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.
40170 Commits
32 Branches
325MB
Tree: c166148409
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c166148409'
${ noResults }
FFmpeg/libavcodec/ppc/dsputil_altivec.c

866 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. 

  28. #include "libavutil/attributes.h"

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

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

  31. #include "libavcodec/avcodec.h"

  32. #include "libavcodec/dsputil.h"

  33. #include "libavcodec/mpegvideo.h"

  34. #include "dsputil_altivec.h"

  35. 

  36. static int sad16_x2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  37.                             int line_size, int h)

  38. {

  39.     int i, s = 0;

  40.     const vector unsigned char zero =

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

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

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

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

  45.     vector signed int sumdiffs;

  46. 

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

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

  49.          * pix1v: pix1[0] - pix1[15]

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

  51.         vector unsigned char pix1v  = vec_ld(0,  pix1);

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

  53.         vector unsigned char pix2r  = vec_ld(16, pix2);

  54.         vector unsigned char pix2v  = vec_perm(pix2l, pix2r, perm1);

  55.         vector unsigned char pix2iv = vec_perm(pix2l, pix2r, perm2);

  56. 

  57.         /* Calculate the average vector. */

  58.         vector unsigned char avgv = vec_avg(pix2v, pix2iv);

  59. 

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

  61.         vector unsigned char t5 = vec_sub(vec_max(pix1v, avgv),

  62.                                           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(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  79.                             int line_size, int h)

  80. {

  81.     int i, s = 0;

  82.     const vector unsigned char zero =

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

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

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

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

  87.     vector signed int sumdiffs;

  88.     uint8_t *pix3 = pix2 + line_size;

  89. 

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

  91.      * iteration becomes pix2 in the next iteration. We can use this

  92.      * fact to avoid a potentially expensive unaligned read, each

  93.      * time around the loop.

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

  95.      * pix2v: pix2[0] - pix2[15]

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

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

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

  99.     vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm);

  100. 

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

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

  103.          * pix1v: pix1[0] - pix1[15]

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

  105.         pix1v = vec_ld(0,  pix1);

  106. 

  107.         pix2l = vec_ld(0,  pix3);

  108.         pix2r = vec_ld(15, pix3);

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

  110. 

  111.         /* Calculate the average vector. */

  112.         avgv = vec_avg(pix2v, pix3v);

  113. 

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

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

  116. 

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

  118.         sad = vec_sum4s(t5, sad);

  119. 

  120.         pix1 += line_size;

  121.         pix2v = pix3v;

  122.         pix3 += line_size;

  123.     }

  124. 

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

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

  127.     sumdiffs = vec_splat(sumdiffs, 3);

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

  129.     return s;

  130. }

  131. 

  132. static int sad16_xy2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  133.                              int line_size, int h)

  134. {

  135.     int i, s = 0;

  136.     uint8_t *pix3 = pix2 + line_size;

  137.     const vector unsigned char zero =

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

  139.     const vector unsigned short two =

  140.         (const vector unsigned short) vec_splat_u16(2);

  141.     vector unsigned char avgv, t5;

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

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

  144.     vector unsigned char pix1v, pix3v, pix3iv;

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

  146.     vector unsigned short avghv, avglv;

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

  148.     vector signed int sumdiffs;

  149. 

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

  151.      * iteration becomes pix2 in the next iteration. We can use this

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

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

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

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

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

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

  158.     vector unsigned char pix2r  = vec_ld(16, pix2);

  159.     vector unsigned char pix2v  = vec_perm(pix2l, pix2r, perm1);

  160.     vector unsigned char pix2iv = vec_perm(pix2l, pix2r, perm2);

  161. 

  162.     vector unsigned short pix2hv  =

  163.         (vector unsigned short) vec_mergeh(zero, pix2v);

  164.     vector unsigned short pix2lv  =

  165.         (vector unsigned short) vec_mergel(zero, pix2v);

  166.     vector unsigned short pix2ihv =

  167.         (vector unsigned short) vec_mergeh(zero, pix2iv);

  168.     vector unsigned short pix2ilv =

  169.         (vector unsigned short) vec_mergel(zero, pix2iv);

  170.     vector unsigned short t1 = vec_add(pix2hv, pix2ihv);

  171.     vector unsigned short t2 = vec_add(pix2lv, pix2ilv);

  172.     vector unsigned short t3, t4;

  173. 

