Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  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 FFmpeg.
  7.  *
  8.  * FFmpeg 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.  * FFmpeg 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 FFmpeg; 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/cpu.h"
  30. #include "libavutil/ppc/cpu.h"
  31. #include "libavutil/ppc/types_altivec.h"
  32. #include "libavutil/ppc/util_altivec.h"
  33. #include "libavcodec/avcodec.h"
  34. #include "libavcodec/mpegvideo.h"
  35. #include "libavcodec/me_cmp.h"
  36.  
  37. #if HAVE_ALTIVEC
  38.  
  39. #if HAVE_BIGENDIAN
  40. #define GET_PERM(per1, per2, pix) {\
  41.     per1 = vec_lvsl(0, pix);\
  42.     per2 = vec_add(per1, vec_splat_u8(1));\
  43. }
  44. #define LOAD_PIX(v, iv, pix, per1, per2) {\
  45.     vector unsigned char pix2l  = vec_ld(0,  pix);\
  46.     vector unsigned char pix2r  = vec_ld(16, pix);\
  47.     v  = vec_perm(pix2l, pix2r, per1);\
  48.     iv = vec_perm(pix2l, pix2r, per2);\
  49. }
  50. #else
  51. #define GET_PERM(per1, per2, pix) {}
  52. #define LOAD_PIX(v, iv, pix, per1, per2) {\
  53.     v  = vec_vsx_ld(0,  pix);\
  54.     iv = vec_vsx_ld(1,  pix);\
  55. }
  56. #endif
  57. static int sad16_x2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  58.                             ptrdiff_t stride, int h)
  59. {
  60.     int i;
  61.     int __attribute__((aligned(16))) s = 0;
  62.     const vector unsigned char zero =
  63.         (const vector unsigned char) vec_splat_u8(0);
  64.     vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
  65.     vector signed int sumdiffs;
  66.     vector unsigned char perm1, perm2, pix2v, pix2iv;
  67.  
  68.     GET_PERM(perm1, perm2, pix2);
  69.     for (i = 0; i < h; i++) {
  70.         /* Read unaligned pixels into our vectors. The vectors are as follows:
  71.          * pix1v: pix1[0] - pix1[15]
  72.          * pix2v: pix2[0] - pix2[15]      pix2iv: pix2[1] - pix2[16] */
  73.         vector unsigned char pix1v  = vec_ld(0,  pix1);
  74.         LOAD_PIX(pix2v, pix2iv, pix2, perm1, perm2);
  75.  
  76.         /* Calculate the average vector. */
  77.         vector unsigned char avgv = vec_avg(pix2v, pix2iv);
  78.  
  79.         /* Calculate a sum of abs differences vector. */
  80.         vector unsigned char t5 = vec_sub(vec_max(pix1v, avgv),
  81.                                           vec_min(pix1v, avgv));
  82.  
  83.         /* Add each 4 pixel group together and put 4 results into sad. */
  84.         sad = vec_sum4s(t5, sad);
  85.  
  86.         pix1 += stride;
  87.         pix2 += stride;
  88.     }
  89.     /* Sum up the four partial sums, and put the result into s. */
  90.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
  91.     sumdiffs = vec_splat(sumdiffs, 3);
  92.     vec_ste(sumdiffs, 0, &s);
  93.  
  94.     return s;
  95. }
  96.  
  97. static int sad16_y2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  98.                             ptrdiff_t stride, int h)
  99. {
  100.     int i;
  101.     int  __attribute__((aligned(16))) s = 0;
  102.     const vector unsigned char zero =
  103.         (const vector unsigned char) vec_splat_u8(0);
  104.     vector unsigned char pix1v, pix3v, avgv, t5;
  105.     vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
  106.     vector signed int sumdiffs;
  107.  
  108.     uint8_t *pix3 = pix2 + stride;
  109.  
  110.     /* Due to the fact that pix3 = pix2 + stride, the pix3 of one
  111.      * iteration becomes pix2 in the next iteration. We can use this
  112.      * fact to avoid a potentially expensive unaligned read, each
  113.      * time around the loop.
  114.      * Read unaligned pixels into our vectors. The vectors are as follows:
  115.      * pix2v: pix2[0] - pix2[15]
  116.      * Split the pixel vectors into shorts. */
  117.     vector unsigned char pix2v = VEC_LD(0, pix2);
  118.  
