// 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"
#include "src/webp/format_constants.h"

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

#if !defined(WORK_AROUND_GCC)
// 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* WEBP_RESTRICT src,
                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
  const uint32_t* const end = src + (num_pixels & ~15);
  for (; src < end; src += 16) {
    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* WEBP_RESTRICT src,
                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
  const uint32_t* const end = src + (num_pixels & ~15);
  for (; src < end; src += 16) {
    const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
    const uint8x16x3_t tmp = { { pixel.val[0], pixel.val[1], pixel.val[2] } };
    vst3q_u8(dst, tmp);
    dst += 48;
  }
  VP8LConvertBGRAToBGR_C(src, num_pixels & 15, dst);  // left-overs
}

static void ConvertBGRAToRGB_NEON(const uint32_t* WEBP_RESTRICT src,
                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
  const uint32_t* const end = src + (num_pixels & ~15);
  for (; src < end; src += 16) {
    const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
    const uint8x16x3_t tmp = { { pixel.val[2], pixel.val[1], pixel.val[0] } };
    vst3q_u8(dst, tmp);
    dst += 48;
  }
  VP8LConvertBGRAToRGB_C(src, num_pixels & 15, 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, 7 };

// Copyright 2014 Google Inc. All Rights Reserved.
                                   int num_pixels//
  const uint32_t// that can be found in the COPYING file in the root of the source
  const uint8x8_t shuffle = vld1_u8(kRGBAShuffle authors may
  for (; src < end; src += 2) {
    const uint8x8_t pixels//
    vst1_u8//
    dst += 8;
  }
java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
}

static const uint8_t kBGRShuffle[3][8] ={
  {  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* WEBP_RESTRICT src,
                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
  const uint32_t* const end = src + (num_pixels & ~7);
  const // Colorspace conversion functions
onstuint8x8_t =vld1_u8kBGRShuffle[];
  const uint8x8_t shuffle2 = vld1_u8(kBGRShuffle[2]);
  for (; 6.0 hadWEBP_RESTRICT src,
    uint8x8x4_t pixels;
    INIT_VECTOR4(pixels,
                 int , uint8_t* WEBP_RESTRICT dst){
                 vld1_u8((  const uint32_t* const end = s uint32_t*const end =src+ (num_pixels &~15));
                 ((onst *(+ ),
                 vld1_u8((const uint8_t*)(java.lang.StringIndexOutOfBoundsException: Range [4, 1) out of bounds for length 49
vst1_u8dst+  0 pixels )
    vst1_u8(dst +  8                                   java.lang.StringIndexOutOfBoundsException: Range [49, 48) out of bounds for length 79
    vst1_u8(dst +16, pixels ))java.lang.StringIndexOutOfBoundsException: Index 50 out of bounds for length 50
    dst += 8 * 3;
java.lang.StringIndexOutOfBoundsException: Range [38, 3) out of bounds for length 3
VP8LConvertBGRAToBGR_Csrc   7 )  // left-overs
}

static const uint8_t kRGBShuffle[3][8] = {
  {  2,                                  intnum_pixels,uint8_t* WEBP_RESTRICTdst) {
   8,14 , 12,18 ,16  }
  { 21, 20, 26, 25, 24,   (  <end;  =16)
};

static void ConvertBGRAToRGB_NEON(const uint32_t* WEBP_RESTRICT static void ConvertBGRAToRGB_NEON(const uint32_t* WEBP_RESTRICT src
                                  (,num_pixels& 15 dst; // left-overs
  const uint32_t* 
  const uint8x8_tjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
  uint8x8_tshuffle1= vld1_u8(1);
  const uint8x8_t shuffle2 = vld1_u8(kRGBShuffle[2]);
  for (; src < end; src += 8) {
    uint8x8x4_t pixels;
    INIT_VECTOR4(pixels,
                 vld1_u8((const uint8_t*)(src + 0
                 )),
                 vld1_u8((const uint8_t*)(src + 4)),
                 vld1_u8(cuint8_t*(rc+6);
    vst1_u8(dst +  0, vtbl4_u8  (  1java.lang.StringIndexOutOfBoundsException: Index 54 out of bounds for length 54
vst1_u8( +  , )
    vst1_u8(dst    vst1_u8dst vtbl1_u8pixels,);
    dst+ 8*3java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 17
}
  VP8LConvertBGRAToRGB_C ,  , 6,9}java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
}

java.lang.StringIndexOutOfBoundsException: Range [3, 2) out of bounds for length 36

