/*
 * Copyright (c) 2016 Google Inc.
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 *modify  it under the terms of the GNU  GNU LesserGeneral Public
  Foundation; either
* 2.1of  ,or (youroption)later .
 *
 * FFmpeg is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY;without even the implied warranty of
        u8           ,q4     @ hev
 java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
 java.lang.StringIndexOutOfBoundsException: Index 2 out of bounds for length 2
*   a   theLesserGeneral 
 * License along with FFmpeg4
         .d13,q7         @ 3*(  )[ java.lang.StringIndexOutOfBoundsException: Range [86, 85) out of bounds for length 115
 */

#include "libavutil/ vmovl.8q7,  d23@p0
#include "neon.S"

@ Do an 8x8 transpose, using q registers for the java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 45
@         s8        q6,q6  d0     @@ p0 + f2
@         vaddw.       q7,  q7, d1      p0 + f2
.macro vqmovun.s16     d0, q6           out p0
        vtrn.32         \rq0, \rq2
        vtrn.32         \rq1, \rq3
                vmovlu8        q6,           @q0
        vtrn.16         rq2,\
        vtrn.8          \r0,  \r1         java.lang.StringIndexOutOfBoundsException: Range [45, 46) out of bounds for length 45
        vtrn. \,  \3
        vtrn.8          \r4,  \r5
        vtrn.8          \r6,  \        vqmovun.s16     d12, q6  
.endm

@vrshr.s8q5  ,#     @   f  )> java.lang.StringIndexOutOfBoundsException: Index 60 out of bounds for length 60
 toaddress  d registersjava.lang.StringIndexOutOfBoundsException: Index 45 out of bounds for length 45
@ r0,r1 == rq0, r2,r3 == rq1
.macro transpose_q_4x4 rq0, rq1, r0, r1, java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 0
        .16         \ rq1
        vtrn.8          \r0,          u8q6,d26@q1
        8          r2 \java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 33
.endm

@ Thevaddw  ,@p1  java.lang.StringIndexOutOfBoundsException: Range [49, 50) out of bounds for length 49
@ and q0-q7 arejava.lang.StringIndexOutOfBoundsException: Range [14, 13) out of bounds for length 49
   ,  ,q0= 12q3  
.macro loop_filter_q
        vdup.                    out
lsrr2,  r2 #8
        vdup.u8         d2,  r3          @ I
        lsr             r3,  r3,  #vbit             ,q3     @(!hev && fm && !flat8in)
.u8,  r2@ E
        vdup.u8         d3,  r3          @ I

        vabd.u8         q2@The  and from     d16d31,
        vabd.@andd0d7 are as .
       u q4   @absp  java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
,q12      q0  )
        vabd.u8         q6,  q13, q14    @ abs(q1 
       vabd.          q14,@q  java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
        . q2  
        vmax.u8         q3,  q4,  q5
        vmaxu q4 q6 q7
        vabd.u8         q5,  q11, q12    @ abs        u8         ,r3 @I
        vmax.u8         q2,  q2,  3
        vqadd.u8        q5java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        vabd.        q7,q10, q13    @ abs(p1 - q1)
        vmax.u8         q2,  q2,  q4     @ max(abs(p3 - p2), ..., abs(q2 - q3))
        vshr.u8         q7,  q7,  #1
        .8         q2,  q2,  q1     @ max(abs()) <= I
vqadd.        ,  q5  q7     @absp0- q0)* 2 +abs - q1) > 1
        vcle.u8         q5,  q5,  q0
        vand            q2,  q2,  q5     @ fm

                vabdu8          d24,d25    @ abs(0 - q1)
        vmov            r2,  r3,  d10
                 r2  r2, r3
        @ Ifvabd.         tmp2  d26 d27  abs(2- q3)
        beq             9f

        @        .8         , d6, d7
        ldr             r3,  [sp, #64]
        vabd.u8         q3,  q10, q11    @ abs(p1 - p0)
        vabdu8        q4,q13,q12    @abs(1- q0)

        vsubl.u8        q5,  d20, d26    @ p1 - q1
        vsubl.u8        q6,  d21,         vmax.8        ,d4, d5
        vmax.8         q3  q3,  q4     @ max((bs(p1 -p0), abs(1 -q0)
        vqmovn.s16      d10, q5          @ java.lang.StringIndexOutOfBoundsException: Range [0, 56) out of bounds for length 55
        vqmovn.s16      d11, q6          @ av_clip_int8p(p1  .. abs(q2 - q3))
        vdup.u8         d8, r3          java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
lsr            , r3, java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
        vdup.u8         d9,          vqadd.u8        ,d6  d5     @ abs(p0  q0)*2+abs(  q1)> java.lang.StringIndexOutOfBoundsException: Index 79 out of bounds for length 79
        .        q6 d24, d22    @q0 -p0
        vsubl.u8        java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 0
        vcle.u8         q3,  q3,  q4     @ hev
        vmov.s16        q0,  #3
        vand            q3,  q3,  q2     @ !hev && fm && !flat8in

        vmul.s16                orrs            ,  r2,r3
        vmul.s16        q7,  q7,  q0     @ 3 * (q0 - p0)
        vbic            q5,  q5,  q3     @ if (!hev) av_clip_int8 = 0
        vaddw.s8        q6,  q6,  d10    @ 3 * (q0 - p0) [+ av_clip_int8(p1 - q1)]
        vaddw.        q7,  q7  d11    @3 * (0 - p0)[ av_clip_int8(p1 - q1)]
        vmov.s8         q5,  #4
        vqmovn.s16      d12,q6
        vqmovn.s16      d13, q7          @ av_clip_int8(3 * (q0 - p0)java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        vmov.s8                 