  119.     for (i = 0; i < h; i++) {
  120.         /* Read unaligned pixels into our vectors. The vectors are as follows:
  121.          * pix1v: pix1[0] - pix1[15]
  122.          * pix3v: pix3[0] - pix3[15] */
  123.         pix1v = vec_ld(0,  pix1);
  124.         pix3v = VEC_LD(0,  pix3);
  125.  
  126.         /* Calculate the average vector. */
  127.         avgv = vec_avg(pix2v, pix3v);
  128.  
  129.         /* Calculate a sum of abs differences vector. */
  130.         t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
  131.  
  132.         /* Add each 4 pixel group together and put 4 results into sad. */
  133.         sad = vec_sum4s(t5, sad);
  134.  
  135.         pix1 += stride;
  136.         pix2v = pix3v;
  137.         pix3 += stride;
  138.     }
  139.  
  140.     /* Sum up the four partial sums, and put the result into s. */
  141.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
  142.     sumdiffs = vec_splat(sumdiffs, 3);
  143.     vec_ste(sumdiffs, 0, &s);
  144.     return s;
  145. }
  146.  
  147. static int sad16_xy2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  148.                              ptrdiff_t stride, int h)
  149. {
  150.     int i;
  151.     int  __attribute__((aligned(16))) s = 0;
  152.     uint8_t *pix3 = pix2 + stride;
  153.     const vector unsigned char zero =
  154.         (const vector unsigned char) vec_splat_u8(0);
  155.     const vector unsigned short two =
  156.         (const vector unsigned short) vec_splat_u16(2);
  157.     vector unsigned char avgv, t5;
  158.     vector unsigned char pix1v, pix3v, pix3iv;
  159.     vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
  160.     vector unsigned short avghv, avglv;
  161.     vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
  162.     vector signed int sumdiffs;
  163.     vector unsigned char perm1, perm2, pix2v, pix2iv;
  164.     GET_PERM(perm1, perm2, pix2);
  165.  
  166.     /* Due to the fact that pix3 = pix2 + stride, the pix3 of one
  167.      * iteration becomes pix2 in the next iteration. We can use this
  168.      * fact to avoid a potentially expensive unaligned read, as well
  169.      * as some splitting, and vector addition each time around the loop.
  170.      * Read unaligned pixels into our vectors. The vectors are as follows:
  171.      * pix2v: pix2[0] - pix2[15]  pix2iv: pix2[1] - pix2[16]
  172.      * Split the pixel vectors into shorts. */
  173.     LOAD_PIX(pix2v, pix2iv, pix2, perm1, perm2);
  174.     vector unsigned short pix2hv  =
  175.         (vector unsigned short) VEC_MERGEH(zero, pix2v);
  176.     vector unsigned short pix2lv  =
  177.         (vector unsigned short) VEC_MERGEL(zero, pix2v);
  178.     vector unsigned short pix2ihv =
  179.         (vector unsigned short) VEC_MERGEH(zero, pix2iv);
  180.     vector unsigned short pix2ilv =
  181.         (vector unsigned short) VEC_MERGEL(zero, pix2iv);
  182.  
  183.     vector unsigned short t1 = vec_add(pix2hv, pix2ihv);
  184.     vector unsigned short t2 = vec_add(pix2lv, pix2ilv);
  185.     vector unsigned short t3, t4;
  186.  
  187.     for (i = 0; i < h; i++) {
  188.         /* Read unaligned pixels into our vectors. The vectors are as follows:
  189.          * pix1v: pix1[0] - pix1[15]
  190.          * pix3v: pix3[0] - pix3[15]      pix3iv: pix3[1] - pix3[16] */
  191.         pix1v  = vec_ld(0, pix1);
  192.         LOAD_PIX(pix3v, pix3iv, pix3, perm1, perm2);
  193.  
  194.         /* Note that AltiVec does have vec_avg, but this works on vector pairs
  195.          * and rounds up. We could do avg(avg(a, b), avg(c, d)), but the
  196.          * rounding would mean that, for example, avg(3, 0, 0, 1) = 2, when
  197.          * it should be 1. Instead, we have to split the pixel vectors into
  198.          * vectors of shorts and do the averaging by hand. */
  199.  
  200.         /* Split the pixel vectors into shorts. */
  201.         pix3hv  = (vector unsigned short) VEC_MERGEH(zero, pix3v);
  202.         pix3lv  = (vector unsigned short) VEC_MERGEL(zero, pix3v);
  203.         pix3ihv = (vector unsigned short) VEC_MERGEH(zero, pix3iv);
  204.         pix3ilv = (vector unsigned short) VEC_MERGEL(zero, pix3iv);
  205.  
