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.
45179 Commits
32 Branches
2.4GB
Tree: a3a501df24
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'a3a501df24'
${ noResults }
FFmpeg/libavcodec/ppc/me_cmp.c

765 lines
34KB
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. 

  25. #include "libavutil/attributes.h"

  26. #include "libavutil/cpu.h"

  27. #include "libavutil/ppc/cpu.h"

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

  29. 

  30. #include "libavcodec/avcodec.h"

  31. #include "libavcodec/mpegvideo.h"

  32. #include "libavcodec/me_cmp.h"

  33. 

  34. #if HAVE_ALTIVEC && HAVE_BIGENDIAN

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

  36.                             ptrdiff_t stride, int h)

  37. {

  38.     int i, s = 0;

  39.     const vector unsigned char zero =

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

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

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

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

  44.     vector signed int sumdiffs;

  45. 

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

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

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

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

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

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

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

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

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

  55. 

  56.         /* Calculate the average vector. */

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

  58. 

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

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

  61.                                           vec_min(pix1v, avgv));

  62. 

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

  64.         sad = vec_sum4s(t5, sad);

  65. 

  66.         pix1 += stride;

  67.         pix2 += stride;

  68.     }

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

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

  71.     sumdiffs = vec_splat(sumdiffs, 3);

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

  73. 

  74.     return s;

  75. }

  76. 

  77. static int sad16_y2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  78.                             ptrdiff_t stride, int h)

  79. {

  80.     int i, s = 0;

  81.     const vector unsigned char zero =

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

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

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

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

  86.     vector signed int sumdiffs;

  87.     uint8_t *pix3 = pix2 + stride;

  88. 

  89.     /* Due to the fact that pix3 = pix2 + stride, the pix3 of one

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

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

  92.      * time around the loop.

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

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

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

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

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

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

  99. 

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

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

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

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

  104.         pix1v = vec_ld(0,  pix1);

  105. 

  106.         pix2l = vec_ld(0,  pix3);

  107.         pix2r = vec_ld(15, pix3);

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

  109. 

  110.         /* Calculate the average vector. */

  111.         avgv = vec_avg(pix2v, pix3v);

  112. 

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

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

  115. 

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

  117.         sad = vec_sum4s(t5, sad);

  118. 

  119.         pix1 += stride;

  120.         pix2v = pix3v;

  121.         pix3 += stride;

  122.     }

  123. 

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

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

  126.     sumdiffs = vec_splat(sumdiffs, 3);

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

  128.     return s;

  129. }

  130. 

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

  132.                              ptrdiff_t stride, int h)

  133. {

  134.     int i, s = 0;

  135.     uint8_t *pix3 = pix2 + stride;

  136.     const vector unsigned char zero =

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

  138.     const vector unsigned short two =

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

  140.     vector unsigned char avgv, t5;

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

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

  143.     vector unsigned char pix1v, pix3v, pix3iv;

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

  145.     vector unsigned short avghv, avglv;

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

  147.     vector signed int sumdiffs;

  148. 

  149.     /* Due to the fact that pix3 = pix2 + stride, the pix3 of one

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

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

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

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

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

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

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

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

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

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

  160. 

  161.     vector unsigned short pix2hv  =

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

  163.     vector unsigned short pix2lv  =

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

  165.     vector unsigned short pix2ihv =

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

  167.     vector unsigned short pix2ilv =

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

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

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

  171.     vector unsigned short t3, t4;

  172. 

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

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

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

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

  177.         pix1v  = vec_ld(0, pix1);

  178. 

  179.         pix2l  = vec_ld(0, pix3);

  180.         pix2r  = vec_ld(16, pix3);

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

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

  183. 

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

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

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

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

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

  189. 

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

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

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

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

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

  195. 

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

  197.         t3 = vec_add(pix3hv, pix3ihv);

  198.         t4 = vec_add(pix3lv, pix3ilv);

  199. 

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

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

  202. 

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

  204.         avgv = vec_pack(avghv, avglv);

  205. 

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

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

  208. 

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

  210.         sad = vec_sum4s(t5, sad);

  211. 

  212.         pix1 += stride;

  213.         pix3 += stride;

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

  215.         t1 = t3;

  216.         t2 = t4;

  217.     }

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

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

  220.     sumdiffs = vec_splat(sumdiffs, 3);

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

  222. 

  223.     return s;

  224. }

  225. 

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

  227.                          ptrdiff_t stride, int h)

  228. {

  229.     int i, s;

  230.     const vector unsigned int zero =

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

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

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

  234.     vector signed int sumdiffs;

  235. 

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

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

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

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

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

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

  242. 

