pixel-a.asm

从服务器上下的x264编码器C源码……希望对大家有帮助……这个是09年4月的

    SCALAR_SUMSUB %2d, %4d, %3d, %5d
    mov         [left_1d+2*%1+0], %2w
    mov         [left_1d+2*%1+2], %3w
    mov         [left_1d+2*%1+4], %4w
    mov         [left_1d+2*%1+6], %5w
%endmacro

%macro SCALAR_HADAMARD_TOP 5 ; x, 4x tmp
    movzx       %2d, byte [r1+%1-FDEC_STRIDE+0]
    movzx       %3d, byte [r1+%1-FDEC_STRIDE+1]
    movzx       %4d, byte [r1+%1-FDEC_STRIDE+2]
    movzx       %5d, byte [r1+%1-FDEC_STRIDE+3]
    SCALAR_SUMSUB %2d, %3d, %4d, %5d
    SCALAR_SUMSUB %2d, %4d, %3d, %5d
    mov         [top_1d+2*%1+0], %2w
    mov         [top_1d+2*%1+2], %3w
    mov         [top_1d+2*%1+4], %4w
    mov         [top_1d+2*%1+6], %5w
%endmacro

%macro SUM_MM_X3 8 ; 3x sum, 4x tmp, op
    pxor        %7, %7
    pshufw      %4, %1, 01001110b
    pshufw      %5, %2, 01001110b
    pshufw      %6, %3, 01001110b
    paddw       %1, %4
    paddw       %2, %5
    paddw       %3, %6
    punpcklwd   %1, %7
    punpcklwd   %2, %7
    punpcklwd   %3, %7
    pshufw      %4, %1, 01001110b
    pshufw      %5, %2, 01001110b
    pshufw      %6, %3, 01001110b
    %8          %1, %4
    %8          %2, %5
    %8          %3, %6
%endmacro

%macro CLEAR_SUMS 0
%ifdef ARCH_X86_64
    mov   qword [sums+0], 0
    mov   qword [sums+8], 0
    mov   qword [sums+16], 0
%else
    pxor  m7, m7
    movq  [sums+0], m7
    movq  [sums+8], m7
    movq  [sums+16], m7
%endif
%endmacro

; in: m1..m3
; out: m7
; clobber: m4..m6
%macro SUM3x4 1
%ifidn %1, ssse3
    pabsw       m4, m1
    pabsw       m5, m2
    pabsw       m7, m3
    paddw       m4, m5
%else
    movq        m4, m1
    movq        m5, m2
    ABS2        m4, m5, m6, m7
    movq        m7, m3
    paddw       m4, m5
    ABS1        m7, m6
%endif
    paddw       m7, m4
%endmacro

; in: m0..m3 (4x4), m7 (3x4)
; out: m0 v, m4 h, m5 dc
; clobber: m6
%macro SUM4x3 3 ; dc, left, top
    movq        m4, %2
    movd        m5, %1
    psllw       m4, 2
    psubw       m4, m0
    psubw       m5, m0
    punpcklwd   m0, m1
    punpcklwd   m2, m3
    punpckldq   m0, m2 ; transpose
    movq        m1, %3
    psllw       m1, 2
    psubw       m0, m1
    ABS2        m4, m5, m2, m3 ; 1x4 sum
    ABS1        m0, m1 ; 4x1 sum
%endmacro

%macro INTRA_SATDS_MMX 1
INIT_MMX
;-----------------------------------------------------------------------------
; void x264_intra_satd_x3_4x4_mmxext( uint8_t *fenc, uint8_t *fdec, int *res )
;-----------------------------------------------------------------------------
cglobal x264_intra_satd_x3_4x4_%1, 2,6
%ifdef ARCH_X86_64
    ; stack is 16 byte aligned because abi says so
    %define  top_1d  rsp-8  ; size 8
    %define  left_1d rsp-16 ; size 8
    %define  t0 r10
%else
    ; stack is 16 byte aligned at least in gcc, and we've pushed 3 regs + return address, so it's still aligned
    SUB         esp, 16
    %define  top_1d  esp+8
    %define  left_1d esp
    %define  t0 r2
%endif

    call load_hadamard
    SCALAR_HADAMARD_LEFT 0, r0, r3, r4, r5
    mov         t0d, r0d
    SCALAR_HADAMARD_TOP  0, r0, r3, r4, r5
    lea         t0d, [t0d + r0d + 4]
    and         t0d, -8
    shl         t0d, 1 ; dc

