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: 0e0d0bce1a
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '0e0d0bce1a'
${ noResults }
OpenBLAS/kernel/power/sgemv_n_8.c

515 lines
12 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. 

  29. /****Note***

  30. UnUsed kernel

  31. This kernel works. But it was not competitive enough to be added in production

  32. It could be used and tested in future or could provide barebone for switching to inline assembly

  33. */

  34. 

  35. #include "common.h"

  36. 

  37. #define NBMAX 4096

  38. 

  39. static void sgemv_kernel_8x8(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, BLASLONG lda4, FLOAT *alpha)

  40. {

  41. 

  42.     BLASLONG i;

  43. 	FLOAT *a0,*a1,*a2,*a3,*b0,*b1,*b2,*b3; 

  44.     FLOAT x0,x1,x2,x3,x4,x5,x6,x7;

  45. 	a0 = ap[0];

  46. 	a1 = ap[1];

  47. 	a2 = ap[2];

  48. 	a3 = ap[3]; 

  49.     b0 = a0 + lda4 ;

  50. 	b1 = a1 + lda4 ;

  51. 	b2 = a2 + lda4 ;

  52. 	b3 = a3 + lda4 ;

  53.     x0 = xo[0] * *alpha;

  54.     x1 = xo[1] * *alpha;

  55.     x2 = xo[2] * *alpha;

  56.     x3 = xo[3] * *alpha;

  57.     x4 = xo[4] * *alpha;

  58.     x5 = xo[5] * *alpha;

  59.     x6 = xo[6] * *alpha;

  60.     x7 = xo[7] * *alpha;

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

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

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

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

  65.     __vector float* vb0 = (__vector float*)b0;

  66.     __vector float* vb1 = (__vector float*)b1;

  67.     __vector float* vb2 = (__vector float*)b2;

  68.     __vector float* vb3 = (__vector float*)b3; 

  69.     

  70.     register __vector float   v_x0 = {x0,x0,x0,x0};

  71.     register __vector float   v_x1 = {x1,x1,x1,x1};

  72.     register __vector float   v_x2 = {x2,x2,x2,x2};

  73.     register __vector float   v_x3 = {x3,x3,x3,x3};

  74.     register __vector float   v_x4 = {x4,x4,x4,x4};

  75.     register __vector float   v_x5 = {x5,x5,x5,x5};

  76.     register __vector float   v_x6 = {x6,x6,x6,x6};

  77.     register __vector float   v_x7 = {x7,x7,x7,x7};

  78.     __vector float* v_y =(__vector float*)y;   

  79.  

  80.     for ( i=0; i< n/4; i+=2)

  81.     {

  82.         register __vector float vy_1=v_y[i];

  83.         register __vector float vy_2=v_y[i+1];

  84.         register __vector float va0_1=va0[i] ; 

  85.         register __vector float va0_2=va0[i+1] ; 

  86.         register __vector float va1_1=va1[i] ; 

  87.         register __vector float va1_2=va1[i+1] ; 

  88.         register __vector float va2_1=va2[i] ; 

  89.         register __vector float va2_2=va2[i+1] ; 

  90.         register __vector float va3_1=va3[i] ; 

  91.         register __vector float va3_2=va3[i+1] ;

  92.         register __vector float vb0_1=vb0[i] ; 

  93.         register __vector float vb0_2=vb0[i+1] ; 

  94.         register __vector float vb1_1=vb1[i] ; 

  95.         register __vector float vb1_2=vb1[i+1] ; 

  96.         register __vector float vb2_1=vb2[i] ; 

  97.         register __vector float vb2_2=vb2[i+1] ; 

  98.         register __vector float vb3_1=vb3[i] ; 

  99.         register __vector float vb3_2=vb3[i+1] ;         

  100.         vy_1   += v_x0 * va0_1  +  v_x1 * va1_1  + v_x2 * va2_1  + v_x3 * va3_1 ;

  101.         vy_1   += v_x4 * vb0_1   +  v_x5 * vb1_1   + v_x6 * vb2_1   + v_x7 * vb3_1 ;

  102.         vy_2   +=  v_x0 * va0_2   +  v_x1 * va1_2   + v_x2 * va2_2   + v_x3 * va3_2 ; 

  103.         vy_2   += v_x4 * vb0_2   +  v_x5 * vb1_2   + v_x6 * vb2_2   + v_x7 * vb3_2 ;

  104.         v_y[i] =vy_1;

  105.         v_y[i+1] =vy_2;   

  106.     }

  107. 