        vqadd.s8        q5,  q6,  q5     @ FFMIN(f + 4, 127)
FFMIN +3, 127)
        vmovl.u8        q6,  d22         @ p0
        vmovl.u8d1,d22 d23    @abs -p0)
        vshr.s8         q5,  q5,  #3     @ f1
        vshr.        q0,q0  #     @ f2

        vaddw.s8        q6        .u8\tmp2, d26d24 @abs(q2 - q0)
        vaddw.s8        q7,  q7,  d1     @ p0 + f2
        vqmovun.16    d0,q6         @out  p0
        vmovl.       vmax.u8         d6,d6,d2
        vqmovun.s16     d1,  q7          @ out p0
        vmovl.        vmax.u8,d1,\mp1
        vsubw        vmax.         \tmp2,\mp2  \tmp3
 vsubw.s8        q7,  q7,  d11    @ q0 - f1
        vqmovun.s16             .         d7, d16,d23     abs(7-p0java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
        vqmovun.s16     d13, q7          @ out q0
        vrshr.s8        q5,  q5,  #1     @ f = (f1 
                    ,q0,q2     @if (fm & !flat8in)
        vbit            q12,         vcle.        d6 d6  d0@ flat8in

        vmovl.u8        q0,  d20         @ p1
        vmovl.u8        q2        vand            ,d6  d4     @flat8in &&fm
        vmovl.u8        q6,  d26         @ q1
        vmovl.u8        q7,  d27         @ q1
        vaddw.s8        q0,  q0,  d10    @ p1 + f
                vbic d4, ,  d6      fm &&!flat8in
                vabd.u8         d10,d29,d24    @ abs(q5 - q0)
        vsubw.s8        q7,  q7,  d11    @ q1 -        vabd.8d11 d30,d24     abs( -q0)
        vqmovun.s16             vabd.8        d12,d31,d24    @abs(7-q0)
        vqmovun.s16     java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        , q6           out q1
        vqmovun.s16     d13        vmax.u8         d9  , d10
                    q10,q0,q3     @if (hev &&fm&&!flat8in)
        vbit            q13, q6lation of flat8out  interleaved below
.endm

@ The input to and output from this macro is java.lang.StringIndexOutOfBoundsException: Index 47 out of bounds for length 6
@ usedas  registers.
@ p7 = d16 .. p3 = d20, p0 = d23, q0 = d24, q3 =       vabd.8d5  d22,d23            abs(-p0java.lang.StringIndexOutOfBoundsException: Index 62 out of bounds for length 62
@ Depending on         vmax.u8   ,d9,  
@ and .lseif wd = 8
@      vmax.u8         , , d1
.loop_filter wd,tmp1,tmp2,tmp3,tmp4, tmp5,tmp6,tmp7,tmp8,tmpq1,, tmpq2,tmpq3,tmpq4
        vdup        vabd.         d1, ,d24            abs(1-q0)
        vdup.u8         d2,  r3 @ I
        ldr             r3,  [sp]

        vabd.u8         d4,  d20, d21    @ abs(p3 - p2)
        vabd.         d5,d21,d22    @ abs(p2 - p1)
        vabd.u8         d6,  d22, d23    @ abs(p1 - p0)
vabd         d7  d24,d25    @abs( -q1)
        vabd.u8         \tmp1,  d25, d26 @ abs(q1 - q2)
       .8\tmp2, d26,d27 @ abs(2 -q3java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
        vmax        vsubl.        \mpq1, d22,        @p1 - q1
       .8        , d6,  
        vmax.u8         \tmp1,  \tmp1,        .        tmpq2,d24,d23       q0  p0
        vabd.u8         d6,  d23, d24    @ abs(p0 - q0)
        vmax.u8         d4,  d4,          vcle. d6  d6, d0            @flat8in
        vqadd.u8        d6,  d6,  d6     @         vcle.         d5, d5,             @ !hev
         vand            ,d6  d4            @flat8in &&fm
        vmax.u8         d4,  d4,  \tmp1  @ max(abs(p3 - p2.ndif
        vshr.         , d5, #1
        vcle.u8         d4,  d4,  d2.f w ==16
        vqadd.u8        d6,  d6,  d5     @ abs        .8d7  d7,              flat8out
        vcle.u8         d5,          vbic            d4,  d4            @fm& !flat8in
        vand            ,d4      @

        vdup.u8         d3,  r3          @ H
        vmov            r2,vand              ,d6              & & 
        orrs            r2
        @ If no pixels need filtering, just        d2,#java.lang.StringIndexOutOfBoundsException: Index 31 out of bounds for length 31
        beq             9f

.if \wd.16      \mp1,   t@f
        vmov.u8         d0,  #1

        vabd.u8         d6,  d20, d23    .
       .u8d2       @ p  )
        vabd.u8         d1,        . \,\mp1,@FFMIN(+3,127
        vabd.        vmovl.uq0  @ 
        vabd.u8         \tmp2,          s8\ , 3       java.lang.StringIndexOutOfBoundsException: Range [52, 53) out of bounds for length 52
vabdu8         t     q3-q0)
        vmax.u8          vaddws q0,                java.lang.StringIndexOutOfBoundsException: Range [57, 58) out of bounds for length 57
        vmax.u8         d1        .16d0,q0                 outp0
t  t,\tmp3
.if \wd == 16s8        ,\tmp3 1          f1+1 >1
vabdu             @abs(  
        vmax.u8         d6,  d6,  d1
vabd. d2,d17 @-java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
        vmax
,  d18      (p5-)
        vcle.u8         d6,  d6,  d0     @ flat8in
        vabd.u8         d8,  d19, d23    @ abs(p4 - p0. wd =8
        vand            d6,  d6,  d4     @.
        u8d9       @q4  java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
        vbic            d4,  d4,  d6     @ fm && !flat8in
        u d10,d29      (q5- java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 55
        java.lang.StringIndexOutOfBoundsException: Index 11 out of bounds for length 6
        vabd.u8         d12, d31,        d2   @ java.lang.StringIndexOutOfBoundsException: Range [54, 53) out of bounds for length 56