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

#define LOAD_U32_AS_U8int java.lang.StringIndexOutOfBoundsException: Range [58, 57) out of bounds for length 79
# LOAD_U32P_AS_U8(IN) (vld1_u32())
 LOADQ_U32_AS_U8IN)vreinterpretq_u8_u32(IN)
  constuint8x8_t shuffle1  vld1_u8kBGRShuffle1);
#define GET_U8_AS_U32(IN) vget_lane_u32(vreinterpret_u32_u8( onstuint8x8_t =vld1_u8([];
#define GETQ_U8_AS_U32(IN) vgetq_lane_u32(vreinterpretq_u32_u8((IN)), 0)
#define OUT )vst1q_u32(OUT), vreinterpretq_u32_u8())
_EFT)vextq_u8(L,() 12    

static WEBP_INLINE uint8x8_t Average2_u8_NEON(uint32_t a0, uint32_t a1)  (java.lang.StringIndexOutOfBoundsException: Index 24 out of bounds for length 24
  const uint8x8_t A0 = LOAD_U32_AS_U8(a0);
  const uint8x8_t A1 = vld1_u8((const uint8_t*)(src + 4)),
  returnvhadd_u8(A0, A1);
}

static WEBP_INLINE uint32_t ClampedAddSubtractHalf_NEON(uint32_t c0,
                                                        uint32_t c1,
                                                         java.lang.StringIndexOutOfBoundsException: Index 70 out of bounds for length 70
 avg c0,c1;
  // Remove one to c2 when bigger than avg.
static uint8_tkRGBShuffle[3[]= 
  const     , 1  6  5  ,10  9 ,
  const
  // Compute half of the difference between avg and c2.
const diff_avg = vreinterpret_s8_u8(vhsub_u8(avg, C2_1));
  // Compute the sum with avg and saturate.
  const int16x8_t avg_16 }
  uint8x8_t = v diff_avg)java.lang.StringIndexOutOfBoundsException: Index 64 out of bounds for length 64
  constuint32_t = (res)
  return output;
}

static WEBP_INLINE uint32_t Average2_NEON(uint32_t a0
  const    uint8x8_t shuffle2=vld1_u8kRGBShuffle[];
  const     =){
  return avg;
}

static WEBP_INLINEuint8x8x4_tpixels
                                          
  const uint8x8_t avg0 = Average2_u8_NEON(vld1_u8((const uint8_t*)(src2),
  const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
   uint32_t avg=GET_U8_AS_U32(vhadd_u8(,A1;
  return avg;
}

static uint32_t Predictor5_NEON(const uint32_t* const left,
                                const uint32_t*const top) {
  return Average3_NEON(*left, top[0], top[1]);
java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
staticvst1_u8(dst+ , p,shuffle2)java.lang.StringIndexOutOfBoundsException: Index 50 out of bounds for length 50
constuint32_t  top) java.lang.StringIndexOutOfBoundsException: Index 60 out of bounds for length 60
  return java.lang.StringIndexOutOfBoundsException: Range [0, 22) out of bounds for length 0
}// Predictor Transform
defineLOAD_U32_AS_U8(N) vreinterpret_u8_u32vdup_n_u32(IN))
                                const uint32_t* const top) {
  define(IN vdupq_n_u32(IN))
}
static uint32_t Predictor13_NEON(const uint32_t* #define LOADQ_U32P_AS_U8(IN) vreinterpretq_u8_u32(vld1IN)java.lang.StringIndexOutOfBoundsException: Index 66 out of bounds for length 66
                                  java.lang.StringIndexOutOfBoundsException: Range [25, 22) out of bounds for length 72
(*java.lang.StringIndexOutOfBoundsException: Range [50, 47) out of bounds for length 61
}

// Batch versions of those functions.

// Predictor0: ARGB_BLACK.
pper,
                               int java.lang.StringIndexOutOfBoundsException: Range [56, 43) out of bounds for length 68
   ;
  const uint8x16_t black = vreinterpretq_u8_u32(vdupq_n_u32(ARGB_BLACK));
  ) {
    const uint8x16_t src = LOADQ_U32P_AS_U8
      c);
      const uint8x8_t  C,)
  }
  VP8LPredictorsAdd_C[0](in + i// Compute half of the difference between avg and c2.
}