  108. }

  109. 	 

  110. static void sgemv_kernel_8x4(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)

  111. {

  112.     BLASLONG i;

  113.     FLOAT x0,x1,x2,x3;

  114.     x0 = xo[0] * *alpha;

  115.     x1 = xo[1] * *alpha;

  116.     x2 = xo[2] * *alpha;

  117.     x3 = xo[3] * *alpha;

  118.     __vector float   v_x0 = {x0,x0,x0,x0};

  119.     __vector float   v_x1 = {x1,x1,x1,x1};

  120.     __vector float   v_x2 = {x2,x2,x2,x2};

  121.     __vector float   v_x3 = {x3,x3,x3,x3};

  122.     __vector float* v_y =(__vector float*)y;      

  123.     __vector float* va0 = (__vector float*)ap[0];

  124.     __vector float* va1 = (__vector float*)ap[1];

  125.     __vector float* va2 = (__vector float*)ap[2];

  126.     __vector float* va3 = (__vector float*)ap[3]; 

  127.  

  128.     for ( i=0; i< n/4; i+=2 )

  129.     {

  130.         register __vector float vy_1=v_y[i];

  131.         register __vector float vy_2=v_y[i+1];

  132.         register __vector float va0_1=va0[i] ; 

  133.         register __vector float va0_2=va0[i+1] ; 

  134.         register __vector float va1_1=va1[i] ; 

  135.         register __vector float va1_2=va1[i+1] ; 

  136.         register __vector float va2_1=va2[i] ; 

  137.         register __vector float va2_2=va2[i+1] ; 

  138.         register __vector float va3_1=va3[i] ; 

  139.         register __vector float va3_2=va3[i+1] ;      

  140.         vy_1   += v_x0 * va0_1  +  v_x1 * va1_1  + v_x2 * va2_1  + v_x3 * va3_1 ;

  141.         vy_2   +=  v_x0 * va0_2   +  v_x1 * va1_2   + v_x2 * va2_2   + v_x3 * va3_2 ;

  142.         v_y[i] =vy_1;

  143.         v_y[i+1] =vy_2;   

  144.     }

  145.   

  146. } 

  147. 

  148. static void sgemv_kernel_8x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)

  149. {

  150. 

  151.     BLASLONG i;

  152.     FLOAT x0,x1;

  153.     x0 = x[0] * *alpha;

  154.     x1 = x[1] * *alpha; 

  155.     __vector float   v_x0 = {x0,x0,x0,x0};

  156.     __vector float   v_x1 = {x1,x1,x1,x1}; 

  157.     __vector float* v_y =(__vector float*)y;      

  158.     __vector float* va0 = (__vector float*)ap[0];

  159.     __vector float* va1 = (__vector float*)ap[1]; 

  160.  

  161.     for ( i=0; i< n/4; i+=2 )

  162.     { 

  163.         v_y[i]   += v_x0 * va0[i]   +  v_x1 * va1[i] ;

  164.         v_y[i+1]  += v_x0 * va0[i+1]   +  v_x1 * va1[i+1] ;     

  165.     }

  166. 

  167. } 

  168.  

  169.  

  170. static void sgemv_kernel_8x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)

  171. {

  172. 

  173.     BLASLONG i;

  174.     FLOAT x0 ;

  175.     x0 = x[0] * *alpha; 

  176.     __vector float   v_x0 = {x0,x0,x0,x0}; 

  177.     __vector float* v_y =(__vector float*)y;      

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

  179.  

