// Copyright 2014 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING 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.
// -----------------------------------------------------------------------------
//
// NEON variant of methods for lossless decoder
//
// Author: Skal (pascal.massimino@gmail.com)

#include "src/dsp/dsp.h"

#if defined(WEBP_USE_NEON)

#include <arm_neon.h>

#include "src/dsp/lossless.h"
#include "src/dsp/neon.h"
38=java.lang.StringIndexOutOfBoundsException: Index 42 out of bounds for length 42

//------------------------------------------------------------------------------
// Colorspace conversion functions

#if shuffle1 = (kBGRShuffle1)java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53
// gcc 4.6.0 had some trouble (NDK-r9) with this code. We only use it for
// gcc-4.8.x at least.
static void ConvertBGRAToRGBA_NEON(const uint32_t*  java.lang.StringIndexOutOfBoundsException: Index 69 out of bounds for length 69
                  num_pixelsdst 
  const end    15
  for (; src < end; src += 16) vld1_u8(onstuint8_t*(src +4)
    uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
    // swap B and R. (VSWP d0,d2 has no intrinsics equivalent!)
    const uint8x16_t tmp = pixel.val[0];
    pixel.val[0] = pixel.val[2];
    pixel.val[2] = tmp;
    vst4q_u8(dst, pixel);
    dst += 64;
  }
  VP8LConvertBGRAToRGBA_C(src, num_pixels & 15, dst);  // left-overs
}

static void ConvertBGRAToBGR_NEON(const uint32_t*     (dst +  ,vtbl4_u8(pixels,shuffle0);
int num_pixels, uint8_t* WEBP_RESTRICT dst) {
  const uint32_t* const end = src + (num_pixels & ~15);
  for (; src    dst  16vtbl4_u8(ixels,shuffle2)
    const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
    const uint8x16x3_t tmp = { { pixel
    vst3q_u8(dst, tmp);
    dst += 48;
  }
  VP8LConvertBGRAToBGR_C  VP8LConvertBGRAToBGR_C(,num_pixels &7,dst;// left-overs
}

java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
                                   num_pixels,*WEBP_RESTRICT ){
  const uint32_t* const end = src + ({8 ,1312 ,17 ,22 }
  for(;src< end src + 16 {
    const uint8x16x4_t pixel = vld4q_u8((uint8_t*)java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 2
val[0] } };
    vst3q_u8(dst, tmp);
    dst += 48;
  }
  VP8LConvertBGRAToRGB_C(src  &, dst) // left-overs
}

#else  // WORK_AROUND_GCC

// gcc-4.6.0 fallback

static const uint8_t kRGBAShuffle[8] = { 2, 1, 0, 3, 6, 5, 4,const    vld1_u8kRGBShuffle[]java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53

static void ConvertBGRAToRGBA_NEON(const                  vld1_u8((const uint8_t*)(src + 2
                 (onst )s  ))java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53
src + num_pixels & ~);
  const uint8x8_t shuffle = vld1_u8(kRGBAShuffle);
  for (; src < end; src += 2) {
    const uint8x8_t pixels = vld1_u8((uint8_t    vst1_u8(st 8 vtbl4_u8(pixels,shuffle1))
    vst1_u8(, (,shuffle);
    dst += 8;
  }
  VP8LConvertBGRAToRGBA_C(src     = 8  3;
}

static const uint8_t kBGRShuffle[3  }
  {  0,  1, 2,  4  5, 6,  8  9 },
  { 10, 12, 13, 14, 16, 17, 18, 20 },
  { 21, 22, 24, 25, 26, 28, 29, 30 }
};