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

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

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

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

  247. 

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

  249.         sad = vec_sum4s(t5, sad);

  250. 

  251.         pix1 += stride;

  252.         pix2 += stride;

  253.     }

  254. 

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

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

  257.     sumdiffs = vec_splat(sumdiffs, 3);

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

  259. 

  260.     return s;

  261. }

  262. 

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

  264.                         ptrdiff_t stride, int h)

  265. {

  266.     int i, s;

  267.     const vector unsigned int zero =

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

  269.     const vector unsigned char permclear =

  270.         (vector unsigned char)

  271.         { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };

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

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

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

  275.     vector signed int sumdiffs;

  276. 

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

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

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

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

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

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

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

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

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

  286.                                           permclear);

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

  288.                                           permclear);

  289. 

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

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

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

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

  294. 

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

  296.         sad = vec_sum4s(t5, sad);

  297. 

  298.         pix1 += stride;

  299.         pix2 += stride;

  300.     }

  301. 

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

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

  304.     sumdiffs = vec_splat(sumdiffs, 3);

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

  306. 

  307.     return s;

  308. }

  309. 

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

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

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

  313.                         ptrdiff_t stride, int h)

  314. {

  315.     int i, s;

  316.     const vector unsigned int zero =

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

  318.     const vector unsigned char permclear =

  319.         (vector unsigned char)

  320.         { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };

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

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

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

  324.     vector signed int sumsqr;

  325. 

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

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

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

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

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

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

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

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

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

  335.                                           permclear);

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

  337.                                           permclear);

  338. 

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

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

  341. 

  342.         /* Calculate abs differences vector. */

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

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

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

  346. 

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

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

  349. 

  350.         pix1 += stride;

  351.         pix2 += stride;

  352.     }

  353. 

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

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

  356.     sumsqr = vec_splat(sumsqr, 3);

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

  358. 

  359.     return s;

  360. }

  361. 

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

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

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

  365.                          ptrdiff_t stride, int h)

  366. {

  367.     int i, s;

  368.     const vector unsigned int zero =

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

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

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

  372.     vector signed int sumsqr;

  373. 

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

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

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

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

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

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

  380. 

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

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

  383. 

  384.         /* Calculate abs differences vector. */

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

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

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

  388. 

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

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

  391. 

  392.         pix1 += stride;

  393.         pix2 += stride;

  394.     }

  395. 

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

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

  398.     sumsqr = vec_splat(sumsqr, 3);

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

  400. 

  401.     return s;

  402. }

  403. 

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

  405.                                      uint8_t *src, ptrdiff_t stride, int h)

  406. {

  407.     int sum;

  408.     register const vector unsigned char vzero =

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

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

  411.                                  temp5, temp6, temp7;

  412.     {

  413.         register const vector signed short vprod1 =

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

  415.         register const vector signed short vprod2 =

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

  417.         register const vector signed short vprod3 =

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

  419.         register const vector unsigned char perm1 =

  420.             (const vector unsigned char)

  421.             { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,

  422.               0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };

  423.         register const vector unsigned char perm2 =

  424.             (const vector unsigned char)

  425.             { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,

  426.               0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };

  427.         register const vector unsigned char perm3 =

  428.             (const vector unsigned char)

  429.             { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,

  430.               0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };

  431. 

  432. #define ONEITERBUTTERFLY(i, res)                                            \

  433.     {                                                                       \

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

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

  436.         register vector unsigned char srcO =                                \

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

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

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

  440.         register vector unsigned char dstO =                                \

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

  442.                                                                             \

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

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

  445.         register vector signed short srcV =                                 \

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

  447.                                              (vector signed char) srcO);    \

  448.         register vector signed short dstV =                                 \

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

  450.                                              (vector signed char) dstO);    \

  451.                                                                             \

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

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

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

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

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

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

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

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

  460.     }

  461.         ONEITERBUTTERFLY(0, temp0);

  462.         ONEITERBUTTERFLY(1, temp1);

  463.         ONEITERBUTTERFLY(2, temp2);

  464.         ONEITERBUTTERFLY(3, temp3);

  465.         ONEITERBUTTERFLY(4, temp4);

  466.         ONEITERBUTTERFLY(5, temp5);

  467.         ONEITERBUTTERFLY(6, temp6);

  468.         ONEITERBUTTERFLY(7, temp7);

  469.     }

  470. #undef ONEITERBUTTERFLY

  471.     {

  472.         register vector signed int vsum;

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

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

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

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

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

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

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

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

  481. 

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

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

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

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

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

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

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

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

  490. 

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

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

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

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

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

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

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

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

  499. 

