falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
39526 Commits
32 Branches
748MB
Tree: 1b51b7ba00
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '1b51b7ba00'
${ noResults }
FFmpeg/libavcodec/ppc/dsputil_altivec.c

965 lines
38KB
Raw Blame History 

  1. /*

  2.  * Copyright (c) 2002 Brian Foley

  3.  * Copyright (c) 2002 Dieter Shirley

  4.  * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>

  5.  *

  6.  * This file is part of Libav.

  7.  *

  8.  * Libav is free software; you can redistribute it and/or

  9.  * modify it under the terms of the GNU Lesser General Public

  10.  * License as published by the Free Software Foundation; either

  11.  * version 2.1 of the License, or (at your option) any later version.

  12.  *

  13.  * Libav is distributed in the hope that it will be useful,

  14.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  15.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  16.  * Lesser General Public License for more details.

  17.  *

  18.  * You should have received a copy of the GNU Lesser General Public

  19.  * License along with Libav; if not, write to the Free Software

  20.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  21.  */

  22. 

  23. #include "config.h"

  24. #if HAVE_ALTIVEC_H

  25. #include <altivec.h>

  26. #endif

  27. #include "libavutil/attributes.h"

  28. #include "libavutil/ppc/types_altivec.h"

  29. #include "libavutil/ppc/util_altivec.h"

  30. #include "libavcodec/avcodec.h"

  31. #include "libavcodec/dsputil.h"

  32. #include "dsputil_altivec.h"

  33. 

  34. static int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  35. {

  36.     int i;

  37.     int s;

  38.     const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);

  39.     vector unsigned char perm1 = vec_lvsl(0, pix2);

  40.     vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));

  41.     vector unsigned char pix2l, pix2r;

  42.     vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;

  43.     vector unsigned int sad;

  44.     vector signed int sumdiffs;

  45. 

  46.     s = 0;

  47.     sad = (vector unsigned int)vec_splat_u32(0);

  48.     for (i = 0; i < h; i++) {

  49.         /* Read unaligned pixels into our vectors. The vectors are as follows:

  50.            pix1v: pix1[0]-pix1[15]

  51.            pix2v: pix2[0]-pix2[15]      pix2iv: pix2[1]-pix2[16] */

  52.         pix1v  = vec_ld( 0, pix1);

  53.         pix2l  = vec_ld( 0, pix2);

  54.         pix2r  = vec_ld(16, pix2);

  55.         pix2v  = vec_perm(pix2l, pix2r, perm1);

  56.         pix2iv = vec_perm(pix2l, pix2r, perm2);

  57. 

  58.         /* Calculate the average vector */

  59.         avgv = vec_avg(pix2v, pix2iv);

  60. 

  61.         /* Calculate a sum of abs differences vector */

  62.         t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));

  63. 

  64.         /* Add each 4 pixel group together and put 4 results into sad */

  65.         sad = vec_sum4s(t5, sad);

  66. 

  67.         pix1 += line_size;

  68.         pix2 += line_size;

  69.     }

  70.     /* Sum up the four partial sums, and put the result into s */

  71.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);

  72.     sumdiffs = vec_splat(sumdiffs, 3);

  73.     vec_ste(sumdiffs, 0, &s);

  74. 

  75.     return s;

  76. }

  77. 

  78. static int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  79. {

  80.     int i;

  81.     int s;

  82.     const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);

  83.     vector unsigned char perm = vec_lvsl(0, pix2);

  84.     vector unsigned char pix2l, pix2r;

  85.     vector unsigned char pix1v, pix2v, pix3v, avgv, t5;

  86.     vector unsigned int sad;

  87.     vector signed int sumdiffs;

  88.     uint8_t *pix3 = pix2 + line_size;

  89. 

  90.     s = 0;

  91.     sad = (vector unsigned int)vec_splat_u32(0);

  92. 

  93.     /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one

  94.        iteration becomes pix2 in the next iteration. We can use this

  95.        fact to avoid a potentially expensive unaligned read, each

  96.        time around the loop.

  97.        Read unaligned pixels into our vectors. The vectors are as follows:

  98.        pix2v: pix2[0]-pix2[15]

  99.        Split the pixel vectors into shorts */

  100.     pix2l = vec_ld( 0, pix2);

  101.     pix2r = vec_ld(15, pix2);

  102.     pix2v = vec_perm(pix2l, pix2r, perm);

  103. 

  104.     for (i = 0; i < h; i++) {

  105.         /* Read unaligned pixels into our vectors. The vectors are as follows:

  106.            pix1v: pix1[0]-pix1[15]

  107.            pix3v: pix3[0]-pix3[15] */

  108.         pix1v = vec_ld(0, pix1);

  109. 

  110.         pix2l = vec_ld( 0, pix3);

  111.         pix2r = vec_ld(15, pix3);

  112.         pix3v = vec_perm(pix2l, pix2r, perm);

  113. 

  114.         /* Calculate the average vector */

  115.         avgv = vec_avg(pix2v, pix3v);

  116. 

  117.         /* Calculate a sum of abs differences vector */

  118.         t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));

  119. 

  120.         /* Add each 4 pixel group together and put 4 results into sad */

  121.         sad = vec_sum4s(t5, sad);

  122. 

  123.         pix1 += line_size;

  124.         pix2v = pix3v;

  125.         pix3 += line_size;

  126. 

  127.     }

  128. 

  129.     /* Sum up the four partial sums, and put the result into s */

  130.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);

  131.     sumdiffs = vec_splat(sumdiffs, 3);

  132.     vec_ste(sumdiffs, 0, &s);

  133.     return s;

  134. }

  135. 