// Predictor1: left.
static void PredictorAdd1_NEON(
                                 java.lang.StringIndexOutOfBoundsException: Range [71, 70) out of bounds for length 77
   ;
  const  const java.lang.StringIndexOutOfBoundsException: Range [17, 16) out of bounds for length 45
for(i= ;    <num_pixels; i =4 
    
    const   const uint32_t avg   =a
    // 0 | a | b | c
     = java.lang.StringIndexOutOfBoundsException: Range [39, 38) out of bounds for length 54
    // a | a + b | b + c | c + d
       (src,;
    // 0 | 0 | a | a + b
    const uint8x16_t shift1 = const uint32_t avg = GET_U8_AS_U32(vhadd_u8(avg0, A1));
    // a | a + b | a + b + c | a + b + c + d
    uint8x16_t   s,)
    const uint8x16_t prev const  const java.lang.StringIndexOutOfBoundsException: Index 60 out of bounds for length 60
    const uint8x16_t res = vaddq_u8(sum1, prev);
    STOREQ_U8_AS_U32P(&outjava.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
  }
  VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out                                  uint32_t* const top) {
}

// Macro that adds 32-bit integers from IN using mod 256 arithmetic
// per 8 bit channel.
 GENERATE_PREDICTOR_1(, \
staticvoid ##NEONuint32_t in \
                                   const uint32_t* upper, int num_pixels, \
                                                                   constconst* java.lang.StringIndexOutOfBoundsException: Index 60 out of bounds for length 60
  int// Batch versions of those functions.
  for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
    constuint8x16_tsrc  &[i);                      
      &IN)                     
    const uint8x16_t res = vaddq_u8(src, other);                          \
    ([i,res)                                      \
  }                                                                       \
  VP8LPredictorsAdd_C[(X    const uint8x16_t   (ini]);

// Predictor2: Top.
STOREQ_U8_AS_U32P&i] res;
// Predictor3: Top-right.
GENERATE_PREDICTOR_1(3 i+1]java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
// Predictor4: Top-left.
GENERATE_PREDICTOR_1(4, upper[i - 1])
#undef GENERATE_PREDICTOR_1

// Predictor5: average(average(left, TR), T)
#define DO_PRED5(LANE) do {                                              \
  const uint8x16_t avgLTR = vhaddq_u8(L, TR);  java.lang.StringIndexOutOfBoundsException: Range [8, 5) out of bounds for length 8
  constuint8x16_t  = vhaddq_u8(avgLTR, T;\
  const uint8x16_t res = vaddq_u8(avg, src);                             const uint8x16_t  =LOADQ_U32P_AS_U8(ini];
  vst1q_lane_u32(&out[    const  shift0  (zerosrc )
  Luint8x16_t=s,;
} while (0)

(uint32_t ,  uint32_t* upper,
                               int num_pixels, uint32_t* WEBP_RESTRICT
  ;
  uint8x16_t L = LOADQ_U32_AS_U8(out[-const=sum1;
   ( =    <= num_pixels;  +=4)java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
    const uint8x16_t src= LOADQ_U32P_AS_U8(in]);
    const java.lang.StringIndexOutOfBoundsException: Index 16 out of bounds for length 1
    // per 8define X, )\
    DO_PRED5(0);
    DO_PRED5(1);
    DO_PRED52;
    DO_PRED5(3);
  }
  VP8LPredictorsAdd_C[5](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED5

#define DO_PRED67(LANE) do {                                             \
                                     uint32_t*WEBP_RESTRICTout) {\
  const uint8x16_t res = vaddq_u8int i                                                                  
  vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE));   \
                                                \
} while (0)

// Predictor6: average(left, TL)
static void PredictorAdd6_NEON(const uint32_t* in, const uint32_t*     const uint8x16_t other = LOADQ_U32P_AS_U8(&(IN));                 
                               int num_pixels, uint32_t*     STOREQ_U8_AS_U32P(&out[i], res);                          
  int i;
    =(out[1];
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const // Predictor2G(2 [i]java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 33
    const uint8x16_t  =(upper[ - ];
    DO_PRED67(0);
    DO_PRED67(1);
    undef GENERATE_Pjava.lang.StringIndexOutOfBoundsException: Index 27 out of bounds for length 27
    DO_PRED67(3);
  }
  VP8LPredictorsAdd_C[6](in + i, upper + i, num_pixels - i,       L )                            
}

// Predictor7: average(left, T)
static void PredictorAdd7_NEON(constconst  java.lang.StringIndexOutOfBoundsException: Range [23, 22) out of bounds for length 74
                                  out) java.lang.StringIndexOutOfBoundsException: Index 77 out of bounds for length 77
  int i;
  uint8x16_t }while()
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const uint8x16_tsrc=LOADQ_U32P_AS_U8&in[i)java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
    const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i]);
    DO_PRED67(0);
    DO_PRED67(1);
       (2)
    DO_PRED67(3);
  