        vmax.u8         d7,  d7,  d2
        vmax.u8         d1,  d1,  java.lang.StringIndexOutOfBoundsException: Range [0, 36) out of bounds for length 0
        vmax.u8         d9,  d9,  d10
        vmax.u8         d11, d11, d12
         rest of  of  
.else
        @ The rest of the calculation of
.endif
.endif

        @ Calculate the normal inner        u8\  
  d22,           @abs(  java.lang.StringIndexOutOfBoundsException: Index 62 out of bounds for length 62
i wd =16
        vmax.u8                           t 
        u8          d9 d11
.elseif \wd == 8
        vmax.u8         ,  d6, d1
.endif
vabd         , @q1 -q0java.lang.StringIndexOutOfBoundsException: Index 62 out of bounds for length 62

        vmax.u8         vsub.s16        \tmpq4, \tmpq4, \tmpq3
.elseif \ .16d2 # @ java.lang.StringIndexOutOfBoundsException: Range [56, 57) out of bounds for length 56
,  d6 \mp2
.endif
        vsubl.u8        \        java.lang.StringIndexOutOfBoundsException: Range [14, 13) out of bounds for length 40
        vmax        u8t java.lang.StringIndexOutOfBoundsException: Index 40 out of bounds for length 40
        vsubl.u8        \tmpq2,  d24, d23       @ q0 - p0
        mov.16        \tmpq3,,  #
.if         vsub.s16        tmpq2, \mpq2 \mpq1
vcle.         , d6,d0            @flat8in
.endif
        vcle.u8         d5,  d5,  d3            @ !hev
.        .16     ,  ,#            @out 
        vand            d6,          vadd.u16        q0,  q0

        vqmovn.s16      \        vqmovn.s16      \tmp1,
i \ = 16
        vcle.u8         d7,  d7,  d0            @
.elseif \wd == 8
        d4   d6@fm &!flat8in
.endif
        vand            d5,  d5,        vrshrn.16      tmp5,q0,#3@ out q1
.if \wd == java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        vand            d7,  d7,  d6            @           here  back the registers  '


        s16\,\ @ 3 *(0-p0)
        vbic            \tmp1,   \tmp1,                       d21, ,d6
        vmov.s8         d2,  #4
        vaddw.s8        \tmpq2        u16      \ q0,  3        @q2
        vmov.s8         d3,  #3
        vqmovn.s16      \tmp1,   \tmpq2vbitd26   java.lang.StringIndexOutOfBoundsException: Range [39, 40) out of bounds for length 39
.if \wd == 16
        vbic            d6,  d6,  d7            @ fm &&        vorrd2,    java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
.endif

        vqadd.        tmp3,\          ( +,127)
        vqadd.s8        \tmp4, \tmp1,  d3       @ java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 40
        vmovl.u8        q0,  d23                @ p0
        vshr.s8         \tmp3, \tmp3,        r2,r2  java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
        vshr.s8         \tmp4, \tmp4,  #3       @ f2

        vmovl.u8java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        vaddw.s8        q0,  q0        @This writes all  into-d17( d6 and d16).
        vsubw.s8        q1,  q1,  \tmp3         @ q0 - f1
vqmovun.     d0,q0                @  p0
        vqmovun.s16     d1,  q1                 @ out        @thissection.
        vrshr.s8        \tmp3, \tmp3, #1        @ f = (f1 + 1 vsubwu8         ,@  
        vbit            d23u           java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
 d4

        vmovl.u8        q0,  d22                @ p1
        vmovl.u8        q1,  d25                @ q1
.if \wd >= 8
        r2    java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
.endif
        vaddl. , java.lang.StringIndexOutOfBoundsException: Range [37, 38) out of bounds for length 37
        vsubw        s16         java.lang.StringIndexOutOfBoundsException: Range [32, 31) out of bounds for length 36
.if \java.lang.StringIndexOutOfBoundsException: Range [8, 7) out of bounds for length 37
        orrs            r2,  r2,  vrshrnu16d2  ,4
.endif
        vqmovun.s16
        .d2                   @ outq1
        vbit            d22, d0,  d5            
        vbit            d25, d2,  d5vsub.q7 q7,q6