  136. static int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  137. {

  138.     int i;

  139.     int s;

  140.     uint8_t *pix3 = pix2 + line_size;

  141.     const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);

  142.     const vector unsigned short two = (const vector unsigned short)vec_splat_u16(2);

  143.     vector unsigned char avgv, t5;

  144.     vector unsigned char perm1 = vec_lvsl(0, pix2);

  145.     vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));

  146.     vector unsigned char pix2l, pix2r;

  147.     vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;

  148.     vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;

  149.     vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;

  150.     vector unsigned short avghv, avglv;

  151.     vector unsigned short t1, t2, t3, t4;

  152.     vector unsigned int sad;

  153.     vector signed int sumdiffs;

  154. 

  155.     sad = (vector unsigned int)vec_splat_u32(0);

  156. 

  157.     s = 0;

  158. 

  159.     /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one

  160.        iteration becomes pix2 in the next iteration. We can use this

  161.        fact to avoid a potentially expensive unaligned read, as well

  162.        as some splitting, and vector addition each time around the loop.

  163.        Read unaligned pixels into our vectors. The vectors are as follows:

  164.        pix2v: pix2[0]-pix2[15]  pix2iv: pix2[1]-pix2[16]

  165.        Split the pixel vectors into shorts */

  166.     pix2l  = vec_ld( 0, pix2);

  167.     pix2r  = vec_ld(16, pix2);

  168.     pix2v  = vec_perm(pix2l, pix2r, perm1);

  169.     pix2iv = vec_perm(pix2l, pix2r, perm2);

  170. 

  171.     pix2hv  = (vector unsigned short) vec_mergeh(zero, pix2v);

  172.     pix2lv  = (vector unsigned short) vec_mergel(zero, pix2v);

  173.     pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv);

  174.     pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv);

  175.     t1 = vec_add(pix2hv, pix2ihv);

  176.     t2 = vec_add(pix2lv, pix2ilv);

  177. 

  178.     for (i = 0; i < h; i++) {

  179.         /* Read unaligned pixels into our vectors. The vectors are as follows:

  180.            pix1v: pix1[0]-pix1[15]

  181.            pix3v: pix3[0]-pix3[15]      pix3iv: pix3[1]-pix3[16] */

  182.         pix1v = vec_ld(0, pix1);

  183. 

  184.         pix2l  = vec_ld( 0, pix3);

  185.         pix2r  = vec_ld(16, pix3);

  186.         pix3v  = vec_perm(pix2l, pix2r, perm1);

  187.         pix3iv = vec_perm(pix2l, pix2r, perm2);

  188. 

  189.         /* Note that AltiVec does have vec_avg, but this works on vector pairs

  190.            and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding

  191.            would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.

  192.            Instead, we have to split the pixel vectors into vectors of shorts,

  193.            and do the averaging by hand. */

  194. 

  195.         /* Split the pixel vectors into shorts */

  196.         pix3hv  = (vector unsigned short) vec_mergeh(zero, pix3v);

  197.         pix3lv  = (vector unsigned short) vec_mergel(zero, pix3v);

  198.         pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);

  199.         pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);

  200. 

  201.         /* Do the averaging on them */

  202.         t3 = vec_add(pix3hv, pix3ihv);

  203.         t4 = vec_add(pix3lv, pix3ilv);

  204. 

  205.         avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);

  206.         avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);

  207. 

  208.         /* Pack the shorts back into a result */

  209.         avgv = vec_pack(avghv, avglv);

  210. 

  211.         /* Calculate a sum of abs differences vector */

  212.         t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));

  213. 

  214.         /* Add each 4 pixel group together and put 4 results into sad */

  215.         sad = vec_sum4s(t5, sad);

  216. 

  217.         pix1 += line_size;

  218.         pix3 += line_size;

  219.         /* Transfer the calculated values for pix3 into pix2 */

  220.         t1 = t3;

  221.         t2 = t4;

  222.     }

  223.     /* Sum up the four partial sums, and put the result into s */

  224.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);

  225.     sumdiffs = vec_splat(sumdiffs, 3);

  226.     vec_ste(sumdiffs, 0, &s);

  227. 

  228.     return s;

  229. }

  230. 

  231. static int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  232. {

  233.     int i;

  234.     int s;

  235.     const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);

  236.     vector unsigned char perm = vec_lvsl(0, pix2);

  237.     vector unsigned char t1, t2, t3,t4, t5;

  238.     vector unsigned int sad;

  239.     vector signed int sumdiffs;

  240. 

  241.     sad = (vector unsigned int)vec_splat_u32(0);

  242. 

  243. 

  244.     for (i = 0; i < h; i++) {

  245.         /* Read potentially unaligned pixels into t1 and t2 */

  246.         vector unsigned char pix2l = vec_ld( 0, pix2);

  247.         vector unsigned char pix2r = vec_ld(15, pix2);

  248.         t1 = vec_ld(0, pix1);

  249.         t2 = vec_perm(pix2l, pix2r, perm);

  250. 

  251.         /* Calculate a sum of abs differences vector */

  252.         t3 = vec_max(t1, t2);

  253.         t4 = vec_min(t1, t2);

  254.         t5 = vec_sub(t3, t4);

  255. 

  256.         /* Add each 4 pixel group together and put 4 results into sad */

  257.         sad = vec_sum4s(t5, sad);

  258. 

  259.         pix1 += line_size;

  260.         pix2 += line_size;

  261.     }

  262. 

  263.     /* Sum up the four partial sums, and put the result into s */

  264.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);

  265.     sumdiffs = vec_splat(sumdiffs, 3);

  266.     vec_ste(sumdiffs, 0, &s);

  267. 