  180.     for ( i=0; i< n/4; i+=2 )

  181.     { 

  182.         v_y[i]   += v_x0 * va0[i]   ;

  183.         v_y[i+1] +=   v_x0 * va0[i+1]   ;        

  184.     }

  185. 

  186. }

  187.  

  188. static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)

  189. {

  190.     BLASLONG i;

  191.         

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

  193.             *dest += *src;

  194.             src++;

  195.             dest += inc_dest;

  196.     }

  197.     return;

  198.      

  199. 

  200. }

  201. 

  202. int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)

  203. {

  204. 	BLASLONG i;

  205. 	FLOAT *a_ptr;

  206. 	FLOAT *x_ptr;

  207. 	FLOAT *y_ptr;

  208. 	FLOAT *ap[4];

  209. 	BLASLONG n1;

  210. 	BLASLONG m1;

  211. 	BLASLONG m2;

  212. 	BLASLONG m3;

  213. 	BLASLONG n2;

  214. 	BLASLONG lda4 =  lda << 2;

  215. 	BLASLONG lda8 =  lda << 3;

  216. 	FLOAT xbuffer[8] __attribute__((aligned(16)));

  217. 	FLOAT *ybuffer;

  218. 

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

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

  221. 

  222. 	ybuffer = buffer;

  223. 	

  224.         if ( inc_x == 1 )

  225. 	{

  226. 		n1 = n >> 3 ;

  227. 		n2 = n &  7 ;

  228. 	}

  229. 	else

  230. 	{

  231. 		n1 = n >> 2 ;

  232. 		n2 = n &  3 ;

  233. 

  234. 	}

  235. 	 

  236.         m3 = m & 7  ;

  237.         m1 = m - m3;

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

  239. 

  240. 

  241. 	y_ptr = y;

  242. 

  243. 	BLASLONG NB = NBMAX;

  244. 

  245. 	while ( NB == NBMAX )

  246. 	{

  247. 		

  248. 		m1 -= NB;

  249. 		if ( m1 < 0)

  250. 		{

  251. 			if ( m2 == 0 ) break;	

  252. 			NB = m2;

  253. 		}

  254. 		

  255. 		a_ptr = a;

  256. 		x_ptr = x;

  257. 		

  258. 		ap[0] = a_ptr;

  259. 		ap[1] = a_ptr + lda;

  260. 		ap[2] = ap[1] + lda;

  261. 		ap[3] = ap[2] + lda;

  262. 

  263. 		if ( inc_y != 1 )

  264. 			memset(ybuffer,0,NB*4);

  265. 		else

  266. 			ybuffer = y_ptr;

  267. 

  268. 		if ( inc_x == 1 )

  269. 		{

  270. 

  271. 

  272. 			for( i = 0; i < n1 ; i++)

  273. 			{

  274. 				sgemv_kernel_8x8(NB,ap,x_ptr,ybuffer,lda4,&alpha);

  275. 				ap[0] += lda8; 

  276. 				ap[1] += lda8; 

  277. 				ap[2] += lda8; 

  278. 				ap[3] += lda8; 

  279. 				a_ptr += lda8;

  280. 				x_ptr += 8;	

  281. 			}

  282. 

  283. 

  284. 			if ( n2 & 4 )

  285. 			{

  286. 				sgemv_kernel_8x4(NB,ap,x_ptr,ybuffer,&alpha);

  287. 				ap[0] += lda4; 

  288. 				ap[1] += lda4; 

  289. 				ap[2] += lda4; 

  290. 				ap[3] += lda4; 

  291. 				a_ptr += lda4;

  292. 				x_ptr += 4;	

  293. 			}

  294. 

  295. 			if ( n2 & 2 )

  296. 			{

  297. 				sgemv_kernel_8x2(NB,ap,x_ptr,ybuffer,&alpha);

  298. 				a_ptr += lda*2;

  299. 				x_ptr += 2;	

  300. 			}

  301. 

  302. 

  303. 			if ( n2 & 1 )

  304. 			{

  305. 				sgemv_kernel_8x1(NB,a_ptr,x_ptr,ybuffer,&alpha); 

  306.                 a_ptr += lda;

  307.                 x_ptr += 1;   

  308. 			}

  309. 