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

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

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

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

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

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

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

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

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

  509.         vsum = vec_splat(vsum, 3);

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

  511.     }

  512.     return sum;

  513. }

  514. 

  515. /*

  516.  * 16x8 works with 16 elements; it allows to avoid replicating loads, and

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

  518.  * inside hadamard8_diff16_altivec.

  519.  *

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

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

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

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

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

  525.  *

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

  527.  * 7450, and its code isn't much faster than gcc-3.3 on the 7450 (but uses

  528.  * 25% fewer instructions...)

  529.  *

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

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

  532.  */

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

  534.                                       uint8_t *src, ptrdiff_t stride, int h)

  535. {

  536.     int sum;

  537.     register vector signed short

  538.         temp0 __asm__ ("v0"),

  539.         temp1 __asm__ ("v1"),

  540.         temp2 __asm__ ("v2"),

  541.         temp3 __asm__ ("v3"),

  542.         temp4 __asm__ ("v4"),

  543.         temp5 __asm__ ("v5"),

  544.         temp6 __asm__ ("v6"),

  545.         temp7 __asm__ ("v7");

  546.     register vector signed short

  547.         temp0S __asm__ ("v8"),

  548.         temp1S __asm__ ("v9"),

  549.         temp2S __asm__ ("v10"),

  550.         temp3S __asm__ ("v11"),

  551.         temp4S __asm__ ("v12"),

  552.         temp5S __asm__ ("v13"),

  553.         temp6S __asm__ ("v14"),

  554.         temp7S __asm__ ("v15");

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

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

  557.     {

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

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

  560. 

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

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

  563. 

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

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

  566. 

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

  568.             (const vector unsigned char)

  569.             { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,

  570.               0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };

  571. 

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

  573.             (const vector unsigned char)

  574.             { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,

  575.               0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };

  576. 

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

  578.             (const vector unsigned char)

  579.             { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,

  580.               0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };

  581. 

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

  583.     {                                                                       \

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

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

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

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

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

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

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

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

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

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

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

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

  596.                                                                             \

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

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

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

  600.                                              (vector signed char) srcO);    \

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

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

  603.                                              (vector signed char) dstO);    \

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

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

  606.                                              (vector signed char) srcO);    \

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

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

  609.                                              (vector signed char) dstO);    \

  610.                                                                             \

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

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

  613.             vec_sub(srcV, dstV);                                            \

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

  615.             vec_sub(srcW, dstW);                                            \

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  638.     }

  639.         ONEITERBUTTERFLY(0, temp0, temp0S);

  640.         ONEITERBUTTERFLY(1, temp1, temp1S);

  641.         ONEITERBUTTERFLY(2, temp2, temp2S);

  642.         ONEITERBUTTERFLY(3, temp3, temp3S);

  643.         ONEITERBUTTERFLY(4, temp4, temp4S);

  644.         ONEITERBUTTERFLY(5, temp5, temp5S);

  645.         ONEITERBUTTERFLY(6, temp6, temp6S);

  646.         ONEITERBUTTERFLY(7, temp7, temp7S);

  647.     }

  648. #undef ONEITERBUTTERFLY

  649.     {

  650.         register vector signed int vsum;

  651. 

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

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

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

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

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

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

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

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

  660. 

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

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

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

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

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

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

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

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

  669. 

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

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

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

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

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

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

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

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

  678. 

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

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

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

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

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

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

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

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

  687. 

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

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

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

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

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

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

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

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

  696. 

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

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

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

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

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

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

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

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

  705. 

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

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

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

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

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

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

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

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

  714. 

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

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

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

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

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

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

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

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

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

  724.         vsum = vec_splat(vsum, 3);

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

  726.     }

  727.     return sum;

  728. }

  729. 

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

  731.                                     uint8_t *src, ptrdiff_t stride, int h)

  732. {

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

  734. 

  735.     if (h == 16) {

  736.         dst   += 8 * stride;

  737.         src   += 8 * stride;

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

  739.     }

  740.     return score;

  741. }

  742. #endif /* HAVE_ALTIVEC && HAVE_BIGENDIAN */

  743. 

  744. av_cold void ff_me_cmp_init_ppc(MECmpContext *c, AVCodecContext *avctx)

  745. {

  746. #if HAVE_ALTIVEC && HAVE_BIGENDIAN

  747.     if (!PPC_ALTIVEC(av_get_cpu_flags()))

  748.         return;

  749. 

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

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

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

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

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

  755. 

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

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

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

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

  760. 

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

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

  763. #endif /* HAVE_ALTIVEC && HAVE_BIGENDIAN */

  764. }