  268.     return s;

  269. }

  270. 

  271. static int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  272. {

  273.     int i;

  274.     int s;

  275.     const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);

  276.     const vector unsigned char permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};

  277.     vector unsigned char perm1 = vec_lvsl(0, pix1);

  278.     vector unsigned char perm2 = vec_lvsl(0, pix2);

  279.     vector unsigned char t1, t2, t3,t4, t5;

  280.     vector unsigned int sad;

  281.     vector signed int sumdiffs;

  282. 

  283.     sad = (vector unsigned int)vec_splat_u32(0);

  284. 

  285.     for (i = 0; i < h; i++) {

  286.         /* Read potentially unaligned pixels into t1 and t2

  287.            Since we're reading 16 pixels, and actually only want 8,

  288.            mask out the last 8 pixels. The 0s don't change the sum. */

  289.         vector unsigned char pix1l = vec_ld(0, pix1);

  290.         vector unsigned char pix1r = vec_ld(7, pix1);

  291.         vector unsigned char pix2l = vec_ld(0, pix2);

  292.         vector unsigned char pix2r = vec_ld(7, pix2);

  293.         t1 = vec_and(vec_perm(pix1l, pix1r, perm1), permclear);

  294.         t2 = vec_and(vec_perm(pix2l, pix2r, perm2), permclear);

  295. 

  296.         /* Calculate a sum of abs differences vector */

  297.         t3 = vec_max(t1, t2);

  298.         t4 = vec_min(t1, t2);

  299.         t5 = vec_sub(t3, t4);

  300. 

  301.         /* Add each 4 pixel group together and put 4 results into sad */

  302.         sad = vec_sum4s(t5, sad);

  303. 

  304.         pix1 += line_size;

  305.         pix2 += line_size;

  306.     }

  307. 

  308.     /* Sum up the four partial sums, and put the result into s */

  309.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);

  310.     sumdiffs = vec_splat(sumdiffs, 3);

  311.     vec_ste(sumdiffs, 0, &s);

  312. 

  313.     return s;

  314. }

  315. 

  316. static int pix_norm1_altivec(uint8_t *pix, int line_size)

  317. {

  318.     int i;

  319.     int s;

  320.     const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);

  321.     vector unsigned char perm = vec_lvsl(0, pix);

  322.     vector unsigned char pixv;

  323.     vector unsigned int sv;

  324.     vector signed int sum;

  325. 

  326.     sv = (vector unsigned int)vec_splat_u32(0);

  327. 

  328.     s = 0;

  329.     for (i = 0; i < 16; i++) {

  330.         /* Read in the potentially unaligned pixels */

  331.         vector unsigned char pixl = vec_ld( 0, pix);

  332.         vector unsigned char pixr = vec_ld(15, pix);

  333.         pixv = vec_perm(pixl, pixr, perm);

  334. 

  335.         /* Square the values, and add them to our sum */

  336.         sv = vec_msum(pixv, pixv, sv);

  337. 

  338.         pix += line_size;

  339.     }

  340.     /* Sum up the four partial sums, and put the result into s */

  341.     sum = vec_sums((vector signed int) sv, (vector signed int) zero);

  342.     sum = vec_splat(sum, 3);

  343.     vec_ste(sum, 0, &s);

  344. 

  345.     return s;

  346. }

  347. 

  348. /**

  349.  * Sum of Squared Errors for a 8x8 block.

  350.  * AltiVec-enhanced.

  351.  * It's the sad8_altivec code above w/ squaring added.

  352.  */

  353. static int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  354. {

  355.     int i;

  356.     int s;

  357.     const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);

  358.     const vector unsigned char permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};

  359.     vector unsigned char perm1 = vec_lvsl(0, pix1);

  360.     vector unsigned char perm2 = vec_lvsl(0, pix2);

  361.     vector unsigned char t1, t2, t3,t4, t5;

  362.     vector unsigned int sum;

  363.     vector signed int sumsqr;

  364. 

  365.     sum = (vector unsigned int)vec_splat_u32(0);

  366. 

  367.     for (i = 0; i < h; i++) {

  368.         /* Read potentially unaligned pixels into t1 and t2

  369.            Since we're reading 16 pixels, and actually only want 8,

  370.            mask out the last 8 pixels. The 0s don't change the sum. */

  371.         vector unsigned char pix1l = vec_ld(0, pix1);

  372.         vector unsigned char pix1r = vec_ld(7, pix1);

  373.         vector unsigned char pix2l = vec_ld(0, pix2);

  374.         vector unsigned char pix2r = vec_ld(7, pix2);

  375.         t1 = vec_and(vec_perm(pix1l, pix1r, perm1), permclear);

  376.         t2 = vec_and(vec_perm(pix2l, pix2r, perm2), permclear);

  377. 

  378.         /* Since we want to use unsigned chars, we can take advantage

  379.            of the fact that abs(a-b)^2 = (a-b)^2. */

  380. 

  381.         /* Calculate abs differences vector */

  382.         t3 = vec_max(t1, t2);

  383.         t4 = vec_min(t1, t2);

  384.         t5 = vec_sub(t3, t4);

  385. 

  386.         /* Square the values and add them to our sum */

  387.         sum = vec_msum(t5, t5, sum);

  388. 

  389.         pix1 += line_size;

  390.         pix2 += line_size;

  391.     }

  392. 

  393.     /* Sum up the four partial sums, and put the result into s */

  394.     sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);

  395.     sumsqr = vec_splat(sumsqr, 3);

  396.     vec_ste(sumsqr, 0, &s);

  397. 

