x264中哈德玛变换simd汇编实现

一 接上一篇哈德玛变换C语言实现

二 哈德玛变换汇编实现

cglobal dct4x4dc, 1,1,5 //参数说明 ,这个函数一个参数  使用了1个寄存器，使用了5个向量寄存器
/*
; PROLOGUE:
; %1 = number of arguments. loads them from stack if needed.
; %2 = number of registers used. pushes callee-saved regs if needed.
; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
; %4 = (optional) stack size to be allocated. The stack will be aligned before
;      allocating the specified stack size. If the required stack alignment is
;      larger than the known stack alignment the stack will be manually aligned
;      and an extra register will be allocated to hold the original stack
;      pointer (to not invalidate r0m etc.). To prevent the use of an extra
;      register as stack pointer, request a negative stack size.
; %4+/%5+ = list of names to define to registers
; PROLOGUE can also be invoked by adding the same options to cglobal
*/
0000000000000000 <s264_8_dct4x4dc_mmx2>: //参数一个数据是2字节，一共32字节。
       0:       0f 6f 5f 18             movq   0x18(%rdi),%mm3 //16 + 8 offset。24 将数组的末尾8字节装入xmm3
       4:       0f 6f 57 10             movq   0x10(%rdi),%mm2
       8:       0f 6f 4f 08             movq   0x8(%rdi),%mm1
       c:       0f 6f 07                movq   (%rdi),%mm0 //依次装载，就是1，2，3，4行，一行存一个向量寄存器