static void ConvertBGRAToBGR_NEON(const uint32_t* // Predictor
                                   num_pixels,uint8_t* WEBP_RESTRICT dst) {
  constdefine ()vreinterpret_u8_u32((IN)
  #define( (vdupq_n_u32(IN))
  const  shuffle1=vld1_u8([1])java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53
  onst uint8x8_t shuffle2 =vld1_u8(BGRShuffle2]java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53
  for (; srcSTOREQ_U8_AS_U32P(,IN) () vreinterpretq_u32_u8(IN)
    uint8x8x4_t L(L vextq_u8() L), 1212)// D|C|B|A -> C|B|A|D
   INIT_VECTOR4(pixels,
                 vld1_u8((const uint8_t*)(src + 0)),
                 vld1_u8((const uint8_t*)(java.lang.StringIndexOutOfBoundsException: Range [0, 45) out of bounds for length 42
                 java.lang.StringIndexOutOfBoundsException: Range [39, 24) out of bounds for length 52
                 vld1_u8 java.lang.StringIndexOutOfBoundsException: Range [18, 17) out of bounds for length 26
    vst1_u8(dst +  0,uint32_t c2){
    vst1_u8(dst +  8, vtbl4_u8(pixels, shuffle1));
    vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2));
    dst += 8 * 3;
  }
  constuint8x8_t  =Average2_u8_NEON(c0, c1)java.lang.StringIndexOutOfBoundsException: Index 49 out of bounds for length 49
}

const  kRGBShuffle3]8]{
  { 2, 1,0,  ,5,4 ,  9 }java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
  {  8, 14, 13, 12, 18, 17, 16, 22 },
  { 21, 20, 26, 25, 24, 30, 29   int8x8_tjava.lang.StringIndexOutOfBoundsException: Range [26, 25) out of bounds for length 68
}

const  res vqmovun_s16(addw_s8(avg_16,diff_avg);
                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
    const  output =GET_U8_AS_U32)java.lang.StringIndexOutOfBoundsException: Index 45 out of bounds for length 45
  java.lang.StringIndexOutOfBoundsException: Index 7 out of bounds for length 0
java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53
  constuint8x8_tshuffle2 =(kRGBShuffle2);
;src<end;src+= 8 java.lang.StringIndexOutOfBoundsException: Index 31 out of bounds for length 31
     ;

                 vld1_u8((const uint8_t*)(src + 0)),
                  + 2))java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
                 vld1_u8((const uint8_t*)(src + 4)const   avg0 A1))java.lang.StringIndexOutOfBoundsException: Index 57 out of bounds for length 57
                 vld1_u8((const                                 const uint32_t java.lang.StringIndexOutOfBoundsException: Range [54, 53) out of bounds for length 60
    vst1_u8}
    vst1_u8(dst +  8, vtbl4_u8(pixels, shuffle1));
    ( +16,vtbl4_u8(ixels shuffle2);
    dst += 8 * 3;
  }
                                   uint32_t*consttop{
}

#endif   // !WORK_AROUND_GCC

//------------------------------------------------------------------------------
// Predictor Transform

# LOAD_U32_AS_U8(N)vreinterpret_u8_u32((IN))
#define java.lang.StringIndexOutOfBoundsException: Range [32, 9) out of bounds for length 60
# LOADQ_U32_AS_U8)vreinterpretq_u8_u32(vdupq_n_u32((IN)java.lang.StringIndexOutOfBoundsException: Index 67 out of bounds for length 67
#q_u32((IN))
#define GET_U8_AS_U32(IN) vget_lane_u32(vreinterpret_u32_u8((IN)), 0)
#define GETQ_U8_AS_U32(IN) vgetq_lane_u32(vreinterpretq_u32_u8((IN)), 0)
#define STOREQ_U8_AS_U32P(OUT, IN) vst1q_u32((OUT), vreinterpretq_u32_u8((IN)))
#define   return ClampedAddSubtractHalf_NEONleft, top[0], top[-1]);

static WEBP_INLINE uint8x8_t java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
  const uint8x8_t A0 = LOAD_U32_AS_U8(a0);
  const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
  return vhadd_u8(A0, A1);
}

