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: e5313f53d5
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'e5313f53d5'
${ noResults }
OpenBLAS/kernel/riscv64/trsm_kernel_RN_rvv_v1.c

793 lines
20 kB
Raw Blame History 

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

  2. Copyright (c) 2022, 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. 

  30. #if !defined(DOUBLE)

  31. #define VSETVL(n) vsetvl_e32m2(n)

  32. #define VSETVL_MAX vsetvlmax_e32m2()

  33. #define FLOAT_V_T vfloat32m2_t

  34. #define VLEV_FLOAT vle32_v_f32m2

  35. #define VLSEV_FLOAT vlse32_v_f32m2

  36. #define VLSEG2_FLOAT vlseg2e32_v_f32m2

  37. #define VSEV_FLOAT vse32_v_f32m2

  38. #define VSSEV_FLOAT vsse32_v_f32m2

  39. #define VSSEG2_FLOAT vsseg2e32_v_f32m2

  40. #define VFMACCVF_FLOAT vfmacc_vf_f32m2

  41. #define VFNMSACVF_FLOAT vfnmsac_vf_f32m2

  42. #else

  43. #define VSETVL(n) vsetvl_e64m2(n)

  44. #define VSETVL_MAX vsetvlmax_e64m2()

  45. #define FLOAT_V_T vfloat64m2_t

  46. #define VLEV_FLOAT vle64_v_f64m2

  47. #define VLSEV_FLOAT vlse64_v_f64m2

  48. #define VLSEG2_FLOAT vlseg2e64_v_f64m2

  49. #define VSEV_FLOAT vse64_v_f64m2

  50. #define VSSEV_FLOAT vsse64_v_f64m2

  51. #define VSSEG2_FLOAT vsseg2e64_v_f64m2

  52. #define VFMACCVF_FLOAT vfmacc_vf_f64m2

  53. #define VFNMSACVF_FLOAT vfnmsac_vf_f64m2

  54. #endif

  55. 

  56. 

  57. static FLOAT dm1 = -1.;

  58. 

  59. #ifdef CONJ

  60. #define GEMM_KERNEL   GEMM_KERNEL_R

  61. #else

  62. #define GEMM_KERNEL   GEMM_KERNEL_N

  63. #endif

  64. 

  65. #if GEMM_DEFAULT_UNROLL_N == 1

  66. #define GEMM_UNROLL_N_SHIFT 0

  67. #endif

  68. 

  69. #if GEMM_DEFAULT_UNROLL_N == 2

  70. #define GEMM_UNROLL_N_SHIFT 1

  71. #endif

  72. 

  73. #if GEMM_DEFAULT_UNROLL_N == 4

  74. #define GEMM_UNROLL_N_SHIFT 2

  75. #endif

  76. 

  77. #if GEMM_DEFAULT_UNROLL_N == 8

  78. #define GEMM_UNROLL_N_SHIFT 3

  79. #endif

  80. 

  81. #if GEMM_DEFAULT_UNROLL_N == 16

  82. #define GEMM_UNROLL_N_SHIFT 4

  83. #endif

  84. 

  85. // Optimizes the implementation in ../arm64/trsm_kernel_RN_sve.c

  86. 

  87. #ifndef COMPLEX

  88. 

  89. #if GEMM_DEFAULT_UNROLL_N == 1

  90. 

  91. static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {

  92. 

  93.     FLOAT aa, bb;

  94.     FLOAT *pb, *pc;

  95.     BLASLONG stride_ldc = sizeof(FLOAT) * ldc;

  96.     int i, j, k;

  97.     size_t vl;

  98.     FLOAT_V_T vb, vc;

  99. 

  100.     for (i = 0; i < n; i++) 

  101.     {

  102.         bb = *(b + i);

  103. 

  104.         for (j = 0; j < m; j ++) 

  105.         {

  106.             aa = *(c + j + i * ldc);

  107.             aa *= bb;

  108.             *a  = aa;

  109.             *(c + j + i * ldc) = aa;

  110.             a ++;

  111. 

  112.             pb = b + i + 1;

  113.             pc = c + j + (i + 1) *ldc;

  114.             for (k = (n - i - 1); k > 0; k -= vl)

  115.             {

  116.                 vl = VSETVL(k);

  117.                 vc = VLSEV_FLOAT(pc, stride_ldc, vl);

  118.                 vb = VLEV_FLOAT(pb, vl);

  119.                 vc = VFNMSACVF_FLOAT(vc, aa, vb, vl);

  120.                 VSSEV_FLOAT(pc, stride_ldc, vc, vl);

  121.                 pb += vl;

  122.                 pc ++;

  123.             }

  124.         }

  125.         b += n;

  126.     }

  127. }

  128. 