    SUM3x4 %1
    SUM4x3 t0d, [left_1d], [top_1d]
    paddw       m4, m7
    paddw       m5, m7
    movq        m1, m5
    psrlq       m1, 16  ; 4x3 sum
    paddw       m0, m1

    SUM_MM_X3   m0, m4, m5, m1, m2, m3, m6, pavgw
%ifndef ARCH_X86_64
    mov         r2,  r2mp
%endif
    movd        [r2+0], m0 ; i4x4_v satd
    movd        [r2+4], m4 ; i4x4_h satd
    movd        [r2+8], m5 ; i4x4_dc satd
%ifndef ARCH_X86_64
    ADD         esp, 16
%endif
    RET

%ifdef ARCH_X86_64
    %define  t0 r10
    %define  t2 r11
%else
    %define  t0 r0
    %define  t2 r2
%endif

;-----------------------------------------------------------------------------
; void x264_intra_satd_x3_16x16_mmxext( uint8_t *fenc, uint8_t *fdec, int *res )
;-----------------------------------------------------------------------------
cglobal x264_intra_satd_x3_16x16_%1, 0,7
%ifdef ARCH_X86_64
    %assign  stack_pad  88
%else
    %assign  stack_pad  88 + ((stack_offset+88+4)&15)
%endif
    ; not really needed on x86_64, just shuts up valgrind about storing data below the stack across a function call
    SUB         rsp, stack_pad
%define sums    rsp+64 ; size 24
%define top_1d  rsp+32 ; size 32
%define left_1d rsp    ; size 32
    movifnidn   r1,  r1mp
    CLEAR_SUMS

    ; 1D hadamards
    xor         t2d, t2d
    mov         t0d, 12
.loop_edge:
    SCALAR_HADAMARD_LEFT t0, r3, r4, r5, r6
    add         t2d, r3d
    SCALAR_HADAMARD_TOP  t0, r3, r4, r5, r6
    add         t2d, r3d
    sub         t0d, 4
    jge .loop_edge
    shr         t2d, 1
    add         t2d, 8
    and         t2d, -16 ; dc

    ; 2D hadamards
    movifnidn   r0,  r0mp
    xor         r3d, r3d
.loop_y:
    xor         r4d, r4d
.loop_x:
    call load_hadamard

    SUM3x4 %1
    SUM4x3 t2d, [left_1d+8*r3], [top_1d+8*r4]
    pavgw       m4, m7
    pavgw       m5, m7
    paddw       m0, [sums+0]  ; i16x16_v satd
    paddw       m4, [sums+8]  ; i16x16_h satd
    paddw       m5, [sums+16] ; i16x16_dc satd
    movq        [sums+0], m0
    movq        [sums+8], m4
    movq        [sums+16], m5

    add         r0, 4
    inc         r4d
    cmp         r4d, 4
    jl  .loop_x
    add         r0, 4*FENC_STRIDE-16
    inc         r3d
    cmp         r3d, 4
    jl  .loop_y

; horizontal sum
    movifnidn   r2, r2mp
    movq        m2, [sums+16]
    movq        m1, [sums+8]
    movq        m0, [sums+0]
    movq        m7, m2
    SUM_MM_X3   m0, m1, m2, m3, m4, m5, m6, paddd
    psrld       m0, 1
    pslld       m7, 16
    psrld       m7, 16
    paddd       m0, m2
    psubd       m0, m7
    movd        [r2+8], m2 ; i16x16_dc satd
    movd        [r2+4], m1 ; i16x16_h satd
    movd        [r2+0], m0 ; i16x16_v satd
    ADD         rsp, stack_pad
    RET

;-----------------------------------------------------------------------------
; void x264_intra_satd_x3_8x8c_mmxext( uint8_t *fenc, uint8_t *fdec, int *res )
;-----------------------------------------------------------------------------
cglobal x264_intra_satd_x3_8x8c_%1, 0,6
    ; not really needed on x86_64, just shuts up valgrind about storing data below the stack across a function call
    SUB          rsp, 72
%define  sums    rsp+48 ; size 24
%define  dc_1d   rsp+32 ; size 16
%define  top_1d  rsp+16 ; size 16
%define  left_1d rsp    ; size 16
    movifnidn   r1,  r1mp
    CLEAR_SUMS

