⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 mpegvideo_mmx_template.c

📁 ffmpeg的完整源代码和作者自己写的文档。不但有在Linux的工程哦
💻 C
📖 第 1 页 / 共 2 页
字号:
            "mov $-128, %%"REG_a"               \n\t"
            ASMALIGN(4)
            "1:                                 \n\t"
            MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
            SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
            MOVQ" (%3, %%"REG_a"), "MM"6        \n\t" // bias[0]
            "paddusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
            MOVQ" (%2, %%"REG_a"), "MM"5        \n\t" // qmat[i]
            "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
            "por "MM"0, "MM"4                   \n\t"
            RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
            MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
            "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
            MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
            MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
            "pandn "MM"1, "MM"0                 \n\t"
            PMAXW(MM"0", MM"3")
            "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
            " js 1b                             \n\t"
            PMAX(MM"3", MM"0")
            "movd "MM"3, %%"REG_a"              \n\t"
            "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
            : "+a" (last_non_zero_p1)
            : "r" (block+64), "r" (qmat+64), "r" (bias+64),
              "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
        );
    }
    asm volatile(
        "movd %1, "MM"1                     \n\t" // max_qcoeff
        SPREADW(MM"1")
        "psubusw "MM"1, "MM"4               \n\t"
        "packuswb "MM"4, "MM"4              \n\t"
#ifdef HAVE_SSE2
        "packuswb "MM"4, "MM"4              \n\t"
#endif
        "movd "MM"4, %0                     \n\t" // *overflow
        : "=g" (*overflow)
        : "g" (s->max_qcoeff)
    );

    if(s->mb_intra) block[0]= level;
    else            block[0]= temp_block[0];

    if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
        if(last_non_zero_p1 <= 1) goto end;
        block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
        block[0x20] = temp_block[0x10];
        if(last_non_zero_p1 <= 4) goto end;
        block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
        block[0x09] = temp_block[0x03];
        if(last_non_zero_p1 <= 7) goto end;
        block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
        block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
        if(last_non_zero_p1 <= 11) goto end;
        block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
        block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
        block[0x0C] = temp_block[0x05];
        if(last_non_zero_p1 <= 16) goto end;
        block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
        block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
        block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
        block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
        if(last_non_zero_p1 <= 24) goto end;
        block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
        block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
        block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
        block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
        if(last_non_zero_p1 <= 32) goto end;
        block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
        block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
        block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
        block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
        if(last_non_zero_p1 <= 40) goto end;
        block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
        block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
        block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
        block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
        if(last_non_zero_p1 <= 48) goto end;
        block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
        block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
        block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
        block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
        if(last_non_zero_p1 <= 56) goto end;
        block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
        block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
        block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
        block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
    }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
        if(last_non_zero_p1 <= 1) goto end;
        block[0x04] = temp_block[0x01];
        block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
        if(last_non_zero_p1 <= 4) goto end;
        block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
        block[0x05] = temp_block[0x03];
        if(last_non_zero_p1 <= 7) goto end;
        block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
        block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
        if(last_non_zero_p1 <= 11) goto end;
        block[0x1C] = temp_block[0x19];
        block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
        block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
        if(last_non_zero_p1 <= 16) goto end;
        block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
        block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
        block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
        block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
        if(last_non_zero_p1 <= 24) goto end;
        block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
        block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
        block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
        block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
        if(last_non_zero_p1 <= 32) goto end;
        block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
        block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
        block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
        block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
        if(last_non_zero_p1 <= 40) goto end;
        block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
        block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
        block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
        block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
        if(last_non_zero_p1 <= 48) goto end;
        block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
        block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
            block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
        block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
        if(last_non_zero_p1 <= 56) goto end;
        block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
        block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
        block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
        block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
    }else{
        if(last_non_zero_p1 <= 1) goto end;
        block[0x01] = temp_block[0x01];
        block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
        if(last_non_zero_p1 <= 4) goto end;
        block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
        block[0x03] = temp_block[0x03];
        if(last_non_zero_p1 <= 7) goto end;
        block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
        block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
        if(last_non_zero_p1 <= 11) goto end;
        block[0x19] = temp_block[0x19];
        block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
        block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
        if(last_non_zero_p1 <= 16) goto end;
        block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
        block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
        block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
        block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
        if(last_non_zero_p1 <= 24) goto end;
        block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
        block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
        block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
        block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
        if(last_non_zero_p1 <= 32) goto end;
        block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
        block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
        block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
        block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
        if(last_non_zero_p1 <= 40) goto end;
        block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
        block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
        block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
        block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
        if(last_non_zero_p1 <= 48) goto end;
        block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
        block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
        block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
        block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
        if(last_non_zero_p1 <= 56) goto end;
        block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
        block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
        block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
        block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
    }
    end:
/*
    for(i=0; i<last_non_zero_p1; i++)
    {
       int j= zigzag_direct_noperm[i];
       block[block_permute_op(j)]= temp_block[j];
    }
*/

    return last_non_zero_p1 - 1;
}

⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -