

  586. 

  587.             src_a0 = LD_DP(pa0);

  588.             pa0 += 2;

  589.             LD_DP2_INC(pb0, 2, src_b0, src_b1);

  590. 

  591.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  592.             res0 = src_a0 * src_b;

  593. 

  594.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  595.             res1 = src_a0 * src_b;

  596. 

  597.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  598.             res2 = src_a0 * src_b;

  599. 

  600.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  601.             res3 = src_a0 * src_b;

  602. 

  603.             for (l = ((temp - 1) >> 1); l--;)

  604.             {

  605.                 src_a0 = LD_DP(pa0);

  606.                 pa0 += 2;

  607.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  608. 

  609.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  610.                 res0 += src_a0 * src_b;

  611. 

  612.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  613.                 res1 += src_a0 * src_b;

  614. 

  615.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  616.                 res2 += src_a0 * src_b;

  617. 

  618.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  619.                 res3 += src_a0 * src_b;

  620. 

  621.                 src_a0 = LD_DP(pa0);

  622.                 pa0 += 2;

  623.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  624. 

  625.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  626.                 res0 += src_a0 * src_b;

  627. 

  628.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  629.                 res1 += src_a0 * src_b;

  630. 

  631.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  632.                 res2 += src_a0 * src_b;

  633. 

  634.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  635.                 res3 += src_a0 * src_b;

  636.             }

  637. 

  638.             if ((temp - 1) & 1)

  639.             {

  640.                 src_a0 = LD_DP(pa0);

  641.                 pa0 += 2;

  642.                 LD_DP2_INC(pb0, 2, src_b0, src_b1);

  643. 

  644.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  645.                 res0 += src_a0 * src_b;

  646. 

  647.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  648.                 res1 += src_a0 * src_b;

  649. 

  650.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);

  651.                 res2 += src_a0 * src_b;

  652. 

  653.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);

  654.                 res3 += src_a0 * src_b;

  655.             }

  656. 

  657. #if defined(TRMMKERNEL)

  658.             dst0 = res0 * v_alpha;

  659.             dst1 = res1 * v_alpha;

  660.             dst2 = res2 * v_alpha;

  661.             dst3 = res3 * v_alpha;

  662. #else

  663.             dst0 = LD_DP(pc0);

  664.             dst1 = LD_DP(pc1);

  665.             dst2 = LD_DP(pc2);

  666.             dst3 = LD_DP(pc3);

  667. 

  668.             dst0 += res0 * v_alpha;

  669.             dst1 += res1 * v_alpha;

  670.             dst2 += res2 * v_alpha;

  671.             dst3 += res3 * v_alpha;

  672. #endif

  673.             ST_DP(dst0, pc0);

  674.             ST_DP(dst1, pc1);

  675.             ST_DP(dst2, pc2);

  676.             ST_DP(dst3, pc3);

  677. 

  678.             pc0 += 2;

  679.             pc1 += 2;

  680.             pc2 += 2;

  681.             pc3 += 2;

  682. 

  683. #if defined(TRMMKERNEL)

  684. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  685.             temp = k - off;

  686. #ifdef LEFT

  687.             temp -= 2; // number of values in A

  688. #else

  689.             temp -= 4; // number of values in B

  690. #endif

  691.             pa0 += temp * 2;

  692.             pb0 += temp * 4;

  693. #endif

  694. 

  695. #ifdef LEFT

  696.             off += 2; // number of values in A

  697. #endif

  698. #endif  // #if defined(TRMMKERNEL)

  699.         }

  700. 

  701.         if (m & 1)

  702.         {

  703. #if defined(TRMMKERNEL)

  704. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  705.             pb0 = B;

  706. #else

  707.             pa0 += off * 1;

  708.             pb0 = B + off * 4;

  709. #endif

  710. 

  711. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  712.             temp = k - off;

  713. #elif defined(LEFT)

  714.             temp = off + 1; // number of values in A

  715. #else

  716.             temp = off + 4; // number of values in B

  717. #endif

  718. #else  // #if !defined(TRMMKERNEL)

  719.             pb0 = B;

  720.             temp = k;

  721. #endif

  722. 

  723.             a0 = pa0[0];

  724.             b0 = pb0[0];

  725.             tmp0 = a0 * b0;

  726. 

  727.             b1 = pb0[1];

  728.             tmp1 = a0 * b1;

  729. 

  730.             b2 = pb0[2];

  731.             tmp2 = a0 * b2;

  732. 

  733.             b3 = pb0[3];

  734.             tmp3 = a0 * b3;

  735. 

  736.             pa0 += 1;

  737.             pb0 += 4;

  738. 

  739.             for (l = ((temp - 1) >> 1); l--;)

  740.             {

  741.                 a0 = pa0[0];

  742.                 b0 = pb0[0];

  743.                 tmp0 += a0 * b0;

  744. 

