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: f9c5023e04
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'f9c5023e04'
${ noResults }
OpenBLAS/kernel/power/cgemv_t.c

572 lines
20 kB
Raw Blame History 

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

  2. Copyright (c) 2019, 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. #define NBMAX 1024 

  31. #include <altivec.h> 

  32. static const unsigned char swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};

  33. 

  34. static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {

  35.     BLASLONG i;

  36.     FLOAT *a0, *a1, *a2, *a3;

  37.     a0 = ap;

  38.     a1 = ap + lda;

  39.     a2 = a1 + lda;

  40.     a3 = a2 + lda;

  41.     __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);

  42.     //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)

  43.     register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};

  44.     register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0};

  45.     register __vector float vtemp1_p = {0.0, 0.0,0.0,0.0};

  46.     register __vector float vtemp1_r = {0.0, 0.0,0.0,0.0};

  47.     register __vector float vtemp2_p = {0.0, 0.0,0.0,0.0};

  48.     register __vector float vtemp2_r = {0.0, 0.0,0.0,0.0};

  49.     register __vector float vtemp3_p = {0.0, 0.0,0.0,0.0};

  50.     register __vector float vtemp3_r = {0.0, 0.0,0.0,0.0};

  51.     __vector float* va0 = (__vector float*) a0;

  52.     __vector float* va1 = (__vector float*) a1;

  53.     __vector float* va2 = (__vector float*) a2;

  54.     __vector float* va3 = (__vector float*) a3;

  55.     __vector float* v_x = (__vector float*) x;

  56. 

  57.     for (i = 0; i < n / 2; i+=2) {

  58.         register __vector float vx_0  = v_x[i]; 

  59.         register __vector float vx_1  = v_x[i+1];         

  60.         register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);

  61.         register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);

  62. 

  63.         vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;

  64.         vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; 

  65.         vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1];

  66.         vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1]; 

  67.         vtemp2_p += vx_0*va2[i] + vx_1*va2[i+1];

  68.         vtemp2_r += vxr_0*va2[i] + vxr_1*va2[i+1]; 

  69.         vtemp3_p += vx_0*va3[i] + vx_1*va3[i+1];

  70.         vtemp3_r += vxr_0*va3[i] + vxr_1*va3[i+1]; 

  71. 

  72.     }

  73. 

  74. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  75. 

  76.     register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];

  77.     register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1] + vtemp0_r[2] + vtemp0_r[3];

  78. 

  79.     register FLOAT temp_r1 = vtemp1_p[0] - vtemp1_p[1] + vtemp1_p[2] - vtemp1_p[3];

  80.     register FLOAT temp_i1 = vtemp1_r[0] + vtemp1_r[1] + vtemp1_r[2] + vtemp1_r[3];

  81. 

  82.     register FLOAT temp_r2 = vtemp2_p[0] - vtemp2_p[1] + vtemp2_p[2] - vtemp2_p[3];

  83.     register FLOAT temp_i2 = vtemp2_r[0] + vtemp2_r[1] + vtemp2_r[2] + vtemp2_r[3];

  84. 

  85.     register FLOAT temp_r3 = vtemp3_p[0] - vtemp3_p[1] + vtemp3_p[2] - vtemp3_p[3];

  86.     register FLOAT temp_i3 = vtemp3_r[0] + vtemp3_r[1] + vtemp3_r[2] + vtemp3_r[3];

  87. 

  88. #else

  89.     register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1] + vtemp0_p[2] + vtemp0_p[3];

  90.     register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1] + vtemp0_r[2] - vtemp0_r[3];

  91. 

  92.     register FLOAT temp_r1 = vtemp1_p[0] + vtemp1_p[1] + vtemp1_p[2] + vtemp1_p[3];

  93.     register FLOAT temp_i1 = vtemp1_r[0] - vtemp1_r[1] + vtemp1_r[2] - vtemp1_r[3];

  94. 

  95.     register FLOAT temp_r2 = vtemp2_p[0] + vtemp2_p[1] + vtemp2_p[2] + vtemp2_p[3];

  96.     register FLOAT temp_i2 = vtemp2_r[0] - vtemp2_r[1] + vtemp2_r[2] - vtemp2_r[3];

  97. 

  98.     register FLOAT temp_r3 = vtemp3_p[0] + vtemp3_p[1] + vtemp3_p[2] + vtemp3_p[3];

  99.     register FLOAT temp_i3 = vtemp3_r[0] - vtemp3_r[1] + vtemp3_r[2] - vtemp3_r[3];

  100. 

  101. #endif    

  102. 

  103. #if !defined(XCONJ)

  104. 