  206.         /* Do the averaging on them. */
  207.         t3 = vec_add(pix3hv, pix3ihv);
  208.         t4 = vec_add(pix3lv, pix3ilv);
  209.  
  210.         avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
  211.         avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
  212.  
  213.         /* Pack the shorts back into a result. */
  214.         avgv = vec_pack(avghv, avglv);
  215.  
  216.         /* Calculate a sum of abs differences vector. */
  217.         t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
  218.  
  219.         /* Add each 4 pixel group together and put 4 results into sad. */
  220.         sad = vec_sum4s(t5, sad);
  221.  
  222.         pix1 += stride;
  223.         pix3 += stride;
  224.         /* Transfer the calculated values for pix3 into pix2. */
  225.         t1 = t3;
  226.         t2 = t4;
  227.     }
  228.     /* Sum up the four partial sums, and put the result into s. */
  229.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
  230.     sumdiffs = vec_splat(sumdiffs, 3);
  231.     vec_ste(sumdiffs, 0, &s);
  232.  
  233.     return s;
  234. }
  235.  
  236. static int sad16_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  237.                          ptrdiff_t stride, int h)
  238. {
  239.     int i;
  240.     int  __attribute__((aligned(16))) s;
  241.     const vector unsigned int zero =
  242.         (const vector unsigned int) vec_splat_u32(0);
  243.     vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
  244.     vector signed int sumdiffs;
  245.  
  246.     for (i = 0; i < h; i++) {
  247.         /* Read potentially unaligned pixels into t1 and t2. */
  248.         vector unsigned char t1 =vec_ld(0, pix1);
  249.         vector unsigned char t2 = VEC_LD(0, pix2);
  250.  
  251.         /* Calculate a sum of abs differences vector. */
  252.         vector unsigned char t3 = vec_max(t1, t2);
  253.         vector unsigned char t4 = vec_min(t1, t2);
  254.         vector unsigned char 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 += stride;
  260.         pix2 += stride;
  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(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  272.                         ptrdiff_t stride, int h)
  273. {
  274.     int i;
  275.     int  __attribute__((aligned(16))) s;
  276.     const vector unsigned int zero =
  277.         (const vector unsigned int) vec_splat_u32(0);
  278.     const vector unsigned char permclear =
  279.         (vector unsigned char)
  280.         { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
  281.     vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
  282.     vector signed int sumdiffs;
  283.  
  284.     for (i = 0; i < h; i++) {
  285.         /* Read potentially unaligned pixels into t1 and t2.
  286.          * Since we're reading 16 pixels, and actually only want 8,
  287.          * mask out the last 8 pixels. The 0s don't change the sum. */
  288.         vector unsigned char pix1l = VEC_LD(0, pix1);
  289.         vector unsigned char pix2l = VEC_LD(0, pix2);
  290.         vector unsigned char t1 = vec_and(pix1l, permclear);
  291.         vector unsigned char t2 = vec_and(pix2l, permclear);
  292.  
  293.         /* Calculate a sum of abs differences vector. */
  294.         vector unsigned char t3 = vec_max(t1, t2);
  295.         vector unsigned char t4 = vec_min(t1, t2);
  296.         vector unsigned char t5 = vec_sub(t3, t4);
  297.  
  298.         /* Add each 4 pixel group together and put 4 results into sad. */
  299.         sad = vec_sum4s(t5, sad);
  300.  
  301.         pix1 += stride;
  302.         pix2 += stride;
  303.     }
  304.  
  305.     /* Sum up the four partial sums, and put the result into s. */
  306.     sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
  307.     sumdiffs = vec_splat(sumdiffs, 3);
  308.     vec_ste(sumdiffs, 0, &s);
  309.  
  310.     return s;
  311. }
  312.  
  313. /* Sum of Squared Errors for an 8x8 block, AltiVec-enhanced.
  314.  * It's the sad8_altivec code above w/ squaring added. */
  315. static int sse8_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  316.                         ptrdiff_t stride, int h)
  317. {
  318.     int i;
  319.     int  __attribute__((aligned(16))) s;
  320.     const vector unsigned int zero =
  321.         (const vector unsigned int) vec_splat_u32(0);
  322.     const vector unsigned char permclear =
  323.         (vector unsigned char)
  324.         { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
  325.     vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
  326.     vector signed int sumsqr;
  327.  
  328.     for (i = 0; i < h; i++) {
  329.         /* Read potentially unaligned pixels into t1 and t2.