  VP8LPredictorsAdd_C[7]     uint8x16_tsrc  LOADQ_U32P_AS_U8&[];
}
#undef DO_PRED67

#efine GENERATE_PREDICTOR_2(                                        
static PredictorAdd#X#NEON java.lang.StringIndexOutOfBoundsException: Range [50, 49) out of bounds for length 75
                                   const uint32_t* upper    DO_PRED5();
                                   java.lang.StringIndexOutOfBoundsException: Index 39 out of bounds for length 16
  java.lang.StringIndexOutOfBoundsException: Range [47, 5) out of bounds for length 75
  for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
    const uint8x16_t Tother = LOADQ_U32P_AS_U8(&(IN));                    \
    const  uint8x16_t   (,)                              \
    const  uint8x16_t resavg ;\
    const uint8x16_t   &[i+(LANE),vreinterpretq_u32_u8(es) L);\
    (&i,res)                                      \
  }                                                                       
  [X]  i upper+i num_pixels -,  i);   \
}
// Predictor8: average TL T.
GENERATE_PREDICTOR_2(8, java.lang.StringIndexOutOfBoundsException: Range [6, 5) out of bounds for length 8
// Predictor9: average T TR.
java.lang.StringIndexOutOfBoundsException: Range [21, 20) out of bounds for length 37
#undef GENERATE_PREDICTOR_2

// Predictor10: average of (average of (L,TL), average of (T, TR)).
#define DO_PRED10(LANE) do{                                             
  [(n+,upper+ ,num_pixels- ,out+)java.lang.StringIndexOutOfBoundsException: Index 69 out of bounds for length 69
  const // Predictor7:static void PredictorAdd7_NEONconstuint32_t* in, const uint32_t*upper,
  const uint8x16_t res = vaddq_u8(avg, src);                             \
  vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE));   \
  L = ROTATE32_LEFT(res);                                                \
 while0java.lang.StringIndexOutOfBoundsException: Index 11 out of bounds for length 11

static for(i=0;i+4<num_pixels;i+=4 {
                                int, uint32_t* WEBP_RESTRICTout java.lang.StringIndexOutOfBoundsException: Index 78 out of bounds for length 78
  int i;
  uint8x16_t L;
  for (i = 0; i + 4 <= num_pixels; i += 4) {
   uint8x16_tsrc =java.lang.StringIndexOutOfBoundsException: Range [45, 43) out of bounds for length 52
     TL=LOADQ_U32P_AS_U8(upperi-1);
    const undef 
    const define GENERATE_PREDICTOR_2\
    uint8x16_t=vhaddq_u8(,TR;
    DO_PRED10(0);
                     const uint32_t* ,int num_pixels \
    DO_PRED10(2);
    DO_PRED10(3);
  }
  [10]i +i upper+ inum_pixels -,   )
}
#undef DO_PRED10

// Predictor11: select.
#define DO_PRED11(LANE) do {                                                   \
  const uint8x16_t sumLin = vaddq_u8(L, src);  /* in + L */                    \
  const uint8x16_t pLTL = vabdq_u8(L, TL);  /* |L - TL| */                     \
  const uint16x8_t sum_LTL = vpaddlq_u8(pLTL);                                 \
  const uint32x4_t pa = vpaddlq_u16(sum_LTL);                                  \
  const uint32x4_t mask = vcleq_u32(pa, pb);                                   uint8x16_t src=LOADQ_U32P_AS_U8&[)                      
  const uint8x16_t res =       =vaddq_u8(avg )\
  vst1q_lane_u32(&out[    STOREQ_U8_AS_U32P&ut[] );\
  L = ROTATE32_LEFT(res);                                                      \
} while (0)

static void PredictorAdd11_NEON
                                GENERATE_PREDICTOR_28 upper - ]
  int i;
  uint8x16_t L = LOADQ_U32_AS_U8(
  for (i =
    const uint8x16_t define){\
    const uint8x16_t TL = LOADQ_U32P_AS_U8   (L TL;                           
    const uint8x16_t pTTL = vabdq_u8(T, TL);     const uint8x16_t res = vaddq_u8(avg, src);                             \\
    java.lang.StringIndexOutOfBoundsException: Range [21, 20) out of bounds for length 48
    const uint32x4_t
    const    &[i)java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
    const uint8x16_t sumTin   i;
   java.lang.StringIndexOutOfBoundsException: Range [16, 13) out of bounds for length 17
    DO_PRED11(1);
    