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

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

  176.          * pix1v: pix1[0] - pix1[15]

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

  178.         pix1v  = vec_ld(0, pix1);

  179. 

  180.         pix2l  = vec_ld(0, pix3);

  181.         pix2r  = vec_ld(16, pix3);

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

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

  184. 

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

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

  187.          * rounding would mean that, for example, avg(3, 0, 0, 1) = 2, when

  188.          * it should be 1. Instead, we have to split the pixel vectors into

  189.          * vectors of shorts and do the averaging by hand. */

  190. 

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

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

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

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

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

  196. 

  197.         /* Do the averaging on them. */

  198.         t3 = vec_add(pix3hv, pix3ihv);

  199.         t4 = vec_add(pix3lv, pix3ilv);

  200. 

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

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

  203. 

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

  205.         avgv = vec_pack(avghv, avglv);

  206. 

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

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

  209. 

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

  211.         sad = vec_sum4s(t5, sad);

  212. 

  213.         pix1 += line_size;

  214.         pix3 += line_size;

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

  216.         t1 = t3;

  217.         t2 = t4;

  218.     }

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

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

  221.     sumdiffs = vec_splat(sumdiffs, 3);

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

  223. 

  224.     return s;

  225. }

  226. 

  227. static int sad16_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  228.                          int line_size, int h)

  229. {

  230.     int i, s;

  231.     const vector unsigned int zero =

  232.         (const vector unsigned int) vec_splat_u32(0);

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

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

  235.     vector signed int sumdiffs;

  236. 

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

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

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

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

  241.         vector unsigned char t1 = vec_ld(0, pix1);

  242.         vector unsigned char t2 = vec_perm(pix2l, pix2r, perm);

  243. 

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

  245.         vector unsigned char t3 = vec_max(t1, t2);

  246.         vector unsigned char t4 = vec_min(t1, t2);

  247.         vector unsigned char t5 = vec_sub(t3, t4);

  248. 

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

  250.         sad = vec_sum4s(t5, sad);

  251. 

  252.         pix1 += line_size;

  253.         pix2 += line_size;

  254.     }

  255. 

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

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

  258.     sumdiffs = vec_splat(sumdiffs, 3);

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

  260. 

  261.     return s;

  262. }

  263. 

  264. static int sad8_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  265.                         int line_size, int h)

  266. {

  267.     int i, s;

  268.     const vector unsigned int zero =

  269.         (const vector unsigned int) vec_splat_u32(0);

  270.     const vector unsigned char permclear =

  271.         (vector unsigned char)

  272.         { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };

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

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

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

  276.     vector signed int sumdiffs;

  277. 

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

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

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

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

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

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

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

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

  286.         vector unsigned char t1 = vec_and(vec_perm(pix1l, pix1r, perm1),

  287.                                           permclear);

  288.         vector unsigned char t2 = vec_and(vec_perm(pix2l, pix2r, perm2),

  289.                                           permclear);

  290. 

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

  292.         vector unsigned char t3 = vec_max(t1, t2);

  293.         vector unsigned char t4 = vec_min(t1, t2);

  294.         vector unsigned char t5 = vec_sub(t3, t4);

  295. 

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

  297.         sad = vec_sum4s(t5, sad);

  298. 

  299.         pix1 += line_size;

  300.         pix2 += line_size;

  301.     }

  302. 

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

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

  305.     sumdiffs = vec_splat(sumdiffs, 3);

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

  307. 

  308.     return s;

  309. }

  310. 

  311. /* Sum of Squared Errors for an 8x8 block, AltiVec-enhanced.

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

  313. static int sse8_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  314.                         int line_size, int h)

  315. {

  316.     int i, s;

  317.     const vector unsigned int zero =

  318.         (const vector unsigned int) vec_splat_u32(0);

  319.     const vector unsigned char permclear =

  320.         (vector unsigned char)

  321.         { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };

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

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

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

  325.     vector signed int sumsqr;

  326. 

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

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

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

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

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

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

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

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

  335.         vector unsigned char t1 = vec_and(vec_perm(pix1l, pix1r, perm1),

  336.                                           permclear);

  337.         vector unsigned char t2 = vec_and(vec_perm(pix2l, pix2r, perm2),

  338.                                           permclear);

  339. 

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

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

  342. 