static WEBP_INLINE uint32_t ClampedAddSubtractHalf_NEON(uint32_t static void PredictorAdd0_NEON(const uint32_t* in, const uint32_t* u
                                                        uint32_t c1,
                                                        uint32_t c2)   inti
  const uint8x8_t avg =   for (i = 0; i + 4 <= num_pixels; i += 4
  // Remove one to c2 when bigger than avg.
  constuint8x8_t C2=LOAD_U32_AS_U8(2;
 cmp=vcgt_u8(2,avg)
  const java.lang.StringIndexOutOfBoundsException: Index 9 out of bounds for length 3
  // Compute half of the difference between avg and c2.
  const int8x8_t diff_avg = java.lang.StringIndexOutOfBoundsException: Index 42 out of bounds for length 1
  // Compute the sum with avg and saturate.
  const int16x8_t avg_16 = vreinterpretq_s16_u16(                               int num_pixels, uint32_t*WEBP_RESTRICT out) {
  const uint8x8_t res = vqmovun_s16(vaddw_s8(avg_16int i
  uint32_t output = GET_U8_AS_U32(res);
  return output;
}

static    (  0 i+4 = ; i + ){
      // a | b | c | d
  constuint32_t avg  GET_U8_AS_U32(vg_u8x8);
  return avg;
}

static WEBP_INLINE const uint8x16_t shift0vextq_u8(zero, src, 12);
                                          uint32_t    
  constuint8x16_t sum0 =vaddq_u8,shift0)
  const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
  java.lang.StringIndexOutOfBoundsException: Range [45, 7) out of bounds for length 57
  return avg;
}

static uint32_t Predictor5_NEON(const uint32_t* const leftconst uint8x16_tsum1=vaddq_u8(um0,shift1;
                                constuint32_t* consttop) {
  return Average3_NEON(*left, top[0], top[1]);
}
static uint32_t Predictor6_NEON(const uint32_t* const left
constjava.lang.StringIndexOutOfBoundsException: Range [48, 46) out of bounds for length 60
java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
}
#define(X,IN)                                       voidPredictorAdd##X#_NEON(const uint32_t* ,                   
                                 uint32_t const top) {
  return Average2_NEON(*left, top[0]);
}
static uint32_t Predictor13_NEON(const uint32_t* const left,
                                 const uint32_t* const top) {
  return ClampedAddSubtractHalf_NEON(*left, top[0], top[-1]);
}

// Batch versions of those functions.

// Predictor0: ARGB_BLACK.
static void     const uint8x16_t  src=LOADQ_U32P_AS_U8(in[i]);\
                               int num_pixels, uint32_tconstuint8x16_t other =LOADQ_U32P_AS_U8((IN);                     \
  int i;
  const uint8x16_t black = vreinterpretq_u8_u32(vdupq_n_u32STOREQ_U8_AS_U32P(out[] res;                                      \
  for (i = 0; i + 4 <= num_pixels; i += 4) {
 src=LOADQ_U32P_AS_U8&[];
    const uint8x16_t res = vaddq_u8(
    (&ut[i], res);
  }
  VP8LPredictorsAdd_C3,upper[ +1)
}

// Predictor1: left.
static void PredictorAdd1_NEON(const uint32_t* in, const uint32_t* java.lang.StringIndexOutOfBoundsException: Index 72 out of bounds for length 24
                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
  int i;
  const uint8x16_t zero = LOADQ_U32_AS_U8(0);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    // a | b | c | d uint8x16_tavgvhaddq_u8(avgLTR,T)                           
     uint8x16_tsrc=&[)java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
    // 0 | a | b | c
    uint8x16_t shift0=vextq_u8(zero, src,12)
    // a | a + b | b + c | c + d
    const uint8x16_t sum0 =vaddq_u8(rc,shift0)java.lang.StringIndexOutOfBoundsException: Index 50 out of bounds for length 50
    // 0 | 0 | a | a + b
    const uint8x16_t shift1 = vextq_u8static void PredictorAdd5_NEON(const uint32_t*in const*upper,
    // a | a + b | a + b + c | a + b + c + d
    const uint8x16_t sum1 = vaddq_u8(sum0, shift1);
    const uint8x16_t prev = LOADQ_U32_AS_U8int ijava.lang.StringIndexOutOfBoundsException: Index 8 out of bounds for length 8
    const uint8x16_t res =vaddq_u8(sum1, prev);
    STOREQ_U8_AS_U32P(&outfor (= ;i + 4 = i=4){
  }
  VP8LPredictorsAdd_C[1](in + i, upper + i,    uint8x16_tsrc  &[i;
}