  745.                 b1 = pb0[1];

  746.                 tmp1 += a0 * b1;

  747. 

  748.                 b2 = pb0[2];

  749.                 tmp2 += a0 * b2;

  750. 

  751.                 b3 = pb0[3];

  752.                 tmp3 += a0 * b3;

  753. 

  754.                 pa0 += 1;

  755.                 pb0 += 4;

  756. 

  757.                 a0 = pa0[0];

  758.                 b0 = pb0[0];

  759.                 tmp0 += a0 * b0;

  760. 

  761.                 b1 = pb0[1];

  762.                 tmp1 += a0 * b1;

  763. 

  764.                 b2 = pb0[2];

  765.                 tmp2 += a0 * b2;

  766. 

  767.                 b3 = pb0[3];

  768.                 tmp3 += a0 * b3;

  769. 

  770.                 pa0 += 1;

  771.                 pb0 += 4;

  772.             }

  773. 

  774.             if ((temp - 1) & 1)

  775.             {

  776.                 a0 = pa0[0];

  777.                 b0 = pb0[0];

  778.                 tmp0 += a0 * b0;

  779. 

  780.                 b1 = pb0[1];

  781.                 tmp1 += a0 * b1;

  782. 

  783.                 b2 = pb0[2];

  784.                 tmp2 += a0 * b2;

  785. 

  786.                 b3 = pb0[3];

  787.                 tmp3 += a0 * b3;

  788. 

  789.                 pa0 += 1;

  790.                 pb0 += 4;

  791.             }

  792. 

  793.             tmp0 = alpha * tmp0;

  794.             tmp1 = alpha * tmp1;

  795.             tmp2 = alpha * tmp2;

  796.             tmp3 = alpha * tmp3;

  797. 

  798. #if defined(TRMMKERNEL)

  799.             pc0[0] = tmp0;

  800.             pc1[0] = tmp1;

  801.             pc2[0] = tmp2;

  802.             pc3[0] = tmp3;

  803. #else

  804.             pc0[0] += tmp0;

  805.             pc1[0] += tmp1;

  806.             pc2[0] += tmp2;

  807.             pc3[0] += tmp3;

  808. #endif

  809.             pc0 += 1;

  810.             pc1 += 1;

  811.             pc2 += 1;

  812.             pc3 += 1;

  813. 

  814. #if defined(TRMMKERNEL)

  815. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  816.             temp = k - off;

  817. #ifdef LEFT

  818.             temp -= 1; // number of values in A

  819. #else

  820.             temp -= 4; // number of values in B

  821. #endif

  822.             pa0 += temp * 1;

  823.             pb0 += temp * 4;

  824. #endif

  825. 

  826. #ifdef LEFT

  827.             off += 1; // number of values in A

  828. #endif

  829. #endif  // #if defined(TRMMKERNEL)

  830.         }

  831. 

  832. #if defined(TRMMKERNEL) && !defined(LEFT)

  833.         off += 4; // number of values in A

  834. #endif

  835. 

  836.         B += (k << 2);

  837.         C += (ldc << 2);

  838.     }

  839. }

  840. 

  841. static void __attribute__ ((noinline))

  842. dgemmkernel_8x4_non_core_msa(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha,

  843.                              FLOAT *A, FLOAT *B, FLOAT *C, BLASLONG ldc

  844. #ifdef TRMMKERNEL

  845.           , BLASLONG offset

  846. #endif

  847.                              )

  848. {

  849.     BLASLONG i, l, temp;

  850. #if defined(TRMMKERNEL)

  851.     BLASLONG off;

  852. #endif

  853.     FLOAT *pc0, *pc1, *pa0, *pb0;

  854.     FLOAT tmp0, tmp1;

  855.     FLOAT a0, b0, b1;

  856.     v2f64 v_alpha = {alpha, alpha};

  857.     v2f64 src_a0, src_a1, src_a2, src_a3, src_b, src_b0;

  858.     v2f64 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;

  859.     v2f64 res0, res1, res2, res3, res4, res5, res6, res7;

  860. 

  861. #if defined(TRMMKERNEL) && !defined(LEFT)

  862.     off = -offset + (4 * (n >> 2));

  863. #endif

  864. 

  865.     if (n & 2)

  866.     {

  867.         pc0 = C;

  868.         pc1 = pc0 + ldc;

  869. 

  870. #if defined(TRMMKERNEL) && defined(LEFT)

  871.         off = offset;

  872. #endif

  873. 

  874.         pa0 = A;

  875. 

  876.         for (i = (m >> 3); i--;)

  877.         {

  878. #if defined(TRMMKERNEL)

  879. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  880.             pb0 = B;

  881. #else

  882.             pa0 += off * 8;

  883.             pb0 = B + off * 2;

  884. #endif

  885. 

