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.
40089 Commits
32 Branches
325MB
Tree: 9a9e2f1c8a
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '9a9e2f1c8a'
${ noResults }
FFmpeg/libavcodec/x86/dsputil_mmx.c

378 lines
15KB
Raw Blame History 

  1. /*

  2.  * MMX optimized DSP utils

  3.  * Copyright (c) 2000, 2001 Fabrice Bellard

  4.  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>

  5.  *

  6.  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>

  7.  *

  8.  * This file is part of Libav.

  9.  *

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

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

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

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

  14.  *

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

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

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

  18.  * Lesser General Public License for more details.

  19.  *

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

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

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

  23.  */

  24. 

  25. #include "config.h"

  26. #include "libavutil/cpu.h"

  27. #include "libavutil/x86/asm.h"

  28. #include "dsputil_x86.h"

  29. #include "inline_asm.h"

  30. 

  31. #if HAVE_INLINE_ASM

  32. 

  33. void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,

  34.                                int line_size)

  35. {

  36.     const int16_t *p;

  37.     uint8_t *pix;

  38. 

  39.     /* read the pixels */

  40.     p   = block;

  41.     pix = pixels;

  42.     /* unrolled loop */

  43.     __asm__ volatile (

  44.         "movq      (%3), %%mm0          \n\t"

  45.         "movq     8(%3), %%mm1          \n\t"

  46.         "movq    16(%3), %%mm2          \n\t"

  47.         "movq    24(%3), %%mm3          \n\t"

  48.         "movq    32(%3), %%mm4          \n\t"

  49.         "movq    40(%3), %%mm5          \n\t"

  50.         "movq    48(%3), %%mm6          \n\t"

  51.         "movq    56(%3), %%mm7          \n\t"

  52.         "packuswb %%mm1, %%mm0          \n\t"

  53.         "packuswb %%mm3, %%mm2          \n\t"

  54.         "packuswb %%mm5, %%mm4          \n\t"

  55.         "packuswb %%mm7, %%mm6          \n\t"

  56.         "movq     %%mm0, (%0)           \n\t"

  57.         "movq     %%mm2, (%0, %1)       \n\t"

  58.         "movq     %%mm4, (%0, %1, 2)    \n\t"

  59.         "movq     %%mm6, (%0, %2)       \n\t"

  60.         :: "r" (pix), "r" ((x86_reg) line_size), "r" ((x86_reg) line_size * 3),

  61.            "r" (p)

  62.         : "memory");

  63.     pix += line_size * 4;

  64.     p   += 32;

  65. 

  66.     // if here would be an exact copy of the code above

  67.     // compiler would generate some very strange code

  68.     // thus using "r"

  69.     __asm__ volatile (

  70.         "movq       (%3), %%mm0         \n\t"

  71.         "movq      8(%3), %%mm1         \n\t"

  72.         "movq     16(%3), %%mm2         \n\t"

  73.         "movq     24(%3), %%mm3         \n\t"

  74.         "movq     32(%3), %%mm4         \n\t"

  75.         "movq     40(%3), %%mm5         \n\t"

  76.         "movq     48(%3), %%mm6         \n\t"

  77.         "movq     56(%3), %%mm7         \n\t"

  78.         "packuswb  %%mm1, %%mm0         \n\t"

  79.         "packuswb  %%mm3, %%mm2         \n\t"

  80.         "packuswb  %%mm5, %%mm4         \n\t"

  81.         "packuswb  %%mm7, %%mm6         \n\t"

  82.         "movq      %%mm0, (%0)          \n\t"

  83.         "movq      %%mm2, (%0, %1)      \n\t"

  84.         "movq      %%mm4, (%0, %1, 2)   \n\t"

  85.         "movq      %%mm6, (%0, %2)      \n\t"

  86.         :: "r" (pix), "r" ((x86_reg) line_size), "r" ((x86_reg) line_size * 3),

  87.            "r" (p)

  88.         : "memory");

  89. }

  90. 