    ; 1D hadamards
    mov         t0d, 4
.loop_edge:
    SCALAR_HADAMARD_LEFT t0, t2, r3, r4, r5
    SCALAR_HADAMARD_TOP  t0, t2, r3, r4, r5
    sub         t0d, 4
    jge .loop_edge

    ; dc
    movzx       t2d, word [left_1d+0]
    movzx       r3d, word [top_1d+0]
    movzx       r4d, word [left_1d+8]
    movzx       r5d, word [top_1d+8]
    add         t2d, r3d
    lea         r3, [r4 + r5]
    lea         t2, [2*t2 + 8]
    lea         r3, [2*r3 + 8]
    lea         r4, [4*r4 + 8]
    lea         r5, [4*r5 + 8]
    and         t2d, -16 ; tl
    and         r3d, -16 ; br
    and         r4d, -16 ; bl
    and         r5d, -16 ; tr
    mov         [dc_1d+ 0], t2d ; tl
    mov         [dc_1d+ 4], r5d ; tr
    mov         [dc_1d+ 8], r4d ; bl
    mov         [dc_1d+12], r3d ; br
    lea         r5, [dc_1d]

    ; 2D hadamards
    movifnidn   r0,  r0mp
    movifnidn   r2,  r2mp
    xor         r3d, r3d
.loop_y:
    xor         r4d, r4d
.loop_x:
    call load_hadamard

    SUM3x4 %1
    SUM4x3 [r5+4*r4], [left_1d+8*r3], [top_1d+8*r4]
    pavgw       m4, m7
    pavgw       m5, m7
    paddw       m0, [sums+16] ; i4x4_v satd
    paddw       m4, [sums+8]  ; i4x4_h satd
    paddw       m5, [sums+0]  ; i4x4_dc satd
    movq        [sums+16], m0
    movq        [sums+8], m4
    movq        [sums+0], m5

    add         r0, 4
    inc         r4d
    cmp         r4d, 2
    jl  .loop_x
    add         r0, 4*FENC_STRIDE-8
    add         r5, 8
    inc         r3d
    cmp         r3d, 2
    jl  .loop_y

; horizontal sum
    movq        m0, [sums+0]
    movq        m1, [sums+8]
    movq        m2, [sums+16]
    movq        m7, m0
    psrlq       m7, 15
    paddw       m2, m7
    SUM_MM_X3   m0, m1, m2, m3, m4, m5, m6, paddd
    psrld       m2, 1
    movd        [r2+0], m0 ; i8x8c_dc satd
    movd        [r2+4], m1 ; i8x8c_h satd
    movd        [r2+8], m2 ; i8x8c_v satd
    ADD         rsp, 72
    RET
%endmacro ; INTRA_SATDS_MMX


%macro ABS_MOV_SSSE3 2
    pabsw   %1, %2
%endmacro

%macro ABS_MOV_MMX 2
    pxor    %1, %1
    psubw   %1, %2
    pmaxsw  %1, %2
%endmacro

%define ABS_MOV ABS_MOV_MMX

; in:  r0=pix, r1=stride, r2=stride*3, r3=tmp, m6=mask_ac4, m7=0
; out: [tmp]=hadamard4, m0=satd
cglobal x264_hadamard_ac_4x4_mmxext
    movh      m0, [r0]
    movh      m1, [r0+r1]
    movh      m2, [r0+r1*2]
    movh      m3, [r0+r2]
    punpcklbw m0, m7
    punpcklbw m1, m7
    punpcklbw m2, m7
    punpcklbw m3, m7
    HADAMARD4_2D 0, 1, 2, 3, 4
    mova [r3],    m0
    mova [r3+8],  m1
    mova [r3+16], m2
    mova [r3+24], m3
    ABS1      m0, m4
    ABS1      m1, m4
    pand      m0, m6
    ABS1      m2, m4
    ABS1      m3, m4
    paddw     m0, m1
    paddw     m2, m3
    paddw     m0, m2
    SAVE_MM_PERMUTATION x264_hadamard_ac_4x4_mmxext
    ret

cglobal x264_hadamard_ac_2x2max_mmxext
    mova      m0, [r3+0x00]
    mova      m1, [r3+0x20]
    mova      m2, [r3+0x40]
    mova      m3, [r3+0x60]
    sub       r3, 8
    SUMSUB_BADC m0, m1, m2, m3, m4
    ABS4 m0, m2, m1, m3, m4, m5
    HADAMARD 0, max, 0, 2, 4, 5
    HADAMARD 0, max, 1, 3, 4, 5
    paddw     m7, m0
    paddw     m7, m1
    SAVE_MM_PERMUTATION x264_hadamard_ac_2x2max_mmxext
    ret