  886. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  887.             temp = k - off;

  888. #elif defined(LEFT)

  889.             temp = off + 8; // number of values in A

  890. #else

  891.             temp = off + 2; // number of values in B

  892. #endif

  893. #else  // #if !defined(TRMMKERNEL)

  894.             pb0 = B;

  895.             temp = k;

  896. #endif

  897. 

  898.             LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  899.             src_b0 = LD_DP(pb0);

  900.             pb0 += 2;

  901. 

  902.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  903.             res0 = src_a0 * src_b;

  904.             res1 = src_a1 * src_b;

  905.             res2 = src_a2 * src_b;

  906.             res3 = src_a3 * src_b;

  907. 

  908.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  909.             res4 = src_a0 * src_b;

  910.             res5 = src_a1 * src_b;

  911.             res6 = src_a2 * src_b;

  912.             res7 = src_a3 * src_b;

  913. 

  914.             for (l = ((temp - 1) >> 1); l--;)

  915.             {

  916.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  917.                 src_b0 = LD_DP(pb0);

  918.                 pb0 += 2;

  919. 

  920.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  921.                 res0 += src_a0 * src_b;

  922.                 res1 += src_a1 * src_b;

  923.                 res2 += src_a2 * src_b;

  924.                 res3 += src_a3 * src_b;

  925. 

  926.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  927.                 res4 += src_a0 * src_b;

  928.                 res5 += src_a1 * src_b;

  929.                 res6 += src_a2 * src_b;

  930.                 res7 += src_a3 * src_b;

  931. 

  932.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  933.                 src_b0 = LD_DP(pb0);

  934.                 pb0 += 2;

  935. 

  936.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  937.                 res0 += src_a0 * src_b;

  938.                 res1 += src_a1 * src_b;

  939.                 res2 += src_a2 * src_b;

  940.                 res3 += src_a3 * src_b;

  941. 

  942.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  943.                 res4 += src_a0 * src_b;

  944.                 res5 += src_a1 * src_b;

  945.                 res6 += src_a2 * src_b;

  946.                 res7 += src_a3 * src_b;

  947.             }

  948. 

  949.             if ((temp - 1) & 1)

  950.             {

  951.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  952.                 src_b0 = LD_DP(pb0);

  953.                 pb0 += 2;

  954. 

  955.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  956.                 res0 += src_a0 * src_b;

  957.                 res1 += src_a1 * src_b;

  958.                 res2 += src_a2 * src_b;

  959.                 res3 += src_a3 * src_b;

  960. 

  961.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  962.                 res4 += src_a0 * src_b;

  963.                 res5 += src_a1 * src_b;

  964.                 res6 += src_a2 * src_b;

  965.                 res7 += src_a3 * src_b;

  966.             }

  967. 

  968. #if defined(TRMMKERNEL)

  969.             dst0 = res0 * v_alpha;

  970.             dst1 = res1 * v_alpha;

  971.             dst2 = res2 * v_alpha;

  972.             dst3 = res3 * v_alpha;

  973.             dst4 = res4 * v_alpha;

  974.             dst5 = res5 * v_alpha;

  975.             dst6 = res6 * v_alpha;

  976.             dst7 = res7 * v_alpha;

  977. #else

  978.             LD_DP4(pc0, 2, dst0, dst1, dst2, dst3);

  979.             LD_DP4(pc1, 2, dst4, dst5, dst6, dst7);

  980. 

  981.             dst0 += res0 * v_alpha;

  982.             dst1 += res1 * v_alpha;

  983.             dst2 += res2 * v_alpha;

  984.             dst3 += res3 * v_alpha;

  985.             dst4 += res4 * v_alpha;

  986.             dst5 += res5 * v_alpha;

  987.             dst6 += res6 * v_alpha;

  988.             dst7 += res7 * v_alpha;

  989. #endif

  990.             ST_DP4_INC(dst0, dst1, dst2, dst3, pc0, 2);

  991.             ST_DP4_INC(dst4, dst5, dst6, dst7, pc1, 2);

  992. 

  993. #if defined(TRMMKERNEL)

  994. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  995.             temp = k - off;

  996. #ifdef LEFT

  997.             temp -= 8; // number of values in A

  998. #else

  999.             temp -= 2; // number of values in B

  1000. #endif

  1001.             pa0 += temp * 8;

  1002.             pb0 += temp * 2;

  1003. #endif

  1004. 

  1005. #ifdef LEFT

  1006.             off += 8; // number of values in A

  1007. #endif

  1008. #endif  // #if defined(TRMMKERNEL)

  1009.         }

  1010. 

  1011.         if (m & 4)

  1012.         {

  1013. #if defined(TRMMKERNEL)

  1014. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1015.             pb0 = B;

  1016. #else

  1017.             pa0 += off * 4;

  1018.             pb0 = B + off * 2;

  1019. #endif

  1020. 

