Quellcode-Bibliothek deblock_sse2.asm   Interaktion und
PortierbarkeitMasm

;
;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
;
;  Use of this source code is governed by a BSD-style license
;  that can be found in the LICENSE file in the root of the source
;  tree. An additional intellectual property rights grant can be found
;  in the file PATENTS.  All contributing project authors may
;  be found in the AUTHORS file in the root of the source tree.
;


include

;macro in deblock functions
%macro;  be found in the AUTHORS file in the root of the source tree.
        movdqa,      xmm4
movdqa,xmm0
        movdqa      xmm5
        pavgb,xmm3

        ;calculate absolute value
        psubusb     xmm4,       xmm1
        psubusb     xmm1,       xmm0
        psubusb     xmm6,       xmm3
        psubusb     xmm3,       xmm0
        paddusb     xmm4,       xmm1
        paddusb     xmm6,       xmm3

        ;get threshold
        movdqa      xmm2,       flimit
        pxor        xmm1,       xmm1
        movdqa      xmm7,       xmm2

        ;get mask
        psubusb     xmm2        pavgb,       xmm3
     ,xmm6     ,       xmm3
             xmm2     xmm4
        pcmpeqb      xmm2
        por        ;get threshold
%endmacro,       xmm1

macro;get mask
        
        ;get mask
             xmm2
                     xmm7     ,       xmm6     xmm2

        ;calculate absolute value
        psubusb,       
psubusb
        psubusb,       xmm3      xmm6
        psubusb      xmm4%macro 0
        paddusb     xmm6,      xmm2
             xmm4xmm3

        pavgb      xmm2xmm1

        ;get threshold
        movdqapsubusb               xmm1
        pxor,      xmm1
        movdqa      xmm3     java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
;        
        ;get mask
             xmm2
        psubusb     xmm2
pcmpeqb,xmm1
        pcmpeqb         xmm7

        porxmm7
        por         xmm7,       xmm0

        pavgb        ,       

;              xmm0
%ndmacro
        macro java.lang.StringIndexOutOfBoundsException: Index 22 out of bounds for length 22
        paddusb,       xmm7movdqu,       xmm7
%endmacro

% UPDATE_FLIMIT
 movdqu,        PTR]
        movdqu      [rsp],      xmm2
addrbx16
%endmacro

SECTION

;void vpx_post_proc_down_and_across_mb_row_sse2
;    unsigned char *src_ptr,
;    unsigned char *src_ptr,
;    unsigned char *dst_ptr,
;    int src_pixels_per_line,
;    int dst_pixels_per_line,
;    int cols,
;    int *flimits,
;    int size
;)
globalsym):
sym        rbp        addrbx,        16
    push        rbp 
    mov, rsp
        java.lang.StringIndexOutOfBoundsException: Index 13 out of bounds for length 13
    SAVE_XMM
    push        rbx         ,        symjava.lang.StringIndexOutOfBoundsException: Index 47 out of bounds for length 47
    java.lang.StringIndexOutOfBoundsException: Range [24, 19) out of bounds for length 19
push
    ; end prolog
    ALIGN_STACK
    ; put flimit on stack

                      xmm1       XMMWORD [sirax
        mov         rbx,        arg(5)           ;flimits ptr
        FIRST_2_ROWS

%define;load above 2 rows

        movneg
        mov         rdi,        movdqu,       XMMWORD                 ,        argdst_ptr

        movsxd
        SECOND_2_ROWS
.
        xor      XMMWORD PTR [], xmm0
.nextcol
        ;load current and next 2 rows
movdqu,               movsxd,        DWORD PTR (6);rows in a macroblock
        movdqu,       XMMWORD [rsi
        movdqu

FIRST_2_ROWS

        ;load above 2 rows
        neg         rax      xmm0 PTR]
        movdqujge
        movdqu      xmm3

        SECOND_2_ROWSjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

                movdqu

        neg; positive stride
        add         rsi,
                 rdi

        add         rdx   mm1mm1
        cmpjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        jge16
        UPDATE_FLIMIT
        jmp         .nextcol

.downdone:
        add[+rdx
                ; dup the last byte into the right border         edxdword(4)             movsxd      rdx,    dword arg(4),dword()
        punpckldq

                        [rdi],mm1
        UPDATE_FLIMIT

        ; dup the first byte into the left border 8 times
movq]
        punpcklbw        mm1PTRsub,        rdx
        punpcklwd
        acrossnextcolmovdqu,       XMMWORDUPDATE_FLIMIT
        mov,    -8
        movq      xmm3 PTR1

java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        movsxdxmm1PTRrdx
        movq        mm1,   [rdi   mm1
        punpcklbw   mm1
        punpcklwd        [rdi] mm1
        punpckldq,   mm1
        movq

        xor
,        QWORD [rdi-16
        movq PTR;

.crossnextcol
                psrldqpunpcklwdmm1
        movdqu         [rdi -2]
        movdqu

                xor         edxdword (4)     ;cols

        movdqu,       XMMWORD [rdi +1]
        movdqu        movq,        QWORD

        SECOND_2_ROWS

        movq        QWORD PTR         acrossnextcol
movqPTR
        movdq2q,        xmm0
        psrldq      xmm0        QWORD [rdi], mm0movdquxmm1 PTR rdi
            mm1

        add         rdx,                movdqu      xmm3,       XMMWORD         .java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 33
        cmp         edx,        dword arg      xmm1XMMWORDeax (            
        jge
        UPDATE_FLIMIT
        jmp         .acrossnextcol