cglobal x264_hadamard_ac_8x8_mmxext
    mova      m6, [mask_ac4 GLOBAL]
    pxor      m7, m7
    call x264_hadamard_ac_4x4_mmxext
    add       r0, 4
    add       r3, 32
    mova      m5, m0
    call x264_hadamard_ac_4x4_mmxext
    lea       r0, [r0+4*r1]
    add       r3, 64
    paddw     m5, m0
    call x264_hadamard_ac_4x4_mmxext
    sub       r0, 4
    sub       r3, 32
    paddw     m5, m0
    call x264_hadamard_ac_4x4_mmxext
    paddw     m5, m0
    sub       r3, 40
    mova [rsp+gprsize+8], m5 ; save satd
%rep 3
    call x264_hadamard_ac_2x2max_mmxext
%endrep
    mova      m0, [r3+0x00]
    mova      m1, [r3+0x20]
    mova      m2, [r3+0x40]
    mova      m3, [r3+0x60]
    SUMSUB_BADC m0, m1, m2, m3, m4
    HADAMARD 0, sumsub, 0, 2, 4, 5
    ABS4 m1, m3, m0, m2, m4, m5
    HADAMARD 0, max, 1, 3, 4, 5
    pand      m6, m0
    paddw     m7, m1
    paddw     m6, m2
    paddw     m7, m7
    paddw     m6, m7
    mova [rsp+gprsize], m6 ; save sa8d
    SWAP      m0, m6
    SAVE_MM_PERMUTATION x264_hadamard_ac_8x8_mmxext
    ret

%macro HADAMARD_AC_WXH_MMX 2
cglobal x264_pixel_hadamard_ac_%1x%2_mmxext, 2,4
    %assign pad 16-gprsize-(stack_offset&15)
    %define ysub r1
    sub  rsp, 16+128+pad
    lea  r2, [r1*3]
    lea  r3, [rsp+16]
    call x264_hadamard_ac_8x8_mmxext
%if %2==16
    %define ysub r2
    lea  r0, [r0+r1*4]
    sub  rsp, 16
    call x264_hadamard_ac_8x8_mmxext
%endif
%if %1==16
    neg  ysub
    sub  rsp, 16
    lea  r0, [r0+ysub*4+8]
    neg  ysub
    call x264_hadamard_ac_8x8_mmxext
%if %2==16
    lea  r0, [r0+r1*4]
    sub  rsp, 16
    call x264_hadamard_ac_8x8_mmxext
%endif
%endif
    mova    m1, [rsp+0x08]
%if %1*%2 >= 128
    paddusw m0, [rsp+0x10]
    paddusw m1, [rsp+0x18]
%endif
%if %1*%2 == 256
    mova    m2, [rsp+0x20]
    paddusw m1, [rsp+0x28]
    paddusw m2, [rsp+0x30]
    mova    m3, m0
    paddusw m1, [rsp+0x38]
    pxor    m3, m2
    pand    m3, [pw_1 GLOBAL]
    pavgw   m0, m2
    psubusw m0, m3
    HADDUW  m0, m2
%else
    psrlw m0, 1
    HADDW m0, m2
%endif
    psrlw m1, 1
    HADDW m1, m3
    movd edx, m0
    movd eax, m1
    shr  edx, 1
%ifdef ARCH_X86_64
    shl  rdx, 32
    add  rax, rdx
%endif
    add  rsp, 128+%1*%2/4+pad
    RET
%endmacro ; HADAMARD_AC_WXH_MMX

HADAMARD_AC_WXH_MMX 16, 16
HADAMARD_AC_WXH_MMX  8, 16
HADAMARD_AC_WXH_MMX 16,  8
HADAMARD_AC_WXH_MMX  8,  8

%macro LOAD_INC_8x4W_SSE2 5
    movh      m%1, [r0]
    movh      m%2, [r0+r1]
    movh      m%3, [r0+r1*2]
    movh      m%4, [r0+r2]
%ifidn %1, 0
    lea       r0, [r0+r1*4]
%endif
    punpcklbw m%1, m%5
    punpcklbw m%2, m%5
    punpcklbw m%3, m%5
    punpcklbw m%4, m%5
%endmacro