  129. #elif GEMM_DEFAULT_UNROLL_N == 2

  130. 

  131. static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {

  132. 

  133.     FLOAT aa0, aa1, bb;

  134.     FLOAT *pb, *pc;

  135.     FLOAT *pa0, *pa1, *pc0, *pc1;

  136.     BLASLONG stride_ldc = sizeof(FLOAT) * ldc;

  137.     int i, j, k;

  138.     size_t vl;

  139.     FLOAT_V_T vb, vc0, vc1;

  140. 

  141.     for (i = 0; i < n; i++) 

  142.     {

  143.         bb = *(b + i);

  144.         pc = c + i * ldc;

  145.         for (j = 0; j < m/2; j ++) 

  146.         {

  147.             pa0 = pc + j * 2;

  148.             pa1 = pc + j * 2 + 1;

  149.             aa0 = *pa0 * bb;

  150.             aa1 = *pa1 * bb;

  151. 

  152.             *pa0    = aa0;

  153.             *pa1    = aa1;

  154.             *a      = aa0;

  155.             *(a + 1)= aa1;

  156.             a  += 2;

  157. 

  158.             pb = b + i + 1;

  159.             pc0 = pa0 + ldc;

  160.             pc1 = pa1 + ldc;

  161.             for (k = (n - i - 1); k > 0; k -= vl)

  162.             {

  163.                 vl = VSETVL(k);

  164.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  165.                 vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);

  166.                 vb = VLEV_FLOAT(pb, vl);

  167.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  168.                 vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);

  169.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  170.                 VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);

  171.                 pb += vl;

  172.                 pc0++;

  173.                 pc1++;

  174.             }

  175.         }

  176.         pc += (m/2)*2;

  177.         if (m & 1)

  178.         {

  179.             pa0 = pc;

  180.             aa0 = *pa0 * bb;

  181.             

  182.             *pa0    = aa0;

  183.             *a      = aa0;

  184.             a  += 1;

  185.            

  186.             pb = b + i + 1;

  187.             pc0 = pa0 + ldc;

  188.             for (k = (n - i - 1); k > 0; k -= vl)

  189.             {

  190.                 vl = VSETVL(k);

  191.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  192.                 vb = VLEV_FLOAT(pb, vl);

  193.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  194.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  195.                 pb += vl;

  196.                 pc0++;

  197.             }

  198.         }

  199.         b += n;

  200.     }

  201. }

  202. 

  203. #elif GEMM_DEFAULT_UNROLL_N == 4

  204. 

  205. static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {

  206. 

  207.     FLOAT bb;

  208.     FLOAT aa0, aa1, aa2, aa3;

  209.     FLOAT *pb, *pc;

  210.     FLOAT *pa0, *pa1, *pa2, *pa3;

  211.     FLOAT *pc0, *pc1, *pc2, *pc3;

  212.     BLASLONG stride_ldc = sizeof(FLOAT) * ldc;

  213.     int i, j, k;

  214.     size_t vl;

  215.     FLOAT_V_T vb, vc0, vc1, vc2, vc3;

  216. 

  217.     for (i = 0; i < n; i++) 

  218.     {

  219.         bb = *(b + i);

  220.         pc = c + i * ldc;

  221.         for (j = 0; j < m/4; j ++) 

  222.         {

  223.             pa0 = pc + j * 4;

  224.             pa1 = pa0 + 1;

  225.             pa2 = pa1 + 1;

  226.             pa3 = pa2 + 1;

  227.             

  228.             aa0 = *pa0 * bb;

  229.             aa1 = *pa1 * bb;

  230.             aa2 = *pa2 * bb;

  231.             aa3 = *pa3 * bb;

  232. 

  233.             *pa0    = aa0;

  234.             *pa1    = aa1;

  235.             *pa2    = aa2;

  236.             *pa3    = aa3;

  237. 

  238.             *a      = aa0;

  239.             *(a + 1)= aa1;

  240.             *(a + 2)= aa2;

  241.             *(a + 3)= aa3;

  242. 

  243.             a  += 4;

  244. 