  330.          * Since we're reading 16 pixels, and actually only want 8,
  331.          * mask out the last 8 pixels. The 0s don't change the sum. */
  332.         vector unsigned char t1 = vec_and(VEC_LD(0, pix1), permclear);
  333.         vector unsigned char t2 = vec_and(VEC_LD(0, pix2), permclear);
  334.  
  335.         /* Since we want to use unsigned chars, we can take advantage
  336.          * of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
  337.  
  338.         /* Calculate abs differences vector. */
  339.         vector unsigned char t3 = vec_max(t1, t2);
  340.         vector unsigned char t4 = vec_min(t1, t2);
  341.         vector unsigned char t5 = vec_sub(t3, t4);
  342.  
  343.         /* Square the values and add them to our sum. */
  344.         sum = vec_msum(t5, t5, sum);
  345.  
  346.         pix1 += stride;
  347.         pix2 += stride;
  348.     }
  349.  
  350.     /* Sum up the four partial sums, and put the result into s. */
  351.     sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
  352.     sumsqr = vec_splat(sumsqr, 3);
  353.     vec_ste(sumsqr, 0, &s);
  354.  
  355.     return s;
  356. }
  357.  
  358. /* Sum of Squared Errors for a 16x16 block, AltiVec-enhanced.
  359.  * It's the sad16_altivec code above w/ squaring added. */
  360. static int sse16_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  361.                          ptrdiff_t stride, int h)
  362. {
  363.     int i;
  364.     int  __attribute__((aligned(16))) s;
  365.     const vector unsigned int zero =
  366.         (const vector unsigned int) vec_splat_u32(0);
  367.     vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
  368.     vector signed int sumsqr;
  369.  
  370.     for (i = 0; i < h; i++) {
  371.         /* Read potentially unaligned pixels into t1 and t2. */
  372.         vector unsigned char t1 = vec_ld(0, pix1);
  373.         vector unsigned char t2 = VEC_LD(0, pix2);
  374.  
  375.         /* Since we want to use unsigned chars, we can take advantage
  376.          * of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
  377.  
  378.         /* Calculate abs differences vector. */
  379.         vector unsigned char t3 = vec_max(t1, t2);
  380.         vector unsigned char t4 = vec_min(t1, t2);
  381.         vector unsigned char t5 = vec_sub(t3, t4);
  382.  
  383.         /* Square the values and add them to our sum. */
  384.         sum = vec_msum(t5, t5, sum);
  385.  
  386.         pix1 += stride;
  387.         pix2 += stride;
  388.     }
  389.  
  390.     /* Sum up the four partial sums, and put the result into s. */
  391.     sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
  392.     sumsqr = vec_splat(sumsqr, 3);
  393.  
  394.     vec_ste(sumsqr, 0, &s);
  395.     return s;
  396. }
  397.  
  398. static int hadamard8_diff8x8_altivec(MpegEncContext *s, uint8_t *dst,
  399.                                      uint8_t *src, ptrdiff_t stride, int h)
  400. {
  401.     int __attribute__((aligned(16))) sum;
  402.     register const vector unsigned char vzero =
  403.         (const vector unsigned char) vec_splat_u8(0);
  404.     register vector signed short temp0, temp1, temp2, temp3, temp4,
  405.                                  temp5, temp6, temp7;
  406.     {
  407.         register const vector signed short vprod1 =
  408.             (const vector signed short) { 1, -1, 1, -1, 1, -1, 1, -1 };
  409.         register const vector signed short vprod2 =
  410.             (const vector signed short) { 1, 1, -1, -1, 1, 1, -1, -1 };
  411.         register const vector signed short vprod3 =
  412.             (const vector signed short) { 1, 1, 1, 1, -1, -1, -1, -1 };
  413.         register const vector unsigned char perm1 =
  414.             (const vector unsigned char)
  415.             { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
  416.               0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };
  417.         register const vector unsigned char perm2 =
  418.             (const vector unsigned char)
  419.             { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
  420.               0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };
  421.         register const vector unsigned char perm3 =
  422.             (const vector unsigned char)
  423.             { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
  424.               0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
  425.  
  426.  