  91. #define put_signed_pixels_clamped_mmx_half(off)             \

  92.     "movq          "#off"(%2), %%mm1        \n\t"           \

  93.     "movq     16 + "#off"(%2), %%mm2        \n\t"           \

  94.     "movq     32 + "#off"(%2), %%mm3        \n\t"           \

  95.     "movq     48 + "#off"(%2), %%mm4        \n\t"           \

  96.     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \

  97.     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \

  98.     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \

  99.     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \

  100.     "paddb              %%mm0, %%mm1        \n\t"           \

  101.     "paddb              %%mm0, %%mm2        \n\t"           \

  102.     "paddb              %%mm0, %%mm3        \n\t"           \

  103.     "paddb              %%mm0, %%mm4        \n\t"           \

  104.     "movq               %%mm1, (%0)         \n\t"           \

  105.     "movq               %%mm2, (%0, %3)     \n\t"           \

  106.     "movq               %%mm3, (%0, %3, 2)  \n\t"           \

  107.     "movq               %%mm4, (%0, %1)     \n\t"

  108. 

  109. void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,

  110.                                       int line_size)

  111. {

  112.     x86_reg line_skip = line_size;

  113.     x86_reg line_skip3;

  114. 

  115.     __asm__ volatile (

  116.         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"

  117.         "lea         (%3, %3, 2), %1        \n\t"

  118.         put_signed_pixels_clamped_mmx_half(0)

  119.         "lea         (%0, %3, 4), %0        \n\t"

  120.         put_signed_pixels_clamped_mmx_half(64)

  121.         : "+&r" (pixels), "=&r" (line_skip3)

  122.         : "r" (block), "r" (line_skip)

  123.         : "memory");

  124. }

  125. 

  126. void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,

  127.                                int line_size)

  128. {

  129.     const int16_t *p;

  130.     uint8_t *pix;

  131.     int i;

  132. 

  133.     /* read the pixels */

  134.     p   = block;

  135.     pix = pixels;

  136.     MOVQ_ZERO(mm7);

  137.     i = 4;

  138.     do {

  139.         __asm__ volatile (

  140.             "movq        (%2), %%mm0    \n\t"

  141.             "movq       8(%2), %%mm1    \n\t"

  142.             "movq      16(%2), %%mm2    \n\t"

  143.             "movq      24(%2), %%mm3    \n\t"

  144.             "movq          %0, %%mm4    \n\t"

  145.             "movq          %1, %%mm6    \n\t"

  146.             "movq       %%mm4, %%mm5    \n\t"

  147.             "punpcklbw  %%mm7, %%mm4    \n\t"

  148.             "punpckhbw  %%mm7, %%mm5    \n\t"

  149.             "paddsw     %%mm4, %%mm0    \n\t"

  150.             "paddsw     %%mm5, %%mm1    \n\t"

  151.             "movq       %%mm6, %%mm5    \n\t"

  152.             "punpcklbw  %%mm7, %%mm6    \n\t"

  153.             "punpckhbw  %%mm7, %%mm5    \n\t"

  154.             "paddsw     %%mm6, %%mm2    \n\t"

  155.             "paddsw     %%mm5, %%mm3    \n\t"

  156.             "packuswb   %%mm1, %%mm0    \n\t"

  157.             "packuswb   %%mm3, %%mm2    \n\t"

  158.             "movq       %%mm0, %0       \n\t"

  159.             "movq       %%mm2, %1       \n\t"

  160.             : "+m" (*pix), "+m" (*(pix + line_size))

  161.             : "r" (p)

  162.             : "memory");

  163.         pix += line_size * 2;

  164.         p   += 16;

  165.     } while (--i);

  166. }

  167. 

  168. /* Draw the edges of width 'w' of an image of size width, height

  169.  * this MMX version can only handle w == 8 || w == 16. */

  170. void ff_draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,

  171.                        int w, int h, int sides)

  172. {

  173.     uint8_t *ptr, *last_line;

  174.     int i;

  175. 

  176.     last_line = buf + (height - 1) * wrap;

  177.     /* left and right */

  178.     ptr = buf;

  179.     if (w == 8) {

  180.         __asm__ volatile (

  181.             "1:                             \n\t"

  182.             "movd            (%0), %%mm0    \n\t"

  183.             "punpcklbw      %%mm0, %%mm0    \n\t"

  184.             "punpcklwd      %%mm0, %%mm0    \n\t"

  185.             "punpckldq      %%mm0, %%mm0    \n\t"

  186.             "movq           %%mm0, -8(%0)   \n\t"

  187.             "movq      -8(%0, %2), %%mm1    \n\t"

  188.             "punpckhbw      %%mm1, %%mm1    \n\t"

  189.             "punpckhwd      %%mm1, %%mm1    \n\t"

  190.             "punpckhdq      %%mm1, %%mm1    \n\t"

  191.             "movq           %%mm1, (%0, %2) \n\t"

  192.             "add               %1, %0       \n\t"

  193.             "cmp               %3, %0       \n\t"