  245.             pb = b + i + 1;

  246.             pc0 = pa0 + ldc;

  247.             pc1 = pa1 + ldc;

  248.             pc2 = pa2 + ldc;

  249.             pc3 = pa3 + ldc;

  250.             for (k = (n - i - 1); k > 0; k -= vl)

  251.             {

  252.                 vl = VSETVL(k);

  253.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  254.                 vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);

  255.                 vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);

  256.                 vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);

  257.                 vb = VLEV_FLOAT(pb, vl);

  258.  

  259.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  260.                 vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);

  261.                 vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);

  262.                 vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);

  263. 

  264.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  265.                 VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);

  266.                 VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);

  267.                 VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);

  268. 

  269.                 pb += vl;

  270.                 pc0++;

  271.                 pc1++;

  272.                 pc2++;

  273.                 pc3++;

  274.             }

  275.         }

  276.         pc += (m/4)*4;

  277. 

  278.         if (m & 2)

  279.         {

  280.             pa0 = pc;

  281.             pa1 = pa0 + 1;

  282.             

  283.             aa0 = *pa0 * bb;

  284.             aa1 = *pa1 * bb;

  285. 

  286.             *pa0    = aa0;

  287.             *pa1    = aa1;

  288. 

  289.             *a      = aa0;

  290.             *(a + 1)= aa1;

  291. 

  292.             a  += 2;

  293. 

  294.             pb = b + i + 1;

  295.             pc0 = pa0 + ldc;

  296.             pc1 = pa1 + ldc;

  297.             for (k = (n - i - 1); k > 0; k -= vl)

  298.             {

  299.                 vl = VSETVL(k);

  300.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  301.                 vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);

  302.                 vb = VLEV_FLOAT(pb, vl);

  303.  

  304.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  305.                 vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);

  306. 

  307.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  308.                 VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);

  309. 

  310.                 pb += vl;

  311.                 pc0++;

  312.                 pc1++;

  313.             }

  314.             pc += 2;

  315.         }

  316. 

  317.         if (m & 1)

  318.         {

  319.             pa0 = pc;

  320.             aa0 = *pa0 * bb;

  321.             

  322.             *pa0    = aa0;

  323.             *a      = aa0;

  324.             a  += 1;

  325.            

  326.             pb = b + i + 1;

  327.             pc0 = pa0 + ldc;

  328.             for (k = (n - i - 1); k > 0; k -= vl)

  329.             {

  330.                 vl = VSETVL(k);

  331.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  332.                 vb = VLEV_FLOAT(pb, vl);

  333.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  334.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  335.                 pb += vl;

  336.                 pc0++;

  337.             }

  338.         }

  339.         b += n;

  340.     }

  341. }

  342. 

  343. #elif GEMM_DEFAULT_UNROLL_N == 8

  344. 

  345. static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {

  346. 

  347.     FLOAT bb;

  348.     FLOAT aa0, aa1, aa2, aa3, aa4, aa5, aa6, aa7;

  349.     FLOAT *pb, *pc;

  350.     FLOAT *pa0, *pa1, *pa2, *pa3, *pa4, *pa5, *pa6, *pa7;

  351.     FLOAT *pc0, *pc1, *pc2, *pc3, *pc4, *pc5, *pc6, *pc7;

  352.     BLASLONG stride_ldc = sizeof(FLOAT) * ldc;

  353.     int i, j, k;

  354.     size_t vl;

  355.     FLOAT_V_T vb, vc0, vc1, vc2, vc3, vc4, vc5, vc6, vc7;

  356. 

  357.     for (i = 0; i < n; i++) 

  358.     {

  359.         bb = *(b + i);

  360.         pc = c + i * ldc;

  361.         for (j = 0; j < m/8; j ++) 

  362.         {

  363.             pa0 = pc + j * 8;

  364.             pa1 = pa0 + 1;

  365.             pa2 = pa1 + 1;

  366.             pa3 = pa2 + 1;

  367.             pa4 = pa3 + 1;

  368.             pa5 = pa4 + 1;

  369.             pa6 = pa5 + 1;

  370.             pa7 = pa6 + 1;

  371.             

  372.             aa0 = *pa0 * bb;

  373.             aa1 = *pa1 * bb;

  374.             aa2 = *pa2 * bb;

  375.             aa3 = *pa3 * bb;

  376.             aa4 = *pa4 * bb;

  377.             aa5 = *pa5 * bb;

  378.             aa6 = *pa6 * bb;

  379.             aa7 = *pa7 * bb;

  380. 