  105.     y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;

  106.     y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;

  107.     y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;

  108.     y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;

  109.     y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;

  110.     y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;

  111.     y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;

  112.     y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;

  113. 

  114. #else

  115. 

  116.     y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;

  117.     y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;

  118.     y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;

  119.     y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;

  120.     y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;

  121.     y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;

  122.     y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;

  123.     y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;

  124. 

  125. #endif

  126. 

  127. }

  128.  

  129. 

  130. static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {

  131.     BLASLONG i;

  132.     FLOAT *a0, *a1;

  133.     a0 = ap;

  134.     a1 = ap + lda; 

  135.     __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);

  136.     //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)

  137.     register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};

  138.     register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0};

  139.     register __vector float vtemp1_p = {0.0, 0.0,0.0,0.0};

  140.     register __vector float vtemp1_r = {0.0, 0.0,0.0,0.0}; 

  141.     __vector float* va0 = (__vector float*) a0;

  142.     __vector float* va1 = (__vector float*) a1; 

  143.     __vector float* v_x = (__vector float*) x;

  144. 

  145.     for (i = 0; i < n / 2; i+=2) {

  146.         register __vector float vx_0  = v_x[i]; 

  147.         register __vector float vx_1  = v_x[i+1];         

  148.         register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);

  149.         register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);

  150. 

  151.         vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;

  152.         vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; 

  153.         vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1];

  154.         vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1];  

  155. 

  156.     }

  157. 

  158. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  159. 

  160.     register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];

  161.     register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1] + vtemp0_r[2] + vtemp0_r[3];

  162. 

  163.     register FLOAT temp_r1 = vtemp1_p[0] - vtemp1_p[1] + vtemp1_p[2] - vtemp1_p[3];

  164.     register FLOAT temp_i1 = vtemp1_r[0] + vtemp1_r[1] + vtemp1_r[2] + vtemp1_r[3];

  165.  

  166. 

  167. #else

  168.     register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1] + vtemp0_p[2] + vtemp0_p[3];

  169.     register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1] + vtemp0_r[2] - vtemp0_r[3];

  170. 

  171.     register FLOAT temp_r1 = vtemp1_p[0] + vtemp1_p[1] + vtemp1_p[2] + vtemp1_p[3];

  172.     register FLOAT temp_i1 = vtemp1_r[0] - vtemp1_r[1] + vtemp1_r[2] - vtemp1_r[3]; 

  173. 

  174. #endif    

  175. 

  176. #if !defined(XCONJ)

  177. 

  178.     y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;

  179.     y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;

  180.     y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;

  181.     y[3] += alpha_r * temp_i1 + alpha_i * temp_r1; 

  182. 

  183. #else

  184. 

  185.     y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;

  186.     y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;

  187.     y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;

  188.     y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1; 

  189. 

  190. #endif

  191.   

  192. }

  193.  

  194. 

  195. static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {

  196.    BLASLONG i;  

  197.     __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);

  198.     //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)

  199.     register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};

  200.     register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0}; 

  201.     __vector float* va0 = (__vector float*) ap; 

  202.     __vector float* v_x = (__vector float*) x;

  203. 

  204.     for (i = 0; i < n / 2; i+=2) {

  205.         register __vector float vx_0  = v_x[i]; 

  206.         register __vector float vx_1  = v_x[i+1];         

  207.         register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);

  208.         register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);

  209. 

  210.         vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;

  211.         vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1];  

  212. 

  213.     }

  214. 

  215. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  216. 

  217.     register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];

  218.     register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1] + vtemp0_r[2] + vtemp0_r[3]; 

  219. 

  220. #else

  221.     register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1] + vtemp0_p[2] + vtemp0_p[3];

  222.     register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1] + vtemp0_r[2] - vtemp0_r[3]; 

  223. 

  224. #endif    

  225. 

  226. #if !defined(XCONJ)

  227. 

  228.     y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;

  229.     y[1] += alpha_r * temp_i0 + alpha_i * temp_r0; 

  230. 

  231. #else

  232. 

  233.     y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;

  234.     y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0; 

  235. 

  236. #endif

  237. 

  238. 

  239. }

  240.  

  241. static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src) {

  242.     BLASLONG i;

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

  244.         *dest = *src;

  245.         *(dest + 1) = *(src + 1);

  246.         dest += 2;

  247.         src += inc_src;

  248.     }

  249. }

  250. 

  251. int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer) {

  252.     BLASLONG i;

  253.     BLASLONG j;

  254.     FLOAT *a_ptr;

  255.     FLOAT *x_ptr;

  256.     FLOAT *y_ptr;

  257. 

  258.     BLASLONG n1;

  259.     BLASLONG m1;

  260.     BLASLONG m2;

  261.     BLASLONG m3;

  262.     BLASLONG n2;

  263. 