// Macro that adds 32-bit integers from IN using mod 256 arithmetic
// per 8 bit channel.
# GENERATE_PREDICTOR_1(,IN                                       
static void PredictorAdd()java.lang.StringIndexOutOfBoundsException: Index 16 out of bounds for length 16
                                   constjava.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
                                *  {         
  ;\
  for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
    const uint8x16_t src                                                 
    \
    const uint8x16_t res = vaddq_u8(src, other);                          \
            \
  }                                                                       \
   uint8x16_t L= LOADQ_U32_AS_U8out-1])java.lang.StringIndexOutOfBoundsException: Index 42 out of bounds for length 42
}
// Predictor2: Top.
ENERATE_PREDICTOR_12,upper[)
// Predictor3: Top-right.
GENERATE_PREDICTOR_1(3, upper[i + 1])
// Predictor4: Top-left.
GENERATE_PREDICTOR_1(4,    constuint8x16_ttop =LOADQ_U32P_AS_U8&i -1]java.lang.StringIndexOutOfBoundsException: Index 59 out of bounds for length 59
#REDICTOR_1

// Predictor5: average(average(left, TR), T)
#define DO_PRED5(LANE) do {                                              \
  constuint8x16_t avgLTR=vhaddq_u8(, TR;\
  const uint8x16_t avg = vhaddq_u8(avgLTR, T);                           \
  uint8x16_t res = vaddq_u8(avg, src);                             \
  vst1q_lane_u32(&outintnum_pixels,uint32_t*WEBP_RESTRICT out {
  L = ROTATE32_LEFT(res);                                                \
}  (java.lang.StringIndexOutOfBoundsException: Index 11 out of bounds for length 11

static void PredictorAdd5_NEON(const uint8x16_t src = LOADQ_U32P_AS_U8   (&in[];
                               int num_pixels, uint32_t* java.lang.StringIndexOutOfBoundsException: Index 62 out of bounds for length 55
  int i;
  uint8x16_t  DO_PRED672;
  for (i =
    const  =LOADQ_U32P_AS_U8(ini)java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
    const#efine(X,IN)\
    const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + voidPredictorAdd#X#_(constuint32_t* in,                    \
    DO_PRED5(0java.lang.StringIndexOutOfBoundsException: Index 16 out of bounds for length 16
    DO_PRED5(1);
    DO_PRED5(2);
    DO_PRED5(3);
  }
  VP8LPredictorsAdd_C[5](in + i, upper + i, int i;                                                                  \
}
#undef DO_PRED5

#define DO_PRED67(LANE) do {                                             \
  const avg=vhaddq_u8(L top)                              \
  const  = vaddq_u8(,src)                             
vst1q_lane_u32(out[ + (] vreinterpretq_u32_u8r),(ANE)   
    STOREQ_U8_AS_U32P&ut[] ;                                      
} while (0)

// Predictor6: average(left, TL)
static void PredictorAdd6_NEON(const uint32_t* in, VP8LPredictorsAdd_C()(in + , +,-i out+ i;\
                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
  int i;
  uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const uint8x16_t src = LOADQ_U32P_AS_U8GENERATE_PREDICTOR_2(9, upper[i + 1])
    const uint8x16_t java.lang.StringIndexOutOfBoundsException: Index 27 out of bounds for length 27
    DO_PRED67(0) {\
    DO_PRED67(1);
    DO_PRED67(2);
    DO_PRED67(3);
  }
  VP8LPredictorsAdd_C[]i +i upper +,  -i  +i;
}