  398.     return s;

  399. }

  400. 

  401. /**

  402.  * Sum of Squared Errors for a 16x16 block.

  403.  * AltiVec-enhanced.

  404.  * It's the sad16_altivec code above w/ squaring added.

  405.  */

  406. static int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)

  407. {

  408.     int i;

  409.     int s;

  410.     const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);

  411.     vector unsigned char perm = vec_lvsl(0, pix2);

  412.     vector unsigned char t1, t2, t3,t4, t5;

  413.     vector unsigned int sum;

  414.     vector signed int sumsqr;

  415. 

  416.     sum = (vector unsigned int)vec_splat_u32(0);

  417. 

  418.     for (i = 0; i < h; i++) {

  419.         /* Read potentially unaligned pixels into t1 and t2 */

  420.         vector unsigned char pix2l = vec_ld( 0, pix2);

  421.         vector unsigned char pix2r = vec_ld(15, pix2);

  422.         t1 = vec_ld(0, pix1);

  423.         t2 = vec_perm(pix2l, pix2r, perm);

  424. 

  425.         /* Since we want to use unsigned chars, we can take advantage

  426.            of the fact that abs(a-b)^2 = (a-b)^2. */

  427. 

  428.         /* Calculate abs differences vector */

  429.         t3 = vec_max(t1, t2);

  430.         t4 = vec_min(t1, t2);

  431.         t5 = vec_sub(t3, t4);

  432. 

  433.         /* Square the values and add them to our sum */

  434.         sum = vec_msum(t5, t5, sum);

  435. 

  436.         pix1 += line_size;

  437.         pix2 += line_size;

  438.     }

  439. 

  440.     /* Sum up the four partial sums, and put the result into s */

  441.     sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);

  442.     sumsqr = vec_splat(sumsqr, 3);

  443.     vec_ste(sumsqr, 0, &s);

  444. 

  445.     return s;

  446. }

  447. 

  448. static int pix_sum_altivec(uint8_t * pix, int line_size)

  449. {

  450.     const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);

  451.     vector unsigned char perm = vec_lvsl(0, pix);

  452.     vector unsigned char t1;

  453.     vector unsigned int sad;

  454.     vector signed int sumdiffs;

  455. 

  456.     int i;

  457.     int s;

  458. 

  459.     sad = (vector unsigned int)vec_splat_u32(0);

  460. 

  461.     for (i = 0; i < 16; i++) {

  462.         /* Read the potentially unaligned 16 pixels into t1 */

  463.         vector unsigned char pixl = vec_ld( 0, pix);

  464.         vector unsigned char pixr = vec_ld(15, pix);

  465.         t1 = vec_perm(pixl, pixr, perm);

  466. 

  467.         /* Add each 4 pixel group together and put 4 results into sad */

  468.         sad = vec_sum4s(t1, sad);

  469. 

  470.         pix += line_size;

  471.     }

  472. 

  473.     /* Sum up the four partial sums, and put the result into s */

  474.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);

  475.     sumdiffs = vec_splat(sumdiffs, 3);

  476.     vec_ste(sumdiffs, 0, &s);

  477. 

  478.     return s;

  479. }

  480. 

  481. static void get_pixels_altivec(int16_t *restrict block, const uint8_t *pixels, int line_size)

  482. {

  483.     int i;

  484.     vector unsigned char perm = vec_lvsl(0, pixels);

  485.     vector unsigned char bytes;

  486.     const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);

  487.     vector signed short shorts;

  488. 

  489.     for (i = 0; i < 8; i++) {

  490.         // Read potentially unaligned pixels.

  491.         // We're reading 16 pixels, and actually only want 8,

  492.         // but we simply ignore the extras.

  493.         vector unsigned char pixl = vec_ld(0, pixels);

  494.         vector unsigned char pixr = vec_ld(7, pixels);

  495.         bytes = vec_perm(pixl, pixr, perm);

  496. 

  497.         // convert the bytes into shorts

  498.         shorts = (vector signed short)vec_mergeh(zero, bytes);

  499. 

  500.         // save the data to the block, we assume the block is 16-byte aligned

  501.         vec_st(shorts, i*16, (vector signed short*)block);

  502. 

  503.         pixels += line_size;

  504.     }

  505. }

  506. 

  507. static void diff_pixels_altivec(int16_t *restrict block, const uint8_t *s1,

  508.         const uint8_t *s2, int stride)

  509. {

  510.     int i;

  511.     vector unsigned char perm1 = vec_lvsl(0, s1);

  512.     vector unsigned char perm2 = vec_lvsl(0, s2);

  513.     vector unsigned char bytes, pixl, pixr;

  514.     const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);

  515.     vector signed short shorts1, shorts2;

  516. 

  517.     for (i = 0; i < 4; i++) {

  518.         // Read potentially unaligned pixels

  519.         // We're reading 16 pixels, and actually only want 8,

  520.         // but we simply ignore the extras.

  521.         pixl = vec_ld( 0, s1);

  522.         pixr = vec_ld(15, s1);

  523.         bytes = vec_perm(pixl, pixr, perm1);

  524. 

  525.         // convert the bytes into shorts

  526.         shorts1 = (vector signed short)vec_mergeh(zero, bytes);

  527. 

  528.         // Do the same for the second block of pixels

  529.         pixl = vec_ld( 0, s2);

  530.         pixr = vec_ld(15, s2);

  531.         bytes = vec_perm(pixl, pixr, perm2);

  532. 

  533.         // convert the bytes into shorts

  534.         shorts2 = (vector signed short)vec_mergeh(zero, bytes);

  535. 