  427. #define ONEITERBUTTERFLY(i, res)                                            \
  428.     {                                                                       \
  429.         register vector unsigned char srcO =  unaligned_load(stride * i, src);  \
  430.         register vector unsigned char dstO = unaligned_load(stride * i, dst);\
  431.                                                                             \
  432.         /* Promote the unsigned chars to signed shorts. */                  \
  433.         /* We're in the 8x8 function, we only care for the first 8. */      \
  434.         register vector signed short srcV =                                 \
  435.             (vector signed short) VEC_MERGEH((vector signed char) vzero,    \
  436.                                              (vector signed char) srcO);    \
  437.         register vector signed short dstV =                                 \
  438.             (vector signed short) VEC_MERGEH((vector signed char) vzero,    \
  439.                                              (vector signed char) dstO);    \
  440.                                                                             \
  441.         /* subtractions inside the first butterfly */                       \
  442.         register vector signed short but0 = vec_sub(srcV, dstV);            \
  443.         register vector signed short op1  = vec_perm(but0, but0, perm1);    \
  444.         register vector signed short but1 = vec_mladd(but0, vprod1, op1);   \
  445.         register vector signed short op2  = vec_perm(but1, but1, perm2);    \
  446.         register vector signed short but2 = vec_mladd(but1, vprod2, op2);   \
  447.         register vector signed short op3  = vec_perm(but2, but2, perm3);    \
  448.         res  = vec_mladd(but2, vprod3, op3);                                \
  449.     }
  450.  
  451.         ONEITERBUTTERFLY(0, temp0);
  452.         ONEITERBUTTERFLY(1, temp1);
  453.         ONEITERBUTTERFLY(2, temp2);
  454.         ONEITERBUTTERFLY(3, temp3);
  455.         ONEITERBUTTERFLY(4, temp4);
  456.         ONEITERBUTTERFLY(5, temp5);
  457.         ONEITERBUTTERFLY(6, temp6);
  458.         ONEITERBUTTERFLY(7, temp7);
  459.     }
  460. #undef ONEITERBUTTERFLY
  461.     {
  462.         register vector signed int vsum;
  463.         register vector signed short line0  = vec_add(temp0, temp1);
  464.         register vector signed short line1  = vec_sub(temp0, temp1);
  465.         register vector signed short line2  = vec_add(temp2, temp3);
  466.         register vector signed short line3  = vec_sub(temp2, temp3);
  467.         register vector signed short line4  = vec_add(temp4, temp5);
  468.         register vector signed short line5  = vec_sub(temp4, temp5);
  469.         register vector signed short line6  = vec_add(temp6, temp7);
  470.         register vector signed short line7  = vec_sub(temp6, temp7);
  471.  
  472.         register vector signed short line0B = vec_add(line0, line2);
  473.         register vector signed short line2B = vec_sub(line0, line2);
  474.         register vector signed short line1B = vec_add(line1, line3);
  475.         register vector signed short line3B = vec_sub(line1, line3);
  476.         register vector signed short line4B = vec_add(line4, line6);
  477.         register vector signed short line6B = vec_sub(line4, line6);
  478.         register vector signed short line5B = vec_add(line5, line7);
  479.         register vector signed short line7B = vec_sub(line5, line7);
  480.  
  481.         register vector signed short line0C = vec_add(line0B, line4B);
  482.         register vector signed short line4C = vec_sub(line0B, line4B);
  483.         register vector signed short line1C = vec_add(line1B, line5B);
  484.         register vector signed short line5C = vec_sub(line1B, line5B);
  485.         register vector signed short line2C = vec_add(line2B, line6B);
  486.         register vector signed short line6C = vec_sub(line2B, line6B);
  487.         register vector signed short line3C = vec_add(line3B, line7B);
  488.         register vector signed short line7C = vec_sub(line3B, line7B);
  489.  
  490.         vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
  491.         vsum = vec_sum4s(vec_abs(line1C), vsum);
  492.         vsum = vec_sum4s(vec_abs(line2C), vsum);
  493.         vsum = vec_sum4s(vec_abs(line3C), vsum);
  494.         vsum = vec_sum4s(vec_abs(line4C), vsum);
  495.         vsum = vec_sum4s(vec_abs(line5C), vsum);
  496.         vsum = vec_sum4s(vec_abs(line6C), vsum);
  497.         vsum = vec_sum4s(vec_abs(line7C), vsum);
  498.         vsum = vec_sums(vsum, (vector signed int) vzero);
  499.         vsum = vec_splat(vsum, 3);
  500.  
  501.         vec_ste(vsum, 0, &sum);
  502.     }
  503.     return sum;
  504. }
  505.  
  506. /*
  507.  * 16x8 works with 16 elements; it can avoid replicating loads, and
  508.  * gives the compiler more room for scheduling. It's only used from
  509.  * inside hadamard8_diff16_altivec.