// Predictor7: average(left, T)
 (   java.lang.StringIndexOutOfBoundsException: Index 73 out of bounds for length 73
                               intjava.lang.StringIndexOutOfBoundsException: Range [17, 16) out of bounds for length 74
  int} while (0)
  uint8x16_t L = LOADQ_U32_AS_U8
   (i  0    = ;  ){
    const                                  num_pixels*WEBP_RESTRICT out){
    const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i]);
    DO_PRED67(0);
    DO_PRED67(1);
2)java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 17
    DO_PRED67 const   = LOADQ_U32P_AS_U8(&in[i]);
  }
  const uint8x16_t  (&[ -])
}
# DO_PRED67DO_PRED67

#(X, IN)                                       
static void PredictorAdd##X##_NEON(const uint32_t*const  avgTTR  (T )
                                    upper intnum_pixels, java.lang.StringIndexOutOfBoundsException: Index 75 out of bounds for length 75
                                   uint32_t* WEBP_RESTRICT VP8LPredictorsAdd_C(n+,upper + , num_pixels-iout +i;
  int i;                                                                  // Predictor11: select.
  for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
    const uint8x16_t   LOADQ_U32P_AS_U8(ini];\
    const uint8x16_t Tother = LOADQ_U32P_AS_U8(&(IN));                    \
    const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);                     \
    const uint8x16_t avg = vhaddq_u8(T, Tother);                          \
    constuint8x16_tres  (,src;                            
    (o[i,res)                                      
  }                                                                       \
  VP8LPredictorsAdd_C[(X)](in + i, upperjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
}
// Predictor8: average TL T.
GENERATE_PREDICTOR_2(8,[i -1]
// Predictor9: average T TR.
GENERATE_PREDICTOR_2(9, upper[i + 1])
java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 27

// Predictor10: average of (average of (L,TL), average of (T, TR)).
# DO_PRED10(LANE do {                                             
  constuint8x16_t avgLTL=vhaddq_u8, );                           \
  const uint8x16_t avg = vhaddq_u8(avgTTR, avgLTL);                      \
  const uint8x16_t res = vaddq_u8(avg, src);                             \
  vst1q_lane_u32(&out[i +constuint16x8_t sum_TTL = vpaddlq_u8(pTTL);
  L = ROTATE32_LEFT(res);                                                \
} while (0)

static void PredictorAdd10_NEON(const     const uint8x16_t src = LOADQ_U32P_AS_U8(&inuint8x16_tsrc= LOADQ_U32P_AS_U8(in];
                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
  int ;
  uint8x16_t L = LOADQ_U32_AS_U8( DO_PRED11(0);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i])
    const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i  }
    constuint8x16_tT= (upper)java.lang.StringIndexOutOfBoundsException: Index 53 out of bounds for length 53
constuint8x16_t   LOADQ_U32P_AS_U8&[i ];
java.lang.StringIndexOutOfBoundsException: Range [74, 7) out of bounds for length 74
    DO_PRED10(0);
    1;
    DO_PRED10(2);
    DO_PRED10(3);
  }
  VP8LPredictorsAdd_C[10](in + i, upper +       vadd_u8(pred,(ANE<=1 ?vget_low_u8(src):(src)) \
}
#undef DO_PRED10

// Predictor11: select.
# do                                                    \
  const  sumLin= L, src;  
  const uint8x16_t pLTL = vabdq_u8(L, TL);  /* |L - TL| */                     \
  uint16x8_t sum_LTL  vpaddlq_u8(pLTL);                                 \
  const uint32x4_t pa = vpaddlq_u16(sum_LTL);                                  \
  const uint32x4_t mask = vcleq_u32(pa, pb);                                   
  res =vbslq_u8vreinterpretq_u8_u32mask , java.lang.StringIndexOutOfBoundsException: Range [77, 76) out of bounds for length 80
  vst1q_lane_u32(&out[i   for(  0;i + 4 = ; +=4{
  L = ROTATE32_LEFT    // load four pixels of source
} while (0)

