x264 反哈德玛变换

一 C语言实现

static void idct4x4dc( dctcoef d[16] )
{
    dctcoef tmp[16];

    for( int i = 0; i < 4; i++ )
    {
        int s01 = d[i*4+0] + d[i*4+1];
        int d01 = d[i*4+0] - d[i*4+1];
        int s23 = d[i*4+2] + d[i*4+3];
        int d23 = d[i*4+2] - d[i*4+3];

        tmp[0*4+i] = s01 + s23;
        tmp[1*4+i] = s01 - s23;
        tmp[2*4+i] = d01 - d23;
        tmp[3*4+i] = d01 + d23;
    }

    for( int i = 0; i < 4; i++ )
    {
        int s01 = tmp[i*4+0] + tmp[i*4+1];
        int d01 = tmp[i*4+0] - tmp[i*4+1];
        int s23 = tmp[i*4+2] + tmp[i*4+3];
        int d23 = tmp[i*4+2] - tmp[i*4+3];

        d[i*4+0] = s01 + s23; 只有这里和哈德玛变换有区别， 没有对2取整
        d[i*4+1] = s01 - s23; 
        d[i*4+2] = d01 - d23;
        d[i*4+3] = d01 + d23;
    }
}

二 SIMD汇编实现

%macro IDCT4x4DC 0
cglobal idct4x4dc, 1,1
    mova   m3, [r0+48] //和哈德玛变换不同，这里一个元素是4字节
    mova   m2, [r0+32]
    mova   m1, [r0+16] 
    mova   m0, [r0+ 0]
//将数据载入xmm寄存器
    WALSH4_1D  d,0,1,2,3,4
    TRANSPOSE4x4D 0,1,2,3,4
    WALSH4_1D  d,0,1,2,3,4
//将数据从xmm寄存器取出来
    mova  [r0+ 0], m0
    mova  [r0+16], m1
    mova  [r0+32], m2
    mova  [r0+48], m3
    RET
%endmacro ; IDCT4x4DC

%macro WALSH4_1D 6 //6个参数
    SUMSUB_BADC %1, %5, %4, %3, %2, %6
    SUMSUB_BADC %1, %5, %3, %4, %2, %6
    SWAP %2, %5, %4
%endmacro

%macro SUMSUB_BADC 5-6
%if %0==6
    SUMSUB_BA %1, %2, %3, %6
    SUMSUB_BA %1, %4, %5, %6
%else
    padd%1  m%2, m%3 //xmm2 + xmm3->xmm2
    padd%1  m%4, m%5 //xmm4 + xmm5->xmm4
    padd%1  m%3, m%3 //xmm3 + xmm3 ->xmm3
    padd%1  m%5, m%5 //xmm5 + xmm5 ->xmm5
    psub%1  m%3, m%2 //xmm3 - xmm2->xmm3
    psub%1  m%5, m%4 //xmm5 - xmm4->xmm5
%endif
%endmacro

%macro SUMSUB_BA 3-4
%if %0==3 //如果是3个参数
    padd%1  m%2, m%3
    padd%1  m%3, m%3
    psub%1  m%3, m%2
%elif avx_enabled
    padd%1  m%4, m%2, m%3
    psub%1  m%3, m%2
    SWAP    %2, %4
%else
    mova    m%4, m%2 //xmm2->xmm4
    padd%1  m%2, m%3 //xmm3 + xmm2->xmm2. paddd
    psub%1  m%3, m%4.//xmm3 - xmm4 ->xmm3
%endif
%endmacro

三 反汇编代码解读

00000000000000b0 <s264_8_idct4x4dc_mmx>:
      b0:       0f 6f 5f 18             movq   0x18(%rdi),%mm3
      b4:       0f 6f 57 10             movq   0x10(%rdi),%mm2
      b8:       0f 6f 4f 08             movq   0x8(%rdi),%mm1
      bc:       0f 6f 07                movq   (%rdi),%mm0
//载入数据
      bf:       0f 6f e3                movq   %mm3,%mm4 //xmm3->xmm4
      c2:       0f fd da                paddw  %mm2,%mm3 //xmm2 + xmm3->xmm3
      c5:       0f f9 d4                psubw  %mm4,%mm2 //xmm4 - xmm2->xmm2
      c8:       0f 6f e1                movq   %mm1,%mm4 //xmm1->xmm4
      cb:       0f fd c8                paddw  %mm0,%mm1 //xmm0 + xmm1->xmm1
      ce:       0f f9 c4                psubw  %mm4,%mm0 //xmm0 - xmm1->xmm0
      d1:       0f 6f e3                movq   %mm3,%mm4 //xmm3->xmm4
      d4:       0f fd d9                paddw  %mm1,%mm3 //xmm3 + xmm1->xmm3
      d7:       0f f9 cc                psubw  %mm4,%mm1 //xmm3 - xmm1->xmm1
      da:       0f 6f e2                movq   %mm2,%mm4 //
      dd:       0f fd d0                paddw  %mm0,%mm2
      e0:       0f f9 c4                psubw  %mm4,%mm0
      e3:       0f 6f e3                movq   %mm3,%mm4
//上面和哈德玛变换的代码一致,左边乘以哈德玛矩阵
      e6:       0f 61 d9                punpcklwd %mm1,%mm3
      e9:       0f 69 e1                punpckhwd %mm1,%mm4
      ec:       0f 6f c8                movq   %mm0,%mm1
      ef:       0f 61 c2                punpcklwd %mm2,%mm0
      f2:       0f 69 ca                punpckhwd %mm2,%mm1
      f5:       0f 6f d3                movq   %mm3,%mm2
      f8:       0f 62 d8                punpckldq %mm0,%mm3
      fb:       0f 6a d0                punpckhdq %mm0,%mm2
      fe:       0f 6f c4                movq   %mm4,%mm0
     101:       0f 62 e1                punpckldq %mm1,%mm4
     104:       0f 6a c1                punpckhdq %mm1,%mm0
//下面一段是右边乘以 哈德玛矩阵
     107:       0f 6f c8                movq   %mm0,%mm1
     10a:       0f fd c4                paddw  %mm4,%mm0
     10d:       0f f9 e1                psubw  %mm1,%mm4
     110:       0f 6f ca                movq   %mm2,%mm1
     113:       0f fd d3                paddw  %mm3,%mm2
     116:       0f f9 d9                psubw  %mm1,%mm3
     119:       0f 6f c8                movq   %mm0,%mm1
     11c:       0f fd c2                paddw  %mm2,%mm0
     11f:       0f f9 d1                psubw  %mm1,%mm2
     122:       0f 6f cc                movq   %mm4,%mm1
     125:       0f fd e3                paddw  %mm3,%mm4
     128:       0f f9 d9                psubw  %mm1,%mm3
//下面是存数据到内存
     12b:       0f 7f 07                movq   %mm0,(%rdi)
     12e:       0f 7f 57 08             movq   %mm2,0x8(%rdi)
     132:       0f 7f 5f 10             movq   %mm3,0x10(%rdi)
     136:       0f 7f 67 18             movq   %mm4,0x18(%rdi)
     13a:       c3                      retq
     13b:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)