%macro LOAD_INC_8x4W_SSSE3 5
    LOAD_DUP_4x8P %3, %4, %1, %2, [r0+r1*2], [r0+r2], [r0], [r0+r1]
%ifidn %1, 0
    lea       r0, [r0+r1*4]
%endif
    HSUMSUB %1, %2, %3, %4, %5
%endmacro

%macro HADAMARD_AC_SSE2 1
INIT_XMM
; in:  r0=pix, r1=stride, r2=stride*3
; out: [esp+16]=sa8d, [esp+32]=satd, r0+=stride*4
cglobal x264_hadamard_ac_8x8_%1
%ifdef ARCH_X86_64
    %define spill0 m8
    %define spill1 m9
    %define spill2 m10
%else
    %define spill0 [rsp+gprsize]
    %define spill1 [rsp+gprsize+16]
    %define spill2 [rsp+gprsize+32]
%endif
%ifnidn %1, sse2
    ;LOAD_INC loads sumsubs
    mova      m7, [hmul_8p GLOBAL]
%else
    ;LOAD_INC only unpacks to words
    pxor      m7, m7
%endif
    LOAD_INC_8x4W 0, 1, 2, 3, 7
%ifidn %1, sse2
    HADAMARD4_2D_SSE 0, 1, 2, 3, 4
%else
    HADAMARD4_V m0, m1, m2, m3, m4
%endif
    mova  spill0, m1
    SWAP 1, 7
    LOAD_INC_8x4W 4, 5, 6, 7, 1
%ifidn %1, sse2
    HADAMARD4_2D_SSE 4, 5, 6, 7, 1
%else
    HADAMARD4_V m4, m5, m6, m7, m1
%endif

%ifnidn %1, sse2
    mova      m1, spill0
    mova      spill0, m6
    mova      spill1, m7
    HADAMARD 1, sumsub, 0, 1, 6, 7
    HADAMARD 1, sumsub, 2, 3, 6, 7
    mova      m6, spill0
    mova      m7, spill1
    mova      spill0, m1
    mova      spill1, m0
    HADAMARD 1, sumsub, 4, 5, 1, 0
    HADAMARD 1, sumsub, 6, 7, 1, 0
    mova      m0, spill1
%endif

    mova  spill1, m2
    mova  spill2, m3
    ABS_MOV   m1, m0
    ABS_MOV   m2, m4
    ABS_MOV   m3, m5
    paddw     m1, m2
    SUMSUB_BA m0, m4; m2
%ifnidn %1, sse2
    pand      m1, [mask_ac4b GLOBAL]
%else
    pand      m1, [mask_ac4 GLOBAL]
%endif
    ABS_MOV   m2, spill0
    paddw     m1, m3
    ABS_MOV   m3, spill1
    paddw     m1, m2
    ABS_MOV   m2, spill2
    paddw     m1, m3
    ABS_MOV   m3, m6
    paddw     m1, m2
    ABS_MOV   m2, m7
    paddw     m1, m3
    mova      m3, m7
    paddw     m1, m2
    mova      m2, m6
    psubw     m7, spill2
    paddw     m3, spill2
    mova  [rsp+gprsize+32], m1 ; save satd
    mova      m1, m5
    psubw     m6, spill1
    paddw     m2, spill1
    psubw     m5, spill0
    paddw     m1, spill0
%ifnidn %1, sse2
    mova  spill1, m4
    HADAMARD 2, amax, 3, 7, 4
    HADAMARD 2, amax, 2, 6, 7, 4
    mova m4, spill1
    HADAMARD 2, amax, 1, 5, 6, 7
    HADAMARD 2, sumsub, 0, 4, 5, 6
%else
    mova  spill1, m4
    HADAMARD 4, amax, 3, 7, 4
    HADAMARD 4, amax, 2, 6, 7, 4
    mova m4, spill1
    HADAMARD 4, amax, 1, 5, 6, 7
    HADAMARD 4, sumsub, 0, 4, 5, 6
%endif
    paddw m2, m3
    paddw m2, m1
    paddw m2, m2
    ABS1      m4, m7
    pand      m0, [mask_ac8 GLOBAL]
    ABS1      m0, m7
    paddw m2, m4
    paddw m0, m2
    mova  [rsp+gprsize+16], m0 ; save sa8d
    SAVE_MM_PERMUTATION x264_hadamard_ac_8x8_%1
    ret