  343.         /* Calculate abs differences vector. */

  344.         vector unsigned char t3 = vec_max(t1, t2);

  345.         vector unsigned char t4 = vec_min(t1, t2);

  346.         vector unsigned char t5 = vec_sub(t3, t4);

  347. 

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

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

  350. 

  351.         pix1 += line_size;

  352.         pix2 += line_size;

  353.     }

  354. 

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

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

  357.     sumsqr = vec_splat(sumsqr, 3);

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

  359. 

  360.     return s;

  361. }

  362. 

  363. /* Sum of Squared Errors for a 16x16 block, AltiVec-enhanced.

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

  365. static int sse16_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  366.                          int line_size, int h)

  367. {

  368.     int i, s;

  369.     const vector unsigned int zero =

  370.         (const vector unsigned int) vec_splat_u32(0);

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

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

  373.     vector signed int sumsqr;

  374. 

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

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

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

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

  379.         vector unsigned char t1 = vec_ld(0, pix1);

  380.         vector unsigned char t2 = vec_perm(pix2l, pix2r, perm);

  381. 

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

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

  384. 

  385.         /* Calculate abs differences vector. */

  386.         vector unsigned char t3 = vec_max(t1, t2);

  387.         vector unsigned char t4 = vec_min(t1, t2);

  388.         vector unsigned char t5 = vec_sub(t3, t4);

  389. 

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

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

  392. 

  393.         pix1 += line_size;

  394.         pix2 += line_size;

  395.     }

  396. 

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

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

  399.     sumsqr = vec_splat(sumsqr, 3);

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

  401. 

  402.     return s;

  403. }

  404. 

  405. static void get_pixels_altivec(int16_t *restrict block, const uint8_t *pixels,

  406.                                int line_size)

  407. {

  408.     int i;

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

  410.     const vector unsigned char zero =

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

  412. 

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

  414.         /* Read potentially unaligned pixels.

  415.          * We're reading 16 pixels, and actually only want 8,

  416.          * but we simply ignore the extras. */

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

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

  419.         vector unsigned char bytes = vec_perm(pixl, pixr, perm);

  420. 

  421.         // Convert the bytes into shorts.

  422.         vector signed short shorts = (vector signed short) vec_mergeh(zero,

  423.                                                                       bytes);

  424. 

  425.         // Save the data to the block, we assume the block is 16-byte aligned.

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

  427. 

  428.         pixels += line_size;

  429.     }

  430. }

  431. 

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

  433.                                 const uint8_t *s2, int stride)

  434. {

  435.     int i;

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

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

  438.     const vector unsigned char zero =

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

  440.     vector signed short shorts1, shorts2;

  441. 

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

  443.         /* Read potentially unaligned pixels.

  444.          * We're reading 16 pixels, and actually only want 8,

  445.          * but we simply ignore the extras. */

  446.         vector unsigned char pixl  = vec_ld(0,  s1);

  447.         vector unsigned char pixr  = vec_ld(15, s1);

  448.         vector unsigned char bytes = vec_perm(pixl, pixr, perm1);

  449. 

  450.         // Convert the bytes into shorts.

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

  452. 

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

  454.         pixl  = vec_ld(0,  s2);

  455.         pixr  = vec_ld(15, s2);

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

  457. 

  458.         // Convert the bytes into shorts.

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

  460. 

  461.         // Do the subtraction.

  462.         shorts1 = vec_sub(shorts1, shorts2);

  463. 

  464.         // Save the data to the block, we assume the block is 16-byte aligned.

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

  466. 

  467.         s1    += stride;

  468.         s2    += stride;

  469.         block += 8;

  470. 

  471.         /* The code below is a copy of the code above...

  472.          * This is a manual unroll. */

  473. 

  474.         /* Read potentially unaligned pixels.

  475.          * We're reading 16 pixels, and actually only want 8,

  476.          * but we simply ignore the extras. */

  477.         pixl  = vec_ld(0,  s1);

  478.         pixr  = vec_ld(15, s1);

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

  480. 

  481.         // Convert the bytes into shorts.

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

  483. 

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

  485.         pixl  = vec_ld(0,  s2);

  486.         pixr  = vec_ld(15, s2);

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

  488. 

  489.         // Convert the bytes into shorts.

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

  491. 

  492.         // Do the subtraction.

  493.         shorts1 = vec_sub(shorts1, shorts2);

  494. 

  495.         // Save the data to the block, we assume the block is 16-byte aligned.