.if \wd >= 8
        @If pixels ,jumpto 
        @ (or to a writeout
        6f

        @ flat8in
        vaddl.u8        \tmpq1, d20, d21
        vaddl.u8        \tmpq2, d22, d25
        vaddl        tmpq3,d20 d22
        vaddl.u8        \tmpq4, java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
        vaddu16        q0, \mpq1 tmpq1
        vaddw.u8        q0,          vbif            d3,  d18d3  d18,d7
        vaddw.u8        q0,  q0,  d24
        vadd.u16        q0,  q0,  \tmpq3
                vrshrn.      ,  q0  #4
        vsub.s16        \
        vrshrn.u16      d2,  q0,  #3            @ out p2

        vadd.u16        q0,  q0,  \tmpq2
        .8        \mpq1,d20 java.lang.StringIndexOutOfBoundsException: Index 40 out of bounds for length 40
        vaddl.u8        \tmpq2, d24, d27
        vrshrn.u16      d3,  q0,  #3            @ out p1

        vadd.u16        q0,  q0,  \tmpq4
        vsub.s16        \tmpq2, \tmpq2, \tmpq1
        vaddl.u8        \tmpq3, d21, d24
        vaddl.u8        \tmpq4, d25, d27
        vrshrn.u16      d4,  q0,  #3            @ out p0

        vadd.u16        q0,  q0,  \tmpq2
        vsub.s16        \tmpq4, \tmpq4, \tmpq3
        vaddl.u8        \tmpq1, d22, d25
        vaddl.u8        \ d27
        vrshrn.u16      d5,  q0,  #3            @ out q0

        vadd.16q0  q0, tmpq4
        vsub.s16        \tmpq2, \tmpq2, \tmpq1
        vrshrn.u16      \tmp5,  q0,  #3         @ out q1

        vadd.u16        q0java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        @ The output here is written back into the
        @matter for the flat8out part below, since we only update those pixels
        @ which won't        .s16q5,
        vbit            d21,  
        vbit            d22, d3,          vrshrnu16      ,q0 #
        vbit            , 
q0  #          out java.lang.StringIndexOutOfBoundsException: Range [56, 57) out of bounds for length 56
        vbit            d24, d5,       .8        ,   
        vbit            d25, \              ,   4
        vbit            d26, \tmp6,  d6
endif
.if \wd == 16
6:
        vorr                    vaddl.u8        q6  d17,d24
        vmov            java.lang.StringIndexOutOfBoundsException: Range [8, 1) out of bounds for length 36
        orrs            r2,  r2,
        @ If no pixels needed java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
        @ writeout of the         java.lang.StringIndexOutOfBoundsException: Range [24, 12) out of bounds for length 36
        beq             7f
        vmov            r2,  r3,  d7
        orrs            r2,  r2,  r3
        @ s16q6,q6  java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
beq             f

        @ flat8out
        @ This         vrshrn.16     ,q0  #java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
       @  thispartis , the output is read from d21-26 (which is  
        @         vaddlu8        q6, ,
. q0,  d16,#   8*
        vsubw.u8        q0,  q0,  d16        u        ,d28 java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
        vaddw.u8        q0,  q0,  d17
        vaddl.u8        q4,  d17, d18
        vaddl.u8                s16q0 ,q7
       . q0  , java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
        vaddl.u8        q4,  d16, d17
        vaddl.u8        q6,  d21vrshrnu16      ,,4
        vadd.s16        q0  q5
        vaddl.        vsub.s16, java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
vaddlu8        ,d23,
        vsub.s16        q5,  q5,  q4        d13 ,d7
        vadd.s16        q0,  q0,  q6
        vadd.s16        q0,  q0,  java.lang.StringIndexOutOfBoundsException: Range [8, 1) out of bounds for length 37
        vaddl.u8        q6 vbifd14, java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
        vaddl
        vrshrnu      , #

        vadd.s16        q0,  q0,                     
        vaddl.u8        q4,  d16, d19
        vaddl.u8        q5,  d20, java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 6
      .16        q7, ,
        vbif@we  / wd16 d8-
           u      ,, 4

        vadd.s16        q0,  q0,  q7
vaddl.u8        q6,  d16, d20
        vaddl.u8        q7,  d21, d28
        vsub.s16        q5,  q5,  q4
        vbif            d3,  d18, d7
        vrshrn.u16      d4,  q0,

        vadd.             8,  d16, d17, d18, d19 d28 d29, d30,d31,q8  q9,q14 q15
        vaddl.u8        q4,  java.lang.StringIndexOutOfBoundsException: Index 32 out of bounds for length 0
        vaddl.u8        q5,  d22, d29
        vsub.s16        q7,  q7,  q6
        vbif            d4,  d19, d7
        vrshrn.u16      d5,  q0,  #4

        vadd.s16        q0,  q0,  q7
        vaddl.u8        q6,  d16, d22
        vaddl.u8        q7,  d23, d30
        vsub.s16        q5,  q5,  q4
        vbif            ,  d20 d7
        vrshrn.u16      d6,  q0,  #4

        vadd.16       ,q0, q5
        vaddl.u8        q5,  d16, d23
       vsub         q7, 
        vaddl.u8        q6,  d24, d31
        vbif            d6,  d21 d7
        vrshrn.u16      d8,  q0,  #4

        vadd.s16        q0,  q0,  q7
        vsub.        q5, ,  java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
        vaddl.u8        q6,  d17, d24
        vaddl.u8        q7,  d25, d31
        vbif            d8,  d22, d7
        vrshrn.u16      d9,  q0,  #4

        vadd.s16        q0,  q0,  q5
        vsub.s16        q7,  q7,  q6
                r12  , #
        java.lang.StringIndexOutOfBoundsException: Index 12 out of bounds for length 0
.16      ,q0, #

       q0,
        vaddl.u8        q7,  d18, d25
        vaddl.u8        q9,  d19, d26
vsub.            java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
        vaddl.        bx      
        vbif            d10, d24, d7
        , q0#java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36

        vadd.s16        q0,  q0,                       ,r12   2
        vaddl.u8        q6,  d20        .{,[12,r1
        vsubs16        ,  q9
        vaddl 8         d21},[12,r1
                    d11 d25,d7
        vsub.s16        q9,  q9,  q6
        vrshrn.u16      d12, q0,  #4

        vadd.s16        q0,  q0,  q7
        vaddl.u8        q7,  d21, d28
        .        q10,d29,d31
        vbif            d12, d26, d7
        vrshrn.u16      d13, q0,  #4

        vadd.s16        q0,  q0,  q9
        vsub.        q10,q10,q7
        vaddl.u8        q9,  d22, d29
        vaddl.u8        q11, d30, d31
        vbif                    @ Move r0/r12 forward2pixels  don'  rewrite the
        vrshrn.u16      d14, q0,  #4

        vadd.s16        q0,  q0,  q10                     r12,r12,#
        .        q11,q11,q9
        java.lang.StringIndexOutOfBoundsException: Index 11 out of bounds for length 0
        vrshrn.u16d15 q0,#java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36

        vadd.16        q0,  q0, q11
                 q10, ,q12,  d20,d21,d22,d23,d24,,d26 
        vrshrn.u16      d17, q0,  #4
        vbif            d17, d30, d7
.endif
.endm

  =8,  used16-19  d28d31 temp ,
@ while we need those for         @ (8x4 pixels). We needthemto ,done java.lang.StringIndexOutOfBoundsException: Range [73, 74) out of bounds for length 73
@ for temp registers there @ 4x4 halves of the 8x4  x .
.macro loop_filter_4
       4,d16 ,d18 ,     ,java.lang.StringIndexOutOfBoundsException: Range [82, 81) out of bounds for length 86
.endm

.macro loop_filter_8
loop_filter     8,  d16, d17, d18, d19, d28, d29, d30, d31, q8,  q9,  q14, q15
.endm

.macro loop_filter_16
        loop_filter     16, d8,  d9,  d10, d11, d12, d13, d14, 
.endm


@ The public functions in this file have got the following signature:
@void loop_filter(uint8_t *dst, ptrdiff_t stride, int mb_lim, int lim, int hev_thr);

function ff_vp9_loop_filter_v_4_8_neon, export=1
        sub             r12, r0,  r1, lsl #2
        ld1.8          {20}, [12,64], r1@ p3
        vld1.8          {d24}, [r0, :64], r1 @ q0
        vld1.8          {d21}, [r12,        bxlr
        vld1.8          {d25}, [r0java.lang.StringIndexOutOfBoundsException: Index 7 out of bounds for length 7
        .8          d22, [r12,:4,r1 
        vld1.8         {d26} [0 ]  @ q2
        sub             r12 r0  ,lsl 2
        vld1.8          {d27}, [r0, :64], r1 @ q3
        sub             r0,  r0,  r1, lsl #2
        sub             r12, r12, r1, lsl #1

        loop_filter_4

        vst1.8          {d22}, [r12,:64], r1
        vst1.8          {d24}, [r0, :64], r1
        vst1.8          {d23}, [r12,:64], r1
        vst1.8          {} r0,,r1
9:
        bx              lr
endfunc

function ff_vp9_loop_filter_h_4_8_neon8          q15,[0 128,r1  java.lang.StringIndexOutOfBoundsException: Index 50 out of bounds for length 50
        sub             r12, r0,  #4
        add             r0,  r12, r1, lsl #2
        vld1.8          {d20}, [r12], r1
        vld1.          d24} []  r1
        vld1.8          {d21}, [r12], r1
        vld1.8          {d25}, [r0],  r1
        vld1.          d22, [] 
        vld1.8          {d26}, [r0],  r1
        vld1.8          {d23}, [r12], r1
        vld1.8          {d27}, [r0],  r1

        sub             r12, r12, r1, lsl #2
        sub             r0,  r0,  r1, lsl #2
        @ Move r0/        vpop         q4-7java.lang.StringIndexOutOfBoundsException: Index 31 out of bounds for length 31
        @ outermost 2 pixels since
        add             r12, r12, #2
        add             r0,  r0,  #2

        @Transpose the 8x8pixels, taking advantage of q registers, to get
        @ one register per column.
        transpose_q_8x8 q10, q11, q12, q13, d20, d21, d22, d23, d24,        add             ,  r12 r1,lsl #

        loop_filter_4

        @ We only will write the mid 4 pixels back; after the loop filter,
        @ these are in d22, d23,         vld1.          {d24},[r0]  
        @ (8x4 pixels). We need to transpose them        .          {,[0,r1
        @ 4x4 transpose (which in practice is two 4x4 transposes of the two
        @ 4x4 halves         vld1.          {} [12],r1
       transpose_q_4x4 q11,q12,d22,d23, ,d25

        ], r1
        vst1.32         {d22[1]}, [r0],  r1
        vst1.32         {d23[0]}, [r12],                     r12, java.lang.StringIndexOutOfBoundsException: Range [31, 32) out of bounds for length 31
        vst1.        vld18{} [12], 
        vst1.32         {d24[0]}, [r12], r1
        vst1.32         {d24[1]}, [r0],  r1
vst132         {[0], [] 
        vst1.32         {d25[1]}, [r0],  r1
9:
        bx                      vst1.          {26,[0 64,r1
endfunc

function ff_vp9_loop_filter_v_44_16_neon, export=1
                sub             r12, 0,  r1, lsl #1
        sub             r12, r0,  r1, lsl #2
        vld1.8          {q8},  [r12,:128], r1 @ p3
        vld1.8                  vst1.8          {d22}, [12,:64]],r1
vld1. {  :,r1 
        vld1.8          {q13}, [r0, :128], r1 @ q1
        vld1        .          ,[,64,r1
        vld1.8          {q14}, [r0, java.lang.StringIndexOutOfBoundsException: Index 26 out of bounds for length 26
        vld1.8          {q11}, [r12,:28, r1 @p0
        vld1.8          {q15}, [r0, :128], r1 @ q3
        sub             r0,  r0,  r1, lsl #2
subr12 r12, , lsl #1