  510.  *
  511.  * Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has
  512.  * a LOT of spill code, it seems gcc (unlike xlc) cannot keep everything in
  513.  * registers by itself. The following code includes hand-made register
  514.  * allocation. It's not clean, but on a 7450 the resulting code is much faster
  515.  * (best case falls from 700+ cycles to 550).
  516.  *
  517.  * xlc doesn't add spill code, but it doesn't know how to schedule for the
  518.  * 7450, and its code isn't much faster than gcc-3.3 on the 7450 (but uses
  519.  * 25% fewer instructions...)
  520.  *
  521.  * On the 970, the hand-made RA is still a win (around 690 vs. around 780),
  522.  * but xlc goes to around 660 on the regular C code...
  523.  */
  524. static int hadamard8_diff16x8_altivec(MpegEncContext *s, uint8_t *dst,
  525.                                       uint8_t *src, ptrdiff_t stride, int h)
  526. {
  527.     int __attribute__((aligned(16))) sum;
  528.     register vector signed short
  529.         temp0 __asm__ ("v0"),
  530.         temp1 __asm__ ("v1"),
  531.         temp2 __asm__ ("v2"),
  532.         temp3 __asm__ ("v3"),
  533.         temp4 __asm__ ("v4"),
  534.         temp5 __asm__ ("v5"),
  535.         temp6 __asm__ ("v6"),
  536.         temp7 __asm__ ("v7");
  537.     register vector signed short
  538.         temp0S __asm__ ("v8"),
  539.         temp1S __asm__ ("v9"),
  540.         temp2S __asm__ ("v10"),
  541.         temp3S __asm__ ("v11"),
  542.         temp4S __asm__ ("v12"),
  543.         temp5S __asm__ ("v13"),
  544.         temp6S __asm__ ("v14"),
  545.         temp7S __asm__ ("v15");
  546.     register const vector unsigned char vzero __asm__ ("v31") =
  547.         (const vector unsigned char) vec_splat_u8(0);
  548.     {
  549.         register const vector signed short vprod1 __asm__ ("v16") =
  550.             (const vector signed short) { 1, -1, 1, -1, 1, -1, 1, -1 };
  551.  
  552.         register const vector signed short vprod2 __asm__ ("v17") =
  553.             (const vector signed short) { 1, 1, -1, -1, 1, 1, -1, -1 };
  554.  
  555.         register const vector signed short vprod3 __asm__ ("v18") =
  556.             (const vector signed short) { 1, 1, 1, 1, -1, -1, -1, -1 };
  557.  
  558.         register const vector unsigned char perm1 __asm__ ("v19") =
  559.             (const vector unsigned char)
  560.             { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
  561.               0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };
  562.  
  563.         register const vector unsigned char perm2 __asm__ ("v20") =
  564.             (const vector unsigned char)
  565.             { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
  566.               0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };
  567.  
  568.         register const vector unsigned char perm3 __asm__ ("v21") =
  569.             (const vector unsigned char)
  570.             { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
  571.               0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
  572.  
  573. #define ONEITERBUTTERFLY(i, res1, res2)                                     \
  574.     {                                                                       \
  575.         register vector unsigned char srcO __asm__ ("v22") =                \
  576.             unaligned_load(stride * i, src);                                    \
  577.         register vector unsigned char dstO __asm__ ("v23") =                \
  578.             unaligned_load(stride * i, dst);\
  579.                                                                             \
  580.         /* Promote the unsigned chars to signed shorts. */                  \
  581.         register vector signed short srcV __asm__ ("v24") =                 \
  582.             (vector signed short) VEC_MERGEH((vector signed char) vzero,    \
  583.                                              (vector signed char) srcO);    \
  584.         register vector signed short dstV __asm__ ("v25") =                 \
  585.             (vector signed short) VEC_MERGEH((vector signed char) vzero,    \
  586.                                              (vector signed char) dstO);    \
  587.         register vector signed short srcW __asm__ ("v26") =                 \
  588.             (vector signed short) VEC_MERGEL((vector signed char) vzero,    \
  589.                                              (vector signed char) srcO);    \
  590.         register vector signed short dstW __asm__ ("v27") =                 \
  591.             (vector signed short) VEC_MERGEL((vector signed char) vzero,    \
  592.                                              (vector signed char) dstO);    \
  593.                                                                             \
  594.         /* subtractions inside the first butterfly */                       \
  595.         register vector signed short but0  __asm__ ("v28") =                \
  596.             vec_sub(srcV, dstV);                                            \