  536.         // Do the subtraction

  537.         shorts1 = vec_sub(shorts1, shorts2);

  538. 

  539.         // save the data to the block, we assume the block is 16-byte aligned

  540.         vec_st(shorts1, 0, (vector signed short*)block);

  541. 

  542.         s1 += stride;

  543.         s2 += stride;

  544.         block += 8;

  545. 

  546. 

  547.         // The code below is a copy of the code above... This is a manual

  548.         // unroll.

  549. 

  550.         // Read potentially unaligned pixels

  551.         // We're reading 16 pixels, and actually only want 8,

  552.         // but we simply ignore the extras.

  553.         pixl = vec_ld( 0, s1);

  554.         pixr = vec_ld(15, s1);

  555.         bytes = vec_perm(pixl, pixr, perm1);

  556. 

  557.         // convert the bytes into shorts

  558.         shorts1 = (vector signed short)vec_mergeh(zero, bytes);

  559. 

  560.         // Do the same for the second block of pixels

  561.         pixl = vec_ld( 0, s2);

  562.         pixr = vec_ld(15, s2);

  563.         bytes = vec_perm(pixl, pixr, perm2);

  564. 

  565.         // convert the bytes into shorts

  566.         shorts2 = (vector signed short)vec_mergeh(zero, bytes);

  567. 

  568.         // Do the subtraction

  569.         shorts1 = vec_sub(shorts1, shorts2);

  570. 

  571.         // save the data to the block, we assume the block is 16-byte aligned

  572.         vec_st(shorts1, 0, (vector signed short*)block);

  573. 

  574.         s1 += stride;

  575.         s2 += stride;

  576.         block += 8;

  577.     }

  578. }

  579. 

  580. 

  581. static void clear_block_altivec(int16_t *block) {

  582.     LOAD_ZERO;

  583.     vec_st(zero_s16v,   0, block);

  584.     vec_st(zero_s16v,  16, block);

  585.     vec_st(zero_s16v,  32, block);

  586.     vec_st(zero_s16v,  48, block);

  587.     vec_st(zero_s16v,  64, block);

  588.     vec_st(zero_s16v,  80, block);

  589.     vec_st(zero_s16v,  96, block);

  590.     vec_st(zero_s16v, 112, block);

  591. }

  592. 

  593. 

  594. static void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {

  595.     register int i;

  596.     register vector unsigned char vdst, vsrc;

  597. 

  598.     /* dst and src are 16 bytes-aligned (guaranteed) */

  599.     for (i = 0 ; (i + 15) < w ; i+=16) {

  600.         vdst = vec_ld(i, (unsigned char*)dst);

  601.         vsrc = vec_ld(i, (unsigned char*)src);

  602.         vdst = vec_add(vsrc, vdst);

  603.         vec_st(vdst, i, (unsigned char*)dst);

  604.     }

  605.     /* if w is not a multiple of 16 */

  606.     for (; (i < w) ; i++) {

  607.         dst[i] = src[i];

  608.     }

  609. }

  610. 

  611. static int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){

  612.     int sum;

  613.     register const vector unsigned char vzero =

  614.                             (const vector unsigned char)vec_splat_u8(0);

  615.     register vector signed short temp0, temp1, temp2, temp3, temp4,

  616.                                  temp5, temp6, temp7;

  617.     {

  618.     register const vector signed short vprod1 =(const vector signed short)

  619.                                                { 1,-1, 1,-1, 1,-1, 1,-1 };

  620.     register const vector signed short vprod2 =(const vector signed short)

  621.                                                { 1, 1,-1,-1, 1, 1,-1,-1 };

  622.     register const vector signed short vprod3 =(const vector signed short)

  623.                                                { 1, 1, 1, 1,-1,-1,-1,-1 };

  624.     register const vector unsigned char perm1 = (const vector unsigned char)

  625.         {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,

  626.          0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};

  627.     register const vector unsigned char perm2 = (const vector unsigned char)

  628.         {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,

  629.          0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};

  630.     register const vector unsigned char perm3 = (const vector unsigned char)

  631.         {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,

  632.          0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};

  633. 

  634. #define ONEITERBUTTERFLY(i, res)                                          \

  635.     {                                                                     \

  636.     register vector unsigned char src1, src2, srcO;                   \

  637.     register vector unsigned char dst1, dst2, dstO;                   \

  638.     register vector signed short srcV, dstV;                          \

  639.     register vector signed short but0, but1, but2, op1, op2, op3;     \

  640.     src1 = vec_ld(stride * i, src);                                   \

  641.     src2 = vec_ld((stride * i) + 15, src);                            \

  642.     srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src));           \

  643.     dst1 = vec_ld(stride * i, dst);                                   \

  644.     dst2 = vec_ld((stride * i) + 15, dst);                            \

  645.     dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst));           \

  646.     /* promote the unsigned chars to signed shorts */                 \

  647.     /* we're in the 8x8 function, we only care for the first 8 */     \

  648.     srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \

  649.            (vector signed char)srcO);                                 \

  650.     dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \

  651.            (vector signed char)dstO);                                 \

  652.     /* subtractions inside the first butterfly */                     \

  653.     but0 = vec_sub(srcV, dstV);                                       \

  654.     op1  = vec_perm(but0, but0, perm1);                               \

  655.     but1 = vec_mladd(but0, vprod1, op1);                              \

  656.     op2  = vec_perm(but1, but1, perm2);                               \

  657.     but2 = vec_mladd(but1, vprod2, op2);                              \

  658.     op3  = vec_perm(but2, but2, perm3);                               \

  659.     res  = vec_mladd(but2, vprod3, op3);                              \

  660.     }

  661.     ONEITERBUTTERFLY(0, temp0);

  662.     ONEITERBUTTERFLY(1, temp1);

  663.     ONEITERBUTTERFLY(2, temp2);

  664.     ONEITERBUTTERFLY(3, temp3);

  665.     ONEITERBUTTERFLY(4, temp4);

  666.     ONEITERBUTTERFLY(5, temp5);

  667.     ONEITERBUTTERFLY(6, temp6);

  668.     ONEITERBUTTERFLY(7, temp7);

  669.     }

  670. #undef ONEITERBUTTERFLY

  671.     {

  672.     register vector signed int vsum;