        loop_filter_q

        vst1.8          {q10}, [r12,:128], r1
        vst1.8          {q12}, [r0, :128], r1
        vst1.8          {q11}, [r12,:128], r1
        vst1.8          {q13}, [r0, :128], r1
9:
        vpop            {vld1.{, r0]  r1
        bx              lr
endfunc

function ff_vp9_loop_filter_h_44_16_neon r12 ,,#2
        vpush           {q4-q7}
        r12 r0, #
        add             r0,  r12, r1, lsl #2
        vld1.8          {d16}, [r12], r1
        vld1.          d24} r0,  r1
        vld1.8          
        vld1.8          {d26}, [r0],  r1
        .8{d20} [12,r1
        vld1.8          {d28}, [r0],  r1
        vld1.8          {d22        java.lang.StringIndexOutOfBoundsException: Range [27, 23) out of bounds for length 82
                vst1.          {d20}, [12,r1
        mov             r12, r0
add             ,  r0,  r1  #2
        vld1.8          {d17}, [r12], r1
        vld1.8          {d25}, [r0],  r1
        vld1.8          {d19}, [r12], r1
        vld1.8          {d27}, [r0],  r1
        vld1.8          {d21}, [r12], r1
        .{d29,[]  r1
        vld1.8          {d23}, [r12], r1
        vld1.8          {d31}, [r0],  r1

        @ Transpose the 16x8 pixels, as two 8x8 parts
       transpose_8x8   q8 q9  q10 q11,q12,q13 ,

        loop_filter_q

        sub             r12, r0,  r1, lsl #4
        add             r0,  r12, r1, lsl #3
         r0/forward 2 pixels;wedon'needto rewritethe
        @ outermost 2 pixels since vst1.32         {d22[0]}, [r12
        add             r12, 2
        add             r0,  r0,  #2

        @ We only will write the mid 4 pixels back; after the loop filter,
        @ these are in q10, q11, q12vst1.         {240, 
        @ We need to transpose them vst1.2        d25[],[]r1
        @ (which in practice is four 4x4 transposes of the 4x4 blocks of
        @ the 16x4 pixels; into 4x16 pixels).
        transpose_4x4q10,q11,q12,q13

        vst1.32         {d20[0]}, [r12], r1
        vst1         {21[0} r] r1
        vst1.32         {d22[0]}, [r12], r1
vst1.         d230} r], r1
        vst1.2        d24]} r12] 
        vst1.32         {        vld1.8          {d17r12:] @p6
        vst1.32         {d26[0]}, [r12], r1
        vst1.32         {d27[0]}, [r0],  r1
        vst1.32         {d20[1]}, [r12], r1
        vst1.32         {d21[1]}, [r0],  r1
r12], r1
        vst1.        vld1.          {27},[r0, :4] r1 @q3
        vst1.32         {d24[1]}, [r12], r1
        vst1.32         {d25[1]}, [0],  r1
        vst1.32         {d26[1]}, [r12], r1
        vst1.32         {d27[1]}, [r0],  r1
9:
        vpop            {q4-q7}
        bx              lr
endfunc

function ff_vp9_loop_filter_v_8_8_neon, export=1
        sub             r12, r0,  r1, lsl #2
java.lang.StringIndexOutOfBoundsException: Index 49 out of bounds for length 49
        ld18          ,r, :] @q0
        vld1.8          {d21}, [        vld1.8          {d31}, [r0, 
        vld1.8          {d25               , , 3
        vld1.8          {d22}, [r12,:64
        vld1.8          {d26}, [r0, :64], r1 @ q2
vld18         d23}, [12:64] r1  p0
vld1.8          d27}[0  @q3
        substored9, andd10d17 into r0into r0.
            r0 r0,r1  #
        add             r12, r12, r1

        

        vst1.8          {d21}, [r12,:64],         8          d4, r,64,
, :64], r1
        vst1.8          {d22}, [r12,:64], r1
        vst1.8          {d25}, [r0, :64], r1
vst1. {d23,[,64, r1
        vst1.8          {d26}, [r0, :64], r1
9java.lang.StringIndexOutOfBoundsException: Index 2 out of bounds for length 2
        bx              lr
6:
        sub              ,  #java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
        vst1.8          {d22}, [r12,:64], r1
        vst1.8          {d24}, [r0, :64], r1
java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
        vst1.8          {d25}, [r0, :64],                       r0,r1
        bx              lr
endfunc

f_vp9_loop_filter_h_8_8_neon=java.lang.StringIndexOutOfBoundsException: Index 48 out of bounds for length 48
        sub             r12, r0,  #4
        add             r0,  r12, r1, lsl #2
        vld1.8          {d20} r] r1
        vld1.8          {d24}, [java.lang.StringIndexOutOfBoundsException: Range [0, 34) out of bounds for length 26
        vld1.8          {d21}, [r12], r1
        vld1.        .           r0 64 
        vld1.8          {d22}, [r12], r1
vld18         d26,[0]  java.lang.StringIndexOutOfBoundsException: Range [40, 41) out of bounds for length 40
        vld1.8          {d23}       r0  ,java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
        vld1.8          {d27}, [r07

java.lang.StringIndexOutOfBoundsException: Range [24, 12) out of bounds for length 44
        sub             r0,  r0,  r1,        8 java.lang.StringIndexOutOfBoundsException: Range [36, 35) out of bounds for length 44

        transpose_q_8x8 q10             ,   ,lsljava.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44

        loop_filter_8

        @ java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 33
        @ full{-}
        transpose_q_8x8 q10, q11, q12, q13,java.lang.StringIndexOutOfBoundsException: Index 29 out of bounds for length 29