  597.         register vector signed short but0S __asm__ ("v29") =                \
  598.             vec_sub(srcW, dstW);                                            \
  599.         register vector signed short op1   __asm__ ("v30") =                \
  600.             vec_perm(but0, but0, perm1);                                    \
  601.         register vector signed short but1  __asm__ ("v22") =                \
  602.             vec_mladd(but0, vprod1, op1);                                   \
  603.         register vector signed short op1S  __asm__ ("v23") =                \
  604.             vec_perm(but0S, but0S, perm1);                                  \
  605.         register vector signed short but1S __asm__ ("v24") =                \
  606.             vec_mladd(but0S, vprod1, op1S);                                 \
  607.         register vector signed short op2   __asm__ ("v25") =                \
  608.             vec_perm(but1, but1, perm2);                                    \
  609.         register vector signed short but2  __asm__ ("v26") =                \
  610.             vec_mladd(but1, vprod2, op2);                                   \
  611.         register vector signed short op2S  __asm__ ("v27") =                \
  612.             vec_perm(but1S, but1S, perm2);                                  \
  613.         register vector signed short but2S __asm__ ("v28") =                \
  614.             vec_mladd(but1S, vprod2, op2S);                                 \
  615.         register vector signed short op3   __asm__ ("v29") =                \
  616.             vec_perm(but2, but2, perm3);                                    \
  617.         register vector signed short op3S  __asm__ ("v30") =                \
  618.             vec_perm(but2S, but2S, perm3);                                  \
  619.         res1 = vec_mladd(but2, vprod3, op3);                                \
  620.         res2 = vec_mladd(but2S, vprod3, op3S);                              \
  621.     }
  622.  
  623.         ONEITERBUTTERFLY(0, temp0, temp0S);
  624.         ONEITERBUTTERFLY(1, temp1, temp1S);
  625.         ONEITERBUTTERFLY(2, temp2, temp2S);
  626.         ONEITERBUTTERFLY(3, temp3, temp3S);
  627.         ONEITERBUTTERFLY(4, temp4, temp4S);
  628.         ONEITERBUTTERFLY(5, temp5, temp5S);
  629.         ONEITERBUTTERFLY(6, temp6, temp6S);
  630.         ONEITERBUTTERFLY(7, temp7, temp7S);
  631.     }
  632. #undef ONEITERBUTTERFLY
  633.     {
  634.         register vector signed int vsum;
  635.  
  636.         register vector signed short line0  = vec_add(temp0, temp1);
  637.         register vector signed short line1  = vec_sub(temp0, temp1);
  638.         register vector signed short line2  = vec_add(temp2, temp3);
  639.         register vector signed short line3  = vec_sub(temp2, temp3);
  640.         register vector signed short line4  = vec_add(temp4, temp5);
  641.         register vector signed short line5  = vec_sub(temp4, temp5);
  642.         register vector signed short line6  = vec_add(temp6, temp7);
  643.         register vector signed short line7  = vec_sub(temp6, temp7);
  644.  
  645.         register vector signed short line0B = vec_add(line0, line2);
  646.         register vector signed short line2B = vec_sub(line0, line2);
  647.         register vector signed short line1B = vec_add(line1, line3);
  648.         register vector signed short line3B = vec_sub(line1, line3);
  649.         register vector signed short line4B = vec_add(line4, line6);
  650.         register vector signed short line6B = vec_sub(line4, line6);
  651.         register vector signed short line5B = vec_add(line5, line7);
  652.         register vector signed short line7B = vec_sub(line5, line7);
  653.  
  654.         register vector signed short line0C = vec_add(line0B, line4B);
  655.         register vector signed short line4C = vec_sub(line0B, line4B);
  656.         register vector signed short line1C = vec_add(line1B, line5B);
  657.         register vector signed short line5C = vec_sub(line1B, line5B);
  658.         register vector signed short line2C = vec_add(line2B, line6B);
  659.         register vector signed short line6C = vec_sub(line2B, line6B);
  660.         register vector signed short line3C = vec_add(line3B, line7B);
  661.         register vector signed short line7C = vec_sub(line3B, line7B);
  662.  
  663.         register vector signed short line0S = vec_add(temp0S, temp1S);
  664.         register vector signed short line1S = vec_sub(temp0S, temp1S);
  665.         register vector signed short line2S = vec_add(temp2S, temp3S);
  666.         register vector signed short line3S = vec_sub(temp2S, temp3S);
  667.         register vector signed short line4S = vec_add(temp4S, temp5S);
  668.         register vector signed short line5S = vec_sub(temp4S, temp5S);
  669.         register vector signed short line6S = vec_add(temp6S, temp7S);
  670.         register vector signed short line7S = vec_sub(temp6S, temp7S);
  671.  