  381.             *pa0    = aa0;

  382.             *pa1    = aa1;

  383.             *pa2    = aa2;

  384.             *pa3    = aa3;

  385.             *pa4    = aa4;

  386.             *pa5    = aa5;

  387.             *pa6    = aa6;

  388.             *pa7    = aa7;

  389. 

  390.             *a      = aa0;

  391.             *(a + 1)= aa1;

  392.             *(a + 2)= aa2;

  393.             *(a + 3)= aa3;

  394.             *(a + 4)= aa4;

  395.             *(a + 5)= aa5;

  396.             *(a + 6)= aa6;

  397.             *(a + 7)= aa7;

  398. 

  399.             a  += 8;

  400. 

  401.             pb = b + i + 1;

  402.             pc0 = pa0 + ldc;

  403.             pc1 = pa1 + ldc;

  404.             pc2 = pa2 + ldc;

  405.             pc3 = pa3 + ldc;

  406.             pc4 = pa4 + ldc;

  407.             pc5 = pa5 + ldc;

  408.             pc6 = pa6 + ldc;

  409.             pc7 = pa7 + ldc;

  410.             for (k = (n - i - 1); k > 0; k -= vl)

  411.             {

  412.                 vl = VSETVL(k);

  413.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  414.                 vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);

  415.                 vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);

  416.                 vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);

  417.                 vc4 = VLSEV_FLOAT(pc4, stride_ldc, vl);

  418.                 vc5 = VLSEV_FLOAT(pc5, stride_ldc, vl);

  419.                 vc6 = VLSEV_FLOAT(pc6, stride_ldc, vl);

  420.                 vc7 = VLSEV_FLOAT(pc7, stride_ldc, vl);

  421.                 vb = VLEV_FLOAT(pb, vl);

  422.  

  423.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  424.                 vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);

  425.                 vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);

  426.                 vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);

  427.                 vc4 = VFNMSACVF_FLOAT(vc4, aa4, vb, vl);

  428.                 vc5 = VFNMSACVF_FLOAT(vc5, aa5, vb, vl);

  429.                 vc6 = VFNMSACVF_FLOAT(vc6, aa6, vb, vl);

  430.                 vc7 = VFNMSACVF_FLOAT(vc7, aa7, vb, vl);

  431. 

  432.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  433.                 VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);

  434.                 VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);

  435.                 VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);

  436.                 VSSEV_FLOAT(pc4, stride_ldc, vc4, vl);

  437.                 VSSEV_FLOAT(pc5, stride_ldc, vc5, vl);

  438.                 VSSEV_FLOAT(pc6, stride_ldc, vc6, vl);

  439.                 VSSEV_FLOAT(pc7, stride_ldc, vc7, vl);

  440. 

  441.                 pb += vl;

  442.                 pc0++;

  443.                 pc1++;

  444.                 pc2++;

  445.                 pc3++;

  446.                 pc4++;

  447.                 pc5++;

  448.                 pc6++;

  449.                 pc7++;

  450.             }

  451.         }

  452.         pc += (m/8)*8;

  453. 

  454.         if (m & 4)

  455.         {

  456.             pa0 = pc;

  457.             pa1 = pa0 + 1;

  458.             pa2 = pa1 + 1;

  459.             pa3 = pa2 + 1;

  460.             

  461.             aa0 = *pa0 * bb;

  462.             aa1 = *pa1 * bb;

  463.             aa2 = *pa2 * bb;

  464.             aa3 = *pa3 * bb;

  465. 

  466.             *pa0    = aa0;

  467.             *pa1    = aa1;

  468.             *pa2    = aa2;

  469.             *pa3    = aa3;

  470. 

  471.             *a      = aa0;

  472.             *(a + 1)= aa1;

  473.             *(a + 2)= aa2;

  474.             *(a + 3)= aa3;

  475. 

  476.             a  += 4;

  477. 

  478.             pb = b + i + 1;

  479.             pc0 = pa0 + ldc;

  480.             pc1 = pa1 + ldc;

  481.             pc2 = pa2 + ldc;

  482.             pc3 = pa3 + ldc;

  483.             for (k = (n - i - 1); k > 0; k -= vl)

  484.             {

  485.                 vl = VSETVL(k);

  486.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  487.                 vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);

  488.                 vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);

  489.                 vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);

  490.                 vb = VLEV_FLOAT(pb, vl);

  491.  

  492.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  493.                 vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);

  494.                 vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);

  495.                 vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);

  496. 

  497.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  498.                 VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);

  499.                 VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);

  500.                 VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);

  501. 