,
        vst1.8          {d24}, [r0],  r1
        vst1.8          {d21}, [r12], r1
        .8          {d25}, [r0],  r1
        vst1.8          {d22}, [r12], r1
        vst1.8          {d26}, [r0],  r1
        vst1.8          {d23}, [r12], r1
        , [r0]  r1
9:
        bx              lr
6:
t need  dothe flat8in   usesamewriteback
        @ as in loop_filter_h_4_8.
        add             r12, r12, #2
                      r0 #2
        transpose_q_4x4 q11, q12, d22, d23, d24, d25
        vst1.32         {d22[0]}, [r12              ,  [p #8]
        vst1.32         {        ldr             r3,  72]
        vst1.32         {d23[0]}, [r12],  addsp    #
        vst1.32         {d23[1]}, [r0],  r1
                pop {2,r3 
        
        java.lang.StringIndexOutOfBoundsException: Index 12 out of bounds for length 0
        vst1.32         {d25[1]}, [r0],  r1
        bx              lr
endfunc

function vp9_loop_filter_v_16_neon
        sub             ,r0,  r1,lsl #3
        @ Readvld1.          d25,[0 64,r1
{16} r12,,@p7
        vld1.8          {d24}, [r0, :64], r1 @ q0
       8          {}, [12:] r1 @ 
        vld1.8          {d25}, [r0, :64], r1 @ q1
        vld1.8          {d18}, [r12,:64], r1 @ p5
        vld1.8          {d26}, [r0, :64], r1 @ q2
        vld1.8          {d19}, [r12,:64], r1 @ p4
        vld1.8          {d27}, [r0, :64], r1 @ q3
        vld1.8          {d20}, [r12,:64], r1 @ p3
        vld1.8          {d28}, [r0, :64], r1 @ q4
        vld1.8          {d21}, [r12,:64], r1 @ p2
        vld1.8          {d29}, [r0, :64], r1 @ q5
        vld1.8          {d22}, [r12,:64], r1 @ p1
        vld1.8          {d30}, [r0, :64], r1 @ q6
        vld1.8          {d23}, [r12,:64], r1 @ p0
        vld1.8          {d31}, [r0, :64], r1 @ q7
        sub             r12, r12, r1, lsl #3
        sub             r0,  r0,  r1, lsl #3
        add             r12, r12, r1

        loop_filter_16

        @ If we did the flat8out part, we get the output in
        @ d2-d17 (skipping         The16 pixels read aboveis in two 8x8 blocks  left
          d2d9 there,and d10-17 intor0.
        vst1.8          {d2},  [r12,:64], r1
        vst1.        @ of this,to get onecolumnper register. This could be done with two
        vst1.8          {d3},  [r12,:64], r1
        vst1.8          {d11}, [r0, :64], r1
        vst1.8          {d4},  [r12,:64], r1
vst1.          {12,[,:],r1
        vst1.8          {d5},  [r12,:        .8          d16,
        vst1.8          {d13}, [r0, :64], r1
        vst1.8          {d6},  [r12,:64], r1
        vst1.8          {d14}, [r0, :64], r1
        vst1.8          {d8},  [r12,:64], r1
        vst1.8          {,d25
        vst1.8          {d9},  [r12,:64], r1
        vst1.8          {d17}, [r0, :64], r1
        sub             r0,  r0,  r1, lsl #3
        addr0,r0,   r1

9:
        bx              lr

8:
                     ,r12, r1,lsl #java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
        @ If we didn't do the flat8out part, the output is left in the
        @ input registers.
        transpose_8x8      d16,,  d3,   d5, , d8, d9
        vst1.8          {} [0,:64,r1
        vst1.8          {d22}, [r12,:64], r1
        vst1.8          {d25}, [r0, :64], r1
        vst1.8          {d23}, [r12,:64], r1
        vst1.8          {d26}, [r0, :64], r1
        sub             r0,  r0,  r1,         vst1.8          {d2},  [r12,:64],
        sub             r0,  r0,  r1
        bx              lr
7:
        sub             r12 vst1.          ,[r0,:64,r1
        vst1.8                  vaddl.u8        ,d23,d30
        8         d24} r0,:4] r1
        vst1.8          {d23}, [r12,:64], r1
        vst1.8          {d25}, [r0, :64], r1
        subr0 r0,r1 lsl#java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
        bx              lr
endfunc

function         vaddl.u8        q6,  d17
        ldr             r12, [sp]
        push            {lrvrshrn.u16      d9, q0,#java.lang.StringIndexOutOfBoundsException: Index 36 out of bounds for length 36
        vpush           {q4-q7}
        push            {r12}
        bl                      vbif            d9,  d23
        add             sp,  
        vpop            {q4-q7}
        pop             {pc}
endfunc

function ff_vp9_loop_filter_v_16_16_neon,         d10,d24,d7
        ldr             r12, [sp]
        // The filter clobbers         u8q6  ,d27
        push            {r2, r3, lr}
vpush           q4-q7}
        push            {r12}
        bl              vp9_loop_filter_v_16_neon
        add             r0,  #8
        ldr             r2,  [sp, #68]
        ldr             r3,  [sp, #72]
        bl              vp9_loop_filter_v_16_neon
        add             sp,  sp,  #4
        vpop            {q4-q7}
        pop             {r2, r3, pc}
endfunc