  673.     register vector signed short line0 = vec_add(temp0, temp1);

  674.     register vector signed short line1 = vec_sub(temp0, temp1);

  675.     register vector signed short line2 = vec_add(temp2, temp3);

  676.     register vector signed short line3 = vec_sub(temp2, temp3);

  677.     register vector signed short line4 = vec_add(temp4, temp5);

  678.     register vector signed short line5 = vec_sub(temp4, temp5);

  679.     register vector signed short line6 = vec_add(temp6, temp7);

  680.     register vector signed short line7 = vec_sub(temp6, temp7);

  681. 

  682.     register vector signed short line0B = vec_add(line0, line2);

  683.     register vector signed short line2B = vec_sub(line0, line2);

  684.     register vector signed short line1B = vec_add(line1, line3);

  685.     register vector signed short line3B = vec_sub(line1, line3);

  686.     register vector signed short line4B = vec_add(line4, line6);

  687.     register vector signed short line6B = vec_sub(line4, line6);

  688.     register vector signed short line5B = vec_add(line5, line7);

  689.     register vector signed short line7B = vec_sub(line5, line7);

  690. 

  691.     register vector signed short line0C = vec_add(line0B, line4B);

  692.     register vector signed short line4C = vec_sub(line0B, line4B);

  693.     register vector signed short line1C = vec_add(line1B, line5B);

  694.     register vector signed short line5C = vec_sub(line1B, line5B);

  695.     register vector signed short line2C = vec_add(line2B, line6B);

  696.     register vector signed short line6C = vec_sub(line2B, line6B);

  697.     register vector signed short line3C = vec_add(line3B, line7B);

  698.     register vector signed short line7C = vec_sub(line3B, line7B);

  699. 

  700.     vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));

  701.     vsum = vec_sum4s(vec_abs(line1C), vsum);

  702.     vsum = vec_sum4s(vec_abs(line2C), vsum);

  703.     vsum = vec_sum4s(vec_abs(line3C), vsum);

  704.     vsum = vec_sum4s(vec_abs(line4C), vsum);

  705.     vsum = vec_sum4s(vec_abs(line5C), vsum);

  706.     vsum = vec_sum4s(vec_abs(line6C), vsum);

  707.     vsum = vec_sum4s(vec_abs(line7C), vsum);

  708.     vsum = vec_sums(vsum, (vector signed int)vzero);

  709.     vsum = vec_splat(vsum, 3);

  710.     vec_ste(vsum, 0, &sum);

  711.     }

  712.     return sum;

  713. }

  714. 

  715. /*

  716. 16x8 works with 16 elements; it allows to avoid replicating loads, and

  717. give the compiler more rooms for scheduling.  It's only used from

  718. inside hadamard8_diff16_altivec.

  719. 

  720. Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has a LOT

  721. of spill code, it seems gcc (unlike xlc) cannot keep everything in registers

  722. by itself. The following code include hand-made registers allocation. It's not

  723. clean, but on a 7450 the resulting code is much faster (best case fall from

  724. 700+ cycles to 550).

  725. 

  726. xlc doesn't add spill code, but it doesn't know how to schedule for the 7450,

  727. and its code isn't much faster than gcc-3.3 on the 7450 (but uses 25% less

  728. instructions...)

  729. 

  730. On the 970, the hand-made RA is still a win (around 690 vs. around 780), but

  731. xlc goes to around 660 on the regular C code...

  732. */

  733. 

  734. static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) {

  735.     int sum;

  736.     register vector signed short

  737.         temp0 __asm__ ("v0"),

  738.         temp1 __asm__ ("v1"),

  739.         temp2 __asm__ ("v2"),

  740.         temp3 __asm__ ("v3"),

  741.         temp4 __asm__ ("v4"),

  742.         temp5 __asm__ ("v5"),

  743.         temp6 __asm__ ("v6"),

  744.         temp7 __asm__ ("v7");

  745.     register vector signed short

  746.         temp0S __asm__ ("v8"),

  747.         temp1S __asm__ ("v9"),

  748.         temp2S __asm__ ("v10"),

  749.         temp3S __asm__ ("v11"),

  750.         temp4S __asm__ ("v12"),

  751.         temp5S __asm__ ("v13"),

  752.         temp6S __asm__ ("v14"),

  753.         temp7S __asm__ ("v15");

  754.     register const vector unsigned char vzero __asm__ ("v31") =

  755.         (const vector unsigned char)vec_splat_u8(0);

  756.     {

  757.     register const vector signed short vprod1 __asm__ ("v16") =

  758.         (const vector signed short){ 1,-1, 1,-1, 1,-1, 1,-1 };

  759.     register const vector signed short vprod2 __asm__ ("v17") =

  760.         (const vector signed short){ 1, 1,-1,-1, 1, 1,-1,-1 };

  761.     register const vector signed short vprod3 __asm__ ("v18") =

  762.         (const vector signed short){ 1, 1, 1, 1,-1,-1,-1,-1 };

  763.     register const vector unsigned char perm1 __asm__ ("v19") =

  764.         (const vector unsigned char)

  765.         {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,

  766.          0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};

  767.     register const vector unsigned char perm2 __asm__ ("v20") =

  768.         (const vector unsigned char)

  769.         {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,

  770.          0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};

  771.     register const vector unsigned char perm3 __asm__ ("v21") =

  772.         (const vector unsigned char)

  773.         {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,

  774.          0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};

  775. 