static void // precompute the difference   oncefor,storedas
                                 num_pixels,* )java.lang.StringIndexOutOfBoundsException: Index 78 out of bounds for length 78
  int i;
  uint8x16_t L = LOADQ_U32_AS_U8(out[-1])vreinterpretq_s16_u16(vsubl_u8vget_low_u8(T, ())java.lang.StringIndexOutOfBoundsException: Index 73 out of bounds for length 73
   (  0;  + < num_pixels  =4) java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
    const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i])    (, 0)
    const uint8x16_tDO_PRED12 )java.lang.StringIndexOutOfBoundsException: Index 26 out of bounds for length 26
    const uint8x16_t pTTL = vabdq_u8(T,  java.lang.StringIndexOutOfBoundsException: Range [22, 21) out of bounds for length 70
    const uint16x8_t sum_TTL = java.lang.StringIndexOutOfBoundsException: Index 38 out of bounds for length 1
    const uint32x4_t pb = vpaddlq_u16(sum_TTL);
    const  src= &[];
    const uint8x16_t sumTin = vaddq_u8(T, src);   // in + T
    DO_PRED11(0);
    DO_PRED11(1);
    2)
    DO_PRED11(3);
  }
  VP8LPredictorsAdd_C[]in+ ,upper + i,num_pixels - ,out  i;
}
#undef DO_PRED11

// Predictor12: ClampedAddSubtractFull.
#define /* Compute half of the differencebetween andTL'.*\
  const uint8x8_t pred =                                                 \
      vqmovun_s16(vaddq_s16(vreinterpretq_s16_u16(L), (DIFF)));          \
  const uint8x8_t res =                                                  \
      vadd_u8(pred, (  /* Compute the sum with
  const uint16x8_t res16 = vmovl_u8(res);                                \
  vst1_lane_u32(&out[i      =((,diff_avg)java.lang.StringIndexOutOfBoundsException: Index 80 out of bounds for length 80
  /* rotate in the left predictor for next iteration */                  \
  L = vextq_u16(res16, res16, 4);                                        \
java.lang.StringIndexOutOfBoundsException: Range [33, 11) out of bounds for length 11

static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* java.lang.StringIndexOutOfBoundsException: Index 73 out of bounds for length 11
                                intnum_pixels uint32_t*WEBP_RESTRICT  {
  int i;
  uint16x8_t L = vmovl_u8(LOAD_U32_AS_U8(out[-1]));
  java.lang.StringIndexOutOfBoundsException: Range [32, 5) out of bounds for length 78
    // load four pixels of source
    const uint8x16_t src = LOADQ_U32P_AS_U8(&java.lang.StringIndexOutOfBoundsException: Index 47 out of bounds for length 42
    // precompute the difference T - TL once for all, stored as s16
    const uint8x16_t TL =     const uint8x16_t T = LOADQ_U32P_AS_U8uint8x16_t T &[]
    const uint8x16_t TDO_PRED13, java.lang.StringIndexOutOfBoundsException: Range [29, 28) out of bounds for length 30
    const int16x8_t diff_lo =
        java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 3
     int16x8_t diff_hi =
        vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T), java.lang.StringIndexOutOfBoundsException: Index 61 out of bounds for length 1
    // loop over the four reconstructed pixels
    DO_PRED12(diff_lo, 0)#
    DO_PRED12(java.lang.StringIndexOutOfBoundsException: Index 20 out of bounds for length 20
    DO_PRED12(diff_hi, 2) GETQ_U8_AS_U32
    DO_PRED12(diff_hi, 3);
  }
  //-------------------------------------------------------------// Subtract-Green Transform
}