  1021. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1022.             temp = k - off;

  1023. #elif defined(LEFT)

  1024.             temp = off + 4; // number of values in A

  1025. #else

  1026.             temp = off + 2; // number of values in B

  1027. #endif

  1028. #else  // #if !defined(TRMMKERNEL)

  1029.             pb0 = B;

  1030.             temp = k;

  1031. #endif

  1032. 

  1033.             LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1034.             src_b0 = LD_DP(pb0);

  1035.             pb0 += 2;

  1036. 

  1037.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1038.             res0 = src_a0 * src_b;

  1039.             res1 = src_a1 * src_b;

  1040. 

  1041.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1042.             res2 = src_a0 * src_b;

  1043.             res3 = src_a1 * src_b;

  1044. 

  1045.             for (l = ((temp - 1) >> 1); l--;)

  1046.             {

  1047.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1048.                 src_b0 = LD_DP(pb0);

  1049.                 pb0 += 2;

  1050. 

  1051.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1052.                 res0 += src_a0 * src_b;

  1053.                 res1 += src_a1 * src_b;

  1054. 

  1055.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1056.                 res2 += src_a0 * src_b;

  1057.                 res3 += src_a1 * src_b;

  1058. 

  1059.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1060.                 src_b0 = LD_DP(pb0);

  1061.                 pb0 += 2;

  1062. 

  1063.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1064.                 res0 += src_a0 * src_b;

  1065.                 res1 += src_a1 * src_b;

  1066. 

  1067.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1068.                 res2 += src_a0 * src_b;

  1069.                 res3 += src_a1 * src_b;

  1070.             }

  1071. 

  1072.             if ((temp - 1) & 1)

  1073.             {

  1074.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1075.                 src_b0 = LD_DP(pb0);

  1076.                 pb0 += 2;

  1077. 

  1078.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1079.                 res0 += src_a0 * src_b;

  1080.                 res1 += src_a1 * src_b;

  1081. 

  1082.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1083.                 res2 += src_a0 * src_b;

  1084.                 res3 += src_a1 * src_b;

  1085.             }

  1086. 

  1087. #if defined(TRMMKERNEL)

  1088.             dst0 = res0 * v_alpha;

  1089.             dst1 = res1 * v_alpha;

  1090.             dst2 = res2 * v_alpha;

  1091.             dst3 = res3 * v_alpha;

  1092. #else

  1093.             LD_DP2(pc0, 2, dst0, dst1);

  1094.             LD_DP2(pc1, 2, dst2, dst3);

  1095. 

  1096.             dst0 += res0 * v_alpha;

  1097.             dst1 += res1 * v_alpha;

  1098.             dst2 += res2 * v_alpha;

  1099.             dst3 += res3 * v_alpha;

  1100. #endif

  1101.             ST_DP2_INC(dst0, dst1, pc0, 2);

  1102.             ST_DP2_INC(dst2, dst3, pc1, 2);

  1103. 

  1104. #if defined(TRMMKERNEL)

  1105. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1106.             temp = k - off;

  1107. #ifdef LEFT

  1108.             temp -= 4; // number of values in A

  1109. #else

  1110.             temp -= 2; // number of values in B

  1111. #endif

  1112.             pa0 += temp * 4;

  1113.             pb0 += temp * 2;

  1114. #endif

  1115. 

  1116. #ifdef LEFT

  1117.             off += 4; // number of values in A

  1118. #endif

  1119. #endif  // #if defined(TRMMKERNEL)

  1120.         }

  1121. 

  1122.         if (m & 2)

  1123.         {

  1124. #if defined(TRMMKERNEL)

  1125. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1126.             pb0 = B;

  1127. #else

  1128.             pa0 += off * 2;

  1129.             pb0 = B + off * 2;

  1130. #endif

  1131. 

  1132. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1133.             temp = k - off;

  1134. #elif defined(LEFT)

  1135.             temp = off + 2; // number of values in A

  1136. #else

  1137.             temp = off + 2; // number of values in B

  1138. #endif

  1139. #else  // #if !defined(TRMMKERNEL)

  1140.             pb0 = B;

  1141.             temp = k;

  1142. #endif

  1143. 

  1144.             src_a0 = LD_DP(pa0);

  1145.             pa0 += 2;

  1146.             src_b0 = LD_DP(pb0);

  1147.             pb0 += 2;

  1148. 

  1149.             src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1150.             res0 = src_a0 * src_b;

  1151. 

  1152.             src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1153.             res1 = src_a0 * src_b;

  1154. 

  1155.             for (l = ((temp - 1) >> 1); l--;)

  1156.             {

  1157.                 src_a0 = LD_DP(pa0);

  1158.                 pa0 += 2;

  1159.                 src_b0 = LD_DP(pb0);

  1160.                 pb0 += 2;

  1161. 

  1162.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1163.                 res0 += src_a0 * src_b;

  1164. 

  1165.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1166.                 res1 += src_a0 * src_b;

  1167. 