  502.                 pb += vl;

  503.                 pc0++;

  504.                 pc1++;

  505.                 pc2++;

  506.                 pc3++;

  507.             }

  508.             pc += 4;

  509.         }

  510. 

  511.         if (m & 2)

  512.         {

  513.             pa0 = pc;

  514.             pa1 = pa0 + 1;

  515.             

  516.             aa0 = *pa0 * bb;

  517.             aa1 = *pa1 * bb;

  518. 

  519.             *pa0    = aa0;

  520.             *pa1    = aa1;

  521. 

  522.             *a      = aa0;

  523.             *(a + 1)= aa1;

  524. 

  525.             a  += 2;

  526. 

  527.             pb = b + i + 1;

  528.             pc0 = pa0 + ldc;

  529.             pc1 = pa1 + ldc;

  530.             for (k = (n - i - 1); k > 0; k -= vl)

  531.             {

  532.                 vl = VSETVL(k);

  533.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  534.                 vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);

  535.                 vb = VLEV_FLOAT(pb, vl);

  536.  

  537.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  538.                 vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);

  539. 

  540.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  541.                 VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);

  542. 

  543.                 pb += vl;

  544.                 pc0++;

  545.                 pc1++;

  546.             }

  547.             pc += 2;

  548.         }

  549. 

  550.         if (m & 1)

  551.         {

  552.             pa0 = pc;

  553.             aa0 = *pa0 * bb;

  554.             

  555.             *pa0    = aa0;

  556.             *a      = aa0;

  557.             a  += 1;

  558.            

  559.             pb = b + i + 1;

  560.             pc0 = pa0 + ldc;

  561.             for (k = (n - i - 1); k > 0; k -= vl)

  562.             {

  563.                 vl = VSETVL(k);

  564.                 vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);

  565.                 vb = VLEV_FLOAT(pb, vl);

  566.                 vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);

  567.                 VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);

  568.                 pb += vl;

  569.                 pc0++;

  570.             }

  571.         }

  572.         b += n;

  573.     }

  574. }

  575. #else

  576. static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {

  577. 

  578.   FLOAT aa, bb;

  579. 

  580.   int i, j, k;

  581. 

  582.   for (i = 0; i < n; i++) {

  583. 

  584.     bb = *(b + i);

  585. 

  586.     for (j = 0; j < m; j ++) {

  587.       aa = *(c + j + i * ldc);

  588.       aa *= bb;

  589.       *a  = aa;

  590.       *(c + j + i * ldc) = aa;

  591.       a ++;

  592. 

  593.       for (k = i + 1; k < n; k ++){

  594. 	*(c + j + k * ldc) -= aa * *(b + k);

  595.       }

  596. 

  597.     }

  598.     b += n;

  599.   }

  600. }

  601. 

  602. #endif

  603. 

  604. #else

  605. 

  606. static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {

  607. 

  608.   FLOAT aa1, aa2;

  609.   FLOAT bb1, bb2;

  610.   FLOAT cc1, cc2;

  611. 

  612.   int i, j, k;

  613. 

  614.   ldc *= 2;

  615. 

  616.   for (i = 0; i < n; i++) {

  617. 

  618.     bb1 = *(b + i * 2 + 0);

  619.     bb2 = *(b + i * 2 + 1);

  620. 

  621.     for (j = 0; j < m; j ++) {

  622.       aa1 = *(c + j * 2 + 0 + i * ldc);

  623.       aa2 = *(c + j * 2 + 1 + i * ldc);

  624. 

  625. #ifndef CONJ

  626.       cc1 = aa1 * bb1 - aa2 * bb2;

  627.       cc2 = aa1 * bb2 + aa2 * bb1;

  628. #else

  629.       cc1 =  aa1 * bb1 + aa2 * bb2;

  630.       cc2 = -aa1 * bb2 + aa2 * bb1;

  631. #endif

  632. 

  633.       *(a + 0) = cc1;

  634.       *(a + 1) = cc2;

  635.       *(c + j * 2 + 0 + i * ldc) = cc1;

  636.       *(c + j * 2 + 1 + i * ldc) = cc2;

  637.       a += 2;

  638. 

  639.       for (k = i + 1; k < n; k ++){

  640. #ifndef CONJ

  641. 	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);

  642. 	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);

  643. #else

  644. 	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);

  645. 	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);

  646. #endif

  647.       }

  648. 

  649.     }

  650.     b += n * 2;

  651.   }

  652. }

  653. 

  654. #endif

  655. 

  656. 