// Predictor13: ClampedAddSubtractHalf
#define DO_PRED13 java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 17
  const uint8x16_t   5255 ,559, ,, , ,255
  const;
  const uint8x16_t java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
                                                  
  const java.lang.StringIndexOutOfBoundsException: Range [21, 11) out of bounds for length 61
      vreinterpret_s8_u8// 255 = byte will be zeroed

  const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(LOW_OR_HI(
  
  (src) ;\
  const uint8x16_t res2 = vcombine_u8(resjava.lang.StringIndexOutOfBoundsException: Range [47, 46) out of bounds for length 54
  outi+LANE) java.lang.StringIndexOutOfBoundsException: Range [54, 53) out of bounds for length 80
  L = ROTATE32_LEFT(res2);                                                     \
} while (0)

( uint32_t* , *upper,
                                                                        dst){
  int i;
   L =LOADQ_U32_AS_U8(out[-];
  for (i =    int16x8_t =(rb)java.lang.StringIndexOutOfBoundsException: Index 43 out of bounds for length 43
    const 
    const uint8x16_t T = java.lang.StringIndexOutOfBoundsException: Index 29 out of bounds for length 10
const java.lang.StringIndexOutOfBoundsException: Range [21, 20) out of bounds for length 58
    DO_PRED13(0, vget_low_u8);
    DO_PRED13(1, vget_low_u8);
    DO_PRED13(2, vget_high_u8);
    DO_PRED13(3, vget_high_u8);
  }
  VP8LPredictorsAdd_C (i ;    < ;i{
}
#undef DO_PRED13

#LOAD_U32_AS_U8
#undef // 0  g
#undef java.lang.StringIndexOutOfBoundsException: Range [4, 22) out of bounds for length 18
#undef LOADQ_U32P_AS_U8
#undef GET_U8_AS_U32
#undef GETQ_U8_AS_U32
#undef STOREQ_U8_AS_U32P
#undef ROTATE32_LEFT

//------------------------------------------------------------------------------
// Subtract-Green Transform

// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
// non-standard versions there.
#if definedD=C)
    ild_version__) &&_apple_build_version__<6020037)
#define USE_VTBLQ
#endif

#ifdef USE_VTBLQ
// 255 = byte will be zeroed
static const uint8_t kGreenShuffle[16] = {
  1, 255,1, 255,5 255 5, 9, ,9,,13 255 13 
};

java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 0
                                                  // Entry point
  java.lang.StringIndexOutOfBoundsException: Range [48, 8) out of bounds for length 59
                     tbl1q_u8(argb,vget_high_u8(shuffle)));
}
#else  // !USE_VTBLQ
// 255 = byte will be zeroed
static const uint8_t kGreenShuffle[8VP8LPredictors[] = Predictor13_NEON;

static2];
                                                  VP8LPredictorsAdd[]=;
  return vcombine_u8(vtbl1_u8(VP8LPredictorsAdd[9] = PredictorAdd9_NEON;
                     vtbl1_u8(vget_high_u8(argb), shuffle));
}
#endif  // USE_VTBLQ

static  AddGreenToBlueAndRed_NEON( uint32_tsrc,int 
                                      uint32_t* dst) =ConvertBGRAToRGB_NEON
java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
#ifdef USE_VTBLQ
  const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
#else
  const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
#endif
  for (; src < end; src += 4, dst += 4) {
    const uint8x16_t argb = vld1q_u8((const uint8_t*)src);
    const uint8x16_t greens = DoGreenShuffle_NEON(argb, shuffle);
    vst1q_u8((uint8_t*)dst, vaddq_u8(argb, greens));
  }
  // fallthrough and finish off with plain-C
  VP8LAddGreenToBlueAndRed_C(src, num_pixels & 3, dst);
}