  1168.                 src_a0 = LD_DP(pa0);

  1169.                 pa0 += 2;

  1170.                 src_b0 = LD_DP(pb0);

  1171.                 pb0 += 2;

  1172. 

  1173.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1174.                 res0 += src_a0 * src_b;

  1175. 

  1176.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1177.                 res1 += src_a0 * src_b;

  1178.             }

  1179. 

  1180.             if ((temp - 1) & 1)

  1181.             {

  1182.                 src_a0 = LD_DP(pa0);

  1183.                 pa0 += 2;

  1184.                 src_b0 = LD_DP(pb0);

  1185.                 pb0 += 2;

  1186. 

  1187.                 src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);

  1188.                 res0 += src_a0 * src_b;

  1189. 

  1190.                 src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);

  1191.                 res1 += src_a0 * src_b;

  1192.             }

  1193. 

  1194. #if defined(TRMMKERNEL)

  1195.             dst0 = res0 * v_alpha;

  1196.             dst1 = res1 * v_alpha;

  1197. #else

  1198.             dst0 = LD_DP(pc0);

  1199.             dst1 = LD_DP(pc1);

  1200. 

  1201.             dst0 += res0 * v_alpha;

  1202.             dst1 += res1 * v_alpha;

  1203. #endif

  1204.             ST_DP(dst0, pc0);

  1205.             ST_DP(dst1, pc1);

  1206. 

  1207.             pc0 += 2;

  1208.             pc1 += 2;

  1209. 

  1210. #if defined(TRMMKERNEL)

  1211. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1212.             temp = k - off;

  1213. #ifdef LEFT

  1214.             temp -= 2; // number of values in A

  1215. #else

  1216.             temp -= 2; // number of values in B

  1217. #endif

  1218.             pa0 += temp * 2;

  1219.             pb0 += temp * 2;

  1220. #endif

  1221. 

  1222. #ifdef LEFT

  1223.             off += 2; // number of values in A

  1224. #endif

  1225. #endif  // #if defined(TRMMKERNEL)

  1226.         }

  1227. 

  1228.         if (m & 1)

  1229.         {

  1230. #if defined(TRMMKERNEL)

  1231. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1232.             pb0 = B;

  1233. #else

  1234.             pa0 += off * 1;

  1235.             pb0 = B + off * 2;

  1236. #endif

  1237. 

  1238. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1239.             temp = k - off;

  1240. #elif defined(LEFT)

  1241.             temp = off + 1; // number of values in A

  1242. #else

  1243.             temp = off + 2; // number of values in B

  1244. #endif

  1245. #else  // #if !defined(TRMMKERNEL)

  1246.             pb0 = B;

  1247.             temp = k;

  1248. #endif

  1249. 

  1250.             a0 = pa0[0];

  1251.             b0 = pb0[0];

  1252.             tmp0 = a0 * b0;

  1253. 

  1254.             b1 = pb0[1];

  1255.             tmp1 = a0 * b1;

  1256. 

  1257.             pa0 += 1;

  1258.             pb0 += 2;

  1259. 

  1260.             for (l = ((temp - 1) >> 1); l--;)

  1261.             {

  1262.                 a0 = pa0[0];

  1263.                 b0 = pb0[0];

  1264.                 tmp0 += a0 * b0;

  1265. 

  1266.                 b1 = pb0[1];

  1267.                 tmp1 += a0 * b1;

  1268. 

  1269.                 pa0 += 1;

  1270.                 pb0 += 2;

  1271. 

  1272.                 a0 = pa0[0];

  1273.                 b0 = pb0[0];

  1274.                 tmp0 += a0 * b0;

  1275. 

  1276.                 b1 = pb0[1];

  1277.                 tmp1 += a0 * b1;

  1278. 

  1279.                 pa0 += 1;

  1280.                 pb0 += 2;

  1281.             }

  1282. 

  1283.             if ((temp - 1) & 1)

  1284.             {

  1285.                 a0 = pa0[0];

  1286.                 b0 = pb0[0];

  1287.                 tmp0 += a0 * b0;

  1288. 

  1289.                 b1 = pb0[1];

  1290.                 tmp1 += a0 * b1;

  1291. 

  1292.                 pa0 += 1;

  1293.                 pb0 += 2;

  1294.             }

  1295. 

  1296.             tmp0 = alpha * tmp0;

  1297.             tmp1 = alpha * tmp1;

  1298. 

  1299. #if defined(TRMMKERNEL)

  1300.             pc0[0] = tmp0;

  1301.             pc1[0] = tmp1;

  1302. #else

  1303.             pc0[0] += tmp0;

  1304.             pc1[0] += tmp1;

  1305. #endif

  1306.             pc0 += 1;

  1307.             pc1 += 1;

  1308. 

  1309. #if defined(TRMMKERNEL)

  1310. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1311.             temp = k - off;

  1312. #ifdef LEFT

  1313.             temp -= 1; // number of values in A

  1314. #else

  1315.             temp -= 2; // number of values in B

  1316. #endif

  1317.             pa0 += temp * 1;

  1318.             pb0 += temp * 2;

  1319. #endif

  1320. 