  264.     FLOAT ybuffer[8], *xbuffer;

  265. 

  266.     if (m < 1) return (0);

  267.     if (n < 1) return (0);

  268. 

  269.     inc_x <<= 1;

  270.     inc_y <<= 1;

  271.     lda <<= 1;

  272. 

  273.     xbuffer = buffer;

  274. 

  275.     n1 = n >> 2;

  276.     n2 = n & 3;

  277. 

  278.     m3 = m & 3;

  279.     m1 = m - m3;

  280.     m2 = (m & (NBMAX - 1)) - m3;

  281. 

  282.     BLASLONG NB = NBMAX;

  283. 

  284.     while (NB == NBMAX) {

  285. 

  286.         m1 -= NB;

  287.         if (m1 < 0) {

  288.             if (m2 == 0) break;

  289.             NB = m2;

  290.         }

  291. 

  292.         y_ptr = y;

  293.         a_ptr = a;

  294.         x_ptr = x;

  295. 

  296.         if (inc_x != 2)

  297.             copy_x(NB, x_ptr, xbuffer, inc_x);

  298.         else

  299.             xbuffer = x_ptr;

  300. 

  301.         if (inc_y == 2) {

  302. 

  303.             for (i = 0; i < n1; i++) {

  304.                 cgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);

  305.                 a_ptr += lda << 2;

  306.                 y_ptr += 8;

  307. 

  308.             }

  309. 

  310.             if (n2 & 2) {

  311.                 cgemv_kernel_4x2(NB, lda, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);

  312.                 a_ptr += lda << 1;

  313.                 y_ptr += 4;

  314. 

  315.             }

  316. 

  317.             if (n2 & 1) {

  318.                 cgemv_kernel_4x1(NB, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);

  319.                 a_ptr += lda;

  320.                 y_ptr += 2;

  321. 

  322.             }

  323. 

  324.         } else {

  325. 

  326.             for (i = 0; i < n1; i++) {

  327.                 memset(ybuffer, 0, sizeof (ybuffer));

  328.                 cgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, ybuffer, alpha_r, alpha_i);

  329. 

  330.                 a_ptr += lda << 2;

  331. 

  332.                 y_ptr[0] += ybuffer[0];

  333.                 y_ptr[1] += ybuffer[1];

  334.                 y_ptr += inc_y;

  335.                 y_ptr[0] += ybuffer[2];

  336.                 y_ptr[1] += ybuffer[3];

  337.                 y_ptr += inc_y;

  338.                 y_ptr[0] += ybuffer[4];

  339.                 y_ptr[1] += ybuffer[5];

  340.                 y_ptr += inc_y;

  341.                 y_ptr[0] += ybuffer[6];

  342.                 y_ptr[1] += ybuffer[7];

  343.                 y_ptr += inc_y;

  344. 

  345.             }

  346. 

  347.             for (i = 0; i < n2; i++) {

  348.                 memset(ybuffer, 0, sizeof (ybuffer));

  349.                 cgemv_kernel_4x1(NB, a_ptr, xbuffer, ybuffer, alpha_r, alpha_i);

  350.                 a_ptr += lda;

  351.                 y_ptr[0] += ybuffer[0];

  352.                 y_ptr[1] += ybuffer[1];

  353.                 y_ptr += inc_y;

  354. 

  355.             }

  356. 

  357.         }

  358.         a += 2 * NB;

  359.         x += NB * inc_x;

  360.     }

  361. 

  362.     if (m3 == 0) return (0);

  363. 

  364.     x_ptr = x;

  365.     j = 0;

  366.     a_ptr = a;

  367.     y_ptr = y;

  368. 

  369.     if (m3 == 3) {

  370. 

  371.         FLOAT temp_r;

  372.         FLOAT temp_i;

  373.         FLOAT x0 = x_ptr[0];

  374.         FLOAT x1 = x_ptr[1];

  375.         x_ptr += inc_x;

  376.         FLOAT x2 = x_ptr[0];

  377.         FLOAT x3 = x_ptr[1];

  378.         x_ptr += inc_x;

  379.         FLOAT x4 = x_ptr[0];

  380.         FLOAT x5 = x_ptr[1];

  381.         while (j < n) {

  382. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  383.             temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;

  384.             temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;

  385.             temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;

  386.             temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;

  387.             temp_r += a_ptr[4] * x4 - a_ptr[5] * x5;

  388.             temp_i += a_ptr[4] * x5 + a_ptr[5] * x4;

  389. #else

  390. 