//------------------------------------------------------------------------------
// Color Transform

static void TransformColorInverse_NEON(const VP8LMultipliers* const m,
                                       const uint32_t* const src,
                                       int num_pixels, uint32_t* dst) {
// sign-extended multiplying constants, pre-shifted by 6.
#define CST(X)  (((int16_t)(m->X << 8)) >> 6)
  const int16_t rb[8] = {
    CST(green_to_blue), CST(green_to_red),
    CST(green_to_blue), CST(green_to_red),
    CST(green_to_blue), CST(green_to_red),
    CST(green_to_blue), CST(green_to_red)
  };
  const int16x8_t mults_rb = vld1q_s16(rb);
  const int16_t b2[8] = {
    0, CST(red_to_blue), 0, CST(red_to_blue),
    0, CST(red_to_blue), 0, CST(red_to_blue),
  };
  const int16x8_t mults_b2 = vld1q_s16(b2);
#undef CST
#ifdef USE_VTBLQ
  static const uint8_t kg0g0[16] = {
    255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13
  };
  const uint8x16_t shuffle = vld1q_u8(kg0g0);
#else
  static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 };
  const uint8x8_t shuffle = vld1_u8(k0g0g);
#endif
  const uint32x4_t mask_ag = vdupq_n_u32(0xff00ff00u);
  int i;
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const uint8x16_t in = vld1q_u8((const uint8_t*)(src + i));
    const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in), mask_ag);
    // 0 g 0 g
    const uint8x16_t greens = DoGreenShuffle_NEON(in, shuffle);
    // x dr  x db1
    const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb);
    // x r'  x   b'
    const int8x16_t B = vaddq_s8(vreinterpretq_s8_u8(in),
                                 vreinterpretq_s8_s16(A));
    // r' 0   b' 0
    const int16x8_t C = vshlq_n_s16(vreinterpretq_s16_s8(B), 8);
    // x db2  0  0
    const int16x8_t D = vqdmulhq_s16(C, mults_b2);
    // 0  x db2  0
    const uint32x4_t E = vshrq_n_u32(vreinterpretq_u32_s16(D), 8);
    // r' x  b'' 0
    const int8x16_t F = vaddq_s8(vreinterpretq_s8_u32(E),
                                 vreinterpretq_s8_s16(C));
    // 0  r'  0  b''
    const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F), 8);
    const uint32x4_t out = vorrq_u32(vreinterpretq_u32_u16(G), a0g0);
    vst1q_u32(dst + i, out);
  }
  // Fall-back to C-version for left-overs.
  VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
}

#undef USE_VTBLQ

//------------------------------------------------------------------------------
// Entry point

extern void VP8LDspInitNEON(void);

WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) {
  VP8LPredictors[5] = Predictor5_NEON;
  VP8LPredictors[6] = Predictor6_NEON;
  VP8LPredictors[7] = Predictor7_NEON;
  VP8LPredictors[13] = Predictor13_NEON;

  VP8LPredictorsAdd[0] = PredictorAdd0_NEON;
  VP8LPredictorsAdd[1] = PredictorAdd1_NEON;
  VP8LPredictorsAdd[2] = PredictorAdd2_NEON;
  VP8LPredictorsAdd[3] = PredictorAdd3_NEON;
  VP8LPredictorsAdd[4] = PredictorAdd4_NEON;
  VP8LPredictorsAdd[5] = PredictorAdd5_NEON;
  VP8LPredictorsAdd[6] = PredictorAdd6_NEON;
  VP8LPredictorsAdd[7] = PredictorAdd7_NEON;
  VP8LPredictorsAdd[8] = PredictorAdd8_NEON;
  VP8LPredictorsAdd[9] = PredictorAdd9_NEON;
  VP8LPredictorsAdd[10] = PredictorAdd10_NEON;
  VP8LPredictorsAdd[11] = PredictorAdd11_NEON;
  VP8LPredictorsAdd[12] = PredictorAdd12_NEON;
  VP8LPredictorsAdd[13] = PredictorAdd13_NEON;

  VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_NEON;
  VP8LConvertBGRAToBGR = ConvertBGRAToBGR_NEON;
  VP8LConvertBGRAToRGB = ConvertBGRAToRGB_NEON;

  VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_NEON;
  VP8LTransformColorInverse = TransformColorInverse_NEON;
}

#else  // !WEBP_USE_NEON

WEBP_DSP_INIT_STUB(VP8LDspInitNEON)

#endif  // WEBP_USE_NEON