  776. #define ONEITERBUTTERFLY(i, res1, res2)                               \

  777.     {                                                                 \

  778.     register vector unsigned char src1 __asm__ ("v22"),               \

  779.                                   src2 __asm__ ("v23"),               \

  780.                                   dst1 __asm__ ("v24"),               \

  781.                                   dst2 __asm__ ("v25"),               \

  782.                                   srcO __asm__ ("v22"),               \

  783.                                   dstO __asm__ ("v23");               \

  784.                                                                       \

  785.     register vector signed short  srcV  __asm__ ("v24"),              \

  786.                                   dstV  __asm__ ("v25"),              \

  787.                                   srcW  __asm__ ("v26"),              \

  788.                                   dstW  __asm__ ("v27"),              \

  789.                                   but0  __asm__ ("v28"),              \

  790.                                   but0S __asm__ ("v29"),              \

  791.                                   op1   __asm__ ("v30"),              \

  792.                                   but1  __asm__ ("v22"),              \

  793.                                   op1S  __asm__ ("v23"),              \

  794.                                   but1S __asm__ ("v24"),              \

  795.                                   op2   __asm__ ("v25"),              \

  796.                                   but2  __asm__ ("v26"),              \

  797.                                   op2S  __asm__ ("v27"),              \

  798.                                   but2S __asm__ ("v28"),              \

  799.                                   op3   __asm__ ("v29"),              \

  800.                                   op3S  __asm__ ("v30");              \

  801.                                                                       \

  802.     src1 = vec_ld(stride * i, src);                                   \

  803.     src2 = vec_ld((stride * i) + 16, src);                            \

  804.     srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src));           \

  805.     dst1 = vec_ld(stride * i, dst);                                   \

  806.     dst2 = vec_ld((stride * i) + 16, dst);                            \

  807.     dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst));           \

  808.     /* promote the unsigned chars to signed shorts */                 \

  809.     srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \

  810.            (vector signed char)srcO);                                 \

  811.     dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \

  812.            (vector signed char)dstO);                                 \

  813.     srcW = (vector signed short)vec_mergel((vector signed char)vzero, \

  814.            (vector signed char)srcO);                                 \

  815.     dstW = (vector signed short)vec_mergel((vector signed char)vzero, \

  816.            (vector signed char)dstO);                                 \

  817.     /* subtractions inside the first butterfly */                     \

  818.     but0 = vec_sub(srcV, dstV);                                       \

  819.     but0S = vec_sub(srcW, dstW);                                      \

  820.     op1 = vec_perm(but0, but0, perm1);                                \

  821.     but1 = vec_mladd(but0, vprod1, op1);                              \

  822.     op1S = vec_perm(but0S, but0S, perm1);                             \

  823.     but1S = vec_mladd(but0S, vprod1, op1S);                           \

  824.     op2 = vec_perm(but1, but1, perm2);                                \

  825.     but2 = vec_mladd(but1, vprod2, op2);                              \

  826.     op2S = vec_perm(but1S, but1S, perm2);                             \

  827.     but2S = vec_mladd(but1S, vprod2, op2S);                           \

  828.     op3 = vec_perm(but2, but2, perm3);                                \

  829.     res1 = vec_mladd(but2, vprod3, op3);                              \

  830.     op3S = vec_perm(but2S, but2S, perm3);                             \

  831.     res2 = vec_mladd(but2S, vprod3, op3S);                            \

  832.     }

  833.     ONEITERBUTTERFLY(0, temp0, temp0S);

  834.     ONEITERBUTTERFLY(1, temp1, temp1S);

  835.     ONEITERBUTTERFLY(2, temp2, temp2S);

  836.     ONEITERBUTTERFLY(3, temp3, temp3S);

  837.     ONEITERBUTTERFLY(4, temp4, temp4S);

  838.     ONEITERBUTTERFLY(5, temp5, temp5S);

  839.     ONEITERBUTTERFLY(6, temp6, temp6S);

  840.     ONEITERBUTTERFLY(7, temp7, temp7S);

  841.     }

  842. #undef ONEITERBUTTERFLY

  843.     {

  844.     register vector signed int vsum;

  845.     register vector signed short line0S, line1S, line2S, line3S, line4S,

  846.                                  line5S, line6S, line7S, line0BS,line2BS,

  847.                                  line1BS,line3BS,line4BS,line6BS,line5BS,

  848.                                  line7BS,line0CS,line4CS,line1CS,line5CS,

  849.                                  line2CS,line6CS,line3CS,line7CS;

  850. 

  851.     register vector signed short line0 = vec_add(temp0, temp1);

  852.     register vector signed short line1 = vec_sub(temp0, temp1);

  853.     register vector signed short line2 = vec_add(temp2, temp3);

  854.     register vector signed short line3 = vec_sub(temp2, temp3);

  855.     register vector signed short line4 = vec_add(temp4, temp5);

  856.     register vector signed short line5 = vec_sub(temp4, temp5);

  857.     register vector signed short line6 = vec_add(temp6, temp7);

  858.     register vector signed short line7 = vec_sub(temp6, temp7);

  859. 