  310. 

  311. 		}

  312. 		else

  313. 		{

  314. 

  315. 			for( i = 0; i < n1 ; i++)

  316. 			{

  317. 				xbuffer[0] = x_ptr[0];

  318. 				x_ptr += inc_x;	

  319. 				xbuffer[1] =  x_ptr[0];

  320. 				x_ptr += inc_x;	

  321. 				xbuffer[2] =  x_ptr[0];

  322. 				x_ptr += inc_x;	

  323. 				xbuffer[3] = x_ptr[0];

  324. 				x_ptr += inc_x;	

  325. 				sgemv_kernel_8x4(NB,ap,xbuffer,ybuffer,&alpha);

  326. 				ap[0] += lda4; 

  327. 				ap[1] += lda4; 

  328. 				ap[2] += lda4; 

  329. 				ap[3] += lda4; 

  330. 				a_ptr += lda4;

  331. 			}

  332. 

  333. 			for( i = 0; i < n2 ; i++)

  334. 			{

  335. 				xbuffer[0] = x_ptr[0];

  336. 				x_ptr += inc_x;	

  337. 				sgemv_kernel_8x1(NB,a_ptr,xbuffer,ybuffer,&alpha);

  338. 				a_ptr += lda;

  339. 

  340. 			}

  341. 

  342. 		}

  343. 

  344. 		a     += NB;

  345. 		if ( inc_y != 1 )

  346. 		{

  347. 			add_y(NB,ybuffer,y_ptr,inc_y);

  348. 			y_ptr += NB * inc_y;

  349. 		}

  350. 		else

  351. 			y_ptr += NB ;

  352. 

  353. 	}

  354. 

  355. 	 

  356. 	if ( m3 & 4 )

  357. 	{

  358. 		a_ptr = a;

  359. 		x_ptr = x;

  360. 		FLOAT temp0 = 0.0;

  361. 		FLOAT temp1 = 0.0;

  362. 		FLOAT temp2 = 0.0;

  363. 		FLOAT temp3 = 0.0;		

  364. 		if ( lda == 4 && inc_x ==1 )

  365. 		{

  366. 

  367. 			for( i = 0; i < ( n & -4 ); i+=4 )

  368. 			{

  369. 

  370. 				temp0 += a_ptr[0] * x_ptr[0] + a_ptr[4] * x_ptr[1];

  371. 				temp1 += a_ptr[1] * x_ptr[0] + a_ptr[5] * x_ptr[1];

  372. 				temp2 += a_ptr[2] * x_ptr[0] + a_ptr[6] * x_ptr[1];

  373. 				temp3 += a_ptr[3] * x_ptr[0] + a_ptr[7] * x_ptr[1];

  374. 

  375. 				temp0 += a_ptr[8] * x_ptr[2] + a_ptr[12]  * x_ptr[3];

  376. 				temp1 += a_ptr[9] * x_ptr[2] + a_ptr[13] * x_ptr[3];

  377. 				temp2 += a_ptr[10] * x_ptr[2] + a_ptr[14] * x_ptr[3];

  378. 				temp3 += a_ptr[11] * x_ptr[2] + a_ptr[15] * x_ptr[3];

  379. 

  380. 				a_ptr += 16;

  381. 				x_ptr += 4;

  382. 			}

  383. 

  384. 			for( ; i < n; i++ )

  385. 			{

  386. 				temp0 += a_ptr[0] * x_ptr[0];

  387. 				temp1 += a_ptr[1] * x_ptr[0];

  388. 				temp2 += a_ptr[2] * x_ptr[0];

  389. 				temp3 += a_ptr[3] * x_ptr[0] ;

  390. 				a_ptr +=4;

  391. 				x_ptr ++;

  392. 			}

  393. 

  394. 		}

  395. 		else

  396. 		{

  397. 