  672.         register vector signed short line0BS = vec_add(line0S, line2S);
  673.         register vector signed short line2BS = vec_sub(line0S, line2S);
  674.         register vector signed short line1BS = vec_add(line1S, line3S);
  675.         register vector signed short line3BS = vec_sub(line1S, line3S);
  676.         register vector signed short line4BS = vec_add(line4S, line6S);
  677.         register vector signed short line6BS = vec_sub(line4S, line6S);
  678.         register vector signed short line5BS = vec_add(line5S, line7S);
  679.         register vector signed short line7BS = vec_sub(line5S, line7S);
  680.  
  681.         register vector signed short line0CS = vec_add(line0BS, line4BS);
  682.         register vector signed short line4CS = vec_sub(line0BS, line4BS);
  683.         register vector signed short line1CS = vec_add(line1BS, line5BS);
  684.         register vector signed short line5CS = vec_sub(line1BS, line5BS);
  685.         register vector signed short line2CS = vec_add(line2BS, line6BS);
  686.         register vector signed short line6CS = vec_sub(line2BS, line6BS);
  687.         register vector signed short line3CS = vec_add(line3BS, line7BS);
  688.         register vector signed short line7CS = vec_sub(line3BS, line7BS);
  689.  
  690.         vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
  691.         vsum = vec_sum4s(vec_abs(line1C), vsum);
  692.         vsum = vec_sum4s(vec_abs(line2C), vsum);
  693.         vsum = vec_sum4s(vec_abs(line3C), vsum);
  694.         vsum = vec_sum4s(vec_abs(line4C), vsum);
  695.         vsum = vec_sum4s(vec_abs(line5C), vsum);
  696.         vsum = vec_sum4s(vec_abs(line6C), vsum);
  697.         vsum = vec_sum4s(vec_abs(line7C), vsum);
  698.  
  699.         vsum = vec_sum4s(vec_abs(line0CS), vsum);
  700.         vsum = vec_sum4s(vec_abs(line1CS), vsum);
  701.         vsum = vec_sum4s(vec_abs(line2CS), vsum);
  702.         vsum = vec_sum4s(vec_abs(line3CS), vsum);
  703.         vsum = vec_sum4s(vec_abs(line4CS), vsum);
  704.         vsum = vec_sum4s(vec_abs(line5CS), vsum);
  705.         vsum = vec_sum4s(vec_abs(line6CS), vsum);
  706.         vsum = vec_sum4s(vec_abs(line7CS), vsum);
  707.         vsum = vec_sums(vsum, (vector signed int) vzero);
  708.         vsum = vec_splat(vsum, 3);
  709.  
  710.         vec_ste(vsum, 0, &sum);
  711.     }
  712.     return sum;
  713. }
  714.  
  715. static int hadamard8_diff16_altivec(MpegEncContext *s, uint8_t *dst,
  716.                                     uint8_t *src, ptrdiff_t stride, int h)
  717. {
  718.     int score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
  719.  
  720.     if (h == 16) {
  721.         dst   += 8 * stride;
  722.         src   += 8 * stride;
  723.         score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
  724.     }
  725.     return score;
  726. }
  727. #endif /* HAVE_ALTIVEC */
  728.  
  729. av_cold void ff_me_cmp_init_ppc(MECmpContext *c, AVCodecContext *avctx)
  730. {
  731. #if HAVE_ALTIVEC
  732.     if (!PPC_ALTIVEC(av_get_cpu_flags()))
  733.         return;
  734.  
  735.     c->pix_abs[0][1] = sad16_x2_altivec;
  736.     c->pix_abs[0][2] = sad16_y2_altivec;
  737.     c->pix_abs[0][3] = sad16_xy2_altivec;
  738.     c->pix_abs[0][0] = sad16_altivec;
  739.     c->pix_abs[1][0] = sad8_altivec;
  740.  
  741.     c->sad[0] = sad16_altivec;
  742.     c->sad[1] = sad8_altivec;
  743.     c->sse[0] = sse16_altivec;
  744.     c->sse[1] = sse8_altivec;
  745.  
  746.     c->hadamard8_diff[0] = hadamard8_diff16_altivec;
  747.     c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
  748. #endif /* HAVE_ALTIVEC */
  749. }
  750.