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

  497. 

  498.         s1    += stride;

  499.         s2    += stride;

  500.         block += 8;

  501.     }

  502. }

  503. 

  504. static int hadamard8_diff8x8_altivec(MpegEncContext *s, uint8_t *dst,

  505.                                      uint8_t *src, int stride, int h)

  506. {

  507.     int sum;

  508.     register const vector unsigned char vzero =

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

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

  511.                                  temp5, temp6, temp7;

  512.     {

  513.         register const vector signed short vprod1 =

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

  515.         register const vector signed short vprod2 =

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

  517.         register const vector signed short vprod3 =

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

  519.         register const vector unsigned char perm1 =

  520.             (const vector unsigned char)

  521.             { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,

  522.               0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };

  523.         register const vector unsigned char perm2 =

  524.             (const vector unsigned char)

  525.             { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,

  526.               0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };

  527.         register const vector unsigned char perm3 =

  528.             (const vector unsigned char)

  529.             { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,

  530.               0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };

  531. 

  532. #define ONEITERBUTTERFLY(i, res)                                            \

  533.     {                                                                       \

  534.         register vector unsigned char src1 = vec_ld(stride * i, src);       \

  535.         register vector unsigned char src2 = vec_ld(stride * i + 15, src);  \

  536.         register vector unsigned char srcO =                                \

  537.             vec_perm(src1, src2, vec_lvsl(stride * i, src));                \

  538.         register vector unsigned char dst1 = vec_ld(stride * i, dst);       \

  539.         register vector unsigned char dst2 = vec_ld(stride * i + 15, dst);  \

  540.         register vector unsigned char dstO =                                \

  541.             vec_perm(dst1, dst2, vec_lvsl(stride * i, dst));                \

  542.                                                                             \

  543.         /* Promote the unsigned chars to signed shorts. */                  \

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

  545.         register vector signed short srcV =                                 \

  546.             (vector signed short) vec_mergeh((vector signed char) vzero,    \

  547.                                              (vector signed char) srcO);    \

  548.         register vector signed short dstV =                                 \

  549.             (vector signed short) vec_mergeh((vector signed char) vzero,    \

  550.                                              (vector signed char) dstO);    \

  551.                                                                             \

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

  553.         register vector signed short but0 = vec_sub(srcV, dstV);            \

  554.         register vector signed short op1  = vec_perm(but0, but0, perm1);    \

  555.         register vector signed short but1 = vec_mladd(but0, vprod1, op1);   \

  556.         register vector signed short op2  = vec_perm(but1, but1, perm2);    \

  557.         register vector signed short but2 = vec_mladd(but1, vprod2, op2);   \

  558.         register vector signed short op3  = vec_perm(but2, but2, perm3);    \

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

  560.     }

  561.         ONEITERBUTTERFLY(0, temp0);

  562.         ONEITERBUTTERFLY(1, temp1);

  563.         ONEITERBUTTERFLY(2, temp2);

  564.         ONEITERBUTTERFLY(3, temp3);

  565.         ONEITERBUTTERFLY(4, temp4);

  566.         ONEITERBUTTERFLY(5, temp5);

  567.         ONEITERBUTTERFLY(6, temp6);

  568.         ONEITERBUTTERFLY(7, temp7);

  569.     }

  570. #undef ONEITERBUTTERFLY

  571.     {

  572.         register vector signed int vsum;

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

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

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

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

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

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

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

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

  581. 

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

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

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

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

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

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

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

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

  590. 

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

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

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

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

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

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

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

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

  599. 

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

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

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

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

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

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

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

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

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

  609.         vsum = vec_splat(vsum, 3);

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

  611.     }

  612.     return sum;

  613. }

  614. 

  615. /*

  616.  * 16x8 works with 16 elements; it allows to avoid replicating loads, and

  617.  * gives the compiler more room for scheduling. It's only used from

  618.  * inside hadamard8_diff16_altivec.

  619.  *

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

  621.  * a LOT of spill code, it seems gcc (unlike xlc) cannot keep everything in

  622.  * registers by itself. The following code includes hand-made register

  623.  * allocation. It's not clean, but on a 7450 the resulting code is much faster

  624.  * (best case falls from 700+ cycles to 550).

  625.  *

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

  627.  * 7450, and its code isn't much faster than gcc-3.3 on the 7450 (but uses

  628.  * 25% fewer instructions...)