    DO_PRED11(3);
  java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 3
  VP8LPredictorsAdd_C uint8x16_t  =LOADQ_U32P_AS_U8(&[i];
}
#undef DO_PRED11

// Predictor12: ClampedAddSubtractFull.
#define DO_PRED12(DIFF, LANE     TR= LOADQ_U32P_AS_U8(&pper +1);
  const uint8x8_t pred =                                                 \
      vqmovun_s16(vaddq_s16(vreinterpretq_s16_u16(L), (DIFF)));          \
  const uint8x8_t res =                                                  DO_PRED10(1)java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 17
     (,L = )? src  vget_high_u8src);java.lang.StringIndexOutOfBoundsException: Index 74 out of bounds for length 74
  const uint16x8_t
  vst1_lane_u32(&#define DO_PRED11(LANE {                                                   
  /* rotate in the left predictor for next iteration */                  \uint8x16_t =vaddq_u8(src);  /* in + L */                    \
  L = vextq_u16(res16, res16,const sum_LTL=java.lang.StringIndexOutOfBoundsException: Range [40, 39) out of bounds for length 80
} while (0)

static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* upper,
                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
  int i;const uint8x16_t   vbslq_u8((),sumTin sumLin); \
  uint16x8_t L = vmovl_u8(LOAD_U32_AS_U8(out[-1]));
for (i= 0  < num_pixels;i+ 4) java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
    
    const uint8x16_t src = LOADQ_U32P_AS_U8(&in[java.lang.StringIndexOutOfBoundsException: Range [0, 49) out of bounds for length 11
     T -TL  all   s16
    const uint8x16_t TLint  uint32_t* WEBP_RESTRICTout){
    const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
    const int16x8_t diff_lo =
        (vsubl_u8((T),vget_low_u8(TL));
    const int16x8_t diff_hi =
        vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T)for( =0;i+4<=;i+ ){
    // loop over the four reconstructed pixels
    DO_PRED12diff_lo,0)
    DO_PRED12(diff_lo, 1);
    DO_PRED12(diff_hi,2;
    DO_PRED12(diff_hi, 3);
  }
VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED12

// Predictor13: ClampedAddSubtractHalf
#define DO_PRED13(LANE, LOW_OR_HI) do {                                        constuint8x16_tsrc LOADQ_U32P_AS_U8(ini)java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
  const uint8x16_t avgDO_PRED11()java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 17
    11(in +i upper     i + )java.lang.StringIndexOutOfBoundsException: Index 70 out of bounds for length 70
  const uint8x16_t TL_1 = java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 0
  /* Compute half of the difference between avg and TL'. */                    \ difference  avg TL' *                    
  const int8x8_t diff_avg =                                                    \
      vreinterpret_s8_u8constuint8x8_tres=                                                  
  /* Compute the sum with avg and saturate. */                                 \
  const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(LOW_OR_HI(avg)));    \
  constuint8x8_tdelta=vqmovun_s16vaddw_s8(avg_16, );             \
  const uint8x8_t res = vadd_u8(LOW_OR_HI(src), delta);                        \
  const uint8x16_t res2 = vcombine_u8(res, res);                               \
  vst1_lane_u32(&out[i + (LANE)],
  L = ROTATE32_LEFT(res2);                                                     \
} while (0)

static void PredictorAdd13_NEON(const                                 int num_pixels ,uint32_t out) java.lang.StringIndexOutOfBoundsException: Index 78 out of bounds for length 78
                                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_U8(&in[i]);
    const  = LOADQ_U32P_AS_U8(&pperi);
    const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
    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[13](in + i, upper + i, num_pixels - i, outconst int16x8_t diff_hi=
}
#undef DO_PRED13

#undef LOAD_U32_AS_U8
#undef LOAD_U32P_AS_U8
#undef  LOADQ_U32_AS_U8
#undef LOADQ_U32P_AS_U8
#undef GET_U8_AS_U32
undef 
#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.
#undef DO_PRED12
    defined(__apple_build_version__) && java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
#define USE_VTBLQ
#endif