  1321. #ifdef LEFT

  1322.             off += 1; // number of values in A

  1323. #endif

  1324. #endif  // #if defined(TRMMKERNEL)

  1325.         }

  1326. 

  1327. #if defined(TRMMKERNEL) && !defined(LEFT)

  1328.         off += 2; // number of values in A

  1329. #endif

  1330. 

  1331.         B += (k << 1);

  1332.         C += (ldc << 1);

  1333.     }

  1334. 

  1335.     if (n & 1)

  1336.     {

  1337.         pc0 = C;

  1338. 

  1339. #if defined(TRMMKERNEL) && defined(LEFT)

  1340.         off = offset;

  1341. #endif

  1342. 

  1343.         pa0 = A;

  1344. 

  1345.         for (i = (m >> 3); i--;)

  1346.         {

  1347. #if defined(TRMMKERNEL)

  1348. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1349.             pb0 = B;

  1350. #else

  1351.             pa0 += off * 8;

  1352.             pb0 = B + off * 1;

  1353. #endif

  1354. 

  1355. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1356.             temp = k - off;

  1357. #elif defined(LEFT)

  1358.             temp = off + 8; // number of values in A

  1359. #else

  1360.             temp = off + 1; // number of values in B

  1361. #endif

  1362. #else  // #if !defined(TRMMKERNEL)

  1363.             pb0 = B;

  1364.             temp = k;

  1365. #endif

  1366. 

  1367.             LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  1368.             src_b[0] = pb0[0];

  1369.             src_b[1] = pb0[0];

  1370. 

  1371.             res0 = src_a0 * src_b;

  1372.             res1 = src_a1 * src_b;

  1373.             res2 = src_a2 * src_b;

  1374.             res3 = src_a3 * src_b;

  1375. 

  1376.             pb0 += 1;

  1377. 

  1378.             for (l = ((temp - 1) >> 1); l--;)

  1379.             {

  1380.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  1381.                 src_b[0] = pb0[0];

  1382.                 src_b[1] = pb0[0];

  1383. 

  1384.                 res0 += src_a0 * src_b;

  1385.                 res1 += src_a1 * src_b;

  1386.                 res2 += src_a2 * src_b;

  1387.                 res3 += src_a3 * src_b;

  1388. 

  1389.                 pb0 += 1;

  1390. 

  1391.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  1392.                 src_b[0] = pb0[0];

  1393.                 src_b[1] = pb0[0];

  1394. 

  1395.                 res0 += src_a0 * src_b;

  1396.                 res1 += src_a1 * src_b;

  1397.                 res2 += src_a2 * src_b;

  1398.                 res3 += src_a3 * src_b;

  1399. 

  1400.                 pb0 += 1;

  1401.             }

  1402. 

  1403.             if ((temp - 1) & 1)

  1404.             {

  1405.                 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);

  1406.                 src_b[0] = pb0[0];

  1407.                 src_b[1] = pb0[0];

  1408. 

  1409.                 res0 += src_a0 * src_b;

  1410.                 res1 += src_a1 * src_b;

  1411.                 res2 += src_a2 * src_b;

  1412.                 res3 += src_a3 * src_b;

  1413. 

  1414.                 pb0 += 1;

  1415.             }

  1416. 

  1417. #if defined(TRMMKERNEL)

  1418.             dst0 = res0 * v_alpha;

  1419.             dst1 = res1 * v_alpha;

  1420.             dst2 = res2 * v_alpha;

  1421.             dst3 = res3 * v_alpha;

  1422. #else

  1423.             LD_DP4(pc0, 2, dst0, dst1, dst2, dst3);

  1424. 

  1425.             dst0 += res0 * v_alpha;

  1426.             dst1 += res1 * v_alpha;

  1427.             dst2 += res2 * v_alpha;

  1428.             dst3 += res3 * v_alpha;

  1429. #endif

  1430.             ST_DP4_INC(dst0, dst1, dst2, dst3, pc0, 2);

  1431. 

  1432. #if defined(TRMMKERNEL)

  1433. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1434.             temp = k - off;

  1435. #ifdef LEFT

  1436.             temp -= 8; // number of values in A

  1437. #else

  1438.             temp -= 1; // number of values in B

  1439. #endif

  1440.             pa0 += temp * 8;

  1441.             pb0 += temp * 1;

  1442. #endif

  1443. 

  1444. #ifdef LEFT

  1445.             off += 8; // number of values in A

  1446. #endif

  1447. #endif  // #if defined(TRMMKERNEL)

  1448.         }

  1449. 

  1450.         if (m & 4)

  1451.         {

  1452. #if defined(TRMMKERNEL)

  1453. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1454.             pb0 = B;

  1455. #else

  1456.             pa0 += off * 4;

  1457.             pb0 = B + off * 1;

  1458. #endif

  1459. 