  629.  *

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

  631.  * but xlc goes to around 660 on the regular C code...

  632.  */

  633. static int hadamard8_diff16x8_altivec(MpegEncContext *s, uint8_t *dst,

  634.                                       uint8_t *src, int stride, int h)

  635. {

  636.     int sum;

  637.     register vector signed short

  638.         temp0 __asm__ ("v0"),

  639.         temp1 __asm__ ("v1"),

  640.         temp2 __asm__ ("v2"),

  641.         temp3 __asm__ ("v3"),

  642.         temp4 __asm__ ("v4"),

  643.         temp5 __asm__ ("v5"),

  644.         temp6 __asm__ ("v6"),

  645.         temp7 __asm__ ("v7");

  646.     register vector signed short

  647.         temp0S __asm__ ("v8"),

  648.         temp1S __asm__ ("v9"),

  649.         temp2S __asm__ ("v10"),

  650.         temp3S __asm__ ("v11"),

  651.         temp4S __asm__ ("v12"),

  652.         temp5S __asm__ ("v13"),

  653.         temp6S __asm__ ("v14"),

  654.         temp7S __asm__ ("v15");

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

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

  657.     {

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

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

  660. 

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

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

  663. 

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

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

  666. 

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

  668.             (const vector unsigned char)

  669.             { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,

  670.               0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };

  671. 

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

  673.             (const vector unsigned char)

  674.             { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,

  675.               0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };

  676. 

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

  678.             (const vector unsigned char)

  679.             { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,

  680.               0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };

  681. 

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

  683.     {                                                                       \

  684.         register vector unsigned char src1 __asm__ ("v22") =                \

  685.             vec_ld(stride * i, src);                                        \

  686.         register vector unsigned char src2 __asm__ ("v23") =                \

  687.             vec_ld(stride * i + 16, src);                                   \

  688.         register vector unsigned char srcO __asm__ ("v22") =                \

  689.             vec_perm(src1, src2, vec_lvsl(stride * i, src));                \

  690.         register vector unsigned char dst1 __asm__ ("v24") =                \

  691.             vec_ld(stride * i, dst);                                        \

  692.         register vector unsigned char dst2 __asm__ ("v25") =                \

  693.             vec_ld(stride * i + 16, dst);                                   \

  694.         register vector unsigned char dstO __asm__ ("v23") =                \

  695.             vec_perm(dst1, dst2, vec_lvsl(stride * i, dst));                \

  696.                                                                             \

  697.         /* Promote the unsigned chars to signed shorts. */                  \

  698.         register vector signed short srcV __asm__ ("v24") =                 \

  699.             (vector signed short) vec_mergeh((vector signed char) vzero,    \

  700.                                              (vector signed char) srcO);    \

  701.         register vector signed short dstV __asm__ ("v25") =                 \

  702.             (vector signed short) vec_mergeh((vector signed char) vzero,    \

  703.                                              (vector signed char) dstO);    \

  704.         register vector signed short srcW __asm__ ("v26") =                 \

  705.             (vector signed short) vec_mergel((vector signed char) vzero,    \

  706.                                              (vector signed char) srcO);    \

  707.         register vector signed short dstW __asm__ ("v27") =                 \

  708.             (vector signed short) vec_mergel((vector signed char) vzero,    \

  709.                                              (vector signed char) dstO);    \

  710.                                                                             \

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

  712.         register vector signed short but0  __asm__ ("v28") =                \

  713.             vec_sub(srcV, dstV);                                            \

  714.         register vector signed short but0S __asm__ ("v29") =                \

  715.             vec_sub(srcW, dstW);                                            \

  716.         register vector signed short op1   __asm__ ("v30") =                \

  717.             vec_perm(but0, but0, perm1);                                    \

  718.         register vector signed short but1  __asm__ ("v22") =                \

  719.             vec_mladd(but0, vprod1, op1);                                   \

  720.         register vector signed short op1S  __asm__ ("v23") =                \

  721.             vec_perm(but0S, but0S, perm1);                                  \

  722.         register vector signed short but1S __asm__ ("v24") =                \

  723.             vec_mladd(but0S, vprod1, op1S);                                 \

  724.         register vector signed short op2   __asm__ ("v25") =                \

  725.             vec_perm(but1, but1, perm2);                                    \

  726.         register vector signed short but2  __asm__ ("v26") =                \