  194.             "jb                1b           \n\t"

  195.             : "+r" (ptr)

  196.             : "r" ((x86_reg) wrap), "r" ((x86_reg) width),

  197.               "r" (ptr + wrap * height));

  198.     } else {

  199.         __asm__ volatile (

  200.             "1:                                 \n\t"

  201.             "movd            (%0), %%mm0        \n\t"

  202.             "punpcklbw      %%mm0, %%mm0        \n\t"

  203.             "punpcklwd      %%mm0, %%mm0        \n\t"

  204.             "punpckldq      %%mm0, %%mm0        \n\t"

  205.             "movq           %%mm0, -8(%0)       \n\t"

  206.             "movq           %%mm0, -16(%0)      \n\t"

  207.             "movq      -8(%0, %2), %%mm1        \n\t"

  208.             "punpckhbw      %%mm1, %%mm1        \n\t"

  209.             "punpckhwd      %%mm1, %%mm1        \n\t"

  210.             "punpckhdq      %%mm1, %%mm1        \n\t"

  211.             "movq           %%mm1,  (%0, %2)    \n\t"

  212.             "movq           %%mm1, 8(%0, %2)    \n\t"

  213.             "add               %1, %0           \n\t"

  214.             "cmp               %3, %0           \n\t"

  215.             "jb                1b               \n\t"

  216.             : "+r" (ptr)

  217.             : "r" ((x86_reg) wrap), "r" ((x86_reg) width),

  218.               "r" (ptr + wrap * height));

  219.     }

  220. 

  221.     /* top and bottom (and hopefully also the corners) */

  222.     if (sides & EDGE_TOP) {

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

  224.             ptr = buf - (i + 1) * wrap - w;

  225.             __asm__ volatile (

  226.                 "1:                             \n\t"

  227.                 "movq (%1, %0), %%mm0           \n\t"

  228.                 "movq    %%mm0, (%0)            \n\t"

  229.                 "movq    %%mm0, (%0, %2)        \n\t"

  230.                 "movq    %%mm0, (%0, %2, 2)     \n\t"

  231.                 "movq    %%mm0, (%0, %3)        \n\t"

  232.                 "add        $8, %0              \n\t"

  233.                 "cmp        %4, %0              \n\t"

  234.                 "jb         1b                  \n\t"

  235.                 : "+r" (ptr)

  236.                 : "r" ((x86_reg) buf - (x86_reg) ptr - w),

  237.                   "r" ((x86_reg) - wrap), "r" ((x86_reg) - wrap * 3),

  238.                   "r" (ptr + width + 2 * w));

  239.         }

  240.     }

  241. 

  242.     if (sides & EDGE_BOTTOM) {

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

  244.             ptr = last_line + (i + 1) * wrap - w;

  245.             __asm__ volatile (

  246.                 "1:                             \n\t"

  247.                 "movq (%1, %0), %%mm0           \n\t"

  248.                 "movq    %%mm0, (%0)            \n\t"

  249.                 "movq    %%mm0, (%0, %2)        \n\t"

  250.                 "movq    %%mm0, (%0, %2, 2)     \n\t"

  251.                 "movq    %%mm0, (%0, %3)        \n\t"

  252.                 "add        $8, %0              \n\t"

  253.                 "cmp        %4, %0              \n\t"

  254.                 "jb         1b                  \n\t"

  255.                 : "+r" (ptr)

  256.                 : "r" ((x86_reg) last_line - (x86_reg) ptr - w),

  257.                   "r" ((x86_reg) wrap), "r" ((x86_reg) wrap * 3),

  258.                   "r" (ptr + width + 2 * w));

  259.         }

  260.     }

  261. }

  262. 

  263. void ff_gmc_mmx(uint8_t *dst, uint8_t *src,

  264.                 int stride, int h, int ox, int oy,

  265.                 int dxx, int dxy, int dyx, int dyy,

  266.                 int shift, int r, int width, int height)

  267. {

  268.     const int w    = 8;

  269.     const int ix   = ox  >> (16 + shift);

  270.     const int iy   = oy  >> (16 + shift);

  271.     const int oxs  = ox  >> 4;

  272.     const int oys  = oy  >> 4;

  273.     const int dxxs = dxx >> 4;

  274.     const int dxys = dxy >> 4;

  275.     const int dyxs = dyx >> 4;

  276.     const int dyys = dyy >> 4;

  277.     const uint16_t r4[4]   = { r, r, r, r };

  278.     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };

  279.     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };

  280.     const uint64_t shift2  = 2 * shift;

  281.     int x, y;

  282. 