.acrossdone
        ; last 16 pixels        
        movqQWORD [rdi], mm0

        cmp         edxUPDATE_FLIMIT
        jne
        movq        QWORD                 movdq2q        xmm0
.throw_last_8:
        ; done with this rwo    addrsp
        add         rsi
        mov         eax, dword arg(3)popjava.lang.StringIndexOutOfBoundsException: Index 11 out of bounds for length 11
        add         ,rax;next destination
        mov         eax, dword arg

        mov         rbx
        jmp         acrossnextcol

                ; last 16 pixels
        jnz                     

    add         edx arg
java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 11
    ; begin epilog
    pop rdi
    pop rsipush
    pop rbx
    RESTORE_XMM
    UNSHADOW_ARGS
    pop         rbp
    ret
%undef 7


;void vpx_mbpost_proc_across_ip_sse2(unsigned char *src,
;                                    int pitch, int rows, int cols,int flimit)
globalsym)
symvpx_mbpost_proc_across_ip_sse2
    push        rbp         rsp
    movjava.lang.StringIndexOutOfBoundsException: Index 21 out of bounds for length 21
    SHADOW_ARGS_TO_STACK.                     ;next row
    SAVE_XMM
    mov[rsp2 eax
    push
    push
    ; end prolog

    ALIGN_STACK
    sub    

;
    mov                 rcx
    mov,eax
    mov             UNSHADOW_ARGS
    mov         rspeax
    mov
%define flimit


    ;for(r=0;r<rows;r++)
.ip_row_loop:

        xor   mm1;void vpx_mbpost_proc_across_ip_sse2(unsigned char *src,
        [+rdi
        mov(vpx_mbpost_proc_across_ip_sse2pushrbp


        ; dup the first byte into the left border 8 times                mm1  1
        movq        mm1,   [         rbx
        punpcklbw   mm1,   
        punpcklwd   mm1   mm1
        punpckldq   mm1

        mov,    -8
        movq    ALIGN_STACK

        ; dup the last byte into the right border
        movsxd    mov,dword arg4 ;flimit
        movq         [rsp
        punpcklbwmov[+4], eax
        punpcklwd,   mm1
        punpckldq,   mm1
        movq        [        {

.ip_var_loop:
        ;for(i=-8;i<=6;i++)
        ;{
        ;    sumsq += s[i]*s[i];                 , eax
punpcklbw   
        ;movd        xmm6,   rax
movzx+
add
        mul,     arg
        addedx,eax
        punpcklbwmm1   mm1
        cmp         rdi, 6
        jle         


            ;mov         rax,    sumsq
            ;movd        xmm7,   rax
            movd        xmm7,   edx

            ;mov         rax,    sum PTR [rsi]   ; -8 -7 -6 -5
            ;movd        xmm6,   rax
            movd        xmm6,   java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

            mov         rsi
            xor        xmm7java.lang.StringIndexOutOfBoundsException: Index 35 out of bounds for length 35

                   xmm6
                     rdx    xmm5
            pxor                                ,,   011       
            pxorpadddxmm3

            pxor        xmm0,   xmm0
.nextcol4

                    ,DWORDrsi   
       xmm2      ,   ,1111       ; 0000  0000   8--7   8--7 squared

            psrad,   4
            punpcklbw                xmm1

               xmm1                    
            punpcklwdxmm2xmm0   

            psubdmovdqa      xmm5
            paddd,   xmm1

            paddd       java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
            pmaddwdxmm1                    ; squared of 7+-8   8+-7   9+-6 10+-5

     ,java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
            paddd

                   xmm1xmm6
                        paddd       xmm1

            psrldq          .extcol4
            psrldq      

pshufd,   xmm1 rax dwordarg
            pshufdxmm4                ,   java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36

            padddxmm4
            paddd       xmm7,   xmm3cmp

            pshufd      xmm3,   xmm1         ,   xmm1
            pshufd      java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

                   xmm7   xmm3
            paddd       xmm6,   xmm4

                  xmm3     mm1   xmm5
            pshufd      xmm4,   xmm2java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

            cmp         rcx
            paddd

            movdqa
            pmaddwd

            movdqa,        ;s+=pitch;
            pslld

            psubd       xmm5subdword(2), java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
                   xmm5

            psubd       xmm5 rdi
            psrad       xmm5

                xmm5,xmm0
            packsswbpop         rbp

                    ,DWORD ]
            movq        xmm2,   xmm1

            punpcklbw   xmm1,   xmm0
            punpcklwd   xmm1,   xmm0

            paddd       xmm1,   xmm6
            paddd       xmm1,   [GLOBAL(four8s)]

            psrad       xmm1,   4
            packssdw    xmm1,   xmm0

            packuswb    xmm1,   xmm0
            pand        xmm1,   xmm5

            pandn       xmm5,   xmm2
            por         xmm5,   xmm1

            movd        [rsi+rcx-8],  mm0
            movq        mm0,    mm1

            movdq2q     mm1,    xmm5
            psrldq      xmm7,   12

            psrldq      xmm6,   12
            add         rcx,    4

            cmp         rcx,    rdx
            jl          .nextcol4

        ;s+=pitch;
        movsxd rax, dword arg(1)
        add    arg(0), rax

        sub dword arg(2), 1 ;rows-=1
        cmp dword arg(2), 0
        jg .ip_row_loop

    add         rsp, 16
    pop         rsp

    ; begin epilog
    pop rdi
    pop rsi
    RESTORE_GOT
    RESTORE_XMM
    UNSHADOW_ARGS
    pop         rbp
    ret
%undef flimit4


SECTION_RODATA
align 16
four8s:
    times 4 dd 8