  657. int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,

  658. #ifdef COMPLEX

  659. 	   FLOAT dummy2,

  660. #endif

  661. 	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){

  662. 

  663.   FLOAT *aa, *cc;

  664.   BLASLONG  kk;

  665.   BLASLONG i, j;

  666. 

  667.   size_t vl = VSETVL_MAX;

  668. 

  669.   //fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  k = %4ld offset = %4ld\n", __FILE__, __FUNCTION__, m, n, k, offset); // Debug

  670. 

  671. 

  672.   j = (n >> GEMM_UNROLL_N_SHIFT);

  673.   kk = -offset;

  674. 

  675.   while (j > 0) {

  676. 

  677.     aa = a;

  678.     cc = c;

  679. 

  680.     i = vl;

  681. 

  682.     if (i <= m) {

  683.       do {

  684. 	if (kk > 0) {

  685. 	  GEMM_KERNEL(vl, GEMM_UNROLL_N, kk, dm1,

  686. #ifdef COMPLEX

  687. 		      ZERO,

  688. #endif

  689. 		      aa, b, cc, ldc);

  690. 	}

  691. 

  692. 	solve(vl, GEMM_UNROLL_N,

  693. 	      aa + kk * vl * COMPSIZE,

  694. 	      b  + kk * GEMM_UNROLL_N * COMPSIZE,

  695. 	      cc, ldc);

  696. 

  697. 	aa += vl * k * COMPSIZE;

  698. 	cc += vl     * COMPSIZE;

  699. 	i += vl;

  700.       } while (i <= m);

  701.     }

  702. 

  703. 

  704.     i = m % vl;

  705.     if (i) {

  706.       if (kk > 0) {

  707.         GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,

  708. #ifdef COMPLEX

  709.             ZERO,

  710. #endif

  711.             aa, b, cc, ldc);

  712.       }

  713.       solve(i, GEMM_UNROLL_N,

  714.           aa + kk * i             * COMPSIZE,

  715.           b  + kk * GEMM_UNROLL_N * COMPSIZE,

  716.           cc, ldc);

  717. 

  718.       aa += i * k * COMPSIZE;

  719.       cc += i     * COMPSIZE;

  720. 

  721.     }

  722. 

  723.     kk += GEMM_UNROLL_N;

  724.     b += GEMM_UNROLL_N * k   * COMPSIZE;

  725.     c += GEMM_UNROLL_N * ldc * COMPSIZE;

  726.     j --;

  727.   }

  728. 

  729.   if (n & (GEMM_UNROLL_N - 1)) {

  730. 

  731.     j = (GEMM_UNROLL_N >> 1);

  732.     while (j > 0) {

  733.       if (n & j) {

  734. 

  735. 	aa = a;

  736. 	cc = c;

  737. 

  738.   i = vl;

  739. 

  740. 	while (i <= m) {

  741. 	  if (kk > 0) {

  742. 	    GEMM_KERNEL(vl, j, kk, dm1,

  743. #ifdef COMPLEX

  744. 			ZERO,

  745. #endif

  746. 			aa,

  747. 			b,

  748. 			cc,

  749. 			ldc);

  750. 	  }

  751. 

  752. 	  solve(vl, j,

  753. 		aa + kk * vl * COMPSIZE,

  754. 		b  + kk * j             * COMPSIZE, cc, ldc);

  755. 

  756. 	  aa += vl * k * COMPSIZE;

  757. 	  cc += vl     * COMPSIZE;

  758. 	  i += vl;

  759. 	}

  760. 

  761.   i = m % vl;

  762.   if (i) {

  763. 	      if (kk > 0) {

  764. 		GEMM_KERNEL(i, j, kk, dm1,

  765. #ifdef COMPLEX

  766. 			    ZERO,

  767. #endif

  768. 			    aa,

  769. 			    b,

  770. 			    cc,

  771. 			    ldc);

  772. 	      }

  773. 

  774. 	      solve(i, j,

  775. 		    aa + kk * i * COMPSIZE,

  776. 		    b  + kk * j * COMPSIZE, cc, ldc);

  777. 

  778. 	      aa += i * k * COMPSIZE;

  779. 	      cc += i     * COMPSIZE;

  780. 

  781.   }

  782. 

  783. 	b += j * k   * COMPSIZE;

  784. 	c += j * ldc * COMPSIZE;

  785. 	kk += j;

  786.       }

  787.       j >>= 1;

  788.     }

  789.   }

  790. 

  791.   return 0;

  792. }