  727.             vec_mladd(but1, vprod2, op2);                                   \

  728.         register vector signed short op2S  __asm__ ("v27") =                \

  729.             vec_perm(but1S, but1S, perm2);                                  \

  730.         register vector signed short but2S __asm__ ("v28") =                \

  731.             vec_mladd(but1S, vprod2, op2S);                                 \

  732.         register vector signed short op3   __asm__ ("v29") =                \

  733.             vec_perm(but2, but2, perm3);                                    \

  734.         register vector signed short op3S  __asm__ ("v30") =                \

  735.             vec_perm(but2S, but2S, perm3);                                  \

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

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

  738.     }

  739.         ONEITERBUTTERFLY(0, temp0, temp0S);

  740.         ONEITERBUTTERFLY(1, temp1, temp1S);

  741.         ONEITERBUTTERFLY(2, temp2, temp2S);

  742.         ONEITERBUTTERFLY(3, temp3, temp3S);

  743.         ONEITERBUTTERFLY(4, temp4, temp4S);

  744.         ONEITERBUTTERFLY(5, temp5, temp5S);

  745.         ONEITERBUTTERFLY(6, temp6, temp6S);

  746.         ONEITERBUTTERFLY(7, temp7, temp7S);

  747.     }

  748. #undef ONEITERBUTTERFLY

  749.     {

  750.         register vector signed int vsum;

  751. 

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

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

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

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

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

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

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

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

  760. 

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

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

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

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

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

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

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

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

  769. 

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

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

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

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

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

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

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

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

  778. 

  779.         register vector signed short line0S = vec_add(temp0S, temp1S);

  780.         register vector signed short line1S = vec_sub(temp0S, temp1S);

  781.         register vector signed short line2S = vec_add(temp2S, temp3S);

  782.         register vector signed short line3S = vec_sub(temp2S, temp3S);

  783.         register vector signed short line4S = vec_add(temp4S, temp5S);

  784.         register vector signed short line5S = vec_sub(temp4S, temp5S);

  785.         register vector signed short line6S = vec_add(temp6S, temp7S);

  786.         register vector signed short line7S = vec_sub(temp6S, temp7S);

  787. 

  788.         register vector signed short line0BS = vec_add(line0S, line2S);

  789.         register vector signed short line2BS = vec_sub(line0S, line2S);

  790.         register vector signed short line1BS = vec_add(line1S, line3S);

  791.         register vector signed short line3BS = vec_sub(line1S, line3S);

  792.         register vector signed short line4BS = vec_add(line4S, line6S);

  793.         register vector signed short line6BS = vec_sub(line4S, line6S);

  794.         register vector signed short line5BS = vec_add(line5S, line7S);

  795.         register vector signed short line7BS = vec_sub(line5S, line7S);

  796. 

  797.         register vector signed short line0CS = vec_add(line0BS, line4BS);

  798.         register vector signed short line4CS = vec_sub(line0BS, line4BS);

  799.         register vector signed short line1CS = vec_add(line1BS, line5BS);

  800.         register vector signed short line5CS = vec_sub(line1BS, line5BS);

  801.         register vector signed short line2CS = vec_add(line2BS, line6BS);

  802.         register vector signed short line6CS = vec_sub(line2BS, line6BS);

  803.         register vector signed short line3CS = vec_add(line3BS, line7BS);

  804.         register vector signed short line7CS = vec_sub(line3BS, line7BS);

  805. 

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

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

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

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

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

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

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

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

  814. 

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

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

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

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

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

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

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

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

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

  824.         vsum = vec_splat(vsum, 3);

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

  826.     }

  827.     return sum;

  828. }

  829. 

  830. static int hadamard8_diff16_altivec(MpegEncContext *s, uint8_t *dst,

  831.                                     uint8_t *src, int stride, int h)

  832. {

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

  834. 

  835.     if (h == 16) {

  836.         dst   += 8 * stride;

  837.         src   += 8 * stride;

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

  839.     }

  840.     return score;

  841. }

  842. 

  843. av_cold void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx,

  844.                                      unsigned high_bit_depth)

  845. {

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

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

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

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

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

  851. 

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

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

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

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

  856. 

  857.     c->diff_pixels = diff_pixels_altivec;

  858. 

  859.     if (!high_bit_depth) {

  860.         c->get_pixels = get_pixels_altivec;

  861.     }

  862. 

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

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

  865. }