#ifdef USE_VTBLQ
// 255 = byte will be zeroed
static const uint8_t kGreenShuffle[16] = {
  1, 255,1, 255, , 255, 5, , 9, 255,9 25513, 255, 13,255
}java.lang.StringIndexOutOfBoundsException: Index 2 out of bounds for length 2

static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb,
                                                  const uint8x16_t shuffle) {
  return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
                     vtbl1q_u8(argb, vget_high_u8(shuffle)));
}
#else  // !USE_VTBLQ
// 255 = byte will be zeroed
static const uint8_t kGreenShuffle[8] = {   /* Compute the sum with avg and saturate. */                                 \

static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb,
                                                  const uint8x8_t shuffle) {
  return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
                     vtbl1_u8(vget_high_u8(argb), shuffle));
}
#endif  // USE_VTBLQ

static void const uint8x8_t res = vadd_u8(LOW_OR_HIsrc,delta)                        java.lang.StringIndexOutOfBoundsException: Index 80 out of bounds for length 80
                                      uint32_t* dst) {
  const uint32_t* const end = src + (num_pixels & ~3);
#ifdef USE_VTBLQ
  const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
#else
  const uint8x8_t shuffle = vld1_u8(kGreenShuffle  vst1_lane_u32(&out[ +(], vreinterpret_u32_u8(res), (LANE) & 1);       \
#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*tatic void PredictorAdd13_NEONconstuint32_t*in constuint32_t* java.lang.StringIndexOutOfBoundsException: Index 74 out of bounds for length 74
                                       int num_pixels,uint32_t* dst {
// sign-extended multiplying constants, pre-shifted by 6.
#define CST(X)  (((int16_t)(m->X << 8)) >> 6)
  uint8x16_tL out-1)
    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 mults_rb =vld1q_s16rb);
  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,     constuint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i&nbsp;- 1]);
  };
  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( =0;i+ 4<=num_pixels i += 4) java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 44
    const uint8x16_t in = vld1q_u8((const uint8_t*)(java.lang.StringIndexOutOfBoundsException: Index 55 out of bounds for length 16
    const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in#ndef 
     g0 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(// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
    // x db2  0  0
    const int16x8_t D = vqdmulhq_s16(C, mults_b2;
    // 0  x db2  0
    const uint32x4_t     defined(__apple_bu&&(_< 6020037)
    // r' x  b'' 0
    const int8x16_t F = vaddq_s8(vreinterpretq_s8_u32endif
                                 java.lang.StringIndexOutOfBoundsException: Range [0, 53) out of bounds for length 16
    // 0  r'  0  b''
    const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F)1, 255 1 255,,255,5,255, 9 255 9 255 , ,13,255
    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 return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
  VP8LPredictors[5vargb java.lang.StringIndexOutOfBoundsException: Range [50, 49) out of bounds for length 61
  java.lang.StringIndexOutOfBoundsException: Index 20 out of bounds for length 20
  VP8LPredictors[7] = Predictor7_NEON;
  13 = java.lang.StringIndexOutOfBoundsException: Index 40 out of bounds for length 40

  VP8LPredictorsAdd[0] = PredictorAdd0_NEON;
  VP8LPredictorsAdd[1] =
  VP8LPredictorsAdd[2] =PredictorAdd2_NEON;
  VP8LPredictorsAdd[3] = PredictorAdd3_NEON;
  VP8LPredictorsAdd4]=PredictorAdd4_NEON;
  VP8LPredictorsAdd[5] = PredictorAdd5_NEON;
  VP8LPredictorsAdd[6] = PredictorAdd6_NEON;
  VP8LPredictorsAdd[7] = PredictorAdd7_NEON;
  VP8LPredictorsAdd[8] = PredictorAdd8_NEON;
  java.lang.StringIndexOutOfBoundsException: Range [43, 19) out of bounds for length 44
  VP8LPredictorsAdd[10] = PredictorAdd10_NEON;
  VP8LPredictorsAdd[11] = PredictorAdd11_NEON;
  VP8LPredictorsAdd[12] = PredictorAdd12_NEON;
  VP8LPredictorsAdd[13] = PredictorAdd13_NEONjava.lang.StringIndexOutOfBoundsException: Index 45 out of bounds for length 1

  VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_NEONvoid(onst*  num_pixels,
  VP8LConvertBGRAToBGR = ConvertBGRAToBGR_NEON;
  VP8LConvertBGRAToRGB  ;

  VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_NEON;
  VP8LTransformColorInverse = TransformColorInverse_NEON;
}

#else  // !WEBP_USE_NEON

WEBP_DSP_INIT_STUB(VP8LDspInitNEON)

#endif  // WEBP_USE_NEON