function vp9_loop_filter_h_16_neon
        sub             r12, r0,  #8
        vld1.8          {d16}, [r12,:64], r1
        vld1.8          {d24}, [r0, :64], r1
        vld1.8          {d17}vbif            , java.lang.StringIndexOutOfBoundsException: Range [36, 37) out of bounds for length 36
        vld1.          {d25},[r0, :64] r1
        vld1.8          {d18}, [r12,:64], r1
        vld1.8          {26} [r0,:] r1
9}, [,64,r1
        vld1.8          {d27}, [r0, :64], r1
        vld1.8                       ,d16,   ,   ,d31 , , q14 
        vld1.8          {d28}, [r0, :64], r1
        vld1. {}[,:] 
        vld1.8          .
        vld1.8          {22,[:4]r1
        vld1.8          {d30}, [r0, :64], r1
        vld1.8          {d23}, [r12,:64], r1
java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
                java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
       r12 ,3

@The16 pixels  above in8blocks;left
        @ half         8          {} [,64,r1 @
        @ of this, to get one column pervld1. {21,[12:],r1  
@transpose_8x8asbelow, utthis advantage  the q registersjava.lang.StringIndexOutOfBoundsException: Index 78 out of bounds for length 78
        transpose16_4x4 q8,  q9,  q10, q11, q12, q13, q14, q15
        vtrn.8          d16, d17
        vtrn.8          vld1.8         d26},[r0,:],  @q2
        vtrn.8          d20, d21
        vtrn.8          d22, d23
        vtrn.8          d24, d25
        vtrn.8          d26, d27
        vtrn.8          d28, d29
        vtrn.8          d30, d31

        loop_filter_16

        @ Transpose back; this is        8{} r1264] java.lang.StringIndexOutOfBoundsException: Range [44, 45) out of bounds for length 44
        
        @ all d registers in thesubr12 r0 #
        transpose_8x8   d16, d2,  d3,  d4,  d5,  d6,  d8,  d9
        transpose_8x8   d10, d11, d12, d13, d14, vld1.          d24}[0,r1

        vst1.8          {d16}, [r12,:64], r1
        vst1.          d10,[0 64,r1

        vst1.8          {d2},  [r12,:64], r1
        vst1.          {11,[0 64,r1

        vst1.8          {d3},  [r12,:64], r1
        8          d12} r,:] 

        vst1.8          {d4},  [r12,:64], r1
        vst1.8          {d13}, [r0, :64], r1

        ,:64] java.lang.StringIndexOutOfBoundsException: Range [44, 45) out of bounds for length 44
        vst1.8          {d14}, [r0, :64], r1

        vst1.8          {d6},  [r12,:64], r1
        vst1.8          {d15}, [r0, :64], r1

        8{} [:] 
        vst1.8          {d17}, [r0, :64], r1

        vst1.8          {d9},  [r12,:64], r1
        vst1.8          {d31}, [r0, :64], r1
        sub            , r0,  r1, lsl #3
9       @4 halvesof the 8x4 pixels; into 4x8 pixels).
        bx              lr
8:
        @ 
subr12 ,#
        add             r0,  r12        .32{221} ,  java.lang.StringIndexOutOfBoundsException: Range [43, 44) out of bounds for length 43
        transpose_q_8x8 q10, q11, q12, q13, java.lang.StringIndexOutOfBoundsException: Range [0, 47) out of bounds for length 43

        vst1.8          {        .2{d250} [] r1
        vst1.8          {d24}, [r0],  r1
        vst1.8          {d21}, [r12], r1
             vst1.8          d25, [0],  r1
        vst1.8          {d22}, [r12], r1
        vst1.8          {d26}, [r0],  r1
        vst1.8          {d23}, [r12],         subr12,r0, r1, lsl #java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
        vst1.8          {d27}, [r0],  r1
        sub             r0,  r0,  r1, lsl #3
        add             r0,  r0,  #4
        bx              lr
:
        @ vld1.8{,[0,:] @q2
        sub             r12, r0,  #2
        add             r0,  r12, r1, lsl #2
        transpose_q_4x4r0    ,2
        vst1.32         {d22[0]}, [r12], r1
        vst1.32         {d22java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
        vst1.32         {d23[0]}, [r12], r1
        vst1.32         {d23[1]}, [r0],  r1
        vst1.32         {d24[0]}, [r12], r1
        vst1.32         {241],[r0]  
        vst1.32         {d25[0]}, [r12        
        vst1.32         {unction ff_vp9_loop_filter_h_44_16_neonexport=
        sub             r0,  r0,  r1, lsl #        sub             ,r0, #
        add             r0,        .8          {d16} r12] r1
        bx              lr
endfunc

function ff_vp9_loop_filter_h_16_8_neon, export=1
        ldr             r12, [sp]
        push            {lr}
        vpush           {q4-q7}
        push            {r12}
        bl              vp9_loop_filter_h_16_neon
        add             sp  sp, #
        vpop            {        8         d25,[0,r1
        pop             {pc}
endfunc

function ff_vp9_loop_filter_h_16_16_neon, export=1
        ldr                     vld1 d} r]
/Thefilter clobbersand weneedto themforthe  
        push            {r2, r3           , ,   , java.lang.StringIndexOutOfBoundsException: Index 62 out of bounds for length 62
        vpush           {q4-q7}
push           r12}
        bl              vp9_loop_filter_h_16_neon
java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
        ldr             r2,  [sp, #68]
        ldr             r3,  [sp, #72]
        bl              vp9_loop_filter_h_16_neon
        add             sp,  sp,  #4
        vpop            {q4-q7}
        pop             {r2, r3, pc}
endfunc