  283.     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);

  284.     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);

  285.     const int dxh = dxy * (h - 1);

  286.     const int dyw = dyx * (w - 1);

  287. 

  288.     if ( // non-constant fullpel offset (3% of blocks)

  289.         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |

  290.          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift) ||

  291.         // uses more than 16 bits of subpel mv (only at huge resolution)

  292.         (dxx | dxy | dyx | dyy) & 15 ||

  293.         (unsigned) ix >= width  - w ||

  294.         (unsigned) iy >= height - h) {

  295.         // FIXME could still use mmx for some of the rows

  296.         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,

  297.                  shift, r, width, height);

  298.         return;

  299.     }

  300. 

  301.     src += ix + iy * stride;

  302. 

  303.     __asm__ volatile (

  304.         "movd         %0, %%mm6         \n\t"

  305.         "pxor      %%mm7, %%mm7         \n\t"

  306.         "punpcklwd %%mm6, %%mm6         \n\t"

  307.         "punpcklwd %%mm6, %%mm6         \n\t"

  308.         :: "r" (1 << shift));

  309. 

  310.     for (x = 0; x < w; x += 4) {

  311.         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),

  312.                             oxs - dxys + dxxs * (x + 1),

  313.                             oxs - dxys + dxxs * (x + 2),

  314.                             oxs - dxys + dxxs * (x + 3) };

  315.         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),

  316.                             oys - dyys + dyxs * (x + 1),

  317.                             oys - dyys + dyxs * (x + 2),

  318.                             oys - dyys + dyxs * (x + 3) };

  319. 

  320.         for (y = 0; y < h; y++) {

  321.             __asm__ volatile (

  322.                 "movq      %0, %%mm4    \n\t"

  323.                 "movq      %1, %%mm5    \n\t"

  324.                 "paddw     %2, %%mm4    \n\t"

  325.                 "paddw     %3, %%mm5    \n\t"

  326.                 "movq   %%mm4, %0       \n\t"

  327.                 "movq   %%mm5, %1       \n\t"

  328.                 "psrlw    $12, %%mm4    \n\t"

  329.                 "psrlw    $12, %%mm5    \n\t"

  330.                 : "+m" (*dx4), "+m" (*dy4)

  331.                 : "m" (*dxy4), "m" (*dyy4));

  332. 

  333.             __asm__ volatile (

  334.                 "movq      %%mm6, %%mm2 \n\t"

  335.                 "movq      %%mm6, %%mm1 \n\t"

  336.                 "psubw     %%mm4, %%mm2 \n\t"

  337.                 "psubw     %%mm5, %%mm1 \n\t"

  338.                 "movq      %%mm2, %%mm0 \n\t"

  339.                 "movq      %%mm4, %%mm3 \n\t"

  340.                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)

  341.                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy

  342.                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy

  343.                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)

  344. 

  345.                 "movd         %4, %%mm5 \n\t"

  346.                 "movd         %3, %%mm4 \n\t"

  347.                 "punpcklbw %%mm7, %%mm5 \n\t"

  348.                 "punpcklbw %%mm7, %%mm4 \n\t"

  349.                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy

  350.                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy

  351. 

  352.                 "movd         %2, %%mm5 \n\t"

  353.                 "movd         %1, %%mm4 \n\t"

  354.                 "punpcklbw %%mm7, %%mm5 \n\t"

  355.                 "punpcklbw %%mm7, %%mm4 \n\t"

  356.                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)

  357.                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)

  358.                 "paddw        %5, %%mm1 \n\t"

  359.                 "paddw     %%mm3, %%mm2 \n\t"

  360.                 "paddw     %%mm1, %%mm0 \n\t"

  361.                 "paddw     %%mm2, %%mm0 \n\t"

  362. 

  363.                 "psrlw        %6, %%mm0 \n\t"

  364.                 "packuswb  %%mm0, %%mm0 \n\t"

  365.                 "movd      %%mm0, %0    \n\t"

  366. 

  367.                 : "=m" (dst[x + y * stride])

  368.                 : "m" (src[0]), "m" (src[1]),

  369.                   "m" (src[stride]), "m" (src[stride + 1]),

  370.                   "m" (*r4), "m" (shift2));

  371.             src += stride;

  372.         }

  373.         src += 4 - h * stride;

  374.     }

  375. }

  376. 

  377. #endif /* HAVE_INLINE_ASM */