  1460. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1461.             temp = k - off;

  1462. #elif defined(LEFT)

  1463.             temp = off + 4; // number of values in A

  1464. #else

  1465.             temp = off + 1; // number of values in B

  1466. #endif

  1467. #else  // #if !defined(TRMMKERNEL)

  1468.             pb0 = B;

  1469.             temp = k;

  1470. #endif

  1471. 

  1472.             LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1473.             src_b[0] = pb0[0];

  1474.             src_b[1] = pb0[0];

  1475. 

  1476.             res0 = src_a0 * src_b;

  1477.             res1 = src_a1 * src_b;

  1478. 

  1479.             pb0 += 1;

  1480. 

  1481.             for (l = ((temp - 1) >> 1); l--;)

  1482.             {

  1483.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1484.                 src_b[0] = pb0[0];

  1485.                 src_b[1] = pb0[0];

  1486. 

  1487.                 res0 += src_a0 * src_b;

  1488.                 res1 += src_a1 * src_b;

  1489. 

  1490.                 pb0 += 1;

  1491. 

  1492.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1493.                 src_b[0] = pb0[0];

  1494.                 src_b[1] = pb0[0];

  1495. 

  1496.                 res0 += src_a0 * src_b;

  1497.                 res1 += src_a1 * src_b;

  1498. 

  1499.                 pb0 += 1;

  1500.             }

  1501. 

  1502.             if ((temp - 1) & 1)

  1503.             {

  1504.                 LD_DP2_INC(pa0, 2, src_a0, src_a1);

  1505.                 src_b[0] = pb0[0];

  1506.                 src_b[1] = pb0[0];

  1507. 

  1508.                 res0 += src_a0 * src_b;

  1509.                 res1 += src_a1 * src_b;

  1510. 

  1511.                 pb0 += 1;

  1512.             }

  1513. 

  1514. #if defined(TRMMKERNEL)

  1515.             dst0 = res0 * v_alpha;

  1516.             dst1 = res1 * v_alpha;

  1517. #else

  1518.             LD_DP2(pc0, 2, dst0, dst1);

  1519. 

  1520.             dst0 += res0 * v_alpha;

  1521.             dst1 += res1 * v_alpha;

  1522. #endif

  1523.             ST_DP2_INC(dst0, dst1, pc0, 2);

  1524. 

  1525. #if defined(TRMMKERNEL)

  1526. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1527.             temp = k - off;

  1528. #ifdef LEFT

  1529.             temp -= 4; // number of values in A

  1530. #else

  1531.             temp -= 1; // number of values in B

  1532. #endif

  1533.             pa0 += temp * 4;

  1534.             pb0 += temp * 1;

  1535. #endif

  1536. 

  1537. #ifdef LEFT

  1538.             off += 4; // number of values in A

  1539. #endif

  1540. #endif  // #if defined(TRMMKERNEL)

  1541.         }

  1542. 

  1543.         if (m & 2)

  1544.         {

  1545. #if defined(TRMMKERNEL)

  1546. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1547.             pb0 = B;

  1548. #else

  1549.             pa0 += off * 2;

  1550.             pb0 = B + off * 1;

  1551. #endif

  1552. 

  1553. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1554.             temp = k - off;

  1555. #elif defined(LEFT)

  1556.             temp = off + 2; // number of values in A

  1557. #else

  1558.             temp = off + 1; // number of values in B

  1559. #endif

  1560. #else  // #if !defined(TRMMKERNEL)

  1561.             pb0 = B;

  1562.             temp = k;

  1563. #endif

  1564. 

  1565.             src_a0 = LD_DP(pa0);

  1566.             src_b[0] = pb0[0];

  1567.             src_b[1] = pb0[0];

  1568. 

  1569.             res0 = src_a0 * src_b;

  1570. 

  1571.             pa0 += 2;

  1572.             pb0 += 1;

  1573. 

  1574.             for (l = ((temp - 1) >> 1); l--;)

  1575.             {

  1576.                 src_a0 = LD_DP(pa0);

  1577.                 src_b[0] = pb0[0];

  1578.                 src_b[1] = pb0[0];

  1579. 

  1580.                 res0 += src_a0 * src_b;

  1581. 

  1582.                 pa0 += 2;

  1583.                 pb0 += 1;

  1584. 

  1585.                 src_a0 = LD_DP(pa0);

  1586.                 src_b[0] = pb0[0];

  1587.                 src_b[1] = pb0[0];

  1588. 

  1589.                 res0 += src_a0 * src_b;

  1590. 

  1591.                 pa0 += 2;

  1592.                 pb0 += 1;

  1593.             }

  1594. 

  1595.             if ((temp - 1) & 1)

  1596.             {

  1597.                 src_a0 = LD_DP(pa0);

  1598.                 src_b[0] = pb0[0];

  1599.                 src_b[1] = pb0[0];

  1600. 

