git.sesse.net Git - ffmpeg/blob - libavcodec/mips/pixblockdsp_mmi.c

   1 /*
   2  * Loongson SIMD optimized pixblockdsp
   3  *
   4  * Copyright (c) 2015 Loongson Technology Corporation Limited
   5  * Copyright (c) 2015 Zhou Xiaoyong <zhouxiaoyong@loongson.cn>
   6  *
   7  * This file is part of FFmpeg.
   8  *
   9  * FFmpeg is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * FFmpeg is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with FFmpeg; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  */
  23
  24 #include "pixblockdsp_mips.h"
  25 #include "libavutil/mips/asmdefs.h"
  26
  27 void ff_get_pixels_8_mmi(int16_t *av_restrict block, const uint8_t *pixels,
  28         ptrdiff_t line_size)
  29 {
  30     double ftmp[6];
  31     mips_reg tmp[2];
  32
  33     __asm__ volatile (
  34         "li         %[tmp1],    0x08                                    \n\t"
  35         "move       %[tmp0],    $0                                      \n\t"
  36         "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]                \n\t"
  37         "1:                                                             \n\t"
  38         "gsldlc1    %[ftmp1],   0x07(%[pixels])                         \n\t"
  39         "gsldrc1    %[ftmp1],   0x00(%[pixels])                         \n\t"
  40         "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]                \n\t"
  41         "punpckhbh  %[ftmp5],   %[ftmp1],       %[ftmp0]                \n\t"
  42         "gssdxc1    %[ftmp2],   0x00(%[block],  %[tmp0])                \n\t"
  43         "gssdxc1    %[ftmp5],   0x08(%[block],  %[tmp0])                \n\t"
  44         PTR_ADDI   "%[tmp1],    %[tmp1],       -0x01                    \n\t"
  45         PTR_ADDIU  "%[tmp0],    %[tmp0],        0x10                    \n\t"
  46         PTR_ADDU   "%[pixels],  %[pixels],      %[line_size]            \n\t"
  47         "bnez       %[tmp1],    1b                                      \n\t"
  48         : [ftmp0]"=&f"(ftmp[0]),            [ftmp1]"=&f"(ftmp[1]),
  49           [ftmp2]"=&f"(ftmp[2]),            [ftmp3]"=&f"(ftmp[3]),
  50           [ftmp4]"=&f"(ftmp[4]),            [ftmp5]"=&f"(ftmp[5]),
  51           [tmp0]"=&r"(tmp[0]),              [tmp1]"=&r"(tmp[1]),
  52           [pixels]"+&r"(pixels)
  53         : [block]"r"((mips_reg)block),      [line_size]"r"((mips_reg)line_size)
  54         : "memory"
  55     );
  56 }
  57
  58 void ff_diff_pixels_mmi(int16_t *av_restrict block, const uint8_t *src1,
  59         const uint8_t *src2, int stride)
  60 {
  61     double ftmp[5];
  62     mips_reg tmp[1];
  63
  64     __asm__ volatile (
  65         "li         %[tmp0],    0x08                                    \n\t"
  66         "xor        %[ftmp4],   %[ftmp4],       %[ftmp4]                \n\t"
  67         "1:                                                             \n\t"
  68         "gsldlc1    %[ftmp0],   0x07(%[src1])                           \n\t"
  69         "gsldrc1    %[ftmp0],   0x00(%[src1])                           \n\t"
  70         "or         %[ftmp1],   %[ftmp0],       %[ftmp0]                \n\t"
  71         "gsldlc1    %[ftmp2],   0x07(%[src2])                           \n\t"
  72         "gsldrc1    %[ftmp2],   0x00(%[src2])                           \n\t"
  73         "or         %[ftmp3],   %[ftmp2],       %[ftmp2]                \n\t"
  74         "punpcklbh  %[ftmp0],   %[ftmp0],       %[ftmp4]                \n\t"
  75         "punpckhbh  %[ftmp1],   %[ftmp1],       %[ftmp4]                \n\t"
  76         "punpcklbh  %[ftmp2],   %[ftmp2],       %[ftmp4]                \n\t"
  77         "punpckhbh  %[ftmp3],   %[ftmp3],       %[ftmp4]                \n\t"
  78         "psubh      %[ftmp0],   %[ftmp0],       %[ftmp2]                \n\t"
  79         "psubh      %[ftmp1],   %[ftmp1],       %[ftmp3]                \n\t"
  80         "gssdlc1    %[ftmp0],   0x07(%[block])                          \n\t"
  81         "gssdrc1    %[ftmp0],   0x00(%[block])                          \n\t"
  82         "gssdlc1    %[ftmp1],   0x0f(%[block])                          \n\t"
  83         "gssdrc1    %[ftmp1],   0x08(%[block])                          \n\t"
  84         PTR_ADDI   "%[tmp0],    %[tmp0], -0x01                          \n\t"
  85         PTR_ADDIU  "%[block],   %[block], 0x10                          \n\t"
  86         PTR_ADDU   "%[src1],    %[src1],        %[stride]               \n\t"
  87         PTR_ADDU   "%[src2],    %[src2],        %[stride]               \n\t"
  88         "bgtz       %[tmp0],    1b                                      \n\t"
  89         : [ftmp0]"=&f"(ftmp[0]),            [ftmp1]"=&f"(ftmp[1]),
  90           [ftmp2]"=&f"(ftmp[2]),            [ftmp3]"=&f"(ftmp[3]),
  91           [ftmp4]"=&f"(ftmp[4]),
  92           [tmp0]"=&r"(tmp[0]),
  93           [block]"+&r"(block),              [src1]"+&r"(src1),
  94           [src2]"+&r"(src2)
  95         : [stride]"r"((mips_reg)stride)
  96         : "memory"
  97     );
  98 }