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

  399. 			{

  400. 				temp0 += a_ptr[0] * x_ptr[0];

  401. 				temp1 += a_ptr[1] * x_ptr[0];

  402. 				temp2 += a_ptr[2] * x_ptr[0];

  403. 				temp3 += a_ptr[3] * x_ptr[0];

  404. 				a_ptr += lda;

  405. 				x_ptr += inc_x;

  406. 

  407. 

  408. 			}

  409. 

  410. 		}

  411. 		y_ptr[0] += alpha * temp0;

  412. 		y_ptr += inc_y;

  413. 		y_ptr[0] += alpha * temp1;

  414. 		y_ptr += inc_y;

  415. 		y_ptr[0] += alpha * temp2;

  416. 		y_ptr += inc_y;

  417. 		y_ptr[0] += alpha * temp3; 

  418. 		y_ptr += inc_y;

  419.         a     += 4;

  420. 	}

  421. 

  422. 

  423. 	if ( m3 & 2 )

  424. 	{

  425. 		a_ptr = a;

  426. 		x_ptr = x;

  427. 		FLOAT temp0 = 0.0;

  428. 		FLOAT temp1 = 0.0;

  429. 		if ( lda == 2 && inc_x ==1 )

  430. 		{

  431. 

  432. 			for( i = 0; i < (n & -4) ; i+=4 )

  433. 			{

  434. 				temp0 += a_ptr[0] * x_ptr[0] + a_ptr[2] * x_ptr[1];

  435. 				temp1 += a_ptr[1] * x_ptr[0] + a_ptr[3] * x_ptr[1];

  436. 				temp0 += a_ptr[4] * x_ptr[2] + a_ptr[6] * x_ptr[3];

  437. 				temp1 += a_ptr[5] * x_ptr[2] + a_ptr[7] * x_ptr[3];

  438. 				a_ptr += 8;

  439. 				x_ptr += 4;

  440. 

  441. 			}

  442. 

  443. 

  444. 			for( ; i < n; i++ )

  445. 			{

  446. 				temp0 += a_ptr[0]   * x_ptr[0];

  447. 				temp1 += a_ptr[1]   * x_ptr[0];

  448. 				a_ptr += 2;

  449. 				x_ptr ++;

  450. 			}

  451. 

  452. 		}

  453. 		else

  454. 		{

  455. 

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

  457. 			{

  458. 				temp0 += a_ptr[0] * x_ptr[0];

  459. 				temp1 += a_ptr[1] * x_ptr[0];

  460. 				a_ptr += lda;

  461. 				x_ptr += inc_x;

  462. 

  463. 

  464. 			}

  465. 

  466. 		}

  467. 		y_ptr[0] += alpha * temp0;

  468. 		y_ptr += inc_y;

  469. 		y_ptr[0] += alpha * temp1;

  470.  		y_ptr += inc_y;

  471.         a     += 2;

  472. 	}

  473. 

  474. 	if ( m3 & 1 )

  475. 	{

  476. 		a_ptr = a;

  477. 		x_ptr = x;

  478. 		FLOAT temp = 0.0;

  479. 		if ( lda == 1 && inc_x ==1 )

  480. 		{

  481. 

  482. 			for( i = 0; i < (n & -4); i+=4 )

  483. 			{

  484. 				temp += a_ptr[i] * x_ptr[i] + a_ptr[i+1] * x_ptr[i+1] + a_ptr[i+2] * x_ptr[i+2] + a_ptr[i+3] * x_ptr[i+3];

  485. 	

  486. 			}

  487. 

  488. 			for( ; i < n; i++ )

  489. 			{

  490. 				temp += a_ptr[i] * x_ptr[i];

  491. 			}

  492. 

  493. 		}

  494. 		else

  495. 		{

  496. 

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

  498. 			{

  499. 				temp += a_ptr[0] * x_ptr[0];

  500. 				a_ptr += lda;

  501. 				x_ptr += inc_x;

  502. 			}

  503. 

  504. 		}

  505. 		y_ptr[0] += alpha * temp;

  506.  

  507.  

  508. 	}

  509. 

  510. 

  511. 	return(0);

  512. }

  513.