  391.             temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;

  392.             temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;

  393.             temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;

  394.             temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;

  395.             temp_r += a_ptr[4] * x4 + a_ptr[5] * x5;

  396.             temp_i += a_ptr[4] * x5 - a_ptr[5] * x4;

  397. #endif

  398. 

  399. #if !defined(XCONJ) 

  400.             y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;

  401.             y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;

  402. #else

  403.             y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;

  404.             y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;

  405. #endif

  406. 

  407.             a_ptr += lda;

  408.             y_ptr += inc_y;

  409.             j++;

  410.         }

  411.         return (0);

  412.     }

  413. 

  414.     if (m3 == 2) {

  415. 

  416.         FLOAT temp_r;

  417.         FLOAT temp_i;

  418.         FLOAT temp_r1;

  419.         FLOAT temp_i1;

  420.         FLOAT x0 = x_ptr[0];

  421.         FLOAT x1 = x_ptr[1];

  422.         x_ptr += inc_x;

  423.         FLOAT x2 = x_ptr[0];

  424.         FLOAT x3 = x_ptr[1];

  425. 

  426.         while (j < (n & -2)) {

  427. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  428.             temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;

  429.             temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;

  430.             temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;

  431.             temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;

  432.             a_ptr += lda;

  433.             temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;

  434.             temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;

  435.             temp_r1 += a_ptr[2] * x2 - a_ptr[3] * x3;

  436.             temp_i1 += a_ptr[2] * x3 + a_ptr[3] * x2;

  437. #else

  438. 

  439.             temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;

  440.             temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;

  441.             temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;

  442.             temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;

  443.             a_ptr += lda;

  444.             temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;

  445.             temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;

  446.             temp_r1 += a_ptr[2] * x2 + a_ptr[3] * x3;

  447.             temp_i1 += a_ptr[2] * x3 - a_ptr[3] * x2;

  448. #endif

  449. 

  450. #if !defined(XCONJ) 

  451.             y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;

  452.             y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;

  453.             y_ptr += inc_y;

  454.             y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;

  455.             y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;

  456. #else

  457.             y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;

  458.             y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;

  459.             y_ptr += inc_y;

  460.             y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;

  461.             y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;

  462. #endif

  463. 

  464.             a_ptr += lda;

  465.             y_ptr += inc_y;

  466.             j += 2;

  467.         }

  468. 

  469.         while (j < n) {

  470. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  471.             temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;

  472.             temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;

  473.             temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;

  474.             temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;

  475. #else

  476. 

  477.             temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;

  478.             temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;

  479.             temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;

  480.             temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;

  481. #endif

  482. 

  483. #if !defined(XCONJ) 

  484.             y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;

  485.             y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;

  486. #else

  487.             y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;

  488.             y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;

  489. #endif

  490. 

  491.             a_ptr += lda;

  492.             y_ptr += inc_y;

  493.             j++;

  494.         }

  495. 

  496.         return (0);

  497.     }

  498. 

  499.     if (m3 == 1) {

  500. 

  501.         FLOAT temp_r;

  502.         FLOAT temp_i;

  503.         FLOAT temp_r1;

  504.         FLOAT temp_i1;

  505.         FLOAT x0 = x_ptr[0];

  506.         FLOAT x1 = x_ptr[1];

  507. 

  508.         while (j < (n & -2)) {

  509. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  510.             temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;

  511.             temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;

  512.             a_ptr += lda;

  513.             temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;

  514.             temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;

  515. #else

  516. 

  517.             temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;

  518.             temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;

  519.             a_ptr += lda;

  520.             temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;

  521.             temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;

  522. #endif

  523. 

  524. #if !defined(XCONJ) 

  525.             y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;

  526.             y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;

  527.             y_ptr += inc_y;

  528.             y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;

  529.             y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;

  530. #else

  531.             y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;

  532.             y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;

  533.             y_ptr += inc_y;

  534.             y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;

  535.             y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;

  536. #endif

  537. 

  538.             a_ptr += lda;

  539.             y_ptr += inc_y;

  540.             j += 2;

  541.         }

  542. 

  543.         while (j < n) {

  544. #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

  545.             temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;

  546.             temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;

  547. #else

  548. 

  549.             temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;

  550.             temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;

  551. #endif

  552. 

  553. #if !defined(XCONJ) 

  554.             y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;

  555.             y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;

  556. #else

  557.             y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;

  558.             y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;

  559. #endif

  560. 

  561.             a_ptr += lda;

  562.             y_ptr += inc_y;

  563.             j++;

  564.         }

  565.         return (0);

  566.     }

  567. 

  568.     return (0);

  569. 

  570. }

  571.