  1601.                 res0 += src_a0 * src_b;

  1602. 

  1603.                 pa0 += 2;

  1604.                 pb0 += 1;

  1605.             }

  1606. 

  1607. #if defined(TRMMKERNEL)

  1608.             dst0 = res0 * v_alpha;

  1609. #else

  1610.             dst0 = LD_DP(pc0);

  1611. 

  1612.             dst0 += res0 * v_alpha;

  1613. #endif

  1614.             ST_DP(dst0, pc0);

  1615. 

  1616.             pc0 += 2;

  1617. 

  1618. #if defined(TRMMKERNEL)

  1619. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1620.             temp = k - off;

  1621. #ifdef LEFT

  1622.             temp -= 2; // number of values in A

  1623. #else

  1624.             temp -= 1; // number of values in B

  1625. #endif

  1626.             pa0 += temp * 2;

  1627.             pb0 += temp * 1;

  1628. #endif

  1629. 

  1630. #ifdef LEFT

  1631.             off += 2; // number of values in A

  1632. #endif

  1633. #endif  // #if defined(TRMMKERNEL)

  1634.         }

  1635. 

  1636.         if (m & 1)

  1637.         {

  1638. #if defined(TRMMKERNEL)

  1639. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1640.             pb0 = B;

  1641. #else

  1642.             pa0 += off * 1;

  1643.             pb0 = B + off * 1;

  1644. #endif

  1645. 

  1646. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1647.             temp = k - off;

  1648. #elif defined(LEFT)

  1649.             temp = off + 1; // number of values in A

  1650. #else

  1651.             temp = off + 1; // number of values in B

  1652. #endif

  1653. #else  // #if !defined(TRMMKERNEL)

  1654.             pb0 = B;

  1655.             temp = k;

  1656. #endif

  1657. 

  1658.             a0 = pa0[0];

  1659.             b0 = pb0[0];

  1660.             tmp0 = a0 * b0;

  1661. 

  1662.             pa0 += 1;

  1663.             pb0 += 1;

  1664. 

  1665.             for (l = ((temp - 1) >> 1); l--;)

  1666.             {

  1667.                 a0 = pa0[0];

  1668.                 b0 = pb0[0];

  1669.                 tmp0 += a0 * b0;

  1670. 

  1671.                 pa0 += 1;

  1672.                 pb0 += 1;

  1673. 

  1674.                 a0 = pa0[0];

  1675.                 b0 = pb0[0];

  1676.                 tmp0 += a0 * b0;

  1677. 

  1678.                 pa0 += 1;

  1679.                 pb0 += 1;

  1680.             }

  1681. 

  1682.             if ((temp - 1) & 1)

  1683.             {

  1684.                 a0 = pa0[0];

  1685.                 b0 = pb0[0];

  1686.                 tmp0 += a0 * b0;

  1687. 

  1688.                 pa0 += 1;

  1689.                 pb0 += 1;

  1690.             }

  1691. 

  1692. #if defined(TRMMKERNEL)

  1693.             pc0[0] = alpha * tmp0;

  1694. #else

  1695.             pc0[0] += alpha * tmp0;

  1696. #endif

  1697.         }

  1698.     }

  1699. }

  1700. 

  1701. int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,

  1702.           FLOAT *C, BLASLONG ldc

  1703. #ifdef TRMMKERNEL

  1704.           , BLASLONG offset

  1705. #endif

  1706.           )

  1707. {

  1708.     if (n >> 2)

  1709.     {

  1710.         if (m >> 3)

  1711. #ifdef TRMMKERNEL

  1712.             dgemmkernel_8x4_core_msa(m, n, k, alpha, A, B, C, ldc, offset);

  1713. #else

  1714.             dgemmkernel_8x4_core_msa(m, n, k, alpha, A, B, C, ldc);

  1715. #endif

  1716. 

  1717.         if (m & 7)

  1718. #ifdef TRMMKERNEL

  1719.             dgemmkernel_7x4_core_msa(m, n, k, alpha, A, B, C, ldc, offset);

  1720. #else

  1721.             dgemmkernel_7x4_core_msa(m, n, k, alpha, A, B, C, ldc);

  1722. #endif

  1723.     }

  1724. 

  1725.     if (n & 3)

  1726.     {

  1727.         B = B + (k << 2) * (n >> 2);

  1728.         C = C + (ldc << 2) * (n >> 2);

  1729. 

  1730. #ifdef TRMMKERNEL

  1731.         dgemmkernel_8x4_non_core_msa(m, n, k, alpha, A, B, C, ldc, offset);

  1732. #else

  1733.         dgemmkernel_8x4_non_core_msa(m, n, k, alpha, A, B, C, ldc);

  1734. #endif

  1735.     }

  1736. 

  1737.     return 0;

  1738. }