       f:       0f 6f 3d 00 00 00 00    movq   0x0(%rip),%mm7        # 16 <s264_8_dct4x4dc_mmx2+0x16> //当前这条指令存入xmm7 , 
      16:       0f 6f e3                movq   %mm3,%mm4 //xmm3->xmm4
      19:       0f fd da                paddw  %mm2,%mm3 // 按16 bit为单位相加 LINE2 + LINE3 ->xmm3
      1c:       0f f9 d4                psubw  %mm4,%mm2 // LINE2 - LINE3 ->xmm2
      1f:       0f 6f e1                movq   %mm1,%mm4 //LINE1 ->xmm4
      22:       0f fd c8                paddw  %mm0,%mm1 //LINE0 + LINE1 ->xmm1
      25:       0f f9 c4                psubw  %mm4,%mm0 //LINE0 - LINE1 ->xmm0
//上面一段是和C语言代码对应的
      28:       0f 6f e3                movq   %mm3,%mm4 //LINE2 + LINE3 ->xmm4
      2b:       0f fd d9                paddw  %mm1,%mm3 //LINE0 + LINE1 + LINE2 + LINE3 ->xmm3
      2e:       0f f9 cc                psubw  %mm4,%mm1 //LINE0 + LINE1 - (LINE2 + LINE3)->xmm1
      31:       0f 6f e2                movq   %mm2,%mm4 //LINE2 - LINE3->xmm4
      34:       0f fd d0                paddw  %mm0,%mm2 //LINE0 - LINE1 + (LINE2 - LINE3 )->xmm2
      37:       0f f9 c4                psubw  %mm4,%mm0 //LINE0 - LINE1 - (LINE2 - LINE3 )->xmm0
      3a:       0f 6f e3                movq   %mm3,%mm4 //LINE0 + LINE1 + LINE2 + LINE3->xmm4
      3d:       0f 61 d9                punpcklwd %mm1,%mm3 //xmm1->低32位->xmm3
      40:       0f 69 e1                punpckhwd %mm1,%mm4 //xmm1高32位 ->xmm4
      43:       0f 6f c8                movq   %mm0,%mm1 //LINE0 - LINE1 - (LINE2 - LINE3 )->xmm1
      46:       0f 61 c2                punpcklwd %mm2,%mm0 //xmm2 低32位->xmm0
      49:       0f 69 ca                punpckhwd %mm2,%mm1 //xmm2高32位xmm1
      4c:       0f 6f d3                movq   %mm3,%mm2 //xmm3->xmm2
      4f:       0f 62 d8                punpckldq %mm0,%mm3 //xmm0低32位->xmm3
      52:       0f 6a d0                punpckhdq %mm0,%mm2 //xmm0高32位->xmm2
      55:       0f 6f c4                movq   %mm4,%mm0 //xmm4->xmm0
      58:       0f 62 e1                punpckldq %mm1,%mm4 //xmm1低32位->xmm4
      5b:       0f 6a c1                punpckhdq %mm1,%mm0 //xmm1高32位->xmm0
      5e:       0f 6f ca                movq   %mm2,%mm1 //xmm2->xmm1
      61:       0f fd d3                paddw  %mm3,%mm2 //tmp[i*4+0] + tmp[i*4+1]; s01
      64:       0f f9 d9                psubw  %mm1,%mm3 //tmp[i*4+0] - tmp[i*4+1]; d01
      67:       0f 6f c8                movq   %mm0,%mm1 //xmm0->xmm1
      6a:       0f fd c4                paddw  %mm4,%mm0 //tmp[i*4+2] + tmp[i*4+3]; s23
      6d:       0f f9 e1                psubw  %mm1,%mm4 //tmp[i*4+2] - tmp[i*4+3]; d23
      70:       0f 6f cf                movq   %mm7,%mm1 //xmm7->xmm1
      73:       0f ef d7                pxor   %mm7,%mm2 // xmm7 ^ xmm2->xmm2 ,再pxor一次就能恢复
      76:       0f f9 c8                psubw  %mm0,%mm1 //xmm1->xmm0 ->xmm1
      79:       0f ef c7                pxor   %mm7,%mm0 //xmm7 ^ xmm0 ->xmm0
      7c:       0f e3 ca                pavgw  %mm2,%mm1 //(xmm2 + xmm1 + 1)/2 取整->xmm1
      7f:       0f e3 c2                pavgw  %mm2,%mm0 // (xmm2 + xmm0 + 1)/2 取整->xmm0
      82:       0f ef cf                pxor   %mm7,%mm1 //xmm7^xmm1->xmm1
      85:       0f ef c7                pxor   %mm7,%mm0 //xmm7->xmm0->xmm0
      88:       0f 6f ef                movq   %mm7,%mm5 //xmm7->xmm5
      8b:       0f ef df                pxor   %mm7,%mm3 //xmm7^xmm3->xmm3
      8e:       0f f9 ec                psubw  %mm4,%mm5 //xmm5-xmm4->xmm5
      91:       0f ef e7                pxor   %mm7,%mm4 //xmm7^xmm4->xmm4
      94:       0f e3 eb                pavgw  %mm3,%mm5 // (xmm3 + xmm5 + 1)/2 能对应上C语言代码，
      97:       0f e3 e3                pavgw  %mm3,%mm4 //(xmm3 + xmm4 + 1)/2
      9a:       0f ef ef                pxor   %mm7,%mm5 //xmm7^xmm5->xmm5
      9d:       0f ef e7                pxor   %mm7,%mm4 //xmm7^xmm4->xmm4
      a0:       0f 7f 07                movq   %mm0,(%rdi) //xmm0->rdi
      a3:       0f 7f 4f 08             movq   %mm1,0x8(%rdi) //xmm1->rdi + 8
      a7:       0f 7f 6f 10             movq   %mm5,0x10(%rdi) //xmm5->rdi + 16
      ab:       0f 7f 67 18             movq   %mm4,0x18(%rdi) //xmm4->rdi + 24
      af:       c3                      retq

因为汇编代码里面有很多宏，为了方便读代码，直接拿编译后的.o 反汇编来看，比较直观。

三 疑惑点

xmm7寄存器的一些运算没太看明白，用rip 指令寄存器存入xmm7 ,然后用xmm7寄存器做各种异或运算来保存xmm0,xmm1,xmm3,xmm5寄存器，这一段没太理解。但是总体上和C语言代码是对得上的。