  860.     register vector signed short line0B = vec_add(line0, line2);

  861.     register vector signed short line2B = vec_sub(line0, line2);

  862.     register vector signed short line1B = vec_add(line1, line3);

  863.     register vector signed short line3B = vec_sub(line1, line3);

  864.     register vector signed short line4B = vec_add(line4, line6);

  865.     register vector signed short line6B = vec_sub(line4, line6);

  866.     register vector signed short line5B = vec_add(line5, line7);

  867.     register vector signed short line7B = vec_sub(line5, line7);

  868. 

  869.     register vector signed short line0C = vec_add(line0B, line4B);

  870.     register vector signed short line4C = vec_sub(line0B, line4B);

  871.     register vector signed short line1C = vec_add(line1B, line5B);

  872.     register vector signed short line5C = vec_sub(line1B, line5B);

  873.     register vector signed short line2C = vec_add(line2B, line6B);

  874.     register vector signed short line6C = vec_sub(line2B, line6B);

  875.     register vector signed short line3C = vec_add(line3B, line7B);

  876.     register vector signed short line7C = vec_sub(line3B, line7B);

  877. 

  878.     vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));

  879.     vsum = vec_sum4s(vec_abs(line1C), vsum);

  880.     vsum = vec_sum4s(vec_abs(line2C), vsum);

  881.     vsum = vec_sum4s(vec_abs(line3C), vsum);

  882.     vsum = vec_sum4s(vec_abs(line4C), vsum);

  883.     vsum = vec_sum4s(vec_abs(line5C), vsum);

  884.     vsum = vec_sum4s(vec_abs(line6C), vsum);

  885.     vsum = vec_sum4s(vec_abs(line7C), vsum);

  886. 

  887.     line0S = vec_add(temp0S, temp1S);

  888.     line1S = vec_sub(temp0S, temp1S);

  889.     line2S = vec_add(temp2S, temp3S);

  890.     line3S = vec_sub(temp2S, temp3S);

  891.     line4S = vec_add(temp4S, temp5S);

  892.     line5S = vec_sub(temp4S, temp5S);

  893.     line6S = vec_add(temp6S, temp7S);

  894.     line7S = vec_sub(temp6S, temp7S);

  895. 

  896.     line0BS = vec_add(line0S, line2S);

  897.     line2BS = vec_sub(line0S, line2S);

  898.     line1BS = vec_add(line1S, line3S);

  899.     line3BS = vec_sub(line1S, line3S);

  900.     line4BS = vec_add(line4S, line6S);

  901.     line6BS = vec_sub(line4S, line6S);

  902.     line5BS = vec_add(line5S, line7S);

  903.     line7BS = vec_sub(line5S, line7S);

  904. 

  905.     line0CS = vec_add(line0BS, line4BS);

  906.     line4CS = vec_sub(line0BS, line4BS);

  907.     line1CS = vec_add(line1BS, line5BS);

  908.     line5CS = vec_sub(line1BS, line5BS);

  909.     line2CS = vec_add(line2BS, line6BS);

  910.     line6CS = vec_sub(line2BS, line6BS);

  911.     line3CS = vec_add(line3BS, line7BS);

  912.     line7CS = vec_sub(line3BS, line7BS);

  913. 

  914.     vsum = vec_sum4s(vec_abs(line0CS), vsum);

  915.     vsum = vec_sum4s(vec_abs(line1CS), vsum);

  916.     vsum = vec_sum4s(vec_abs(line2CS), vsum);

  917.     vsum = vec_sum4s(vec_abs(line3CS), vsum);

  918.     vsum = vec_sum4s(vec_abs(line4CS), vsum);

  919.     vsum = vec_sum4s(vec_abs(line5CS), vsum);

  920.     vsum = vec_sum4s(vec_abs(line6CS), vsum);

  921.     vsum = vec_sum4s(vec_abs(line7CS), vsum);

  922.     vsum = vec_sums(vsum, (vector signed int)vzero);

  923.     vsum = vec_splat(vsum, 3);

  924.     vec_ste(vsum, 0, &sum);

  925.     }

  926.     return sum;

  927. }

  928. 

  929. static int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){

  930.     int score;

  931.     score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);

  932.     if (h==16) {

  933.         dst += 8*stride;

  934.         src += 8*stride;

  935.         score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);

  936.     }

  937.     return score;

  938. }

  939. 

  940. av_cold void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx)

  941. {

  942.     const int high_bit_depth = avctx->bits_per_raw_sample > 8;

  943. 

  944.     c->pix_abs[0][1] = sad16_x2_altivec;

  945.     c->pix_abs[0][2] = sad16_y2_altivec;

  946.     c->pix_abs[0][3] = sad16_xy2_altivec;

  947.     c->pix_abs[0][0] = sad16_altivec;

  948.     c->pix_abs[1][0] = sad8_altivec;

  949.     c->sad[0]= sad16_altivec;

  950.     c->sad[1]= sad8_altivec;

  951.     c->pix_norm1 = pix_norm1_altivec;

  952.     c->sse[1]= sse8_altivec;

  953.     c->sse[0]= sse16_altivec;

  954.     c->pix_sum = pix_sum_altivec;

  955.     c->diff_pixels = diff_pixels_altivec;

  956.     c->add_bytes= add_bytes_altivec;

  957.     if (!high_bit_depth) {

  958.     c->get_pixels = get_pixels_altivec;

  959.     c->clear_block = clear_block_altivec;

  960.     }

  961. 

  962.     c->hadamard8_diff[0] = hadamard8_diff16_altivec;

  963.     c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;

  964. }